Issue 41 · Oct 5–11, 2026
Every candidate
All 4,125 papers were scored from their abstracts by gpt-5.6-luna; 2,105 were not skipped. Shown here: the top 150 of those, in score order. Picks are marked.
-
Base Models Can Reason By Taking a Cue From Training Data
This paper shows that some base-model tokens act as learned cues for entering different reasoning modes: forcing cues such as “.\n\nOkay” or “Alright,” substantially improves math and coding performance, sometimes approaching RL-trained models. The authors trace these effects to training data, causally editing data so arbitrary words become reasoning cues or existing cues lose their effect, and show that cue-induced hidden states correspond to different document types.
It offers a non-obvious causal account of much of the apparent benefit of RL reasoning training and demonstrates that training-data associations can be deliberately edited to control reasoning behavior, with large reported gains across models and tasks.
-
Long-WAM: Scaling the Context of World-Action Models
Long-WAM studies how causal world-action models can use long visual histories for real-time robot control. It finds that longer context substantially improves performance only when the underlying video model was pretrained autoregressively: extending history to 19.2 seconds raises RoboCasa GR-1 success from 63.3% to 78.7%, with further gains from robot-video pretraining. The system runs in roughly 107 ms per action chunk and achieves 95% success on dynamic cup stacking, where two comparison systems achieve zero successes.
The paper identifies a non-obvious interaction between pretraining direction and useful control context, then supports it with multi-benchmark results and real-time deployment on physical robots, including a large dynamic-manipulation gain.
-
ARC: A Reasoning Recipe for Robot Foundation Models
ARC adds action-grounded reasoning traces to existing robot foundation models: each trace explains why the next action is appropriate and what effect it should have. The traces are automatically generated from existing DROID demonstrations, then used to adapt different model architectures without collecting new robot data or doing foundation-scale training. The resulting models reportedly gain up to 50 percentage points on RoboLab-Reasoning-50 and 82.2 points in real-robot task success for π0.5.
The combination of automatically generated causal action reasoning, architecture-specific adaptation, and very large zero-shot and real-robot gains suggests a potentially general way to improve RFMs without scaling data or models, though the abstract leaves important baseline and evaluation details unspecified.
-
VioLA: Learning Generalist Humanoid Control Policies from Human Data
VioLA trains a generalist humanoid policy on mostly human motion data by predicting body and hand-motion latents rather than robot joint commands; pretrained controllers convert those latents into executable robot behavior. Using 140.6 million frames, it reports zero-shot real-robot locomotion at 100% success and manipulation at 88.6%, substantially outperforming the cited baselines without task-specific teleoperation. The notable idea is aligning human and robot motion in a shared latent action space so abundant human recordings become usable demonstrations.
This is a potentially important cross-embodiment learning direction, and the reported large zero-shot gains on a real humanoid directly address the data and action-space bottlenecks rather than offering an incremental policy tweak.
-
UNREAL: Unifying Retrieval and Long-Context with a Single Model
UNREAL uses a small trainable module attached to a frozen LLM to encode corpus chunks and generate evidence-selection queries from the LLM’s own internal representations. The same mechanism reportedly improves multi-hop retrieval substantially on a 3B-token Wikipedia index and removes distractors in long-context inference, with large gains on HotpotQA, 2WikiMultiHopQA, NoLiMa, and other long-context tasks while reducing inference cost at sufficiently long contexts.
The potentially important contribution is a single model-native selection mechanism that spans conventional corpus retrieval and 128K–256K-context evidence filtering, with unusually large reported gains and efficiency improvements rather than a minor retriever tweak.
-
HLA-WM: Hybrid Linear Attention for Long-Horizon Video World Models
HLA-WM is a training-free memory mechanism for long-horizon video world models that combines geometry-based retrieval of past scene chunks with recurrent linear-attention state recomputation. On 60-second rollouts, it improves revisit consistency and camera control, including a 0.74 dB PSNR gain and 28.5% lower rotation error, while using 12× less historical-state memory than full KV caching with only up to a 1.6% throughput cost. The gains also transfer across another benchmark and downstream refinement settings.
It offers a technically distinct solution to long-range forgetting in efficient recurrent attention, with substantial memory savings and consistent multi-benchmark improvements rather than a single marginal score gain.
-
The Robot Is Not Its Description: GaugeBench for Representation Robustness in Morphology-Aware Policies
GaugeBench tests whether morphology-aware robot policies are robust to physically equivalent changes in a robot’s description, such as reversing joint axes or reordering joints, while keeping the mechanism fixed. MetaMorph performance collapses from 4030.6 to 51.6 under equivalent re-descriptions, with axis reversal alone causing most of the damage; the effect also appears in another framework, but can be mitigated by coordinate-aware transport or training across conventions.
It identifies a major, previously hidden failure mode in cross-embodiment robotics—representation sensitivity that can be worse than changing the robot itself—and supports it with large-scale tests, cross-framework replication, causal ablations, and an effective remedy.
-
A Shortcut to Structure in AlphaFold 3
The authors use causal interventions and layer-by-layer probing to show that AlphaFold 3 obtains much of its global fold information from the MSA pathway, while the Pairformer rapidly propagates that signal into long-range geometry. Removing the MSA preserves local secondary structure but usually destroys topology; remarkably, a structurally aligned MSA from a different protein can restore much of the fold prediction, raising median TM-score from 0.44 to 0.72. This suggests the model reads alignments partly as transferable descriptions of a fold, not simply as evolutionary evidence specific to the query sequence.
The causal evidence supports a genuinely non-obvious mechanistic claim—that AlphaFold’s MSA representations encode transferable fold information and act as a shortcut to global structure—challenging the usual interpretation of MSAs as query-specific evolutionary context.
-
Subspace Uncertainty and Sharp Sampling Thresholds on the Boolean Cube
This paper characterizes the worst-case number of random Boolean-cube samples needed to achieve the usual parametric regression rate when the target is known to lie in a degree-at-most- subspace. It proves a sharp exponential threshold governed by a Boolean uncertainty principle: noisy regression can require roughly $(m+t)\exp(2k)$ samples when $k\ll d$, despite noiseless identification requiring only about $(m+t)2^k$. The work also gives near-optimal concentration constructions and shows that the remaining subexponential gap is genuinely of Airy-kernel scale.
It provides a sharp, assumption-challenging explanation of why known low-degree structure can still make noisy learning exponentially harder than noiseless recovery, backed by matching upper/lower bounds and an uncertainty-principle construction.
-
KL Before and After Mixing: A Round-Complexity Separation in Parallel Sampling
This paper studies two ways to measure the error of randomized parallel samplers: averaging the KL divergence of each scheduled sampler, versus measuring the KL divergence of the overall output distribution. For fixed-count binary targets, it proves a striking gap: achieving fixed accuracy takes Θ((log n)^2) expected rounds under the first metric but only about log log n rounds under the second, with extensions to weighted targets and a full-support Beta–Bernoulli model. It also gives an exact dynamic-programming characterization of the optimal output divergence at every round budget.
The paper identifies a large, formally proved separation between two seemingly natural sampling-quality criteria, including sharp asymptotics and optimality results rather than a minor algorithmic improvement.
-
Extended amygdala function is governed by threat probability, not uncertainty
Using a formal decomposition of threat uncertainty, the authors argue that anxiety-related self-report, physiology, and brain activity track threat probability more than uncertainty itself. Human fMRI implicates both the bed nucleus of the stria terminalis (BST) and central amygdala in probability, with BST especially responsive to low-probability threats; lesion experiments in rhesus monkeys further suggest a causal role for BST in defensive behavior in those contexts.
This directly challenges a canonical account of extended-amygdala function, identifies a probability-based alternative that reconciles prior results, and supports it with computational analysis, human fMRI, and causal cross-species lesion evidence.
-
A low-adhesion mechanism enables axon growth in the developing brain
The study finds that developing cortical axons can grow rapidly in living brain tissue without strongly gripping or pulling on the surrounding extracellular matrix. Using live imaging, mouse genetics, perturbations, and super-resolution imaging, the authors argue that stabilized, protruding microtubules drive axon extension, while myosin II mainly shapes the growth cone rather than powering forward growth.
It presents a potentially assumption-changing mechanism for rapid axon navigation in vivo—low-adhesion, microtubule-driven growth rather than traction-based pulling—supported by complementary imaging and perturbation experiments.
-
The Onset of Breathing at Birth Switches Off Regenerative Capacity
The study reports that the onset of breathing causes a rapid fall in mTOR activity in newborn mouse neurons and several other tissues. The proposed mechanism is CO2 loss, which raises intracellular pH, activates SNRK, and suppresses mTOR; surprisingly, the process remains reversible in adults, since CO2 inhalation or SNRK deletion reactivates mTOR and promotes retinal axon regeneration after injury.
It proposes an unusually simple, developmentally timed mechanism linking the first breath to the loss of regenerative capacity, and supports reversibility with live imaging, genetic manipulation, and regeneration experiments, although the evidence is currently preclinical and largely in mice.
-
Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight
The paper proposes Self-Retrospection Distillation (SRD), which uses completed agent trajectories to train the same policy to predict useful information and failure modes from its initial, pre-interaction state. This provides a learning signal even when group-relative RL rewards are uniform; across 10 tool-use and long-horizon tasks, it reportedly improves results by up to 24.2 percentage points, including raising a 2B model from 0% to 60.6% success when 98% of groups are all-failure.
The hindsight-to-foresight training formulation is a non-obvious way to address reward-sparse RLVR, and the reported gains in nearly all-failure settings are large, though the abstract does not establish how robustly they generalize beyond the evaluated tasks.
-
Lower Bounds for Parallel Diffusion Sampling
This paper proves that parallelizing score-function evaluations cannot always reduce diffusion sampling to a small number of sequential rounds. For smooth near-isotropic Gaussian mixtures, it gives a roughly d^(1/3) round lower bound, and for anisotropic axis-aligned boxes it gives a tight linear-in-d bound, even when algorithms make polynomially many adaptive queries per round.
It appears to establish the first general polynomial lower bounds on parallel diffusion sampling, showing a fundamental sequential bottleneck rather than merely proposing another sampler or reporting an efficiency tradeoff.
-
Rubix: Global Correspondence-Free Point Set Alignment through Assignment Geometry
Rubix gives a globally optimal solution to planar point-set alignment when correspondences are unknown. It represents every possible matching as a point in a complex-valued correlation polygon, proves that this polygon has at most n(n−1) vertices, and recovers the best matching and rotation using assignment queries; extensions handle 3D rotations and partial matching approximately. The method reportedly matches reference optima much faster than rotation-grid search and improves several shape and scan-matching tasks over alternating minimization.
The combination of an exact global-alignment formulation, a sharp resolution of an open geometric-combinatorial problem, and substantial speedups over practical baselines is unusually novel and potentially broadly useful.
-
Three-dimensional imaging of isolated membrane-protein complexes in vacuo with an X-ray laser
The authors use aerosolized, isolated photosystem I membrane-protein trimers and an X-ray free-electron laser to reconstruct their 3D electron density from single-particle diffraction patterns in vacuum. Helium gas exchange reduced background scattering by 80%, enabling a 3.8 nm reconstruction from 32,788 patterns; the measured shape agrees with simulations and native-mass-spectrometry observations. This demonstrates single-particle XFEL imaging of a relatively small membrane-protein complex rather than only viruses or larger particles.
It demonstrates a previously unavailable experimental capability—3D single-particle X-ray imaging of an isolated 1-MDa membrane complex in vacuo—with substantial data and instrumental validation, although the resolution is still coarse and the result is a proof of concept.
-
Femtosecond Three-Dimensional Imaging of Single-Protein with Hard X-ray Laser
The authors demonstrate single-particle imaging of an individual giant-hemoglobin complex using femtosecond hard-X-ray free-electron-laser pulses. They reconstruct 2D diffraction classes to about 15 Å and a 3D image at roughly 20 Å resolution, extending XFEL single-particle imaging beyond the previously demonstrated regime of large viruses. The result establishes a possible route to ultrafast, room-temperature structural imaging of individual biomolecules, though sample heterogeneity currently limits the reconstruction.
This is a genuinely new experimental capability—3D single-particle imaging of an individual molecular complex with an XFEL—supported by quantitative reconstructions rather than a proposed future method, despite current resolution and heterogeneity limitations.
-
Designing Protein Structures from Natural Language
Prompteina is an all-atom protein structure generator conditioned on natural-language descriptions of structural and functional properties, rather than requiring a separate input format for each design objective. Trained on 1.8 million structures with over 50 million captions, it outperforms baselines on most tested properties and binder-design targets, and experimental testing found nine binders to GM2A, including one with an apparent 107 nM affinity. The main new capability is using language as a flexible interface for specifying diverse protein-design goals, including residue-level constraints.
