Features as Rewards: Using Interpretability to Reduce Hallucinations
In our recent essay on intentional design, we described a vision for using interpretability to guide model training — moving from guess-and-check to closed-loop control. Our new paper introduces our first concrete demonstration of that vision: using lightweight probes on a model's internal representations as reward signals for reinforcement learning, applied to the problem of reducing hallucinations. We call this approach RLFR: Reinforcement Learning from Feature Rewards. Our results show that RLFR is both effective in terms of reducing hallucinations without off-target effects, and retains the ability to use the probe as a monitor at test time, which unlocks strong test-time scaling performance. Quantitatively, RLFR reduces hallucinations in Gemma-3-12B-IT by 58% (when run with our probing harness), at ~90× lower cost per intervention than the LLM-as-judge alternative, with no degradation on standard benchmarks. You can explore some samples in our interactive viewer here. Teaching a l
Features as Rewards: Using Interpretability to Reduce Hallucinations Research Features as Rewards: Using Interpretability to Reduce Hallucinations Authors Aaditya Prasad * † Connor Watts * † Jack Merullo † Dhruvil Gala † Owen Lewis † Thomas McGrath † Ekdeep Singh Lubana † * Equal contribution † Goodfire Blog post by Tom McGrath and Michael Byun Published February 11, 2026 Full Paper Read on arXiv → Your browser does not support the video tag. Contents Why is it hard to fix hallucinations? Our approach Probe pipeline: detecting hallucinations in complex long-form responses RL setup Results: red
saved by
related reading
- Reward Hacking in Reinforcement Learning | Lil'Loglilianweng.github.io
- State of RL for reasoning LLMs | A. Weersaweers.de
- RLHF: Reinforcement Learning from Human Feedbackhuyenchip.com
- Training a Misaligned Reward Seekeralignment.anthropic.com
- [2604.16812] Introspection Adapters: Training LLMs to Report Their Learned Behaviorsarxiv.org
- Reinforcement learning towards broadly and persistently beneficial modelsalignment.openai.com
- [2604.22082] Removing Sandbagging in LLMs by Training with Weak Supervisionarxiv.org
- Unfamiliar Finetuning Examples Control How Language Models Hallucinatearxiv.org
- Auditing language models for hidden objectives — LessWronglesswrong.com
- [2607.18966] Measuring Reward-Seeking via Contrastive Belief Updatesarxiv.org
- [2602.15515] The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probesarxiv.org
- [2506.19733] Breaking Barriers: Do Reinforcement Post Training Gains Transfer To Unseen Domains?arxiv.org