Issue 31 · Jul 27 – Aug 2, 2026
Week 2026-W31
3,172 papers scanned 150 shortlisted 10 picked $10.55 spent
This week's strongest signal is on the boundary between brains and learning machines: two independent, striking results on how the brain does credit assignment (a cerebellar loss-gradient signal and millisecond/single-neuron precision in songbirds), plus a genetic study that overturns the standard perineuronal-net account of critical-period closure. On the engineering side, the headline is efficiency and architecture — a scaled diffusion LLM, stable FP4 training, and a generalizable speech BCI — alongside several papers that cleanly isolate failure modes in LLM agents and evaluators. Note that many arXiv IDs and model names here reference near-future artifacts; discount the promotional framing and read for the mechanisms.
-
DiffusionGemma Technical Report
A discrete-diffusion LLM converted from a Gemma MoE at under 10% of the original training budget, decoding 256-token blocks in parallel at ~1,500 tok/s on one H100 while retaining multimodality, long context, and even AR generation. This is the most credible large-scale challenge yet to the autoregressive decoding bottleneck.
Look for Check the actual quality/benchmark tables versus the AR base, not just tokens/sec — the abstract is thin on capability retention and the real question is Pareto position, not peak speed.
-
Climbing fibers encode the gradient of a loss function for the cerebellum
Direct physiological evidence that cerebellar climbing fibers deliver a signal proportional to the dot product of a reward-weighted error with each Purkinje cell's motor influence — i.e., a per-neuron loss gradient. This is a concrete brain implementation of gradient-based credit assignment and a genuine bridge to ML.
Look for Scrutinize how the per-cell 'potent vector' is estimated from spike-triggered suppression and whether the dot-product claim holds per-cell or only on average across the population.
-
Millisecond-scale, single-neuron credit assignment in a songbird
Neuron-specific, activity-contingent perturbations show songbird vocal learning can assign credit with ~3.2 ms temporal and single-neuron spatial precision, challenging the view that dopamine-mediated reinforcement is slow and diffuse. Pairs naturally with the cerebellum result to reframe how precise biological credit assignment can be.
Look for It's a preprint with limited reported sample size/replication — check n, controls for the targeted-vs-neighbor comparison, and how contingency was enforced.
-
A generalizable speech neuroprosthesis
A transformer speech decoder trained jointly across six intracortical participants beats per-user models (>50% relative WER reduction) and hits <7% WER for a held-out user after fewer than 200 calibration sentences. Cross-subject pooling that slashes calibration is a major step toward deployable speech neuroprostheses.
Look for Absolute WERs and how much of the gain depends on the phoneme-to-word language model versus the neural decoder; also whether transfer holds across differing electrode placements and etiologies.
-
Neuropil aggrecan, not perineuronal nets, closes the critical period for visual plasticity
Cell-type-specific aggrecan deletions show that removing perineuronal nets from PV interneurons does NOT reopen visual plasticity, while removing aggrecan from excitatory/neuropil sources does. A clean genetic result overturning a widely cited PNN-based model of critical-period closure.
Look for Whether the excitatory/neuropil manipulation truly dissociates from residual PNN effects, and how completely PNNs were eliminated in the inhibitory-deletion condition.
-
Stable FP4 Training via Transposition-Invariant Block Quantization
Identifies transposition-induced scale inconsistency as a root cause of FP4 training instability and fixes it with 2D transposition-invariant block quantization, demonstrated on dense models to 7B and a 30B MoE over 100B tokens within ~1.3% of BF16. A concrete explanation plus a simple fix that could push training below FP8.
Look for The mixed-precision caveat (attention stays MXFP8) and whether the <1.3% gap holds on downstream tasks and larger token budgets, not just perplexity.
-
Bicycle Acrobatics with Reinforcement Learning
An underactuated bicycle robot learns and composes jumps, front flips, wheelies, and kip-ups via RL with a state-triggered orchestrator, performing 15+ consecutive maneuvers and generalizing to unseen table layouts on hardware. A real-world capability result on a platform not usually associated with acrobatics.
Look for How much the orchestrator/hand-tuning of triggers carries the long-horizon sequences versus the learned policies themselves, and robustness of the flips across trials.
-
Keep It InMind: Benchmarking the Implicit-Association Blind Spot in Agent Memory
Cleanly isolates the 'implicit-association blind spot' in agent memory: six vector/graph/agentic systems recall facts on demand at up to 100% but answer only ~14% of indirect queries where the bridging knowledge is implicit, versus 84% with memory in context. Strong cross-system evidence that scaling embeddings doesn't fix retrieval-time routing.
Look for Whether the benchmark's 'implicit' links are genuinely retrieval problems rather than reasoning gaps, and how a reasoning-augmented retrieval step (not just bigger embeddings) fares.
