Issue 38 · Sep 14–20, 2026
Week 2026-W38
3,476 papers scanned 150 shortlisted 10 picked $14.12 spent
This week leans heavy on genuinely new framings rather than benchmark bumps: analog neural computation in radio hardware, mechanistic evidence that LLMs build Bayesian belief states in-context, and several assumption-breaking neuroscience/BCI results. Robotics contributes a striking new physical capability (low-inertia dynamic manipulation) and a surprising failure mode in action-conditioned world models. Note that many of the flashiest agent/efficiency papers carry model names from apparently future systems and single-benchmark evidence, so discount their headline numbers accordingly.
-
Radio-Frequency Convolutional Neural Networks
Repurposing the frequency mixer already present in every wireless radio to perform CNN convolution in the analog frequency domain is a genuinely new hardware direction, not another analog-accelerator variant. The reported scale (26M params, 9 layers), task diversity, and ~0.72 fJ/MAC energy make it potentially transformative for edge inference.
Look for Check how much energy and latency the 'data preparation and readout' actually cost end-to-end, and whether weights-over-the-air assumptions hold outside a controlled setup.
-
Large Language Models Develop Belief State Geometry In-Context
Direct representational and causal evidence that production LLMs form linearly-decodable Bayesian belief states during in-context learning, with patching/steering showing the subspace is functionally used. This is exactly the kind of result that connects mechanistic interpretability to how computation might work more generally.
Look for The evidence is confined to controlled HMM prompts; be skeptical about whether the belief-state geometry generalizes to naturalistic ICL tasks.
-
Long-horizon autoformalization of a core theorem underlying MIP* = RE
A coordinated multi-agent system produced a 126k-line machine-checked Lean proof of a core theorem underlying MIP*=RE, correcting side conditions along the way. This is a concrete demonstration that AI proving agents can handle deep, long-horizon formalization rather than isolated lemmas.
Look for Assess how much human supervision the 63-day run actually required and whether the workflow generalizes beyond this particular proof structure.
-
AthenaZero: A low-inertia, bimanual robot for dynamic manipulation
A bimanual robot with an order-of-magnitude lower endpoint inertia (roughly human-like) achieves throwing above 30 m/s and catching/batting above 14 m/s, including human-robot play. This is a real new physical capability in dynamic manipulation, not a simulation or demo tweak.
Look for Watch how much the low-inertia mechanical design constrains payload and precision for ordinary quasi-static manipulation tasks.
-
iTBS induces a suppression-rebound pattern in human cortical excitability
Intracranial recordings in 20 neurosurgical patients show iTBS produces a minute of suppression followed by rebound excitability, directly challenging the standard LTP/excitation interpretation of a widely used stimulation protocol. Directly relevant to the reader's interest in transcranial stimulation mechanisms.
Look for Rebound findings rest on a small, specialized epilepsy-surgery cohort and only three TMS cases; broader replication is needed before revising clinical assumptions.
-
A single cortical circuit implements fast and slow visual search
Intracranial EEG evidence that easy (fast) and hard (slow) visual search recruit the same circuit at different speeds—a time-stretched trajectory—challenges four decades of the 'distinct circuits' account of behavioral flexibility. A clean, mechanistically specific reframing.
Look for Only two tasks in 19 patients; the temporal-scaling claim may not extend to other fast/slow behavioral dissociations.
-
Robot World Models Are Not Invariant to How the Actions Are Written
Shows action-conditioned robot world models catastrophically diverge when the same commands are re-written as absolute vs. relative deltas, despite the encodings being mutually reconstructible. A non-obvious, well-quantified failure mode with a concrete invariance test and repair.
Look for Verify whether the effect persists in modern large VLA/world-model architectures and affects real robot performance, not just latent retrieval metrics.
-
AESSI: An Around-Ear Silent Speech Interface for Cross-Day Online Reuse without Test-Day Calibration
An around-ear silent-speech interface achieving cross-day online reuse without test-day calibration (92% offline, 98% in later online tests) targets one of the hardest practical problems in wearable neuro-interfaces. Directly in the reader's ear-EEG/EMG wheelhouse.
Look for Task is a fixed 25-sentence set with only five participants in the long-delay online test; scaling to open vocabulary is unproven.
-
How Model Growth, Recursion, and Boundary Operators Influence Scaling Exponents
Claims that architectural interventions (looped/recursive depth, model growth) can change pretraining scaling exponents—not just constants—with a 7.4B growth model matching GPT-3 13B at ~20x less compute. If robust, this would materially revise how we think about scaling.
Look for Be skeptical of the scaling-exponent claim: check baseline fairness, the compute range probed, and whether CORE is the only benchmark showing the effect.