This introduces a broadly useful natural-language conditioning paradigm for protein structure generation and supports it with large-scale evaluations, iterative design comparisons, and experimental binder validation rather than only benchmark results.
-
TokenRouter: Efficient Serving System for Token-Level LLM Routing
TokenRouter is a serving architecture for LLM systems that choose a different model at each generated token. It separates request-level routing logic from model-specific execution, using asynchronous per-model subservers and delayed batching whose settings are derived from a throughput model. Across several routing methods, workloads, and model pairs, it reports 2.01–64.15× higher decoding throughput than prior systems, with released code.
Token-level routing is difficult to serve efficiently, and the proposed request-centric/model-centric design plus very large reported throughput gains could be broadly useful, but the abstract does not give enough experimental detail to validate the unusually wide 2–64× improvement claim.
-
SCOPE: Certified Theorem Proving with a Language Model as the Policy Planner
SCOPE separates Lean proof generation into high-level operator planning by a language model, exact numerical execution by a symbolic engine, and final proof rendering by a compiler. A 135M model certifies 191/218 problems (87.6%), substantially outperforming the reported 7B baselines while using far fewer tokens and less time; experiments also identify stale engine state as a major failure mode and show 60.3% success on a public Lean-Workbook subset.
The unusually large, independently rechecked gap over much larger theorem-proving models, together with the operator-planning/exact-execution decomposition and diagnostic ablation, suggests a potentially important alternative to simply scaling proof-generation models.
-
TestJack: Should you trust the results in coding benchmarks? Agentic Coding Benchmarks Auditing via Evaluator Evolution
TestJack adaptively audits coding-agent trials by generating tests aimed at prompt requirements the submitted patch may have missed, while filtering tests against the ground-truth solution. Across six model backends and five coding benchmarks, it finds that roughly 34.4% of trials labeled correct by fixed tests violate task requirements, reducing the reported resolution rate from 50.6% to 33.2%.
The paper presents a concrete adaptive-evaluation direction with broad benchmark evidence and a striking result that fixed-test coding benchmarks may substantially overestimate agent capability.
-
AgentGarten: Code Worlds for Evolving Agents
AgentGarten combines code-defined simulators with a shared, real-time neural renderer to create persistent interactive worlds whose rules and visuals can be varied together. It also has agents distill experience into reusable playbooks, and reports learning after roughly four interaction rounds rather than the millions of episodes used by a conventional reinforcement-learning baseline. The potentially novel contribution is the combination of programmable world state, learned visual observations, and cross-round experience inheritance, though the abstract gives few details about tasks or comparisons.
The framework targets a real bottleneck in training interactive agents and claims an unusually large sample-efficiency gain, but the abstract provides insufficient quantitative and experimental detail to justify a strong recommendation.
-
Reflex: Visually Grounded Reactive Humanoid Control
Reflex trains a humanoid to catch thrown boxes using only egocentric RGB-D video. It separates the problem into whole-body catching, estimating object dynamics from partial observations, and visually inferring those dynamics; it reaches 85.7% success in simulation and 65% at 2.5 m and 40% at 5 m on a real Unitree G1, though performance remains below humans.
The paper demonstrates a meaningful real-world capability—visually grounded, whole-body reactive catching on a humanoid—with strong simulation and hardware results, but the contribution is a task-specific staged framework rather than a clearly general breakthrough in reactive control.
-
SGF+: Decoupling Gradient Flows for Autoregressive Video Generation
The paper identifies a conflict between two jobs in autoregressive video models: denoising the current frames and writing context representations for future frames. SGF+ gives these jobs separate parameters while retaining causal interaction, and reports better visual quality and temporal consistency without extra data or longer training; notably, a model trained on 5-second rollouts reportedly generates continuously for up to 24 hours.
Role-specific parameterization guided by observed gradient interference is a plausible new design principle, and the claimed long-horizon extrapolation is notable, but the abstract provides no quantitative results or detail sufficient for a strong recommendation.
-
LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches
LoGRA compresses reinforcement-learning gradients for LLM post-training into low-rank sketches, using them for both parameter updates and policy synchronization. Combined with a predicted-KL controller that limits overly large policy changes, it reportedly cuts average training memory by up to 45.7% while preserving performance and enables 27B-parameter RL training for over 1,100 steps on a single eight-GPU node where dense Adam cannot fit.
The combination of low-rank gradient retention and predictive step control addresses a major practical bottleneck and enables a notably larger training setup, but the abstract provides limited detail on task breadth, quality tradeoffs, and comparisons against other memory-efficient optimizers.
-
DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models
DiffusionART is an online reinforcement-learning method for fine-tuning diffusion image and video generators with black-box or pairwise rewards. It aggregates rewards across samples for each prompt to create a shared update direction, and uses a trust-region-based restoration step instead of EMA tuning; on HunyuanVideo-13B, it reportedly achieves comparable late-stage rewards with about 12× fewer GPU hours and higher mean training rewards than two existing methods.
The aggregation-first credit assignment and EMA-free stabilization appear meaningfully new, and the reported 12× efficiency gain on a large video model is potentially important, but the abstract provides too little detail about baselines, tasks, and breadth of evidence for a strong recommendation.
-
Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks
Memento 3 gives a frozen LLM agent a persistent natural-language rulebook describing hypothesized environment dynamics, which it compiles into executable prediction and planning code. The agent revises rules using prediction errors and accepts code updates only after replay-based verification; it reportedly solves all 25 public ARC-AGI-3 games with human-level action efficiency and learns a Pong controller that wins 21:0 across three openings without further LLM calls. The combination of explicit, revisable world models and verified executable updates is the main new direction.
The explicit rulebook-to-code self-improvement loop is more substantial than a routine agent-memory tweak, but the strongest claims rest on a narrow public benchmark and only three Pong episodes, so the evidence is not yet strong enough for a must-read verdict.
-
VisionWeave: Weaving Elastic Visual Representations as a Native Capability of MLLMs
VisionWeave trains multimodal language models to represent different image or video regions at different token granularities, rather than using uniformly sized visual patches or a fixed token-pruning budget. On Qwen-based models, it reportedly saves 43% of visual tokens while retaining 98.9% of native performance across eight benchmarks, and improves serving throughput by 2.3× with lower latency.
The end-to-end, content-adaptive allocation of coarse and fine visual representations, together with substantial reported serving gains, is more than a routine pruning tweak, but the abstract does not establish that the approach generalizes beyond this model family or beats the broader state of the art decisively.
-
Stateless Language Agents: Scaling Long-Horizon Automated Research
The paper proposes keeping durable research state in an external harness rather than in agents’ conversation histories: each invocation gets a fresh, role-specific context, while an Advisor assigns experiments to parallel Workers based on summarized evidence. Across software engineering, kernel optimization, and algorithm-design tasks, this approach reportedly achieves the best final results and matches a strong kernel baseline with over 84% fewer tokens, with ablations attributing gains to focused contexts and explicit task assignments.
The explicit separation of persistent search state from stateless agent invocations, combined with long-horizon and token-efficiency results across three task types, is a meaningful and somewhat counterintuitive alternative to ever-growing agent memory, though the abstract does not establish broad generality beyond these benchmarks.
-
AffordCraft: Scalable Construction of Task-Ready Simulation Assets from Single Images
AffordCraft reconstructs simulation-ready articulated objects from a single RGB image and a task instruction by retrieving and fitting an existing asset, rather than generating geometry and joints from scratch. Across 2,000 images, it produces assets that pass its physical-validity tests for 1,703 images, while using 10–78× less GPU time per valid asset than compared generative methods; it also works on cluttered scenes and supports downstream manipulation-policy training. The main tradeoff is dependence on a growing asset library: expanding it improves coverage substantially, but only 51% of selections have the requested part label in the largest setting.
The paper presents a practically important shift from per-image generative reconstruction to scalable retrieval-and-fitting of physically valid articulated assets, with broad quantitative evidence for much higher simulation validity and far lower cost, though library coverage limits generality.
-
Cascadia: Resident 975B MoE Inference on Eleven AI PCs
Cascadia runs a 975-billion-parameter sparse MoE model with 41 billion active parameters across eleven consumer AI PCs, using a custom resident execution engine that maps expert layers onto integrated GPUs and shared memory. It achieves 46.87 complete-serving decode tokens/s, peaking at 60.29 aggregate tokens/s across 88 streams, while supporting contexts up to 64k tokens; first-token latency remains several seconds and long-context processing is limited mainly by CPU attention.
The paper demonstrates a practically important capability—serving a nearly trillion-parameter model on a small cluster of commodity AI PCs—rather than merely improving a benchmark, with concrete throughput, latency, context-length, and scaling measurements.
-
Readout Stability in Prefill-Only Decision Models:Zero-Label Prediction and Inference-Time Compute Allocation
This paper studies prefill-only decision models that score all candidate answers in one forward pass, and finds that after changing only the candidate menu, the new prediction can often be recovered by renormalizing and ranking the cached first-pass scores—without labels or another model call. Across seven model families and ten datasets, this ranking-based estimator predicts post-intervention accuracy within 4.2 points, while generative models do not show the same stability; a confidence cascade and curated menus also outperform repeated inference or simply using a larger model at matched cost.
The paper identifies a non-obvious, broadly tested readout property that enables zero-label inference-time compute allocation and reports a striking efficiency result, including a small model with a curated menu substantially outperforming a much larger model on the full label set.
-
Have I Seen Enough? Frozen Video-Language Models Encode Evidence Readiness
This paper finds that frozen video-language models already encode a question-dependent signal indicating whether enough visual evidence has arrived to answer, even though they were not trained for this task. A simple linear probe decodes this signal across seven models, generalizes across video families, remains informative when the model answers incorrectly, and enables a streaming answer-timing policy that improves accuracy by up to 9.75 percentage points at matched video duration. The signal is distinct from uncertainty and from existing learned streaming triggers.
It identifies a potentially general latent capability in existing VideoLLMs—question-conditioned evidence readiness—and demonstrates a low-cost intervention with substantial latency-matched accuracy gains across many configurations.
-
Beyond Reconstruction: What Matters in Action Tokenization for Robot Policies?
The paper argues that action tokenizers for robot policies should not be optimized only to reconstruct continuous actions. It introduces ProAct, which additionally encourages token sequences to be predictable from observations and robust when the policy produces imperfect or unseen tokens. Across Robomimic, LIBERO, RoboTwin, vision-language-action policies, and real-world manipulation, it reports average success gains of 11.3, 21.8, and 36.7 percentage points, respectively, across varied tokenizer architectures.
The paper presents a useful new framing of action tokenization as a policy interface rather than mere reconstruction, backed by unusually large gains across benchmarks, VLA policies, and real-robot experiments.
-
Consistent Distribution Matching for Data-Free Diffusion Distillation
The paper proposes a data-free, simulation-free method for distilling diffusion or flow models into one-step or few-step students. A single student network handles both sample generation and score estimation, avoiding curated distillation data, teacher rollouts, and auxiliary proxy models; the authors also connect the objective to Wasserstein convergence of the student’s distributions. On ImageNet 256×256, it reports FID 2.04 at one function evaluation and 1.37 at four, apparently exceeding prior data-free distillation methods.
Data-free one-step diffusion distillation with a unified student and a convergence argument is a meaningful methodological direction, and the reported ImageNet FIDs suggest a potentially large practical improvement if the comparisons and training protocol are sound.
-
A Deafening Silence: Catastrophic Forgetting Lives in the Output Embeddings of Tokens the Data Never Speaks
The paper finds that continual pretraining forgets old knowledge mainly through output-embedding rows for tokens that are rare or absent in the new corpus, rather than through the model body. It traces this to Adam’s normalization of persistent gradients on those tokens and shows that increasing Adam’s epsilon only for the output projection reduces forgetting by 39.4–67.9% across models from 160M to 12B parameters, while preserving learning on the new data.
It offers a specific, mechanistically motivated explanation for catastrophic forgetting and a remarkably simple intervention with broad reported validation, including compatibility with replay and LoRA.
-
Universal Construction and Exact Self-Reproduction in Ternary McCulloch-Pitts Networks
The paper constructs a fixed ternary McCulloch–Pitts threshold network that stores descriptions of other networks, builds them in a fresh state bank, copies their descriptions, and transfers control to the result. Initialized with its own description, it exactly reconstructs itself—including its stored network weights—generation after generation; a larger SUBLEQ-based construction demonstrates universality and the authors formally verify the schemes in Rocq. It also analyzes robustness margins, distinguishing resistance to pre-activation noise from vulnerability to output flips.
