Search papers, labs, and topics across Lattice.
This paper introduces LLM-as-a-Tutor, a novel framework that enhances reinforcement learning (RL) for non-verifiable instruction following by adapting training prompts dynamically based on the evolving policy's capabilities. By enabling a single LLM to serve both as an examiner that identifies non-challenging prompts and as a generator that appends constraints, the method ensures a continuous alignment between prompt difficulty and policy performance. The results demonstrate significant improvements across three complex instruction-following benchmarks, highlighting the importance of prompt adaptation in achieving effective policy-awareness in RL settings.
Static prompts in RL training can hinder performance, but LLM-as-a-Tutor dynamically adapts them to match policy capabilities, leading to superior outcomes.
Reinforcement learning (RL) for non-verifiable instruction following increasingly relies on LLM judges with prompt-specific rubrics as reward signals. While recent methods adapt these rubrics to the evolving policy during training, the training prompts themselves remain static, drawn from fixed corpora. This static approach often results in a critical misalignment between prompt difficulty and policy capability, leaving the judge unable to recover a discriminative reward signal when prompts fail to elicit quality variance among rollouts. To address this misalignment, we introduce LLM-as-a-Tutor, a framework that extends the LLM's role from judge to tutor: a single model serves as an examiner that pairwise-compares policy rollouts to detect non-challenging prompts, and as a generator that appends atomic constraints to them. This append-only design monotonically raises difficulty in step with the policy's capability, producing a self-calibrating training signal without external difficulty schedules. On three complex instruction-following benchmarks, our method consistently outperforms both policy-unaware baselines and prior policy-adaptive methods that adapt rubrics or rewrite prompts, suggesting prompt adaptation as a missing axis of policy-awareness in non-verifiable RL.
CMU ML