-
Taming Long-form Text-to-Speech
An inference-only rollback-and-regenerate mechanism cuts worst-of-10 long-form TTS WER from 35.2% to 3.4% and dramatically improves voice-cloning consistency, plus a diagnostic metric that exposes failures hidden by global similarity scores. Broadly useful and immediately deployable for autoregressive TTS.
Look for Gains are worst-of-N; confirm typical-case improvement and overhead on models beyond Qwen3-TTS.
Also notable
-
An Eternal Irradiance Camera
Tech
A battery-free omnidirectional 'irradiance camera' that reconstructs illumination from 49 detectors using low-order spherical harmonics—an unusually different, privacy-preserving sensor design with a working wireless prototype.
-
Observing and evading quantum back-action on a kilogram-scale oscillator
Tech
Quantum back-action evasion demonstrated at kilogram scale using LIGO's 40 kg mirrors, suppressing observed motion near resonance by ~47%—macroscopic quantum reservoir engineering.
-
The Output-Space Hypothesis: Enumerative Equivalence Checking for Tensor Programs
AI / ML
An 'output-space' verification strategy for tensor programs that flips the quantifiers and found 600 buggy CUDA kernels among 6,988 previously judged correct by differential testing.
-
Using OCR Heads to Verbalize Image Semantics
AI / ML
Identifies OCR-critical attention heads in VLMs that also verbalize semantics of arbitrary image regions, yielding an interpretable 'verbalization lens' and causal concept edits from layer 0.
-
Reinforcement Learning for Real-Time Vision-Language-Action Policies
Robotics
Decouples slow expressive VLA planning from a fast RL-trained action-editor, lifting dynamic real-robot manipulation from 42% to 97% with only 10 minutes of online data—addresses the VLA latency bottleneck.
-
A Language-Guided Multimodal Foundation Model for Zero-Shot and Multi-Task Brain Signal Analysis
Non-invasive BCI
A language-guided brain-signal foundation model trained on 70k+ hours claiming large zero-shot cross-task gains; potentially important but the abstract lacks modality, leakage-control, and baseline detail.
-
An electronically steerable epidural ultrasound interface for deep brain neuromodulation in freely moving rats
BCI
An electronically steerable implantable epidural ultrasound array with a custom ASIC delivering multi-site deep-brain neuromodulation in freely moving rats—an interesting chronic, steerable FUS hardware direction.
-
Transcranial Random Aperiodic Stimulation Improves Working Memory Precision
Non-invasive BCI
Introduces transcranial random aperiodic stimulation (tRAS) and shows steepening aperiodic activity causally improves working-memory precision—a novel handle on the often-dismissed 1/f component.
-
Occupancy Network-Guided Autonomous Robotic Partial Nephrectomy
Robotics
First closed-loop autonomous robotic tumor resection (partial nephrectomy) using occupancy networks for deforming tissue, with negative margins across eight phantom trials.
-
The cerebellum exploits a thalamic carrier frequency to refine motor behaviors
Neuroscience
Links a CaV3.1-dependent thalamic delta 'carrier frequency' to cerebellum-dependent offline motor learning via recordings, closed-loop stimulation, and cell-type knockout.
-
Scaling Discovery through Test-Time Communication
AI / ML
Test-time communication among agents lets a team of k match ~4k independent attempts and solve tasks no single agent can—an interesting scaling result for multi-agent inference, if the compute accounting holds.
-
Self-excited actuation enables adaptive and resilient flapping-wing flight
Robotics
First flight-capable flapping-wing robot using insect-like asynchronous (self-excited) actuation, enabling sensorless instantaneous collision response—a genuinely new embodied-control strategy.
Projects
A heavy week for inference efficiency (ternary weights, disk-streamed MoE) and omni-modal/real-time interaction, plus a striking new trend: a dozen-plus 'typed decision' System-One-style projects that replace token generation with calibrated single-pass decisions. No notable BCI or neurotech releases this week; the connectome demos are the closest neuroscience touchpoint.
-
JustVugg/colibri
A zero-dependency C engine that treats NVMe, RAM, and VRAM as one inference hierarchy and streams MoE experts from disk, claiming usable inference of 744B–2.8T models on consumer hardware. If the throughput holds up, this meaningfully changes who can run frontier-scale sparse models, and the traction (37k stars in weeks) suggests it works well enough for many.
Look for Check actual tokens/sec on realistic prompts and quality at the quantizations used — disk-streamed MoE latency can be far worse than headline numbers imply.