Exact von Neumann-style self-reproduction in a fixed, discrete threshold network is a genuinely unusual theoretical construction with formal verification, though its practical impact and broader relevance beyond computability theory are unclear.
-
Slot3R: Set-Associative Spatial Memory for Streaming 3D Reconstruction
Slot3R adds a set-associative spatial memory to the pretrained Point3R streaming 3D reconstruction system without retraining its backbone. Instead of merging all observations at the same 3D location, it stores multiple state entries and uses bounded reads for the decoder, preserving views and visibility conditions that may later prove useful. On 7Scenes and NeuralRGBD, it reports roughly 57–72% lower point-cloud accuracy error, improved pose results, and continued operation at 600–1000 frames where competing systems run out of memory, while running at about 19 FPS.
The separation of spatial addressing from state identity is a genuinely useful memory-design idea, and the reported large accuracy gains plus substantially longer streaming context suggest a meaningful advance rather than a minor reconstruction tweak.
-
MagicBeanStalk: Escaping the RLVR Zero Advantage Trap with Recursive Data Synthesis
The paper argues that most existing RLVR math data provides little learning signal because a target policy either solves nearly every rollout or none of them. MagicBeanStalk recursively decomposes seed problems and recombines their reasoning patterns into new problems calibrated to the policy and rollout budget, producing mixed success/failure outcomes. On eight math benchmarks it raises performance by 8.6 points for gpt-oss-120b and 5.4 points for DeepSeek v4Flash, substantially outperforming training on the full source corpus or filtered subsets; it also shows gains in finance from very few seed solutions.
The combination of an empirically identified RLVR zero-advantage failure mode with recursive, policy- and budget-calibrated synthesis yields large gains across math and finance, suggesting a potentially general recipe for making verifiable training data effective rather than merely larger.
-
Oído: Open-Vocabulary Speech Recognition on a $5 Microcontroller without a Neural Accelerator
Oído runs a 13-million-parameter Conformer-CTC speech recognizer on a $5 ESP32-S3 microcontroller with no neural accelerator, using int8 computation and a memory/bandwidth-aware execution schedule. On LibriSpeech it reaches 3.70%/8.23% WER, with an optional language model or transducer head improving this to 2.99%/6.67%, but inference is about 2× slower than real time; streaming produces partial transcripts. The released implementation also supports an int4 model and Spanish recognition, making open-vocabulary ASR feasible on hardware normally limited to keyword spotting.
The notable result is not merely quantization but credible open-vocabulary ASR on an extremely constrained, accelerator-free $5 MCU with strong error rates and released hardware measurements, although the lack of real-time performance limits immediate usefulness.
-
Humanize MLE Agents: Multi-Agent Session Alternation Beyond Single-Agent Limits
The paper studies why long-running LLM agents lose effectiveness over time, arguing that reasoning quality declines within extended sessions and that custom context-management harnesses become stale as models improve. It proposes alternating short sessions between two agents sharing a workspace, with context reset between sessions, and gives a semi-Markov analysis of when this can outperform one long session. On 75 MLE-bench Kaggle tasks, the method achieves a 78.2% any-medal rate—8 points above the strongest open-source baseline using one quarter of its time budget—and improves over its constituent agents across five agent pairings.
The combination of a simple context-reset/session-alternation workflow, a formal explanation for its benefit, and substantial gains across 75 heterogeneous long-horizon engineering tasks makes this more than an incremental agent tweak.
-
S-Space: Exploring Spatial Workspace in Multimodal Language Models
The paper identifies a linear subspace in multimodal language models where object-token activations encode horizontal position, vertical position, and distance. Through readouts and causal interventions, it argues that this subspace acts as an internal spatial workspace: perception writes locations into it, reasoning can transform them across viewpoints, and language selectively reads them out. The analysis also suggests that failures often arise not from missing spatial representations but from unreliable spatial transformations.
This is a potentially valuable mechanistic account of how MLLMs represent and manipulate space, with causal evidence and a non-obvious perception-versus-transformation failure decomposition, but the abstract gives too few quantitative details to justify a stronger verdict.
-
Engineered mRNA-packaging virus-like particles enable efficient and precise mitochondrial base editing in vitro and in vivo
The authors engineer virus-like particles that package mRNA encoding mitochondrial base editors, allowing the editors to act transiently rather than persist from plasmid or viral DNA. The particles reached over 50% editing at several mtDNA sites in cultured cells, achieved 41% correction of a pathogenic mutation in a patient-derived line, and produced about 12% editing after a single mouse retinal injection while reducing genome-wide mitochondrial off-target editing by 23–41-fold versus plasmid delivery.
This is a genuinely useful delivery direction for mitochondrial gene editing, combining substantial editing efficiency with a large specificity improvement and initial in vivo validation, although the in vivo efficacy and preclinical scope remain limited.
-
Longitudinal imaging of single cells in opaque tissue with cell chirp ultrasound
The authors introduce “cell chirp,” which uses genetically encoded gas vesicles and frequency-coded ultrasound to detect and image individually labeled cells deep inside optically opaque tissue. They report quantitative single-cell detection, suppression of background by more than 100-fold, multi-day tracking of reporter expression, and week-scale tracking of rare tumor-cell populations during T-cell engager therapy.
This appears to enable a genuinely new capability—longitudinal, real-time imaging of genetically defined single cells in living opaque tissue—using a potentially scalable ultrasound approach rather than optical access or destructive sampling.
-
GRACE: Generation-aware latent compression for efficient video generation
GRACE compresses a pretrained video autoencoder without discarding compatibility with the diffusion transformer that already uses its latent space. It preserves the original latent as a base, learns a residual for lost information, aligns the compressed representation in the frozen DiT’s feature space, and uses lightweight DiT adaptation with asymmetric denoising. On Wan2.1-I2V-14B at 480×832×81, it reportedly reduces tokens 8× and latency 11.1× while matching the uncompressed pipeline on VBench.
The combination of generation-aware latent compression and compatibility with a pretrained video DiT addresses a costly practical bottleneck, with an unusually large reported latency reduction, but the abstract gives evidence from only one model and benchmark setting.
-
When Do We Need On-Policy Distillation? Distilling on Offline Student Rollouts Is Often Better
The paper compares standard on-policy distillation, where the student is trained on its evolving rollouts, with Semi-OPD, which uses offline rollouts from the initial student while still incorporating teacher preferences. Across 17 teacher–student pairs from 1.5B to 235B parameters, Semi-OPD wins in 14 cases, with reported gains up to 13.6% accuracy and 11.4× faster training. The key explanation is that on-policy distillation can become increasingly off-policy relative to a poorly aligned teacher as contexts grow, while fixed student rollouts provide shorter, more stable distillation contexts.
It challenges the assumption that on-policy rollouts are generally preferable, with broad teacher–student comparisons, substantial efficiency gains, and a proposed alignment-based criterion for when OPD actually helps.
-
Optimizing the Optimizer: Language Models Discover Faster Molecular Relaxation Algorithms
The authors use an autoresearch language-model agent to modify the Sella molecular geometry optimizer, with evaluation gates intended to prevent premature convergence and overfitting to the search molecules. The resulting AutoSella variants reportedly use 22.8–59.8% fewer force evaluations than Sella on held-out molecules and unseen potential models, including r2SCAN-3c DFT, despite the search agent never receiving DFT gradients.
This is a genuinely interesting example of LMs discovering a broadly useful numerical optimization algorithm rather than merely tuning a model, but the abstract does not provide enough benchmark detail or evidence to establish that the gains are robust and widely applicable.
-
SoloQ: Calibration-Free Quantization for Diffusion Language Models
SoloQ is a post-training quantization method for diffusion language models that avoids calibration data by rotating and rescaling weights and activations into a predictable distribution. It also quantizes committed KV-cache states in block-diffusion models, and reports preserved accuracy at 4 bits, up to 2.61× lower peak memory, and up to 2.24× faster inference across several dLLM families.
Calibration-free quantization tailored to shifting diffusion denoising states, combined with KV-cache compression and multi-model results, is a meaningful deployment advance, though the abstract does not establish a major capability or broadly general breakthrough.
-
Learning to Learn a Language
The paper introduces a 300M-parameter byte-level transformer trained only on samples from a distribution of synthetic sequence-generating processes, with no exposure to real language. At inference, its weights remain frozen while it uses a long prefix to infer the generating process, achieving low bits-per-byte on Wikipedia in six languages and modeling numerals, algorithmic sequences, code, and speech better than gzip and PPMd. The central claim is that the model learns a general procedure for identifying and predicting structured sequences rather than memorizing a language.
This is a genuinely unusual meta-learning framing with broad claimed transfer from synthetic priors to unseen natural and non-text sequences, but the abstract does not establish how strong the comparisons are or whether the synthetic prior is sufficiently representative to support the generality claims.
-
I-BFM: Reward-Conditioned Robust Humanoid Interaction via Unsupervised Reinforcement Learning
I-BFM trains one humanoid controller to interact with objects through carrying, pushing, and kicking, while also supporting recovery, motion tracking, and chained tasks. It learns reusable interaction dynamics without task-specific policy optimization, and reports strong robustness to large failures: carrying success stays at 89.3% after the robot falls versus 1.3% for a planning baseline, with demonstrations on a Unitree G1.
The combination of a shared, reward-conditioned humanoid interaction policy, unsupervised representation learning, and substantial real-world recovery and task-chaining results points to a potentially meaningful step beyond isolated loco-manipulation skills, although the abstract does not establish how broad or reproducible the gains are.
-
CUSP: CUSUM-Governed Survival Hazard Alarms at the Perception Onset for Off-Road Navigation
CUSP learns an early warning signal for off-road driving hazards from intervention-terminated logs, where the robot usually does not actually fail. It separately labels the earlier point when a human first perceives driving as unsafe, trains a discrete-time survival hazard model, and uses CUSUM risk accumulation to trigger alarms. Across five unseen sites and 142 events, it detected 85 events at a fixed false-alarm rate, versus 26 for the best of nine adapted baselines, including hazards invisible to the navigation model’s existing signals.
The paper targets a genuine deployment gap—detecting impending off-road danger before intervention—and combines a useful treatment of censored intervention data with a large reported gain across unseen sites, though the evidence is still limited to the abstract-scale evaluation.
-
Rehearse Everything, Remember Nothing: Attic-KV Rehearses What Will Be Read
The paper argues that compressing a KV cache by rereading the entire context can be counterproductive at very small budgets, because attention gets spread across too many entries. Attic-KV instead has the model generate self-quiz question–answer pairs that quote relevant context, plus a content-adaptive number of anchor tokens; this reportedly improves both training-free and gradient/trained cache compressors, with gains as large as 41.9 points over full rereading at a 3% keep ratio.
The self-testing rather than rereading framing is a genuinely non-obvious KV-compression strategy, and the abstract reports large gains across two benchmarks and multiple cache-compression methods, though the evidence still needs verification beyond the summarized experiments.
-
PhysEvo: Astra Can Act, Let It
PhysEvo lets a frozen robot-control model improve through experience rather than weight updates: a meta-agent analyzes failed trajectories, revises tools and skills, and tests those revisions, including improvements to its own diagnostic process. On 42 simulated RoboDojo tasks it reports 62% success versus 47.17% for a one-shot baseline, and on five real-world tasks it reports 84% success across 25 trials after simulation-based evolution and continued revision.
The persistent, recursively improved tool-and-skill layer around a frozen model is a genuinely interesting alternative to finetuning, but the evidence is limited by task-specific versions, unclear baseline comparability, and a small real-robot evaluation.
-
SanSi: A Looped Typed Decision Model for System 1.5 Thinking
SanSi converts a pretrained looped language model into a typed decision model that repeatedly updates its hidden state and emits option probabilities after each loop, rather than generating explicit reasoning tokens. Across 10,027 decisions, it reports a 13.5-point accuracy gain over a same-size single-pass model, some extrapolation to deeper reasoning than seen in training, and improved reinforcement-learning outcomes when used as a learned judge.
The combination of recurrent latent computation, anytime typed decisions, and depth extrapolation is a substantive alternative to token-based reasoning, but the abstract does not establish how broadly the gains generalize or how much comes from the specific training setup.
-
ProgramGym: Scaling Training Environments for Long-Horizon Program Reproduction Agents
ProgramGym builds a pipeline for turning existing repositories into verifiable, sandboxed training environments with repaired builds and black-box tests, producing 1,648 environments across C, C++, Go, and Rust. Its accompanying agent uses separated analysis, implementation, and review stages plus revision and failure recovery; fine-tuning and agent workflow changes substantially improve program-reproduction results, with reported gains from 34.87% to 42.84% and up to 56.25%.
