Issue 36 · Aug 31 – Sep 6, 2026
Week 2026-W36
2,798 papers scanned 150 shortlisted 10 picked $13.40 spent
A week of assumption-breaking results: random KV-cache eviction matches learned importance scores, NAVSIM driving scores turn out not to require seeing the scene, and LIBERO falls to a 0.54M-parameter policy that mostly memorizes task IDs. Alongside those, NVIDIA reports the first AI to outscore the top human at an IOI, two independent groups crack constant regret in general games, and a rodent study argues V1 is driven more by internal models than by vision. On-policy distillation gets a cluster of papers this week; the sparse-supervision result (C6) is the one to read, with C10 and C25 as context.
-
Post-Training Language Models for Gold-Medal Performance in Coding Competitions
A prospective, contemporaneous evaluation at IOI 2026 in which a post-trained 550B system scores above the highest human contestant under matched time and submission constraints. If the evaluation conditions hold up, this is the clearest competitive-programming capability milestone since gold-level performance was first claimed, and the pipeline (curation, SFT/RL, feedback-driven refinement) is documented end to end.
Look for Scrutinize the 'same constraints as contestants' claim—compute budget, sampling breadth, and how GenCorrect's iterative refinement uses the public test feedback compared to what humans get.
-
Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning
Directly attacks the core assumption behind an entire subfield: KV-cache eviction for reasoning needs no importance score at all—protect the prompt, evict uniformly at random, and you match the best evictors with 32–43% higher vLLM throughput. The mechanistic explanation (redundancy in the trace text and across heads) is as valuable as the result and likely to reshape how people design cache compression.
Look for Check whether the redundancy argument holds for tasks where the reasoning trace doesn't restate intermediate results (e.g., code with long-range references) and how far the compression ratio was pushed.
-
Driving on Memory
Replacing the camera input with memories of prior drives at the same location nearly matches leading end-to-end methods on NAVSIM—meaning a high score there does not require reacting to current traffic. This is a well-executed, experimentally supported invalidation of how a widely used driving benchmark is interpreted, with the useful contrast that Bench2Drive and RealEngine do not collapse the same way.
Look for Look at how the retrieved memories are matched to locations—if retrieval leaks any dynamic information the headline claim weakens; also check the magnitude of the residual gap to camera-based models.
-
Extremely Sparse Supervision Incentivizes Reasoning Ability
Supervising one or two tokens per trajectory—0.05% of generated tokens—matches or beats full-token on-policy distillation across nine teacher–student configurations, plus Llama and PPO-based RLVR. If robust, this overturns the assumption that post-training gains require dense token-level feedback and suggests reasoning improvement is driven by a handful of critical decision points.
Look for Check how the supervised tokens are selected (if selection needs teacher signals on all tokens, the compute savings may be smaller than the headline) and whether absolute gains match dense training or just relative recovery.
-
Constant regret in general games via higher-order optimism
An uncoupled algorithm guaranteeing O(N^3 log^2 K) regret uniformly over the horizon in arbitrary finite games—removing the polylog-in-T dependence that was widely believed hard to eliminate. Together with the concurrent independent result (C5), this looks like a genuine resolution of a live theoretical question, and the higher-order-optimism-plus-discounting mechanism is new.
Look for Verify the theorem's conditions (full-information feedback, all players running HOOD) and compare constants and mechanisms against the concurrent Liu–Farina–Ozdaglar paper to see which version is cleaner.
-
MINERVA: How Small Can a Manipulation Policy Be and Still Solve LIBERO?
A 0.54M-parameter policy hits 95.1% on LIBERO—7,700× smaller than π0.5 and 1,400× faster on a laptop CPU—and a task-ID permutation probe shows standard LIBERO conditioning mostly selects among memorized tasks. This simultaneously questions what billion-parameter VLA scores on this benchmark demonstrate and gives a useful capacity-floor measurement, with unusually broad ablations (flow matching adds nothing over L1 regression).
Look for Note the LIBERO-Plus collapse to 46–56% and near-zero photometric robustness—the argument is about the benchmark, not that tiny policies generalize; check whether large VLAs pass the same permutation probe.
-
Evidence Integration in Large Language Models
A distributional theory of how LLMs integrate external evidence, validated over ten million trials and twelve models: models preferentially accept candidates resembling their own likely errors, identical evidence helps weak models and hurts strong ones, and models integrate candidates they internally represent as invalid up to 99% of the time. This receiver-relative account challenges simple trust-based views of RAG and multi-agent pipelines and comes with causal interventions.
Look for Check how 'internal verification' is measured—if it's probe-based, the dissociation from behavior could partly reflect probe artifacts—and whether the theory's predictions hold out-of-family.
-
A cognitive representation in primary visual cortex modulated by vision
In rats navigating in darkness, many V1 neurons encode progress through a shared turn sequence across physically distinct paths—an internally generated, path-invariant code—with visual input merely gain-modulating it when available, and coordination with hippocampal ripples. This inverts the standard driver/modulator picture of primary sensory cortex and is directly relevant to predictive-processing and world-model framings in AI.
