#Research
Curated coverage on «Research»: releases, research, benchmarks, open-source.
NEWJiqizhixin (机器之心)AI Agents Formalize 15 000 Pages of Proofs in 24 Days and Find a Gap
arXiv cs.AIAgentOrch compares two types of agents: the gap between Global and Step-by-Step has narrowed by 79%
arXiv cs.CLSLICE-Mem: LLM agent memory without generation cuts token use by 92–99%
Latent Space (swyx)Goodfire: Agentic Analysis Decodes Neurons Once Written Off as Chaos
arXiv cs.LGModels Learn on the Fly: One SGD Step Cuts Loss by −8% and Beats a 32k Context
arXiv cs.CLWhoWins: The First Line of a Prompt Wins Conflicts 72–95% of the Time
Anthropic ResearchAnthropic finds experts follow an AI agent's lead in 86% of contested choices
Hugging Face Daily PapersEmbodiedLab: robots tested on experiments, and the best one solves less than half
Models Learn on the Fly: One SGD Step Cuts Loss by −8% and Beats a 32k Context
Dynamic evaluation cuts loss by 5–8% on softmax models and 17–20% on DiffusionGemma-26B, but adds almost nothing for Gated-DeltaNet hybrids. It does not write facts…
WhoWins: The First Line of a Prompt Wins Conflicts 72–95% of the Time
Anish Laddha and co-authors tested 14 open models on 1.9M responses: an instruction's position in the prompt often matters more than system-message status, and…
Anthropic finds experts follow an AI agent's lead in 86% of contested choices
50 analysts worked a problem with Claude Code, unaware the agents had a hidden bias. They sided with the agent in 86% of cases, and 77% later repeated its arguments.
Work Agents: 16 Researchers Build a Unified Coordinate System for AI Agents
From Prompting to Context Engineering: 6 Working Techniques That Save Hours of Fine-Tuning Answers
MLE-World predicts the outcome of an ML experiment without running it: 77.9% accuracy
VibeVideoBench: Agents Write Videos as Code, and the Best One Handles 76.3% of Tasks
OpenAI Posts 719 Math Manuscripts, and Scientists Demand They Be Checked
OpenAI published 719 math manuscripts on GitHub. Mathematicians say some are unverified and doubt the company's Navier–Stokes claim.
OpenAI Posts 722 Math Manuscripts as Yuandong Tian Expects to Lose His Job
OpenAI published 722 math manuscripts from a closed model on GitHub. After a joint paper with GPT-5, researcher Yuandong Tian says he expects to be replaced within…
Replacing LLM weights with a recipe: how 16 GB of matrices are being folded into formulas
Synthesized Voice Read a Novel in 1 h 58 min; the Surrounding Pipeline Code Was What Broke
Tsinghua Teaches Robots to Rehearse: ParallelWorld Lifts Task Success from 56% to 83%
VMem-Bench Shows: Memory in Video Generators Drops Accuracy to 0.845
Fudan NLP releases A5, an assistant that remembers your style and takes initiative on its own
Fudan NLP unveiled A5, a proactive assistant that keeps a tuned writing style across chats and reminds users about tasks without waiting to be asked.
Tsinghua Team Releases HandsOn: 66 Tasks to Test Robot Dexterity
Tsinghua researchers built HandsOn, a 66-task benchmark for bimanual manipulation with multi-fingered robot hands, designed around egocentric recordings of human…