The scalable creation of long-horizon, verifiable program-reproduction environments is a meaningful infrastructure direction with sizable reported gains, but the abstract does not establish how broad, reliable, or benchmark-specific those results are.
-
TouchScale: 500 Hours of Human Vision and Touch for Visual-Tactile Learning
TouchScale is a 500-hour dataset of synchronized egocentric RGB-D video, wrist video, and dense bimanual tactile measurements collected with one consistent wearable setup across roughly 2,000 everyday and manipulation tasks. Scaling this data substantially improves zero-shot tactile prediction on an unseen sensor and increases real-robot success on four contact-rich tasks from 22.5% to 57.5%, with performance generally improving as more data is added.
The unusually large, consistently collected visual-tactile dataset and the sizable cross-sensor and real-robot gains make this more than a routine dataset release, though its main contribution is a resource rather than a fundamentally new learning method.
-
NeMo-DCR: Bit-Exact Delta-Compressed Refit for Scalable Agentic RL at Trillion-Parameter Scale
NeMo-DCR synchronizes agentic-RL policies by transmitting only changed parameter bits rather than full checkpoints, while preserving bit-exact equivalence with a dense refit. It uses delta compression, native serving-runtime placement, in-place application, failure recovery, and relay/object-storage streaming; on 30B–1T models it reports 12–40× faster refits, including 150 seconds versus 87.5 minutes for a 1T model at a 3% change rate.
The combination of bit-exact sparse synchronization, shard-to-checkpoint placement, and failure-safe cross-cluster refitting is a substantial systems contribution with striking scale results, but it is a specialized infrastructure optimization rather than a new learning capability or broadly demonstrated paradigm.
-
WorldSonus: Bringing Sound to Worlds
WorldSonus generates synchronized stereo sound from video in a streaming, interactive setting. It uses a causal autoregressive diffusion model for low-latency chunked generation, schedules text-based sound controls during playback, and trains on stereo/ambisonic data to align audio with scene geometry and camera motion; the authors report an RTF of 0.41 and competitive results on open-domain video-to-audio benchmarks.
The combination of real-time streaming, mid-generation sound control, and spatial alignment addresses important missing capabilities in video-to-audio world models, but the abstract provides limited quantitative evidence beyond latency and broad claims of competitiveness.
-
DreamTrue: Action-Faithful Robot World Model with Counterfactual Post-Training
DreamTrue trains a multi-view robot video world model to predict how different actions will affect objects across robot embodiments. It combines geometric calibration of action-conditioned image trajectories with counterfactual action edits, then uses a human-labeled defect model as a reward for reinforcement-learning post-training; on AgiBot it reports much better action following and a reduction in interaction defects from 48.12% to 6.25%.
The combination of cross-embodiment action calibration, counterfactual interaction augmentation, and defect-based reward modeling is a substantive world-model training direction, but the abstract provides limited detail on baselines, evaluation breadth, and whether the gains transfer to real robot control rather than video prediction.
-
When to Switch: Reliable Action-Chunk Extension for Vision-Language-Action Models
The paper identifies subskill transitions as the main source of errors when vision-language-action models execute long action chunks. RACE predicts when such a transition will occur using an auxiliary denoising pass, then conditions action generation on that timing; this supports 2–4× longer chunks, with improved simulated success and a real-robot reduction in stop-and-go idle time of about 5×.
The transition-timing framing targets a real bottleneck in VLA inference and is supported by both simulation and hardware results, but the abstract lacks absolute success rates, task scale, and comparisons detailed enough to establish a major advance.
-
Hybrid Latent Attention for Looped Language Models
The paper reduces the KV-cache overhead of looped language models by keeping exact keys and values for recent tokens while representing older tokens with compact latents that each loop reads directly. On 1.4B and 2.6B-parameter models, this cuts cache size by 10.7x and improves decoding throughput by 2.5x at 1K context and up to 7.4x at 16K, while retaining roughly 96–100% of long-context retrieval performance and comparable fine-tuned math performance.
This is a technically distinct approach to the unusually expensive KV caching of recurrent/looped transformers, with large measured memory and throughput gains and relatively small quality loss across multiple model sizes and task types.
-
A Self-Pruning Transformer: Extreme KV-Cache Compression with Universal Attention
The paper introduces Universal Attention, a trainable attention mechanism whose learned decay factors identify and remove low-value tokens from the KV cache while retaining RoPE and softmax attention. It reports 10× cache compression on language and synthetic tasks with improved downstream performance, and up to 25× compression at 16k context with better long-context generalization than prior methods and an unpruned model.
Adaptive, highly aggressive KV-cache pruning that reportedly improves rather than merely preserves quality would be a major deployment advance, but the unusually strong claims need verification across tasks, models, and latency/memory measurements before warranting a strong recommendation.
-
Language Models as AI Research World Models
This paper treats language models as “research world models” that learn from past AI experiments to predict which new interventions will work, helping research agents choose experiments under limited budgets. Using more than 2,600 records across nine training and inference environments, the authors find that accumulated in-environment experience improves ranking of unseen interventions and transfers across environments; in one cross-environment test it reduced selection regret by 78%.
It presents a useful new framing for AI research agents and supports it with unusually broad experimental-history data, including substantial cross-environment transfer and gains in multi-round experiment selection.
-
Clinician use of language models diverges from how the models are evaluated
The authors analyzed 127,833 queries from 6,342 clinicians using an institutional LLM assistant, then compared the observed tasks with 58 public clinical-AI benchmarks. Real-world use was dominated by documentation, administration, and knowledge retrieval, while diagnosis was only 3.7%; over a third of queries were not answerable as posed, and benchmark task distributions overlapped only 31% with actual use. The main contribution is large-scale evidence that clinical benchmarks evaluate a substantially different workload from the one clinicians give LLM assistants.
This is a large, concrete, and non-obvious demonstration that widely used clinical LLM evaluations may be poorly aligned with deployment, though it is primarily an evaluation study rather than a new model or capability.
-
From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation
The paper proposes Trace2Env, which turns historical interaction traces into a reusable “worldbook” that an agentic language model uses to simulate an unavailable environment. The simulator maintains episodic state and predicts both immediate observations and lasting effects of actions; across nine environments, it reportedly improves next-step fidelity, long-horizon consistency, and replay validity compared with prompt-only language world models.
The trace-to-agentic-environment framing is a meaningful alternative to rebuilding executable environments, but the abstract gives no quantitative results or evidence that the approach generalizes beyond the tested settings.
-
RealtimeWAM: One-Step Asynchronous World Action Models
RealtimeWAM makes visual World Action Models faster by combining one-step action generation with asynchronous execution of the video and action experts. Its Teacher-Anchored Consistency Distillation trains the action model against both local consistency targets and the teacher’s final multi-step rollout, while Cross-Expert Wavefront Pipelining overlaps computation through shared video KV caches. On LIBERO, LIBERO-Plus, and RoboTwin, it reportedly keeps performance within 1% of the original models while achieving about 25× end-to-end speedup on an H100.
The combination of endpoint-aware one-step distillation and fine-grained asynchronous video/action pipelining could substantially improve real-time robot control, but the abstract gives limited detail about absolute latency, hardware comparisons, and real-robot validation, so the unusually large speedup needs scrutiny.
-
Learning Representations to Guide Pixel Diffusion Transformers
The paper argues that representations useful for visual understanding are not automatically useful for pixel-level generation, because semantic information can be lost during synthesis. It introduces Representation Guidance: a compact bottleneck forms a generation-oriented internal representation, which then guides the pixel decoder by changing the flow-matching target. On ImageNet, the method reports strong FID scores with fewer parameters than larger pixel diffusion transformers, and it also extends to text-to-image generation.
The explicit use of a learned bottleneck representation to guide pixel diffusion, together with the claim that representation quality and generative utility diverge, is a meaningful direction, but the abstract provides limited comparative and experimental detail for the unusually strong results.
-
Mechanics of Long-Context Hybrid Models Part 1.1: From Hybrid Attention to Hybrid Position
This paper analyzes why hybrids of full attention with sliding-window or linear attention behave differently when extending context. It identifies distinct positional effects—including a tradeoff between continual long-context training and length extrapolation—and proposes Sliding-Window Linear Attention, which reportedly extrapolates context 16× without retraining while retaining perfect accuracy on one 64k needle-in-a-haystack task.
The positional-bias analysis and the reported 16× training-free extrapolation are potentially useful, non-obvious contributions, but the abstract provides limited evidence beyond a synthetic retrieval benchmark and does not establish broad gains over strong long-context methods.
-
Recurrent Looped Transformer
The paper proposes a Transformer that combines a parallel causal encoder with a recurrent decoder whose state is fed back at every token, giving computation depth that grows with sequence length without increasing per-token cost. On small algorithmic tasks, this recurrent feedback lets models extrapolate far beyond training lengths—for example, perfect parity on 256-bit inputs and 97% accuracy on long permutation tracking, where a standard Transformer fails—but chunked feedback harms tasks requiring token-level updates.
The recurrent depth and strong length-extrapolation results are a genuinely interesting architectural direction, but the evidence is limited to six synthetic algorithmic tasks and small models, so practical generality remains unestablished.
-
MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement
This technical report describes MiMo-V2.6, an omni-modal model trained with substantially scaled reinforcement learning. It increases RL batch and token throughput, uses diverse long-horizon agent environments and more expensive group-based grading, and adds infrastructure for asynchronous rollouts, mixed-task training, and reward-hacking defenses; the abstract does not provide concrete capability or benchmark results.
The scale and infrastructure for multimodal, long-context agentic RL could be important and reusable, but the abstract offers mainly engineering claims and no quantitative evidence that scaling produced a major capability jump.
-
DAEDALUS: Bootstrapping Agent Memory from Self-Generated Tasks
DAEDALUS builds reusable agent memory without human-written tasks or an oracle verifier: an explorer generates challenging tasks, a solver attempts them, and heuristics from failures are retained only after repeated success. Across three tool-use benchmarks, it reports gains of up to 15.9 percentage points in success and 2.2× in pass^5, with benefits from small exploration budgets and transfer across model families.
The self-play-like bootstrapping of procedural memory from generated tasks without task supervision or an oracle is a genuinely interesting direction, supported by multi-benchmark gains, but the abstract does not establish a major capability or cost breakthrough.
-
ALIVE: Interaction-Aligned Object Insertion for First-Frame-Guided Video Editing
ALIVE trains a video editor to insert an object while making it participate coherently in the existing scene—for example, being picked up or manipulated—using only an edited first frame and an object-level instruction. It introduces interaction-focused training pairs, a model that predicts interaction guidance automatically, and a benchmark for interaction fidelity and source preservation; it reports large gains over the strongest baseline on its interaction benchmark and smaller gains on general object insertion.
Interaction-aware object insertion is a meaningful capability beyond ordinary video editing, but the evidence is based largely on a newly curated dataset and VLM-based evaluation, with insufficient absolute results or human validation in the abstract to warrant a strong verdict.
-
Towards Looped Models Done Right, Part II: Rethinking at Fixed Points
This paper studies looped language models whose recurrent hidden states approach fixed points, showing that this behavior can support truncated training, shared terminal KV caches, faster student prefilling, and cheaper RL updates. It introduces a prediction-feedback-learned depth prior and an orthogonal input-injection method; across 100M–1.6B models these reduce perplexity, and at 1.6B the learned prior achieves comparable downstream performance with a 3× smaller KV cache than fixed-depth training.
The fixed-point framing yields several concrete efficiency benefits, and the learned depth prior plus orthogonal injection are substantive methods, but the reported scale and gains do not yet establish a major shift beyond the looped-model setting.
-
SWE-Duel: Towards a Self-Scaling Arena for Coding Agents
SWE-Duel proposes a coding-agent arena where agents create feature-and-bug tasks for one another at evaluation time, rather than solving a fixed benchmark. Across 1,500 turns on 15 repositories and five languages, it separates solver ability from task-authoring ability and finds that newer model versions generate substantially harder defects for a fixed solver (42.9% versus 4.0% defeat rates).
The self-generated, adversarial evaluation setup is a genuinely useful alternative to saturating and contaminated coding benchmarks, with some empirical evidence that task difficulty scales with agent capability, but the tournament is still relatively small and the benchmark claims are preliminary.
-
TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language Models
TRACE trains FP4-quantized MoE language models so that the quantized training and rollout paths behave more similarly, using rollout-side rounding information to guide training-side decisions. It also caches selected quantization metadata to limit communication and storage costs. Across four MoE models and reasoning, coding, and long-horizon RL tasks, the method reportedly matches BF16 rollout performance while providing up to 5.4× rollout speedup and better final performance than post-hoc FP4 quantization.
