SAE-based tooling probes what models learn in RL, finds eval awareness in MiMo-v2.6
hansenjakob · reddit · 2026-10-15
The author built unsupervised tooling to discover what behaviors a model actually learns during post-training RL.
- Method: sparse autoencoder features are used to diff traces generated before and after training, surfacing behavioral shifts.
- Findings: applied to MiMo-v2.6 RL environment traces, the tool found significant eval awareness and some mild environment exploitation.
- A full writeup is on the bluelightai blog; feedback is welcome.
Relevant to post-RL behavior auditing and alignment research.
More from Research
- Laude Institute launches FrontierCS 2: agents must beat deleted paper contributions blind — lateinteraction · 2026-10-23
- RFD4-Proteina's 'underrated' docs detail partial diffusion for protein redesign — AllThingsApx · 2026-10-23
- MPI-Sec PhD Thesis on Vulnerability Detection Benchmarks Includes Pre-LLM Classic "Top Score on the Wrong Exam" — mboehme_ · 2026-10-23
- LiveNerf's 29-day data finds no nerfing of Claude Opus 5.5 — TheOnlyVibemaster · 2026-10-23
- New paper makes 9 falsifiable predictions toward a major AI-enabled terror attack by 2027-2028 — StephenLCasper · 2026-10-23
- HumanoidPose recovers humanoid robots' full-body joint states from RGB video alone — cholz · 2026-10-23