-
Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-design
Reframes robot design from optimizing controllers for fixed bodies to jointly generating embodiments and control via a diffusion transformer over unified body/state/action tokens, with dynamics self-guidance for unseen rewards — plus a fabricated ALOHA redesign that cut tracking error >70%. A genuinely new framing with a hardware payoff.
Look for How well the single learned dynamics model generalizes across truly different embodiment classes, and whether the fabricated result depended on hand-selected reward shaping.
-
Judging Is Not Enumerating: Silent Omissions in LLM-Authored Acceptable Sets
Distinguishes judging membership from enumerating a complete acceptable set, showing LLMs author badly incomplete test suites/answer keys/reward criteria (omission-dominated, hard to audit) even when they judge well and can state the predicate at ~0.99 F1. Directly relevant to anyone using LLMs as verifiers or RLVR reward authors.
Look for How much execution-based repair (the 3.3–10.6x yield gain) closes the gap in practice, and whether the effect persists under test-time reasoning rather than one-shot greedy authoring.
Also notable
-
The Kinetics of Training: A Driven-Nucleation Rate Law for Emergence, Plasticity Loss, and Circuit Control in Language Models
AI / ML
A 'driven-nucleation' rate law framing capability emergence as no-partial-credit joint circuit alignment, with a QK-reinit cure for plasticity loss — intriguing theory of emergence, worth reading if the preregistered predictions hold.
-
A Change of Frame Makes Balance Observable: Distillation-Free Humanoid Single-Leg Stance
Robotics
A simple support-foot frame change makes the capture point observable from onboard sensors, enabling clean single-leg humanoid balance (89/90 vs 0/90 for prior policies) with real Unitree G1 transfer.
-
HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone
Robotics
Robot-free UMI data made high-fidelity enough (~3mm) to deploy policies directly on hardware without a real-robot anchor, matching teleoperation across three VLA/world-action backbones (158 upvotes).
-
Pictura: Perspective-View Self-Play at Scale for Driving
Robotics
Perspective-view self-play at 50B agent steps trains a driving policy directly from egocentric cameras that matches (and zero-shot beats) privileged-state agents, questioning the need for privileged self-play.
-
Instruction-Tuned Language Models Cannot Sample from Distributions They Can Describe
AI / ML
Argues instruction-tuned LLMs collapse to a single output rather than sampling a persona distribution they can describe — a KNOWS/DOES split that undermines 'silicon sampling' for survey emulation.
-
Collusion with Competitive Marginals: Price-Level Audits Are Blind by Construction
AI / ML
Proves price/bid-level collusion audits are blind by construction when agents couple only through unexplained-component correlations, with evidence the mechanism appears in LLM bidding agents.
-
Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory
AI / ML
Scales a parametric long-term memory module to 6.9B/300B tokens, showing memory parameters buy more than base-model scaling (410M+6.9B memory matches Pythia-12B with 39% fewer params).
-
TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM
Robotics
A non-LLM-centric VLA (V+L→A) hits 97.7% on LIBERO at 0.2B params, 31ms latency, 0.9GB VRAM — promising real-time control design, but evidence is one simulated benchmark.
-
Parallel construction of object motion from retina to cortex
Neuroscience
Retina and cortex compute object motion in parallel with distinct tuning (fine jitter vs coarse drift) rather than cortex inheriting the retinal computation — challenges the strict hierarchical view of vision.
-
Cryoaerosolization Enables Scalable Vitrification-Based Cell Cryopreservation
Tech
Cryoaerosolization scales vitrification cell cryopreservation by ~two orders of magnitude in throughput while keeping >90% post-thaw viability across iPSCs, fibroblasts, and pig RBCs.
-
Deep conservation of head direction circuits in bees, ants and flies
Neuroscience
Synaptic-resolution maps show insect head-direction circuits are conserved in cell type and layout across ~300M years yet differ fundamentally in synaptic wiring, with both implementations modeling as ring attractors.
-
Context Is King: How In-Context Specification Shapes the Geometry of Concepts
AI / ML
Activation-patching evidence that LLMs build conceptual geometry (cycles vs trees, even on meaningless tokens) from in-context specification and causally use it, with larger models overriding pretrained priors.
Projects
A strong systems week: two independent projects prove frontier-scale MoE models can run locally by streaming experts from disk, alongside substantive world-model releases (physical-language reasoning, stereo video, streaming training recipes) and a concrete recursive self-improvement stack.
-
sqliteai/warp
Running the full 2.78T-parameter Kimi K3 on a 64 GB MacBook at ~0.6 tok/s via expert-aware NVMe streaming is a genuinely new local-inference capability, not a distillation trick. The dependency-free C engine with lookahead prefetching and detailed measurements makes this the strongest of this week's two disk-streaming efforts and a pattern likely to be widely copied.