The direct alignment of train- and rollout-side FP4 behavior, combined with joint weight/activation/KV-cache quantization, addresses a real systems bottleneck and claims a substantial speedup, but the abstract does not provide enough quantitative detail to establish a field-changing advance.
-
Lost in the bf16 Cast: Exporting Ternary Language Models Can Revert Most Low-Learning-Rate Code Changes
The paper finds that exporting fine-tuned ternary language models through a bf16 intermediate can change a small fraction of thresholded weight codes, yet those changes can destroy downstream accuracy—e.g., reducing strict GSM8K accuracy from 58.79% to 0.78% in one model. It traces this to bf16 rounding and ties-to-even behavior near quantization thresholds, and shows two export-compatible fixes that preserve online accuracy across three model families.
This exposes a surprising and potentially serious deployment failure mode for ternary models, with large measured effects and practical remedies, though its importance is concentrated in a still-narrow model family and pipeline.
-
Rethinking Cross-Tokenizer On-Policy Distillation: From Alignment Coverage to Supervision Reliability
This paper studies on-policy distillation when teacher and student use different tokenizers. It finds that strict one-to-one token alignment already captures most probability mass, while adding supervision for tokenizer-mismatch spans can hurt accuracy; a student-selected top-16 shared-vocabulary loss performs about as well as full shared-vocabulary supervision across math and code tasks. The authors argue that reliable, well-aligned supervision matters more than maximizing coverage, supported by gradient-alignment diagnostics.
The counterintuitive result that broader cross-tokenizer supervision degrades learning, together with a concrete reliability-based explanation, is worth noting, but the evidence is limited to three teacher–student pairs and appears more like a useful distillation insight than a major new direction.
-
DSReg: Provably Recovering Individual World Latents without Reconstruction
This paper proposes DSReg, a regularizer intended to separate individual latent factors from a representation that is already identifiable only up to a linear transformation, without using reconstruction, a decoder, or labels. It introduces “Structural Diversity,” the condition that different latent variables have distinct dependency patterns with observations, and claims a proof of recovery up to signed permutation. Experiments on synthetic data, visual encoders, world-model probes, and external renderers reportedly show better factor recovery while preserving dense prediction, but the abstract gives no quantitative results.
The combination of JEPA-style prediction, post-hoc latent disentanglement, and an identifiability guarantee without reconstruction is a genuinely interesting theoretical direction, but the abstract provides insufficient empirical detail to justify a stronger verdict.
-
SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models
SpaceCast-Bench evaluates whether vision-language models can predict how a scene changes after an intervention, rather than merely identify spatial relations visible in an image. Across 3,862 questions and 182 real-world scenes, the best model scores 58.0% versus 87.2% for humans; analyses suggest that intermediate “bridge” views and explicit 3D evidence matter, while generated outcome images or videos help less reliably. Fine-tuning a small Qwen3-VL model on synthetic data raises performance substantially and transfers to six out-of-domain benchmarks.
The benchmark targets a relatively under-tested capability—predictive spatial state updating—and reports a large model-human gap plus non-obvious evidence about view integration and 3D representations, though it is still primarily an evaluation and dataset paper rather than a demonstrated capability breakthrough.
-
REMORY: Learning Residual Memory for Context Compaction
REMORY adds a bounded set of learned soft tokens to a textual summary of an agent’s history. These residual memory tokens are trained so a frozen language model more closely matches the behavior it would have had with the full history; experiments report better source attribution and fewer repeated tool outputs and errors while using about 5.2% of the original context positions.
The residual-memory-token formulation is a meaningful alternative to summary-only context compression, but the abstract gives limited quantitative detail and does not yet establish broad gains beyond the reported benchmarks.
-
One Model to Loop Them All: Elastic Latent Recurrence for Language Models
The paper introduces Elastic Latent Recurrence, a way to train a language model that reuses the same layers for multiple passes while preserving nearly normal-model accuracy after just one pass. On 4B dense and 30B sparse-MoE models, additional loops improve ten-task scores by 7.25 and 4.75 points, while techniques such as late mixed-depth training and retaining only the final KV cache reduce training and inference costs. The result is a single checkpoint whose quality can be traded against compute at serving time.
This is a substantive and well-supported approach to making elastic-compute language models practical, with sizable reported gains, but the evidence is still limited to two model scales and the abstract does not establish broad superiority over strong adaptive-depth or test-time-compute alternatives.
-
AgentSWE: Can Coding Agents Build the Agent You Actually Want?
AgentSWE benchmarks coding agents that are asked to create, modify, or optimize other software agents from natural-language requirements. Across 25 tasks, the authors report that agents commonly overfit to visible examples, produce code with nonfunctional runtimes, and claim success without satisfying held-out requirements; demonstrated behavior is often much better than the frozen delivered artifact.
The benchmark targets an important and relatively neglected capability—reliably engineering agents to meet requirements—and reports a non-obvious gap between apparent demonstrations and held-out behavior, but the abstract gives few quantitative details and the task set is small.
-
HuatuoGPT-3: RL-Only Domain Adaptation from Base Models
The paper proposes OnePO, an RL-only method for adapting a base language model to a domain without supervised fine-tuning. It uses teacher outputs early in training, emphasizes informative low-probability tokens, and retires the teacher when the policy improves; on medical QA, it reports better HealthBench scores than SFT+RL and pure RL, including a 27B HuatuoGPT-3 model.
RL-only domain adaptation with explicit mechanisms for avoiding early gradient starvation and later teacher anchoring is a meaningful idea, but the evidence is mostly benchmark claims in one domain and the abstract does not establish broad generality or independently credible frontier-model comparisons.
-
A self-learning scientific agent for X-ray diffraction
The paper presents Gan Jiang, an agent for powder X-ray diffraction that uses several physics-based analysis tools and improves its own executable analysis skills by diagnosing failures and revising instructions and code, without retraining the language model. It reports better refinement scores across FullProf, GSAS-II, and PyWPEM, and substantially higher phase-identification accuracy than the strongest comparator on several simulated and experimental datasets, alongside demonstrations on cosmetics, batteries, and catalysts.
The combination of validated self-revision with physics-constrained scientific tools is a meaningful agent design direction and produces large reported accuracy gains, but the evidence is concentrated in one specialized XRD workflow and the abstract does not establish broad transfer beyond that domain.
-
RLHND: Video Foundation Models as Physically Grounded Hand Trackers for Robot Learning
RLHND repurposes a pretrained video diffusion model as a deterministic clip-level feature extractor for monocular egocentric hand tracking. It predicts anatomically plausible hand pose together with dense surface contact and force, and reports state-of-the-art results on tracking and tactile-estimation benchmarks, plus retargeting and real-robot demonstrations. The main novelty is bringing video-model priors about motion and hand–object interaction into physically grounded tracking rather than estimating pose frame by frame.
The combination of video-foundation-model priors with pose, contact, and force estimation could make human-video-to-robot learning substantially more useful, but the abstract provides no quantitative margins or evidence that it beats simpler temporal and tactile-prediction approaches beyond benchmark leadership claims.
-
OuroWorld: Bringing Any 3D World Alive as Diverse, Endlessly Looping 3D Cinemagraphs
OuroWorld converts static 3D Gaussian Splatting scenes into dynamic, endlessly looping 3D cinemagraphs. It uses a vision-language model and video generator to propose motion, then fits a periodic 4D Gaussian Splatting deformation model that supports object motion, general deformation, and lighting changes while handling inconsistent multi-view supervision. Across 39 scenes, it reports better visual quality than prior methods and strong user-study preferences, though the evaluation is largely perceptual and ground-truth-free.
The combination of video-generated motion, periodic 4DGS, and cross-view inconsistency handling is a meaningful direction for animating static 3D worlds, but the evidence is limited to a modest scene set and mostly subjective comparisons rather than demonstrating a major capability jump.
-
One Block, Multiple Depths: Recurrent Vision Transformers with Depth-Programmed Experts
The paper reuses one Transformer block recurrently across depth, but gives each recurrence a different FFN by mixing a small bank of shared experts according to a continuous depth coordinate. This reportedly matches full-depth vision encoders at similar compute, cuts stored parameters by about 70% for a DeiT-III-level model, preserves most DINOv2 transfer performance, and lets one checkpoint run at multiple depths through resampling.
The combination of recurrent depth, weight-space expert mixing, and resampleable depth provides a meaningful parameter-efficiency and elastic-deployment idea, but the abstract lacks detailed quantitative comparisons and broad evidence needed for a strong verdict.
-
SepGen: Multi-Stem Audio-Video Separation and Generation in a Single Model
SepGen modifies a pretrained audio-video generator so that one sampling pass produces the video, the mixed soundtrack, and separate audio stems for captioned sound sources. It also works as a caption-guided separator for existing mixtures, reportedly outperforming language-conditioned separators—especially for speech—even when the captions omit the spoken words.
Producing source-specific waveforms jointly with generated video and audio, while also supporting text-guided separation, is a meaningful and relatively distinctive capability, but the abstract gives limited quantitative evidence about the size and breadth of the gains.
-
Scaling Agentic Data for Long-Horizon Terminal Intelligence
TerminalHorizon is an automated procedure for creating long-horizon terminal-agent training environments. It independently executes professional tasks to find work that a synthesized environment would otherwise shortcut, reconstructs environments around that evidence, and progressively expands tasks with new requirements; fine-tuning Qwen3.5-27B on 1.5K such environments reportedly improves performance across 12 agent and reasoning benchmarks. The abstract also reports 12-hour case studies where the model iteratively experiments and improves its final outputs, but gives no quantitative results.
The combination of evidence-driven environment reconstruction and progressive long-horizon task expansion is a substantive data-generation idea for agent training, but the abstract provides too few numbers or details to establish that the gains are broad or durable.
-
Learning More by Training Less: Uncertainty-Guided Progressive Contraction for Efficient Reasoning Post-Training
The paper proposes compressing reasoning post-training data in two stages: first by using Monte Carlo reasoning trees to identify uncertain, informative parts of teacher-generated chains of thought, and then by routing training online based on the student's evolving learning state. It claims to preserve or improve reasoning performance while using as little as 0.22% of the optimization exposure of full-data supervised fine-tuning, across several model families and teacher–student settings.
The claimed 450x reduction in training exposure would be highly consequential, but the abstract provides no concrete benchmark deltas or accounting for the potentially substantial cost of constructing Monte Carlo reasoning trees, so the result needs verification before being treated as a major advance.
-
Multi-Agent Egocentric World Model with Fine-Grained Embodied Interaction
The paper models several agents’ first-person video streams together while they perform fine-grained actions in a shared environment. ME-World jointly generates the streams, conditions them on all agents’ poses, and uses shared environment memory to maintain consistent interactions and state changes; experiments on real and synthetic data report better consistency, control, identity preservation, and video quality than prior methods.
Jointly enforcing cross-agent consistency for fine-grained embodied interactions is a meaningful world-modeling direction, but the abstract gives no quantitative results or evidence of real downstream robotic capability.
-
Later Is Better: Token Reduction for ViTs Under Distribution Shift
The paper studies where in a ViT’s depth token pruning should happen when inputs are distribution-shifted. It proposes a simple late-concentrated power-law schedule and finds that, at the same compute budget, delaying reduction substantially recovers the accuracy gap versus the uncompressed model across multiple pruning methods, backbones, corruption and shift suites, and even video and vision-language tasks. The proposed explanation is that early pruning errors are propagated through more subsequent layers.
A simple scheduling change yields a broad, compute-neutral robustness improvement and challenges the implicit assumption that token reduction can be distributed uniformly across depth, but it remains an incremental inference-strategy advance rather than a major new capability.
-
NEMORA: Neural Equivariant Multipole Operators for Long-Range Atomistic Learning
NEMORA extends the Fast Multipole Method into a learned, equivariant operator for propagating tensorial information across long ranges in atomistic graph networks. It preserves FMM-style hierarchical structure and radial factors while learning many-body angular couplings, with claimed linear time and memory scaling; experiments report large reductions in energy and force errors on non-local benchmarks and applicability to systems with hundreds of thousands of atoms.
The learned equivariant FMM formulation is a genuinely interesting architectural direction with potentially important scaling benefits, but the abstract gives few concrete benchmark details and the very large improvement claims need careful validation.
-
Tell Robot What Not to Do: A Negation Understanding Perspective
The paper studies whether vision-language-action models can obey instructions that include explicit exclusions, such as completing a task without touching a specified object. NegaAlign adds small trainable transformation and visual-alignment modules while freezing the pretrained VLA, and reports improvements in negated-task success from 2.6% to 88.45% in simulation and from 12.4% to 88.8% on real-world tasks, while preserving affirmative-task performance. The main new contribution is a parameter-efficient way to retrofit negation understanding using supervision derived from existing demonstrations.