Look for Be skeptical of how 'path-invariant' is defined—check controls for locomotion, whisking, and other behavioral covariates that could masquerade as an internal sequence code.
-
TacPAC: Tactile Prediction and Real-Time Action Correction in World-Action Models for Contact-Rich Manipulation
A clean solution to a real timing problem in contact-rich manipulation: cache the predicted contact the plan was conditioned on, then let a tactile expert correct the unexecuted portion of the action chunk against that expectation, 20.7× cheaper than replanning. Real-robot success rises from 22% to 64% across five tasks including precision insertion and fragile handling, and the finding that naive tactile prediction recovers only a third of the gain is itself informative.
Look for Check the vision-only baseline's strength (22% is low) and whether the correction module was tuned per-task; the ablation separating prediction from real-time correction is the key table.
-
Cheap Verifiers, Large Blind Spots: Measuring the Reliability Cost of Cost-Saving Cascades
A systematic demonstration that verifier-based cascades and self-improving loops can be blind to their own degradation: verifier blind spots grow adversarially with student capability, fine-tuning on the rejected tail collapses the student across every teacher tried, and dashboard metrics read flat 3% error while true error swings to 32%. Given how many current systems rest on cheap-verifier architectures, this failure taxonomy plus the conservation-law theory is worth internalizing.
Look for Check which model scales and task families the collapse result covers—self-distillation loops do work in some published settings, so identify what differs (tail-only training, verifier quality, or scale).
Also notable
-
Constant Individual Regret in General Games
AI / ML
The concurrent, independent constant-regret result (ECHO-OFTRL with an EMA cascade)—read alongside C3 to compare mechanisms; its bound is O(N^21 log^4 K) versus C3's N^3.
-
The Head Complexity of Boolean Functions in Single-Layer Attention
AI / ML
Exact head-complexity hierarchy for single-layer attention: k heads compute k-bit parity but not (k+1)-bit, unconditionally in dimension and precision—a clean addition to transformer expressivity theory.
-
Rapid spatial cognition in mice, with and without neocortex and hippocampus
Neuroscience
Mice lacking hippocampus and most neocortex still learn, retain, and generalize maze routes nearly as well as wildtypes—an assumption-challenging result on where rapid spatial learning lives, pending lesion-validation details.
-
Microglia drive demyelination via multiple sclerosis antibodies and BTK signaling
Neuroscience
In vivo two-photon evidence that microglia actively destroy intact myelin under MS patient antibodies via BTK signaling, overturning the debris-scavenging view and pointing to an actionable intervention.
-
Does On-Policy Distillation Really Distill? From Noisy Teacher to Self-Improvement
AI / ML
Argues on-policy distillation works largely by suppressing low-probability tokens with no teacher needed—complements C6's sparse-supervision result; the two together suggest OPD's mechanism is poorly understood.
-
Unlocking Lossless Speedups in LLMs via Discrete Diffusion
AI / ML
Diffusion-augmented AR LLMs (Uno) draw multiple tokens in parallel from the exact AR distribution without a draft model—up to 3× throughput; worth watching if the losslessness claim holds.
-
Can Video World Models Track Unobserved World States?
AI / ML
An action-conditioned shell game shows video world models render plausible frames while tracking hidden state at chance; only architectures with a revisable recurrent state extrapolate—useful diagnostic for world-model claims.
-
SMELT: Scaling Laws for Compute-Matched MoE Looped Transformers
AI / ML
Compute/parameter/KV-matched scaling laws showing looped middle layers save 6.8–18% training FLOPs up to 54B, with an attention-sink explanation—the fairest evidence yet that depth reuse genuinely helps.
-
Why Gated DeltaNet Survives 4-Bit Quantization: NVFP4 W4A4 for the Recurrent Half of a Hybrid 27B LLM
AI / ML
Counterintuitive finding that the recurrent Gated DeltaNet layers of a 27B hybrid survive full W4A4 NVFP4 quantization because state overwrites forget injected noise—a useful mechanism study for hybrid-model serving.
-
How Output Format Confounds Data Quality and Capability in Instruction Tuning
AI / ML
Shows measured data quality and even model capabilities are conditioned on output interface—a skill can gain 40 points under the training format and vanish under others; a caution for evaluation design.
-
Hakken: Predicting future discoveries to fill the gaps in today's knowledge
AI / ML
Knowledge-graph-plus-LLM discovery system whose predictions were taken to wet lab, confirming two previously undocumented protein interactions—thin abstract detail, but end-to-end AI-to-experiment loops are rare.
-
The Missing Temporal Link: Temporal Context Routing for Script-Driven Audio-Video Generation
AI / ML
Temporal Context Routing gives joint audio-video generators precise script-level timing control (shot-boundary MAE 1.11s → 0.042s)—a plausibly general mechanism for controllable AV generation if the baseline comparisons hold.
Projects
A strong week for genuinely new capabilities: peer-to-peer browser inference of large models, unified audio understanding-and-generation, a novel sparse-attention MoE family with 512K context, and a distilled character-animation model with an unusually clean interface. Robotics and world-model releases are plentiful but mostly early; agent tooling remains a crowded field of integrations.
