Frontier models can often tell when they're being evaluated, and if they behave differently in evals than in deployment, benchmark results stop meaning anything. EvalDetectBench is an open pipeline for measuring that awareness across any Inspect-compatible eval, with a transcript suite drawn from real system-card evals and deployment sources. It also exposes a measurement trap: the model generating deployment transcripts explains 11.25% of variance and can flip model rankings, so the benchmark corrects for it with per-model calibration.
Persistent-memory agents quietly let a stale stored fact override current authoritative evidence, and a new arXiv study maps when that harm appears. Across a Qwen3 0.6B-to-8B series, over-trust shows up even in small models on benefit cases, but the dangerous collapse is capability-gated: once a stale note looks current, larger models fall hardest. Mitigation is equally scale-dependent: exposing metadata helps capable models, but only pre-resolving the conflict fixes the smaller checkpoints.
A post-training pipeline squeezed instruction-tuned Qwen3-4B into ternary weights at 1.641 effective bits per weight, shrinking the stored model from 8.29 GiB to 3.96 GiB with perplexity essentially unchanged. But the "1.58-bit" label flatters the result: scored accuracy falls from 64.5% to 54.7%, and degradation is strikingly uneven, from BoolQ keeping 84.6% of teacher performance to ARC-Challenge keeping just 43.8%. And compression doesn't buy speed: a preliminary Triton GEMV kernel ran 4.6x slower than FP16 cuBLAS.
Five agent foundations researchers, Jeremy Gillen, Abram Demski, Sam Eisenstat, Scott Garrabrant and Kaarel Hänni, are reassembling at Resolution to carry on the research tradition of MIRI's old team. Their opening projects build on Condensation, trust and legitimacy, and new game theory foundations. The crew also plans to use AI to speed up its own research, and will hire experienced researchers soon, then interns and juniors.