Look for Verify sustained throughput on longer contexts and NVMe wear/latency behavior; check license and whether it supports models beyond Kimi K3 and GLM-5.3-Flash.
3 min read ·GitHub ↗ ·C·Apache-2.0
-
LUMIA-Group/MemoryDecoder-at-Scale
A separately pretrained, swappable parametric memory scaled to 6.9B params and 300B tokens, where a 410M backbone plus memory beats Pythia-12B, is a substantive architectural result on the memory-vs-parameters question. Domain-specific memories improving frozen Qwen3 backbones suggest a practical alternative to both fine-tuning and retrieval wrappers.
Look for Check that gains hold on generation quality (not just perplexity/kNN-aligned metrics) and what inference-time compute/memory overhead the decoder adds.
3 min read ·GitHub ↗ ·Python·Apache-2.0
-
yaoyao-jpg/PhiZero
PhiZero's explicit separation of physical-dynamics prediction (in a discrete 'Physical Language') from video rendering is a genuinely interesting world-model design that addresses the physics-incoherence failure of pixel-space video models. Released reasoner and tokenizer checkpoints plus runnable inference and motion-transfer pipelines make it testable, not paper-vapor.
Look for Confirm the checkpoints are actually up (README says 'coming soon') and test whether physical predictions generalize beyond the curated evaluation examples.
3 min read ·GitHub ↗ ·Python
-
FrontisAI/OpenRSI
OpenRSI turns recursive self-improvement from a slogan into a measurable engineering loop with executable MLE environments, RL, and evolutionary search, and reports raising MLE-Bench Lite Medal Average from 39% to 61%. That is a large jump on a benchmark that directly measures research automation, a direction the reader should track closely.
Look for Scrutinize whether the MLE-Bench gain reproduces independently and how much of it comes from search compute rather than genuine model improvement.
3 min read ·GitHub ↗ ·Python
-
NIMI-research/Tycho
Tycho's approach of having a multimodal model induce executable deterministic world models from observations, then falsify and plan against them, reportedly yields perfect scores on all 25 public ARC-AGI-3 games. The empiricist program-induction loop is a meaningfully different take on open-ended visual reasoning, with reproducible policies and replay tooling.
Look for Check frontier-model API costs per game and whether the approach transfers beyond the 25 public games or overfits their mechanics.
3 min read ·GitHub ↗ ·Python·Apache-2.0
-
H-EmbodVis/TurboVLA
97.7% on LIBERO at 31 ms latency and 0.9 GB VRAM by ditching the LLM-centric pathway is a real efficiency jump for VLA policies, making 30+ Hz closed-loop control feasible on commodity GPUs. Released code and checkpoints make it directly usable for robotics experimentation.
Look for Evidence is simulation-heavy; look for real-robot results and whether skipping the LLM pathway hurts language generalization outside benchmark instructions.
3 min read ·GitHub ↗ ·Python·Apache-2.0
Also notable
-
FareedKhan-dev/kimi-k3-in-c
GitHub
The same disk-streaming idea as WARP in portable C99, hitting 8.24 GB peak RAM for Kimi K3 — impressive proof of concept, but at 26–33 s/token WARP is the more usable variant.
-
alexiglad/XM
GitHub
Explorative Modeling's best-of-K training objective as a 'third pretraining axis' is a genuinely distinctive idea across modalities, but the compute-vs-gain tradeoff against strong baselines is not yet established.
-
Pan-Chera/Multi-Agent-CAD
GitHub
Decoupled multi-agent text-to-CAD with 116× fewer tokens and working print-in-place assemblies is a neat constrained test-time-compute result, though benchmarked on a small task set.
-
MiniMax-AI/MiniMax-H3
GitHub
MiniMax H3's unified conditioning on text, up to nine images, video, and audio references with native stereo audio is notable omni-modal capability, but hosted-only with little inspectable substance.
-
Zhao-Yian/MiniWorld
GitHub
MiniWorld releases complete code, weights, and recipes for training streaming causal video world models from scratch — a useful democratization even if the modeling advance needs independent validation.
-
VAST-AI-Research/StereoWorld
GitHub
StereoWorld (CVPR) generates controllable binocular video from a single image with released teacher and distilled student weights — a genuinely new controllable-stereo capability worth a look.
-
zju3dv/INTACT-JEPA
GitHub
INTACT-JEPA replaces CEM/MPPI action search in JEPA world models with a direct intent-to-action predictor, with code and checkpoints, though evidence is still project-reported benchmarks.
-
Audio8-AI/Audio8_TTS
GitHub
A 0.6B multilingual zero-shot-cloning TTS with INT4 CPU deployment and streaming is unusually deployable at compact scale, but the preview release lacks comparative quality evidence.
The shortlist: top candidates that survived triage · Archive