Simulated Control Evaluations for Monitors
Simulated Control Evaluations (SCEs) run attack policies against LLM environment simulators that fake tool outputs, letting monitors be tested without executing real attacks. On LinuxArena tests where execution is available, SCEs reproduce attack and monitor behavior and predict safety with about a 3.6 percentage-point RMSE. — lesswrong.com
Qwen3.5‑9B Instruct Model Published
Hugging Face hosts a Qwen3.5 9B instruct-style model provided as quantized GGUF files, with documentation and instructions for local and server use. The listing shows tooling and examples for running the model on common inference stacks. — huggingface.co
Huihui Qwen3.5 0.8B Release
Hugging Face lists a Huihui Qwen3.5 0.8B quantized model available for download in GGUF formats. The model page includes quant tables and usage notes for local deployment. — huggingface.co
Rank‑1 LoReFT Fixes Reward Hacking
A single-layer rank‑1 LoReFT used as the FORGET adapter matches or improves on GRAFT’s reward-hacking suppression while using only about 0.029% of the parameters and yielding higher clean-solve rates. The method removes the need for an anchor environment in the reported experiments. — lesswrong.com
SIRA: Fellowship for Social Science
SIRA is launching a mentored program to bring social scientists into AI-safety research, aiming to enroll about 15–20 fellows to study the societal impacts of AI. The initiative intends to pair fellows with mentors and produce applied research on human and institutional effects. — lesswrong.com