FRIDAY · SEPTEMBER 11, 2026 · ISSUE 11 · SINCE MONDAY No. 11
Daily News — 2026-09-11
1,446 papers indexed on arXiv·~11,000 packages released on PyPI·120 papers surfaced by Hugging Face
None of it is in your agent's weights.
Spotlight
Swapping continuous motor regression for a small vocabulary of deterministic semantic units lets a 2B VLM outperform fine-tuned VLAs on sim-to-real transfer without retraining.
The central bet in Show-Harness is that the bottleneck between VLMs and robots is not model capability but interface design. Rather than fine-tuning a VLM to regress continuous motor signals, the system exposes a small…
A read-only WeChat intelligence layer that's genuinely useful once you clear the non-trivial database access hurdle — and admirably honest about that hurdle.
WeChat Intelligence Hub treats your local WeChat message database as a personal intelligence source rather than a chat archive. The core idea is straightforward: given authorized, read-only access to the encrypted…
Papers
Swapping continuous motor regression for a small vocabulary of deterministic semantic units lets a 2B VLM outperform fine-tuned VLAs on sim-to-real transfer without retraining.
Show-Harness routes VLM outputs through discrete semantic actions to embodiment-specific modules, enabling zero-shot deployment across robots and GUIs.
Replacing expensive sandbox execution with a corrected world model cuts RL training cost three-to-fourfold while matching or beating full-execution baselines—a result backed by convergence proofs.
World Model RL cuts post-training cost for research agents by replacing live environment execution with a learned world model plus debiasing and denoising.
Routing-harness logs already contain the capability-demand signal needed for curriculum training—NeoHorse-1 shows the loop works once; whether it compounds across iterations is still open.
NeoHorse-1 targets recursive self-improvement by combining agentic post-training with routing, feedback loops, and curriculum distillation.
Sequential intervention finally gives multi-agent prompt optimization a principled way to assign blame before writing a gradient.
A prompt optimization method for multi-agent systems that uses sequential intervention to identify target agents and clusters gradients semantically.
Separating executable world state from visual generation closes a gap that prompt-switching never could—and the benchmark margins make that concrete.
A world model that separates state evolution from video generation using executable rules and 3D bounding boxes for controllable environments.
Separating perception pretraining from RL-trained action, then uniting them via distillation, lets DriveZero outperform the human driver baseline without ever learning from one.
An end-to-end autonomous driving system combining a vision foundation model with closed-loop reinforcement learning to go beyond human demonstrations.
10 more tool picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
A full-stack RL post-training system that takes token-fidelity and train-rollout mismatch seriously, with a 744B-parameter agentic case study to show the design under real load.
Miles is an open-source system for large-scale reinforcement learning and post-training, supporting diverse backends, LoRA, and diffusion models.
A disciplined from-scratch world-action model that turns data composition into a measurable, falsifiable predictor of zero-shot manipulation capability.
GE-Act 2.0 trains a world-action model from scratch with knowledge-aligned optimization for scalable zero-shot manipulation across diverse skills.
Tools & packages
A read-only WeChat intelligence layer that's genuinely useful once you clear the non-trivial database access hurdle — and admirably honest about that hurdle.
A local-first WeChat intelligence system offering a read-only CLI, searchable chat history, daily briefings, and opportunity tracking.
A credential-borrowing usage monitor that reads from the same internal sources your coding tools do — honest about the fragility that entails.
macOS menubar app that keeps Claude Code, Cursor, Codex, and Antigravity usage limits pinned to a screen edge at all times.
A Codex skill that generates interactive holographic trading cards with real parallax and laser shimmer, delivering editable Blender scenes alongside the browser viewer.
A tool that converts a description or uploaded reference into an editable Blender card and an interactive Three.js page.
A structured agent skill that turns a topic into a fully sourced, code-rendered explainer video — honest about its costs, constraints, and fixed visual style.
A Claude Code skill that turns any topic into a motion-graphics explainer video with TTS voiceover, subtitles, and a chapter progress bar.
Git-native agent memory that trades semantic search for auditability, zero API cost, and a knowledge layer any developer can inspect with standard tools.
A pure-Go, git-native memory layer for AI coding agents using Google OKF v0.2 with sub-300µs BM25 search and an embedded MCP server.
A lunch randomizer that earns attention not for what it does but for how clearly it explains every architectural trade-off it made.
A local web app for deciding what to eat for lunch, using automatic cookie storage with no login or backend required.
8 more paper picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
A locally-run YouTube-to-shorts tool whose explicit virality scoring rubric makes its clip selection more auditable than any closed SaaS alternative.
An open-source tool that converts YouTube videos into short clips with highlight detection, subtitles, translation, and voiceover.
A four-engine translation gauntlet designed to erase LLM fingerprints, with a plausible mechanism but no published benchmark to back the evasion claim.
An open-source tool that rewrites AI-generated text to bypass detectors such as Turnitin or GPTZero.
A community mod that runs NVIDIA's neural denoiser on any supported game, with per-pass tuning and a multipass option the upstream project doesn't offer.
OptiScaler fork adding multipass DLSS with pre-SR upscaling, plus per-pass controls for lighting, detail, and colour strength.
Offline token accounting for Claude Code, Codex, and Gemini CLI, built entirely on standard-library Python with one dependency.
A CLI that parses Claude Code, Codex, and Gemini CLI session logs to calculate costs broken down by model, project, and day.