The very large improvement and cross-model, real-world results suggest a meaningful missing capability in current VLAs, but the evidence is weakened by reliance on a new benchmark and limited abstract-level detail about task scale and baselines.
-
Same Query, Different Answer: Run-to-Run Variance in Semantic Query Processing
This study shows that LLM-based semantic query engines can produce different row-level judgments across repeated runs even at temperature 0. Across 1.2 million calls and several benchmarks, small per-row variation often becomes large answer-level variation, with up to two thirds of UDA-Bench answers changing; voting reduces but does not eliminate the problem, and the authors propose benchmark and engine-design rules.
The broad, multi-benchmark finding that deterministic-looking semantic query systems can have substantial hidden run-to-run variance—and that query plans amplify it—is both non-obvious and supported by unusually extensive measurements, though it is primarily a reliability and benchmarking analysis rather than a new core modeling capability.
-
RSI-Anything: Unlocking Recursive Self-Improvement with World Models
RSI-Anything uses a predictive world model to choose actions and learning data, then selectively retain updated policy–model pairs after checking their compatibility. Across code repair, GUI tasks, laboratory planning, and reaction optimization, it reports improvements such as fewer GUI actions, laboratory-planning scores rising from 21.8% to 46.0%, and 96.15% versus 82.16% yield against Bayesian optimization under a 40-experiment budget.
The domain-general combination of predictive selection, conditional world-model adaptation, and guarded retention is a meaningful RSI direction, but the abstract gives limited detail about baselines, scale, and whether the broad cross-domain gains reflect a genuinely general mechanism rather than tailored workflows.
-
InstantFusion: A Shared Latent Interface for Heterogeneous Image Generators
InstantFusion introduces a shared latent space for connecting the intermediate denoising states of independently trained image generators with different architectures and representations. Lightweight model-specific autoencoders align both latent states and local denoising velocities, allowing generation to switch between frozen models for acceleration, preference composition, distillation, and cross-model image composition without training a separate mapper for each model pair.
The potentially important result is that heterogeneous generators appear to share enough latent spatial structure for a common transition interface, but the abstract gives no quantitative results or evidence that the claimed versatility is more than competitive integration across tasks.
-
Learning to Read the Contextual Tokens in Diffusion Transformers
The paper probes the evolving text-token representations inside multimodal diffusion transformers by mapping them into a frozen LLM, which can answer natural-language questions about the image being generated. It finds that these tokens encode image-specific semantics early in denoising—even with an empty prompt—and proposes Contextual Alignment, a training objective intended to strengthen this representation and improve generation quality and coverage.
The finding that nominally textual contextual tokens become an early, readable representation of the emerging image, plus using that representation as a training target, is genuinely interesting, but the abstract provides few quantitative details and the probing setup may not establish causal interpretability.
-
U-Space: Uncovering When and Why Uncertainty Arises in Language Models
The paper proposes U-Space, a low-dimensional residual-stream subspace built from semantic “certainty” and “doubt” directions. U-Lens projects token states into this space to show how uncertainty changes during generation and to derive a scalar confidence score, without correctness labels, extra sampling, or a separately trained estimator. The authors report better uncertainty prediction than existing baselines on reasoning tasks, including when controlling for answer length, and better transfer than supervised methods.
The interpretable, training-free token-level uncertainty representation is a meaningful idea, but the abstract gives no quantitative results or details establishing that it generalizes beyond the tested models and benchmarks.
-
MIMESIS: Learning User Simulators as Training Environments for Interactive Agents
MIMESIS trains a 9B language model specifically to simulate realistic, less-cooperative users, using human conversation data, reasoning supervision, and 13 behavioral patterns. Agents trained through multi-turn reinforcement learning with this simulator reportedly generalize better across nine unseen user simulators than agents trained with GPT-5.5; a coaching method using the simulator’s reasoning and responses provides additional dense supervision.
The purpose-built, behaviorally diverse user simulator and its reported improvement in cross-simulator agent generalization are genuinely relevant, but the abstract gives limited detail about task gains, evaluation validity, and how much comes from simulator design versus training setup.
-
Can AI Agents Make Open-Ended Scientific Discovery? Evidence from Station
The paper evaluates multi-agent AI systems on open-ended research tasks in Station, an environment designed to support persistent exploration through supervision and periodic reflection. Given only the questions from three ICLR papers, Station agents reportedly rediscovered 62.7% of the papers’ decomposed findings, versus 14.4–20.6% for AI Scientist-v2 and 15.4% for Codex Multiagent-v2; two additional tasks produced some findings later reported by humans. The main contribution is the environment and its exploration mechanisms, though the evaluation is small and partly judged against known oracle results.
The high rediscovery rate and attempt to study discovery without predefined metrics are genuinely interesting, but evidence from only three curated tasks and an oracle-based criterion makes this a promising evaluation rather than convincing evidence of open-ended scientific autonomy.
-
DLoop: Looped Speculative Decoding
DLoop lets a draft model perform several speculative-decoding rounds before asking the large target model to verify, continuing only while its predictions remain confident. A loop-aware training procedure helps the draft model handle its own unverified hidden states; across EAGLE-3, DFlash, Domino, DSpark, and multi-token prediction systems, it reports 5–41% higher wall-clock speedups with lossless generation.
The method addresses a real limitation of speculative decoding and reports broad speedups across several draft architectures, but it is still an optimization of an established inference paradigm rather than a clearly transformative result.
-
Q-Learning with Scalar Adjoint Matching
This paper simplifies reinforcement-learning fine-tuning of flow-based policies by replacing expensive per-step vector–Jacobian products with a scalar adjoint derived from the observed near-diagonal structure of the flow Jacobian. The resulting SQAM method adds value regularization on policy-generated actions, outperforming the strongest baselines by 18–35 percentage points on four difficult OGBench domains and improving supervised fine-tuning on three real bimanual-robot tasks.
The scalar-adjoint approximation is a substantive efficiency-oriented idea for RL fine-tuning of flow policies, supported by both benchmark and real-robot experiments, but the abstract does not establish broad generality or a decisive capability jump.
-
OneSearch-VL: Unified Multimodal Deep Research Agent for Image and Video
OneSearch-VL is a multimodal research agent for answering questions about single images, multiple images, and videos using visual grounding, web retrieval, and evidence composition. Its Visually Grounded Evidence Graph links visual regions and entities to retrieved facts and answer operations, and is used to build training data and an evidence-aware RL reward; the 8B model reportedly outperforms a tool-augmented Qwen3-VL-8B by 20.2 and 17.6 points on its new multi-image and video benchmarks, with gains on existing evaluations as well.
The evidence-graph formulation for preserving visual-to-source-to-answer dependencies is a meaningful, potentially reusable approach to multimodal research, but the strongest results are on newly introduced benchmarks and the abstract gives limited detail about evaluation breadth or independent validation.
-
MC-Sparse: Deconstructing and Closing the Dense-Sparse Attention Gap in Diffusion Transformers
The paper analyzes why sparse attention loses quality in diffusion transformers, attributing the gap to coarse token grouping, inaccurate key-value selection, and discarded attention contributions. It proposes MC-Sparse, a training-free method that selects individual KV tokens, groups queries for GPU efficiency, and caches selection metadata and residual corrections across denoising steps; it reports 1.80× and 2.32× denoising speedups on video and 3D generation with negligible quality loss.
The combination of an explicit error decomposition, token-level sparse selection, and cross-step metadata/residual caching appears practically useful for accelerating diffusion transformers, though the abstract provides limited comparative and quality evidence and the approach is primarily an engineering advance.
-
TestPrism: Rethinking Test Evaluation Beyond a Single Reference
TestPrism evaluates generated tests against many valid and invalid implementations rather than a single reference solution. On 300 tasks and 3,000 implementations, standard evaluation reports 59.67% success while the stricter joint criterion reaches only 28.00%; their TestHelix procedure improves this by about 9 percentage points over native baselines.
The multi-implementation evaluation exposes a substantial, non-obvious overestimation of test quality in coding-agent benchmarks, although the proposed improvement and benchmark remain fairly domain-specific.
-
Embodied Turing Machines: Stateful Code for Robot Recursive Self-Improvement
The paper proposes Code-Only-as-Policy (COAP), where persistent code explicitly tracks robot, environment, and task state from vision and proprioception, then controls the robot without a VLM or VLA during execution. A shared code library is developed and extended across tasks, reaching 70.24% success on 42 bimanual RoboDojo tasks; the authors argue this enables fast, controllable execution and recursive self-improvement, though performance depends heavily on state estimation and code robustness.
The code-as-a-persistent, extensible policy and explicit-state framing is a genuinely interesting alternative to putting a large model in the control loop, but the abstract provides limited evidence about comparisons, generalization, and how much perception or task-specific engineering is hidden in the library.
-
Bellman-Centric Learning: Near-Optimal Regret for Linear Bandits with Memory
This paper studies linear bandits where past actions affect future rewards through a known finite-memory dynamical map. Its RSM-LinUCB algorithm combines bandit estimation with Bellman-style planning and obtains a regret bound in which memory length contributes additively, rather than through the multiplicative dependence seen in prior analyses; a matching minimax lower bound supports near-optimality, with an extension to generalized linear rewards.
The additive treatment of memory, backed by a matching lower bound and broader reward-model extension, is a substantive theoretical result that challenges the apparent necessity of prior memory-horizon coupling, though its impact is mainly within a specialized bandit setting.
-
Ultrasound Operator Guidance Using World Modeling and Retrieval Based Action Planning
The paper treats ultrasound probe guidance as planning in a learned latent world model. It retrieves anatomically similar ultrasound states and uses physically observed probe transitions from a reference database, avoiding both a learned dynamics model and probe-tracking hardware. On carotid ultrasound, it reaches target views in 86% of retrospective episodes versus 52% and 43% for baselines, and 83% in a real-time prospective study on unseen volunteers; it can also navigate back to previously acquired patient-specific views.
This is a genuinely useful combination of image-only sensing, retrieval-based world modeling, and receding-horizon probe planning, with unusually credible prospective real-time evidence, though the demonstration remains limited to carotid ultrasound and a feasibility-scale study.
-
Sharp dimensional analysis of midpoint methods for Langevin sampling
This paper gives sharper complexity bounds for deterministic and randomized midpoint discretizations of Langevin sampling under strongly log-concave targets. It shows deterministic Heun and underdamped midpoint methods achieve improved dimension and accuracy dependence, while matching lower bounds establish optimality in those exponents; surprisingly, randomized midpoint variants suffer a worse high-dimensional lower bound even from favorable initialization.
The stationarity-cancellation and Malliavin-smoothing analysis yields unusually sharp rates and a non-obvious result that deterministic discretization can outperform randomized methods in high dimensions, though the scope is limited to strongly log-concave Langevin sampling.
-
Fine-Tuning a 3B-Parameter LLM on a Smartphone: Characterizing Sustained Training
This paper measures complete LoRA fine-tuning runs of a 3B-parameter language model on an iPhone 17 Pro, including memory, speed, thermals, and energy. It finds that a phone can complete a typical personalization run within one battery charge and achieve comparable personalization to server-trained adapters, but sustained training roughly halves throughput; fixing an MLX backward-pass kernel improved training speed by 1.47× while reducing energy use by a third.
The sustained, end-to-end demonstration of practical multi-billion-parameter on-device fine-tuning, together with evidence that runtime support for backward passes is the main bottleneck, is a meaningful systems result, but it does not yet establish a broad capability or cost breakthrough beyond one phone/model setup.
-
COMPASS 2.0: psychometric representational similarity analysis distinguishes symptom structure from personal signal
Introduces psychometric representational similarity analysis to separate agreement with questionnaire wording and symptom structure from evidence that a language model’s speech-derived scores track the individual speaker. In preregistered discovery and confirmation analyses of 275 clinical-interview participants, derived symptom geometry matched item wording more than self-report, and this apparent agreement persisted even when participants were assigned other people’s answers; person-paired scores reflected general distress more than specific symptoms.
The paper presents a useful, assumption-challenging evaluation framework and credible preregistered evidence that apparent validity of language-based psychiatric assessment can be driven by questionnaire wording rather than person-specific signal, though it is primarily a methodological critique rather than a new model or demonstrated capability.
-
Video Encoders Built on Image Representations
The paper builds a compact video pathway from frozen per-frame image representations. It selects a fixed number of frame tokens using relevance, diversity, and temporal correspondence, then applies lightweight neighboring-frame refinement only to the selected tokens. Across 13 benchmarks and three vision-language backbones, it reportedly matches full image-based video encoding with 28–35% as many visual tokens, and gives a 2.16× end-to-end speedup on Qwen3-VL-32B.
