FRIDAY · SEPTEMBER 25, 2026 · ISSUE 18 · YESTERDAY No. 18
Daily News — 2026-09-25
366 papers indexed on arXiv·~14,000 packages released on PyPI·29 papers surfaced by Hugging Face
None of it is in your agent's weights.
Spotlight
Inserting a learned rubric between query and scorer — then training both jointly on just 480 pairs — is a practical fix for the score instability that makes video reward models unreliable for RL.
Scalar drift is the core problem this work attacks: when a multimodal language model scores a generated video on a 1–5 scale without any explicit anchor, its internal standard wanders. Scores collapse into a narrow high…
Runtime abliteration that preserves ternary weights exactly — the right solution to a real quantization constraint, with honest evaluation caveats.
Abliteration — the technique of removing a model's refusal behavior by orthogonalizing its weight matrices against a learned refusal direction — breaks down when the model is aggressively quantized. You cannot save the…
Papers
Inserting a learned rubric between query and scorer — then training both jointly on just 480 pairs — is a practical fix for the score instability that makes video reward models unreliable for RL.
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.
Namespace renaming alone cuts SWE-bench Pass@1 by up to 7.4 points, exposing how much current agent scores reflect repository familiarity rather than reasoning.
An evaluation framework that dynamically transforms test repositories to detect whether coding agents rely on memorized cues rather than genuine reasoning.
A credible sparse MoE release that trades total parameter count for real inference efficiency, with long-context stability and agent tool-calling as its clearest practical strengths.
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.
A fixed-size episodic memory that beats compressed-history baselines on recall tasks while cutting inference latency by over 90% — the efficiency gap is the real result.
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.
Solving the math before building the environment is a small ordering change with large consequences for reward reliability and transfer.
VHD-Play generates agentic RL environments by solving a mathematical model first, then rendering its dynamics as stateful tools with built-in scoring.
Bidirectional cross-attention in joint video generators is often one-way in practice; RecCAR fixes that with a KL regularizer that costs almost nothing to add.
RecCAR is a KL regularizer for joint video diffusion transformers that aligns weak modality-to-video attention toward the stronger video-to-modality direction.
Tools & packages
Runtime abliteration that preserves ternary weights exactly — the right solution to a real quantization constraint, with honest evaluation caveats.
Runtime behavioral ablation for a 27B compressed LLM without weight modification or re-quantization.
A watermark-free, locally-run alternative to clip-generation SaaS tools, with content-aware highlight scoring built in.
A pipeline that generates vertical short videos using LLMs, Whisper transcription, highlight detection, and automated editing.
A four-step multilingual laundering pipeline that uses Turkish and Japanese as structural disruption layers to defeat AI-detection classifiers.
An open-source tool that rewrites AI-generated text into more human-sounding output using a multilingual LLM-powered pipeline.
A rigorously sequenced free curriculum that earns its breadth by actually following through on the math, from backprop to speculative decoding to agent loop engineering.
A free course covering AI engineering topics from machine learning and neural networks to LLMs, RAG, AI agents, and AI system design.
A window manager that hides rather than closes, remembers per-screen layouts, and makes no network connections - honest about its limits and worth trying.
A Mac utility that saves and restores per-project window layouts, each accessible by a single shortcut.
A lean, honest reference for effort-adaptive reasoning in long coding tasks — the lease mechanism is the idea worth stealing.
Uses Jev to tune GPT-6 reasoning effort per Codex task, reducing token usage without manual configuration.
6 more paper picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
A logo-generation platform that hasn't generated a logo yet — but the infrastructure honesty and the construction-over-sampling premise are both unusually clear.
Separate AI models for brand analysis, typography, and symbol design feed a composition engine that outputs real logo lockups with clear-space rules.
A well-extracted scaffold that shows how pre-built character APIs and written style rules let an agent compose animation without reinventing the drawing system each time.
A starter kit for animating hand-painted cartoons with Claude, bundling p5.js, p5.brush, a character, and 31 acted emotions.
A disciplined separation of LLM reasoning from runtime execution, with honest numbers and real constraints stated plainly.
JevHarness lets LLMs author task-specific eval harnesses, with optional full-trajectory reward reflection and GEPA evolutionary refinement.
A continuously auto-generated user-agent corpus — useful for test fixtures and naive rotation, but not a substitute for real browser fingerprints.
A database of user-agent strings for 100+ browsers on 100+ operating systems, updated every 5 minutes via GitHub Actions.