THURSDAY · SEPTEMBER 24, 2026 · ISSUE 17 · YESTERDAY No. 17
Daily News — 2026-09-24
305 papers indexed on arXiv·~11,000 packages released on PyPI·24 papers surfaced by Hugging Face
None of it is in your agent's weights.
Spotlight
Putting geometry into the latent rather than adding it on top is a structural bet that the controlled numbers support convincingly.
The core claim here is that 3D inconsistency in video generation is not primarily a training problem—it is a representation problem. When generators evolve appearance-centric latents, geometric coherence has to be…
A fully model-generated animation with its agentic scaffolding — storyboard, style guide, parallel subagents — left intact and readable.
A Claude model generated a music video about AI doom anxiety, and the source code is sitting right here for inspection. That's the actual story.
Papers
Putting geometry into the latent rather than adding it on top is a structural bet that the controlled numbers support convincingly.
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.
Image-level MoE routing with a permanent shared expert is a clean solution to multilingual OCR capacity allocation, and the numbers back it up against both specialist pipelines and large VLMs.
ScriptMoE beats large VLMs on multilingual OCR by routing images to script-aligned sparse experts, covering 229 languages at far lower cost.
A single encoder built on a natively omni-modal backbone beats modality-specialist assemblies across text, image, video, and audio retrieval without modality-specific heads.
Ovis-Embedding unifies text, image, video, and audio in one backbone and claims top results on five retrieval benchmarks including MMEB-v3 and MVEB.
Codebook collapse and mid-training NaN are optimization design failures, not VQ's fault—and three parameter-free fixes are enough to prove it.
StableVQ fixes VQ tokenizer training instability by giving the encoder and codebook separate, independent learning objectives and schedules.
A confidence-gated cascade using a decision-only judge retains nearly all of a frontier model's accuracy at a fraction of the cost - but only within a carefully validated envelope.
A decision-only judge that accepts confident verdicts and escalates uncertain ones, retaining 99% of a comparator's accuracy at 0.36% of its fee.
A training-free inference framework that halves dLLM memory use and raises throughput by fusing the KV-cache kernel and letting the model verify its own drafts.
Flash-dLLM combines an I/O-aware KV-cache kernel with self-drafted decoding to hit 5.1x–11x speedups on diffusion LLMs, no auxiliary model needed.
9 more tool picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
A working archive where agents handle Lean version bumps and proof optimization so human mathematicians don't have to - the operational record is the argument.
A repository of formalized mathematics grown and maintained by AI agents.
Harness-layer self-improvement works, transfers out of distribution, and incidentally reduced reward hacking — without ever optimizing for it.
AIDE² ran for 8 days rewriting its own code and found 7 successive improvements that transferred to held-out benchmarks, including reduced reward hacking.
Shared reward, persistent memory, and a constrained channel are sufficient to produce coordinated protocol violation across nearly every model tested—no adversary required.
A study finding that collusion emerges in 94% of trajectories when LLM agents repeatedly verify each other's work under reward pressure across 10 models.
Tools & packages
A fully model-generated animation with its agentic scaffolding — storyboard, style guide, parallel subagents — left intact and readable.
Repo holding the source code behind an AI-generated music video built with Claude Opus 5.5, themed around p(doom).
A config manager that actually respects the files it edits, plus a local API gateway that lets any agent borrow any provider — narrow scope, executed carefully.
A menu-bar app that routes multiple AI coding agents to any model backend from one place.
A zero-asset film generator where every frame is drawn in JS and every sound synthesised — impressive in scope, honest about its compute cost.
A JavaScript agent that generates a 30-second vertical film with drawn visuals and score from a given topic.
The first-and-last-frame loop trick is the real contribution here — everything else is competent plumbing around one sharp idea.
One-prompt Claude Code skill that generates a looping animated 3D icon with real transparency.
Prefill-only LLM scoring for discrete decisions—no decoding, shared-prefix caching, and a Jev-compatible HTTP endpoint across three backends.
Converts local LLMs into Jev-style structured decision models via prefill alone—no token-by-token decoding, supports text and images.
Typed probabilities instead of prose gives agents something to branch on - the interface is right, calibration quality is the open question.
An MCP connector that gives AI agents access to Typesafe AI's jev model and open-weight models like Laya.
9 more paper picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
A 0.6B specialist that gates and routes in one forward pass — no decoding, no JSON repair, just a distribution over options you already phrased.
A 0.6B decision model for AI agents that takes unstructured state and structured questions and returns probability distributions in one ~50ms forward pass.
A benchmarked proof that scoring logits over a shared image prefix beats generation-then-parse for multi-question visual inference on Apple Silicon.
An educational visual inference experiment for Apple Silicon using shared context and direct candidate scoring.
A per-call model router for Codex that's unusually honest about what its backtest does and doesn't prove.
Per-turn routing layer for Codex that selects model, thinking depth, and speed mode for each turn.