WEDNESDAY · SEPTEMBER 30, 2026 · ISSUE 21 No. 21
781 papers indexed on arXiv·~28,000 packages released on PyPI·50 papers surfaced by Hugging Face
None of it is in your agent's weights.
Writing a score before rendering audio turns out to matter. YuE2's central claim is that making melody, harmony, rhythm, and form explicit in a readable intermediate representation—before any acoustic generation…
The central claim here is architectural, not cosmetic: web agents fail at the browser layer, not the model layer. Most automation frameworks bolt anti-detection measures onto a standard browser after the fact—JavaScript…
SEPTEMBER 30 — 9 PAPERS · 9 TOOLS
DN-MOPD rescales each specialist's feedback by its measured spread, fixing the imbalance that lets instruction-following dominate multi-teacher distillation.
HexaAnything wraps VLA/WAM policies in code-represented task state so verified execution traces feed back as training data, beating direct VLA on unseen tasks.
Duplex-MPE is a benchmark of 2,000 scenarios testing whether a full-duplex speech assistant answers, stays silent, or stops in multi-party conversations.
CoWA splits causal-history access across KV heads for full collective coverage, cutting 128K training latency 7-8x while tracking FullAttn accuracy at scale.
Scaling-laws analysis predicts encoder-free MLLMs close the multimodal gap with encoder-based models at ~10^22 FLOPs, within practical pretraining budgets.
TaH2 adds a learned iteration decider to looped transformers, lifting the accuracy-compute slope 53% over a non-looped baseline on AIME benchmarks.
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
A method that distills a strong agent's robot manipulation experience into a reusable playbook, improving success from 37.3% to 64.0% in real-world tasks.
FlyBy trains 4B/8B models to detect knowledge bottlenecks mid-reasoning and query stronger models, letting FlyBy-4B beat Qwen3-14B at 2.7x lower serving cost.
Curated list of uncensored and cybersecurity-fine-tuned models for offensive security and red-teaming tasks.
MCP server that connects an AI assistant to a live SolidWorks instance to sketch, extrude, fillet, export STEP/STL, and generate macros.
MCP extensions for ChatGPT that make plugins feel like first-class native features rather than third-party add-ons.
A toolkit of skills, tools, and MCP integrations that lets Claude Code recon, reverse-engineer, and mod PC games with generated art, 3D, and audio.
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
Motion-graphics skills repo for Claude Code and Codex - covers animations, transitions, and visual effects.
An animated blueprint drawing that walks through a before-and-after UX redesign step by step.
DeepSeek releases a clean, efficient GEMM kernel library targeting Huawei Ascend NPUs.
Because a 2007 Nokia can't Google anymore, this unofficial J2ME app + Go server gives it Claude instead.
SEPTEMBER 29 — 8 PAPERS · 7 TOOLS
FuseReg trains over random encoder-layer subsets so one decoder handles any fusion without retraining, cutting unguided gFID by up to 29%.
Disaggregated quantization uses separate formats for prefill and decode, yielding 32+ point accuracy gains and 1.78x faster first-token latency on 27B models.
PISA cuts block-sparse attention's quadratic block-selection bottleneck to O(N log N) via a coarse-to-fine pyramid Top-K strategy with fused Triton kernels.
InternW0-Δ unifies visual dynamics, 4D geometry, and action generation in a Mixture-of-Transformers WAM trained on 20K+ hours of open robot and egocentric data.
SEPTEMBER 26 — 9 PAPERS · 7 TOOLS
WROP pairs a 1.5M-sample object-permanence corpus with a 300-question blind Elo exam; their 16B PWM-WROP ranks first among continuation video world models.
AEWM edits noisy reasoning and actions in task history rather than predicting tool responses, improving agent scores by 3.2–6.7 points across six benchmarks.
A study finding that part-of-speech distinctions are recoverable from SAE activations via compact latent groups, not one-to-one latent-to-category mappings.
Qwen-Planner-Agent uses an agentic data flywheel and CARE reward shaping to build mobile planning agents in a closed AI-for-AI loop.
SEPTEMBER 25 — 6 PAPERS · 10 TOOLS
RewardVerse uses generated rubrics as an intermediate step before scoring to reduce scalar drift in video reward models trained with a two-stage RL algorithm.
An evaluation framework that dynamically transforms test repositories to detect whether coding agents rely on memorized cues rather than genuine reasoning.
Hunyuan-A13B is an open-source 80B/13B-active MoE model with a dual-mode CoT that switches between fast and slow reasoning by task complexity.
MemBodied adds fixed-size episodic memory to Vision-Language-Action models, achieving 7.81× the success rate of a stateless policy on memory-dependent tasks.
SEPTEMBER 24 — 9 PAPERS · 9 TOOLS
GAE is an autoencoder whose latent space decodes jointly to appearance, depth, cameras, and point maps, cutting FVD by 12.7% and 23.1% on two benchmarks.
ScriptMoE beats large VLMs on multilingual OCR by routing images to script-aligned sparse experts, covering 229 languages at far lower cost.
Ovis-Embedding unifies text, image, video, and audio in one backbone and claims top results on five retrieval benchmarks including MMEB-v3 and MVEB.
StableVQ fixes VQ tokenizer training instability by giving the encoder and codebook separate, independent learning objectives and schedules.