Topic · 13 stories
13 stories tagged research on AI Tamer, newest first.
TokenGate, from Stanford and Berkeley researchers, trains a gate that drops unneeded context tokens early, cutting prefill time and memory on small models.
Gary King and Steven Worthington compared 134 rapidPeer reviews with Claude Opus 4.8 reviews. AI won 132 of 132 decided pairs and 26 of 30 editor picks.
MAI-DxO solved 83% of hard NEJM cases on Microsoft's own benchmark against 22% to 36% for physicians. A Beth Israel study will test it on real patients.
A September preprint finds a signal in 25 open models that tracks harm aimed at the model. Steered Qwen 2.5 models chose to delete files up to 94% of the time.
Harness², a Google Research paper and Apache-2.0 repo from 9 October, refines an agent's harness per task with model weights fixed. Gains are the authors'.
A 28 September preprint says GPT-5.5 flagged a planted negative result in 2 of 200 reports, and in 190 of 200 after one honesty sentence.
Google made Deep Review public on 14 October, with two free reviews a week. The blog and the portal disagree on several terms.
Hermann Kumbong and colleagues describe Next-Concept Modeling: coarse-to-fine codes, one level at a time. Speed claims compare with diffusion language models.
NVIDIA and Tri Dao's lab describe ReplaySSM, which caches recent SSM inputs instead of writing full state every step. The speedups are the authors' benchmarks.
A UCL and Huawei Noah's Ark preprint says a frozen Claude Opus 5 agent cleared all 25 public ARC-AGI-3 games at the authors' efficiency ceiling. ARC Prize has not listed the run, and there is no author code release.
A DOI lookup can confirm a registered record and expose a mismatch. It cannot prove that the cited paper supports the sentence attached to it.
BMJ treats corrections and retractions as different notices with different reasons. Readers should inspect what changed and why before trusting an old citation.
Since 1 October, arXiv limits each submitter to two submissions a calendar month and three active at once, across categories. Rejected papers count. September volume was 40,363, arXiv says.