3 min read ·GitHub ↗ ·C·Apache-2.0
-
prism-ml/Ternary-Bonsai-2-27B-gguf
End-to-end ternary weights on a 27B multimodal model at ~1.7–2.1 bits/weight with claimed near-FP16 benchmark retention, via rotated ternary quantization and custom kernels rather than standard PTQ. If the reasoning/agentic retention claims survive independent testing, this is a genuine jump in local-inference economics (sub-8 GB for a capable 27B).
Look for Independently verify the near-FP16 claims on reasoning tasks — extreme-low-bit releases often degrade precisely where benchmarks don't look.
4 min read ·Hugging Face ↗ ·apache-2.0
-
inclusionAI/Realtime-Venus
Open full-duplex audiovisual interaction with released 9B models: the assistant keeps listening and speaking while background reasoning or tool work runs asynchronously, then folds results back into the live conversation. This is exactly the real-time interactive speech direction the reader tracks, and it's a working system with code, weights, and browser examples rather than a demo video.
Look for Test the actual turn-taking and interruption quality in the browser demo — full-duplex claims often break down under overlapping speech; weights also at P87.
3 min read ·GitHub ↗ ·Python·Apache-2.0
-
XiaomiMiMo/MiMo-V2.6-Pro-RL
A 1.02T-parameter sparse MoE with native text/image/video/audio input, 1M-token context, and an unusually detailed agentic RL recipe (groupwise reward synthesis, advantage redistribution, multi-prefix on-policy distillation) reportedly approaching proprietary systems on agent benchmarks. An open native-omnimodal model at this scale is worth knowing about regardless of whether the RL claims fully hold.
Look for Benchmarks are vendor-reported and the model is impractical to self-host; check for independent evals and whether the RL recipe details are actually reproducible; the 309B Flash variant (P11) may be the more usable release.
3 min read ·Hugging Face ↗ ·mit
-
Mapika/decider
The strongest representative of this week's biggest emergent trend: models that emit typed, calibrated decisions in one forward pass instead of generating and parsing text, with released 0.8B–35B weights, training code, and browser-agent evaluations. Fast calibrated decision layers could become a standard component in agent stacks, and this release is the most complete of the ~15 similar projects this week.
Look for Verify the calibration claims out-of-distribution and compare against simple constrained-decoding baselines (P50, P59) before adopting — much of the ecosystem is days old.
3 min read ·GitHub ↗ ·Python·Apache-2.0
-
netease-youdao/Confucius4-R2T2
Streaming ASR that commits an append-only transcript via longest-stable-prefix decoding, claiming 200–600 ms latency with near-offline accuracy — attacking the unstable-partial-hypothesis problem that plagues every real-time voice pipeline. Ships with vLLM and WebSocket inference, making it directly usable.
Look for Test stability and accuracy tradeoff at the shortest chunk sizes and on noisy/accented audio; check which languages actually hit the claimed latency.
3 min read ·Hugging Face ↗ ·other
Also notable
-
Davinchy/macuda
GitHub
A userspace C NVIDIA driver plus CUDA shims running unmodified llama.cpp on an RTX 5090 over Thunderbolt from an Apple Silicon Mac at ~91–101% of Windows throughput — a genuinely surprising systems feat, though limited to validated kernels.
-
moondream/parakeet-redux
HF model
Ternary (1.58-bit) Parakeet ASR at 178 MB with claimed near-original accuracy and 113× real-time on 8 CPU cores — extreme compression coming for speech models too.
-
mlabonne/chessfly
HF space
The full 138k-neuron fruit-fly connectome running in-browser via WebGPU as a fixed substrate trained to evaluate chess positions — a concrete connectome-as-computation demo (see also FlyDrones, P64), though the readout does the heavy lifting.
-
Gensyn/open-1b-sft
HF model
A 1.6B model whose entire training trajectory is independently replayable via per-step state hashes — a genuinely new release property for trust and reproducibility (harness at P73).
-
eidon-ai/tracker-pov
HF dataset
1,274 hours of real-home egocentric video synchronized with seven upper-body IMUs across 13k household tasks — unusually substantial data for learning human manipulation, if you can stomach the 9 TB download.
-
TencentARC/WorldCrafter
GitHub
Camera-controlled video world model with implicit 3D-aware memory for long-rollout scene consistency, with released Base and distilled Fast weights.
-
XGEN-Labs/XGEN-JING
GitHub
Egocentric interactive world model generating action-conditioned first-person video with synchronized audio and reference-image composition — a meaningful direction, but currently bidirectional and needing six H100s (weights at P47).
-
Video-Rebirth/hyperflow
GitHub
Data-free flow self-distillation LoRA cutting MiniMax-H3 video+audio generation from 49 to 8 forwards with released weights — a large practical inference-cost reduction if quality holds.
The shortlist: top candidates that survived triage · Archive