The separation of frame representation, token allocation, and temporal interaction yields a notable efficiency result and challenges the assumption that video must mix frames early, but the abstract does not establish a fundamentally new capability beyond strong benchmark and speed improvements.
-
Towards Decentralized Formation of Minimum-Length Communication Networks Using Robot Swarms
The paper presents a decentralized swarm algorithm that builds and continuously repairs communication networks using only local interactions, without global positioning or centralized topology planning. Robots move and reconfigure links to shorten the network while preserving connectivity; formal analysis gives connectivity and length guarantees, and simulations plus physical experiments report networks close to centrally computed Euclidean Steiner trees, including adaptation to moving targets.
The combination of local-only control, formal network-length guarantees, adaptive deployment, and physical swarm validation is a substantial robotics result, but the abstract does not provide enough quantitative scale or comparisons to establish a must-read breakthrough.
-
Scale-Invariant Training for Time Series Foundation Models
The paper identifies a training mismatch in time-series foundation models: reversing per-series normalization before computing the loss makes high-magnitude series contribute disproportionately large gradients. It proposes computing the loss in normalized space instead, proving that this makes training invariant to arbitrary rescaling for common scalers and losses. Across four model architectures and multiple benchmarks, the method reportedly reduces MASE by about 19–22% in pretraining, with similar gains in supervised forecasting, and requires only a one-line implementation change.
The paper offers a clean theoretical explanation for a likely widespread but underexamined training failure and reports sizable, cross-architecture gains from a trivial change, though the abstract alone does not establish how robust the results are across datasets and implementations.
-
ScanSTL: Parallel Robustness Evaluation for Signal Temporal Logic
ScanSTL computes exact Signal Temporal Logic robustness traces and gradients using parallel scans and compact reductions rather than sequential temporal recurrences. The JAX implementation reports large speedups over STLCG++—243× for forward evaluation and 104× for gradients on CPU across nine specifications—and sub-millisecond evaluation of long Until formulas on an RTX 5090, with simulated robot planning showing faster repair and greater MPC capacity.
The combination of exact differentiable STL evaluation with logarithmic parallel depth could materially expand optimization-based robot planning, but the evidence is primarily operator benchmarks and simulations rather than demonstrated real-robot or end-to-end planning gains.
-
Two-Sample Testing for Random Graphs without Vertex Correspondence
This paper analyzes how many independently sampled graphs are needed to distinguish two graph populations when their vertices cannot be matched. In a degree-preserving planted-block alternative, it proves that signed triangle statistics need roughly t^-3 graphs per group, versus t^-1 with vertex alignment; tree-based statistics, including degree features and finite message-passing GNN representations, can be asymptotically powerless. It also shows that when triangle signals cancel, 4-cycles determine a still worse t^-4 rate, while generic differences under a non-constant null can be detected at the aligned rate.
The theoretically sharp result that common tree-based and message-passing features miss certain unaligned graph differences, while small motifs recover them at substantially different sample-complexity rates, is a useful and non-obvious warning for graph comparison and GNN evaluation, though the setting is specialized and the evidence is primarily theoretical.
-
Is $\sqrt{d}$ Separation Necessary for Gradient EM to Learn Gaussian Mixtures in High Dimensions?
This paper proves a near-tight lower bound for learning high-dimensional Gaussian mixtures with gradient EM. In a worst-case construction, even using extra mixture components and population-level updates, random initialization cannot guarantee convergence in sub-exponential time when component separation is only on the order of d^(0.5−ε); separation scaling like √d is therefore essentially necessary for this class of methods.
The result gives a theoretically meaningful, nearly matching barrier that challenges the hope that over-parameterization alone removes the √d separation requirement, though it is a worst-case population result for a particular hybrid EM variant rather than a broad practical learning result.
-
High-dimensional online calibration from harmonic weights
This paper gives an online forecasting algorithm for multidimensional outcomes that achieves calibration in roughly d^{O(1/ε)} rounds, improving prior bounds that had substantially worse dependence on dimension and accuracy. The method uses harmonically weighted, smoothed past outcomes and is claimed to work uniformly across convex forecast sets and error norms, with the guarantees tied to an optimal discrepancy property of the discrete Hilbert transform.
The claimed improvement from roughly exponential-in-1/ε-squared dependence to d^{O(1/ε)}, together with a norm-independent harmonic-weight construction, is a substantive theoretical advance, though the abstract alone does not establish how broadly practical or foundational the result will be.
-
Seeing Through the Displaced Frame: Privileged Noise Distillation for Vision-Force Precision Assembly
The paper addresses a failure mode in precision assembly where pose noise shifts both the robot’s observations and the action coordinate frame, making the error unidentifiable before contact. It trains a deployable vision-force policy using privileged access to the frame offset, or by relabelling demonstrations into the displaced frame; on FORGE it retains 92–99% success under 5 mm noise versus 32–60% for the state-based policy, and transfers to a Franka with 83.3% success versus 34.4%.
The compelling contribution is not another sensor or policy architecture but showing that supervision encoding compensation for a latent action-frame error is crucial, with large robustness gains and a real-robot result, though evidence is concentrated on one benchmark and task family.
-
Stochastic Gradient Descent Ascent is Suboptimal for Nonconvex-PL Min-Max Games
This paper proves tight lower bounds for two-timescale stochastic gradient descent ascent (SGDA) on smooth nonconvex–Polyak–Łojasiewicz min-max games, matching known upper bounds and showing a worse condition-number dependence than Smoothed-AGDA. It also shows that SGDA may fail to reach a stationary point when the inner-loop timescale is too slow, specifically for ratios below order κ².
The tight lower bound and explicit failure regime establish a meaningful limitation of a standard min-max optimizer and motivate alternative algorithms, though the impact is primarily theoretical rather than a demonstrated capability gain.
-
Visual Abstention in Unified Multimodal Models
The paper studies whether unified vision-language models can recognize when a requested image transformation is impossible and decline instead of hallucinating a result. On the new Draw-or-Decline benchmark, existing models almost never abstain from infeasible edits; the proposed paired feasible/infeasible training method raises refusal accuracy to 93.0% while maintaining 74.3% feasible-edit accuracy and only 0.8% false refusals.
The large, previously under-measured gap between visual editing ability and feasibility recognition, plus strong abstention gains without an editing tradeoff, is genuinely interesting, but the evidence is still centered on one benchmark and one trained model.
-
Detecting a Shift Is Not Enough: Exact Minimax Limits of Linear Representation Repair
This paper formulates removal of a mean distribution shift as a minimax decision problem: learn one linear transformation from noisy paired calibration data while obeying a distortion budget. It derives an exact finite-sample residual-shift risk and shows a sharp detection–repair gap: detecting a shift requires signal-to-noise ratio only on the order of sqrt(d), while removing a fixed fraction requires order d, essentially because its direction must be estimated. The result also predicts the limits of standard linear concept-erasure methods and is illustrated on participant/device shifts in sleep EEG.
The exact minimax characterization and detection-versus-repair separation provide a useful, non-obvious theory for when representation correction is fundamentally data-limited rather than algorithm-limited, though the practical evidence is limited and the setting is specialized.
-
Context-Conditioned Hamilton-Jacobi Reachability for Adaptive Safety Filtering
The paper learns a single Hamilton–Jacobi reachability value function conditioned on local boundary geometry, tire friction, and disturbance magnitude, rather than synthesizing a separate safety certificate for each deployment context. On 11 unseen geometries and friction coefficients from 0.4 to 2.0, it matches geometry-specific certificates closely, and a full-scale vehicle maintained lateral containment near the handling limit during adversarial driving and autonomous racing, reaching 0.99 g acceleration.
The combination of context-conditioned reachability, transfer to unseen geometries and grip conditions, and full-scale near-limit vehicle validation is a meaningful step beyond conventional per-context HJ synthesis, but the demonstrated scope remains limited to one vehicle model and a modest set of geometries.
-
Compact Robot Policies Need Fine-Grained Visual Representations
The paper studies whether large multimodal or generative priors are actually necessary for multitask robot manipulation, and introduces CoRP, a 48.9M-parameter policy using a pretrained, trainable visual encoder plus a flow-matching action generator. CoRP reportedly matches much larger systems on LIBERO and RoboTwin 2.0, while controlled ablations show that pretrained initialization, task adaptation, and a hard budget of 48 visual tokens matter more than retaining all patch tokens or adding a variational bottleneck. It also finds that language helps mainly when the image does not uniquely specify the goal.
The controlled evidence that compact policies can match much larger robot policies—and that hard visual compression can outperform richer token streams—is a non-obvious, practically important finding, though it is still based mainly on benchmark success rates rather than broad real-world validation.
-
Neural Petri flows for chemical reactions
The paper derives a neural architecture that preserves Petri-net semantics—local enabling, nonnegative marking, and conservation—for every setting of its learned weights, while learning only reaction rate laws or readouts. Its Neural Petri Flow model uses a single firing vector for atom mapping, reaction classification, and product prediction, reporting strong results across reaction datasets, including 87.7% product prediction on USPTO-480K and 90.5% valid top-1 elementary-step predictions on FlowER without post-hoc filtering.
The combination of a formal proof that the architecture retains chemical Petri-net semantics and a unified, constraint-valid reaction model is genuinely interesting, but the abstract does not establish that the reported gains generalize beyond curated benchmarks or materially outperform modern reaction models across broader settings.
-
CARE: Certifying Acceleration for Vision-Language-Action Inference
CARE selects accelerated VLA inference methods using paired closed-loop rollouts, certifying with finite-sample confidence that the accelerated policy rarely fails on episodes the original policy would solve. It chooses the fastest candidate within a user-specified failure budget and reports 9.0–10.8× certified speedups on LIBERO while preserving at least 85.8% of reference-solved episodes, with similar results across other agents and tasks.
The notable contribution is a practical statistical certification framework for acceleration-induced closed-loop failures, with substantial measured speedups and cross-model validation, though it is primarily a selection and evaluation methodology rather than a new VLA capability or accelerator.
-
Spatial Induction Heads: In-Context Learning of Multidimensional Cellular Automata
This paper studies how transformers learn local rules in multidimensional cellular automata when the grid is flattened into a sequence, so spatial neighbors are far apart in token order. It proposes a two-layer “spatial induction head” circuit: one layer gathers the relevant spatial neighborhood and another matches it to prior examples, implementing Bayesian prediction; trained two-layer transformers generalize to unseen rules with near-perfect deterministic rollouts and near-Bayes-optimal stochastic predictions.
It offers a concrete mechanistic extension of induction-head theory to noncontiguous spatial contexts, with explicit constructions, Bayesian guarantees, and strong synthetic experiments, but its broader importance beyond cellular automata is not yet established.
-
LayerRoPE: Dynamic Depth-wise Magnitude & Angular Superposition
LayerRoPE treats the growth and rotation of Transformer residual-stream normalization weights across depth as a learned positional signal, analogous to positional encoding across sequence positions. It replaces separate per-layer normalization gains with a shared vector plus depth-dependent magnitude and angle, and reports better optimization, lower compute to reach comparable loss, improved learning-rate robustness, and scaling to very deep networks, looped latent models, and Vision Transformers.
The paper proposes a genuinely different interpretation and parameterization of depth normalization, backed by broad cross-architecture and depth-scaling experiments, but the unusually strong claims need verification from detailed results and ablations before warranting a strong recommendation.
-
On-Policy Distillation Teaches New Skills but Not New Knowledge
The paper separates factual knowledge transfer from multi-step reasoning skill in on-policy distillation using controlled synthetic tasks and tests the result on factual QA and competition mathematics. It finds that reverse-KL distillation transfers compositional reasoning across unseen structures but adds little new factual knowledge; forward KL improves factual transfer, while student rollouts mainly help execute multi-step reasoning.
The controlled separation of knowledge acquisition from reasoning organization, together with the finding that KL direction determines what distillation transfers, is a useful and non-obvious result, though the evidence is still primarily model- and benchmark-based rather than a demonstrated broad training principle.
-
RealtimeWAM: How Fast Can I Run My World Action Model?
RealtimeWAM is a training-free inference framework for speeding up visual world-action models during closed-loop robot control. It overlaps observation and prediction processing, caches features and Transformer residuals when observations or predicted actions change little, and reports 8.9–10.7× lower latency on FastWAM and OpenWAM while nearly preserving simulated success; in five real-world tasks, it even improves success over native inference.
The combination of adaptive computation and parallel execution produces unusually large latency reductions with reported real-robot gains, but the evidence is limited to two WAMs, three simulation suites, and five real-world tasks, so it is promising rather than clearly a field-changing result.