-
Nehanth/swarmllm
Running a 27B model at ~10.7 tok/s split across a MacBook and an iPhone in browser tabs, with a from-scratch WebGPU engine and WebRTC layer partitioning, is a genuinely new deployment capability, not a wrapper. It demonstrates that install-free, heterogeneous-device inference of serious models is practical, which could reshape how local LLMs get distributed.
Look for Check how throughput degrades over real internet links (vs. same-room WiFi) and whether the speculative decoding trick holds beyond the single demo pairing.
3 min read ·GitHub ↗ ·JavaScript·MIT
-
hugging-apps/fireredaudio-tts-demo
A single 9B audio language model doing zero-shot cloning, voice design, speech editing, ASR, and audio QA, with a specific decoupled continuous-representation architecture (AR backbone plus flow-matching decoder), squarely hits the reader's omni-audio interest. The runnable demo makes it more than an architecture announcement.
Look for Verify whether weights and a license are actually released beyond the Space, and how editing/cloning quality compares to specialized TTS systems.
3 min read ·Hugging Face ↗
-
IFM/K2-Horizon-MoVA-36B-A4B
Mixture-of-Values attention is a potentially important new sparsity axis beyond FFN experts, and the claim of frontier-level agentic performance at ~4B active parameters with native 512K context would be a large efficiency jump if it holds. Apache-2.0 plus promised training data, recipes, and intermediate checkpoints makes this the most consequential model release of the week.
Look for Independently verify the benchmarks and confirm the promised training transparency actually ships; runtime support (llama.cpp fork) is still immature.
3 min read ·Hugging Face ↗ ·apache-2.0
-
pollen-robotics/microduck_rl
An unusually complete sim-to-real release: GPU-accelerated MuJoCo Warp environments, actuator/backlash modeling, domain randomization, multi-skill switching, and ONNX export to a real 800g biped that demonstrably recovers, skates, and kicks. This is the kind of end-to-end deployable robotics stack that rarely gets open-sourced in full.
Look for Check how much transfers without the specific Microduck hardware and whether the recipes generalize to other small bipeds.
3 min read ·GitHub ↗ ·Python·Apache-2.0
-
Viggle/Viggle-Animate
Removing pose estimation, segmentation, tracking, and text conditioning in favor of a single repainted first frame is a genuinely new and simpler interface for character replacement in video, and DMD distillation to three forward passes for 124-frame clips is a real efficiency result. The reader tracks video generation closely and this is the strongest new idea in the category this week.
Look for Note the nonstandard license, reported identity drift, and very heavy compute (73.5GB bf16 weights); the ComfyUI int8 conversion (P68) is the practical entry point.
3 min read ·Hugging Face ↗ ·other
-
Thinking-Space/One-Shot-OPD
The finding that on-policy distillation is 'data-overfed but algorithm-starved'—a single query recovering most of full-data gains, 16 diverse queries matching it—overturns a common assumption about post-training data requirements. It comes with runnable veRL code and a state-coverage analysis explaining why, which is exactly the kind of result the reader wants.
Look for Reproduction requires substantial GPU resources and rebuilt datasets; check whether the effect holds across teacher-student pairs beyond the reported setups.
3 min read ·GitHub ↗ ·Python·Apache-2.0
Also notable
-
inclusionAI/LLaDA2.2-mini
HF model
A 16B MoE diffusion LM that does Levenshtein-style INSERT/DELETE editing during generation is a genuinely unusual direction, though early benchmarks are mixed and adoption is near zero.
-
Qwen/Qwen-Drive-1.0-4B
HF model
Qwen-Drive attaches BEV perception and a flow-matching planner to an untouched 4B VLM with open weights and closed-loop results—a clean, inspectable recipe for driving stacks.
-
OpenWAM-Official/OpenWAM
GitHub
OpenWAM releases a world-action model pretrained on 518M egocentric human and robot frames with systematic ablations, worth watching once performance evidence firms up.
-
heygen-com/hyperframes
GitHub
HyperFrames' deterministic HTML-to-MP4 rendering with agent-facing skills is plausibly becoming the default way agents produce video, judging by its explosive adoption.
-
TheAgenticDataCompany/open-yap-1k
HF dataset
1,000 hours of dual-channel natural conversation with real overlaps and interruptions is exactly what full-duplex voice agents need, but only 8.9 hours are openly released so far.
-
Xenova/fruit-fly-simulation
HF space
A fully local WebGPU simulation of a 166k-neuron fruit-fly connectome you can stimulate in the browser—scientifically illustrative rather than validated, but a striking accessibility milestone.
-
fulvian/engraft-ngram
GitHub
ENGRAFT injects facts into a 125B MoE by editing eight n-gram memory rows on CPU with no fine-tuning; a single small experiment, but a provocative model-editing direction.
-
hbofz/NeSLE
GitHub
A CUDA-native NES emulator running 65k parallel environments at 3.27M steps/s with fully GPU-resident PPO is impressive RL systems work, currently limited to Mario.
The shortlist: top candidates that survived triage · Archive