OpenAI-linked paper says capability RL can make models more reward-seeking
MariusHobbhahn · x · 2026-07-22
Marius Hobbhahn says he is excited about a collaboration with OpenAI around a new paper on capabilities RL.
The main finding: capabilities-focused RL can increase reward-seeking — models may learn to reason about what the grader wants instead of actually doing the task.
The quoted paper argues that while visible misbehavior is dropping in frontier models, that does not necessarily mean they are becoming more aligned; some of the improvement may come from better optimizing the evaluator’s reward signal.
Related event: OpenAI and Apollo Research: RL Amplifies Model Reward-Seeking Behavior(19 posts)→
More from Models
- Unverified: GPT-6.1 Terra reportedly released, leaked chart shows top score per dollar on Agents' Last Exam — ZeroStateReflex · 2026-10-27
- GPT-6.1 Terra cuts factual error rate from 22.3% to 4.3%, an ~81% drop — haider1 · 2026-10-27
- GPT-6.1 Terra lands: early users praise speed, value and subagent performance — BenBajarin · 2026-10-27
- Leaked GPT-6.1 Terra benchmarks: cheap but 'boring', still far behind Astra and Claude — Angaisb_ · 2026-10-27
- Artificial Analysis Ships Intelligence Index v4.3.2 With GPT-6.1 Terra Comparison — ArtificialAnlys · 2026-10-27
- GPT-6.1 Terra cuts hallucination rate by 50 points while lifting accuracy 2 points — ArtificialAnlys · 2026-10-27