#Research
Curated coverage on «Research»: releases, research, benchmarks, open-source.
LEASH hands hard steps to a strong agent and cuts Opus 5 costs by 19.2%
Researchers personalize three MoE LLMs and find the experts responsible for adaptation
Why You Can't Trust an AI Agent to Grade Its Own Work: A Report Claiming -9,500 Lines When the Codebase Actually Grew by +21,489
HeirCommit Filters Out Up to 83% of Harmful Updates in Self-Evolving LLMs
EvoHeir checks an updated model on fresh questions before it is inherited, raising average math scores 5.3% over Agent0 across four base models.
ACT3D teaches a VLM to explore 3D scenes and beats Gemini 3.6 Flash: 53.2% across six benchmarks
ACT3D, a Qwen-based model, learns to move through a 3D scene and pick its own viewpoints via RL: 53.2% average across six benchmarks with 50–70% fewer frames.
Librarian Agents: LLMs Pre-Build the Workspace and Gain 33 Points on BrowseComp-Plus
A builder agent lays out a document corpus as a file system in advance, and a second agent answers questions. On BrowseComp-Plus the gain is 33 points; on BRIGHT it…
"Wait" in Reasoning Models Turns Out to Be Inherited from Data, Not Discovered by RL
OpenAI's Interpretability Founder Says 90% of Neurons Have a Single Function
Claude Traces the Origin of Five NIST Curve Seeds in Half an Hour
An LMS and CRM at 205,000 Lines of Code, Built by AI Agents: The Real Cost of Running It, Three Lead States, and One \r Character
Changing the definition of "unsafe" moves a guardrail score from 0 to 100 with 0.03% of weights
The authors labeled ~53 thousand prompts twice under different policies and trained 20 guardrail models. Benchmark spread reached 23x, and some test sets were…
Researchers Propose Anti-Persistence: LLM Agent Memory Without Stored Rules
A new memory design for LLM agents stores trajectories and outcomes, then rebuilds guidance from scratch for each task. It uses the fewest LLM calls of the designs…
POPE audit: the choice of "yes/no" tokens shifts LLaVA-1.5 F1 by 6.13 points
Parallel Refinement Instead of Diffusion: DRLMs Hit 93.9% on Sudoku-Extreme
TokenWorth Teaches AI Agents to Forget: Best Results on Three Environments With the Fewest Tokens
Ropedia W0 trained on 10,000 hours of human video and 3,000 hours of robot data
Naive RAG Beat Agentic Memory Systems in Six of Seven Tests
Raw-RAG, which uses dense retrieval and reranking over raw conversation history, took first place in six of seven setups and beat the best agentic memory by…
Cola2 beats Llama-3-8B on HellaSwag with 43 times fewer tokens
Hierarchical model Cola2 compresses text into continuous latents and scores 60.10 across eight tasks after training on 346B tokens, over 43x less than Llama-3's…