Self-rewriting: a 27B model edits its own weights without training
A new preprint shows Qwen3.5-27B gaining 6.6 points on unseen maths by editing its own weights, with no gradients, and only at the largest size tested.
Read articleInsights
One notable recent AI research paper or result, explained clearly for practitioners.
A new preprint shows Qwen3.5-27B gaining 6.6 points on unseen maths by editing its own weights, with no gradients, and only at the largest size tested.
Read articleA new METR note finds the latest AI agents can reliably pick out staged safety tests, so clean eval results may say less than they seem to.
Read articleA new benchmark hides most requirements with simulated colleagues, and the best coding agents fully complete fewer than 8% of its 101 tasks.
Read articleA Berkeley and UIUC paper shows that trimming what an agent harness sends to the model cut costs by 23% on average, with a 3% relative drop in accuracy.
Read articleA NeurIPS 2026 paper shows votes, chosen prompts and extra decoy models can lift a model up to 45 places on arena leaderboards.
Read articleAmazon and Columbia researchers turn logged agent runs into code workflows that finish in about a second, at a small cost in accuracy.
Read articleGoogle researchers built SynthCorp to simulate company data estates for testing data agents, and the top agent still misread changes in history.
Read articleA Villanova preprint traced 95 failed multi-agent runs and found most errors entered early, often with contrary evidence already in the trace.
Read articleMicrosoft Research showed that editing a few dormant neurons can make an agent model run attacker code on a trigger. Here's how it works and what to check.
Read articleA new arXiv paper finds ten small agent teams beat 100 solo agents and one 100-agent team, provided a verifier can score the results.
Read articleA new technical report finds AI agents keep most of their accuracy on short, frequently compacted context, at much lower estimated cost.
Read articleA November arXiv study finds learned routers barely beat the best single model, though they can cut costs sharply once a quality floor is certified.
Read article