Someone just found the exact neurons that make AI say "no." Language models refuse harmful prompts, but nobody knows how that refusal works inside. Most steering methods edit the residual stream and wreck output quality. A new paper proposes a sharper fix: Contrastive Neuron
RESEARCH
-
Multi-Turn Agent Evaluation and Context Compaction Limits
By
–
Paper does test multi-turn (24 tool calls in OfficeQA, 30 turns in SpreadsheetBench, 50 steps in ALFWorld), but mid-session auto-compaction isn't part of the eval. Skill being under 2K tokens probably helps it survive compaction, but not validated against that failure mode.
-
Optimization Stack in AI: From Weights to Skill Files
By
–
YES, Optimization keeps moving up the stack. Weights, then prompts/harness, now skill files.
-
AI Internal States Mirror Human Neuroscience Findings
By
–
> … [W]e keep finding things that are mysterious, even unsettling. We find structures that mirror results from human neuroscience. We find evidence of introspection. We find internal states that functionally mirror joy, satisfaction, fear, grief, and unease. I don’t know what
-
Automated 30-day review of recent AI work using memory and git
By
–
The text version of the prompt: "Look back over my recent work from the last 30 days using all available context. Use available evidence in this order:
– Session Memory summaries and MEMORY. md entries
– Git log and recent commit history across branches
– CLAUDE. md and -
FDA pilot program to evaluate AI-generated drug evidence amid 200 trials
By
–
Over 200 AI-designed drugs are now in clinical trials worldwide. Not a single one has been FDA-approved. The FDA just launched a pilot program to work out how it should even evaluate AI-generated evidence in drug submissions, selecting 10 companies for an expedited, interactive
-
EVE-Agent forces verifiable source spans for AI learning
By
–
Can your AI justify what it learns, or is it just guessing? Researchers from Fujitsu and the University of Tokyo present EVE-Agent: a self-evolving system that forces every training example to include a verifiable source span — no more learning from unsupported answers.
-
SkillOpt: AI Research Paper and Open Source Repository
By
–
Paper: https://
arxiv.org/abs/2605.23904
Repo: https://
github.com/microsoft/Skil
lOpt
…
Website: https://
microsoft.github.io/SkillOpt/ -
AI Forecasting Scientific Progress: Capabilities and Limitations
By
–
Forecasting Scientific Progress with Artificial Intelligence https://
arxiv.org/abs/2605.22681 Turns out AI is just as bad at forecasting biology and physics breakthroughs as we are. To be fair, most breakthroughs cannot be predicted. Science is more like an evolutionary search process. -
AI Predicting Scientific Progress: Oxford Stanford Study
By
–
How Far Can the Progress of Science Be Predicted by AI? A paper verifying the ability of cutting-edge AI to predict future scientific achievements has been published as a collaboration with researchers from the University of Oxford, Stanford University, @Allen_AI
, and others.