-
ResidualQuant: KV Cache Quantization for Looped Transformers with 2-Bit Residuals
ResidualQuant compresses KV caches in looped Transformers by storing the final loop’s KV states as a reference and encoding earlier loops as low-precision residuals. With rotations, least-squares scaling, and loop-dependent precision, it reaches 2-bit residual quantization while staying near BF16 accuracy, reporting 80.7% theoretical KV-storage reduction and up to 2.73x decode or 4.15x peak throughput improvements on an RTX 5090.
The cross-loop residual representation is a specific but genuinely useful idea for reducing the unusual KV-cache bottleneck of looped Transformers, with sizable reported memory and throughput gains, though the benefit appears specialized and the evidence is limited to the abstract’s benchmarks and hardware results.
-
RFPO: Rectified Flow Policy Optimization for Embodied Control
RFPO trains rectified-flow robot policies specifically to remain reliable when their action trajectories are integrated with very few numerical steps. Across four robot platforms, including real-robot locomotion, a one-step policy stayed within 2.4% of 64-step performance; on Go2 it achieved 98.5% of the reward with 54.9× lower inference latency.
The combination of reward-aware online reflow and auxiliary PPO supervision addresses a real deployment bottleneck, and the large, cross-platform one-step speedup with real-robot validation is substantially more notable than a routine policy optimization tweak.
-
LLA-MPPI: Rapidly Adaptive Whole-body Control of Legged Robots with GPU-Accelerated Parallel Simulations
The paper introduces an adaptive whole-body MPPI controller that selects among GPU-parallelized contact simulators by comparing recent prediction errors, then plans using the selected model. It adapts online without offline training and handles changing payloads, altered dynamics, and a disabled leg; in simulation it reaches 97.5% task success versus 74% for the strongest baseline, with hardware demonstrations on a Unitree Go2.
Online, physically interpretable model selection for contact-rich whole-body control, backed by a large simulation gap and several hardware adaptations, is meaningfully more than a routine controller tweak, though the evidence is still limited to four simulated tasks and demonstrations rather than broad real-world evaluation.
-
Why Forget-Only Unlearning Needs Memorization
This paper studies whether a model can support deletion requests when the unlearning algorithm sees only the trained model and the examples to remove, not the retained training data. It proves that this is sometimes impossible because different training sets can yield the same model but require different post-deletion models, and shows that successful forget-only unlearning may require memorizing nearly the entire training set—even for simple threshold learners that normally retain only a single boundary point.
The paper gives a clear information-theoretic explanation for why forget-only unlearning can require substantially more retained information than ordinary training, with sharp lower bounds and a striking threshold-learning example.
-
Listen-to-Reason: Listen with Experts, Retrieve over a Graph, Reason with LLMs
Listen-to-Reason (L2R) routes audio to a text-only LLM through explicit, human-readable semantic nodes retrieved from frozen speech, music, and environmental-sound experts, rather than coupling the audio encoder directly to the LLM. The nodes support tracing and causal intervention: replacing the deciding node changes 78% of correct SAKURA answers. L2R uses about 1,400× fewer trained parameters and much less audio data than comparable LALMs, wins on SAKURA, trails by 6–12 points on MMAU/MMAR, and adapts to a new domain using only a small head and five labeled clips per species.
The explicit semantic retrieval interface offers a genuinely different, interpretable and highly data-efficient alternative to end-to-end audio-language models, but its mixed benchmark results and limited abstract-level evidence make this promising rather than a must-read.
-
NavGPT-3: Harnessing Context in a Hierarchical Navigation Runtime
NavGPT-3 introduces an OS-like runtime that coordinates a long-horizon reasoning model, a fast visual-language-action policy, and monitoring/tool threads. The runtime can interrupt reasoning and switch control to the action policy, reducing reaction latency from several seconds to roughly 0.5–1 second; on R2R-CE and RxR-CE, the full system reports 81.51 SR and human-level RxR-CE performance, respectively. The paper also proposes scene-change-based visual token allocation for its 8B action policy and provides extensive ablations.
The explicit runtime interface between deliberative language-model reasoning and reactive control, together with reported human-level results on challenging navigation benchmarks, is genuinely interesting, though the evidence appears benchmark-based and the gains may partly reflect scale and engineering rather than a fundamentally new control paradigm.
-
Learning How to Search for Plans with Exponentially Less Space
The paper proposes learning domain-specific search-control policies with registers, modes, and explicit choice points. If the policy is structurally terminating, execution takes polynomial space and the only exponential factor is the number of genuine choices, avoiding storage of visited states; an LLM learns such policies through counterexample-guided refinement. On 1,890 planning tasks, the resulting solver solves 1,709, outperforming several established planners while usually staying under one second and 100 MiB.
The combination of a polynomial-space planning framework, a meaningful notion of choice depth, and LLM-guided policy synthesis is a substantial and nonstandard direction, but the abstract does not establish how broadly the formal guarantees or empirical gains generalize beyond the reported planning suites.
-
Fluid-Gen-Zero: Grounding Pretrained Video Generators in Physics without Training
Fluid-Gen-Zero uses a physics simulator to determine fluid and object motion, then injects those signals into pretrained video generators without retraining them. A VLM organizes the simulation and video-generation process, while region-aware latent guidance aligns denoising with the simulated dynamics. Across three video backbones, it substantially reduces object-trajectory and fluid-flow errors while mostly preserving visual quality.
The training-free separation of physics simulation from appearance generation is a credible and fairly general direction, supported by improvements across multiple backbones, but the contribution appears specialized to fluid-object videos and may rely on a newly introduced benchmark.
-
Read What Matters: Query-Adaptive Quantization for KV Caches
ReadKV stores KV-cache entries in progressive codes and lets each decoding query choose where to spend a limited bit-read budget. It first allocates precision to keys based on the query, uses the resulting attention to allocate value precision, and proves that query-dependent reading can outperform any query-independent reader in a constructed setting. Across six models, it reports low perplexity degradation with substantially fewer logical reads and, in a limited 8K-token A10G test, 39% lower latency than TurboQuant.
Query-adaptive access to immutable, progressively coded KV caches is a genuinely interesting direction with theory and promising efficiency results, but the systems evidence is still narrow—especially the single-layer, batch-one latency evaluation—so it falls short of strong.
-
WAM-Cache: Staleness-Bounded KV Reuse for Efficient World Action Models
WAM-Cache accelerates closed-loop World Action Models by reusing layerwise key-value representations from prior video observations and selectively refreshing only some tokens. It finds that refreshing tokens based on visual motion—even with oracle knowledge of KV drift—is a poor proxy for preserving action quality; selecting tokens using the action expert’s attention, visual surprise, and an age bound reduces video-DiT prefill FLOPs by 32–42% while keeping performance close to the dense baseline across simulation and a real robot.
The action-conditioned refresh criterion and the finding that visual KV drift does not predict downstream control quality are non-obvious, with substantial compute savings and validation across several benchmarks plus a real robot, but the gain is an optimization of an emerging architecture rather than a clearly transformative capability jump.
-
Reliability-Aware Future Conditioning for Temporally Robust Robot Manipulation
The paper shows that generated future videos can hurt robot manipulation when their timing is misaligned with the robot’s actual phase: on CALVIN, small shifts reduce performance to below a policy that uses no future. Reliability-Aware Future Conditioning evaluates nearby temporal hypotheses at each step, downweights unreliable clips, and can fall back toward a static policy without alignment labels; it improves robustness under artificial shifts and raises success from 26.7% to 56.7% on a Franka under naturally occurring timing mismatch.
It identifies a consequential failure mode for video-conditioned robot control and provides a reward-trained reliability mechanism with substantial simulated and real-robot recovery, though the evidence is limited to the evaluated task sets and builds on an existing future-conditioning system.
-
Training on the Future: A Delay-Aware Audit of Test-Time Adaptation for Time-Series Forecasting
This paper audits test-time adaptation for time-series forecasting under realistic label delays, enforcing that an H-step forecast cannot be updated until its target is actually available. Across four published TTA methods and five benchmarks, it finds that apparent gains are often inflated by leaky next-step updates, simple recursive least-squares or linear correctors frequently outperform learned adapters at far lower cost, and delayed adaptation can become harmful on drift-heavy data. It also shows that the backbone training recipe can affect frozen-model error far more than adaptation does, and releases a leakage-free evaluation harness and cached results.
The work provides a broad, reproducible reality check that challenges common TTA evaluation practice: causal delays, simple baselines, and backbone choices dominate many claimed adaptation benefits, though it is primarily an audit rather than a new modeling direction.
-
Rounding in Preconditioner Space: Redesigning 4-bit AdamW Optimizer-State Quantization
This paper argues that quantizing AdamW’s second moment in ordinary state space can produce disproportionately large errors in the adaptive preconditioner, especially near zero. It introduces ZIP-SR, which performs stochastic rounding in preconditioner space while preserving zero, and ZE-EDEN, which calibrates a zero-excluding codebook; across 130M–2.7B-parameter GPT- and Llama-style training runs, these methods reduce the validation-loss gap between 4-bit and full-precision AdamW, with a maximum reported reduction of 70%.
The preconditioner-space analysis and corresponding 4-bit optimizer-state recipes are a substantive, plausibly general improvement to memory-efficient training, supported by experiments across several model scales, but the abstract does not establish a decisive capability or efficiency breakthrough over all relevant baselines.
-
Critical Mass: Structure-Disjoint Splits Drift With the Libraries They Are Built On
The paper scales the standard MCES-based structure-disjoint split for molecular machine learning from roughly 29,000 to 227,690 molecules using admissible bounds and union-find shortcuts, while preserving the exact partition. At this scale, the similarity graph percolates: one connected component contains 95.5% of molecules, and the remaining held-out compounds are systematically larger and more complex, exposing that the fixed distance threshold mixes structural novelty with size extrapolation.
It identifies a substantial, non-obvious failure mode in a widely used OOD evaluation convention and provides an exact scalable remedy, but the result is mainly specific to molecular benchmark design rather than a broad new ML capability.
-
In-Context Learning Can Help Vision-Language Models Overcome Training Priors
The paper tests whether vision-language models fail on images that conflict with learned statistical priors because they lack visual ability, or because they are not using available visual evidence. Across five visual tasks and five VLM families, a few controlled visual demonstrations substantially improve performance on counterfactual, prior-conflicting examples—by 9.7–30.3% overall and 27.2% for Qwen3-VL 235B—while barely changing prior-aligned accuracy. Representation analyses suggest demonstrations shift attention toward grounding cues, but they cannot compensate for genuinely missing visual capabilities.
This provides a broad, non-obvious diagnosis that some apparent VLM capability failures are failures to override training priors, with a simple in-context intervention producing large gains, though it is primarily an empirical finding rather than a fundamentally new model or capability.
-
Tabular Foundation Models Are Effectively Shallow
The paper tests whether tabular foundation models really need their many transformer blocks. Using a closed-form, label-free affine replacement, the authors remove up to 96% of the blocks across eight models and 171 classification datasets, finding that half-depth models usually lose almost no AUROC and can be substantially faster and smaller. The results also suggest that different model families concentrate useful computation in different parts of the network, with some retaining most performance from a single block.
This is a broad, non-obvious efficiency result with a simple compression method and unusually extensive cross-model evaluation, but it is still focused on tabular models and does not clearly establish a general principle beyond that setting.
-
Taking Tokenization off the Host for Agentic LLM Serving
GraphTok moves the full text-tokenization pipeline from the CPU to a GPU, implementing pre-tokenization, BPE merging, and token-ID assembly in a replayable CUDA graph while preserving exact reference outputs. It reports 4.7–6.0× faster tokenization than gigatoken on Llama 3 and 2.1–2.4× lower time-to-first-token for cached-prefix requests in vLLM, with much higher frontend capacity under tail-latency constraints.
The paper identifies and substantially improves a usually neglected serving bottleneck, showing large end-to-end capacity and latency gains rather than merely a tokenizer microbenchmark, but it remains a systems optimization with results limited to the reported workloads and hardware.
-
Dream to Remember: Continual Steering of Generalist Robot Policies with World Models
D2R continually adapts a frozen generalist robot policy to new tasks by training a steering module, while a frozen world model generates imagined rollouts of earlier tasks to prevent forgetting. On CALVIN, it improves eight tasks from 63.7% to 74.8% without new environment data, versus 11.7% after LoRA fine-tuning; it also adapts to unseen LIBERO tasks and demonstrates sequential improvement on a real Franka robot while preserving prior skills.
The combination of world-model rehearsal with frozen-policy steering offers a credible and non-obvious solution to continual learning without replay data, with results spanning simulation, unseen tasks, and a real robot, but it remains a specialized method whose broader robustness depends on world-model fidelity.