BlackSwan (arXiv:2611.02950) shows confirmation-style action verification lets wrong writes through. A refute-first expect-suspect-test pass lifts AppWorld challenge TGC by 14.9–25.2 points and WebShop by 12.7.
CircleCI's October 9 outage was scheduled three days earlier: Temporal's 3-day retention queued the October 6 history for deletion, and those deletes filled a 1TB Postgres WAL volume. Deletes are writes, and TTL jobs put them on the calendar.
On a 90-case benchmark, harnesses that let the model choose what to read beat five published root cause analysis agents. The published agents missed a service that simply went quiet.
GitHub's 19 October Actions outage came from a retry storm. The connection limiter that would have capped it was already in production, in log-only shadow mode.
Search-augmented models often answer from snippets without opening a single page. A new Imperial College study finds that is where most of their stale answers and contradicted citations come from.
Autoresearch agents plateau as history shrinks edits. Forgetting sessions and keeping only the best candidate unblocks discovery under the same budget.
A new bias study requested thinking budgets from 0 to 8,192 tokens. GPT-5.1 used about 117 at every positive setting, and DeepSeek-R1 and Qwen3 thought even when told not to. Measured by what the models actually spent, more thinking did not reliably reduce bias.
A new Seoul National University paper finds LLMs repeat the original answer under negation 37 to 71% of the time, and more often when they're confident. The fix that looks best on standard metrics quietly teaches one stock wrong answer.
TasteVal finds AI experimental taste doubling every 3 months since Dec 2025. The same paper labeled 2,000+ model experiments and found zero invented ideas.
OpenAI’s Sep 30 writeup shows protected reasoning was extracted without breaking encryption: by replaying encrypted CoT across chats until it surfaced in the clear.
FTA freezes failed tool observations before generation. Across six models and 3,600 human-annotated responses, false-success falls from 22.8% baseline to 0.8% with a structured evidence contract, while useful responses rise from 74.9% to 98.8%.
Anthropic's Project Swap put Claude agents into a book barter market. Live efficiency hit ~0.55 vs a ~0.89 optimum; ~85% of the shortfall was preference representation, not the trading floor.