Search papers, labs, and topics across Lattice.
This paper introduces SKALD, an on-policy self-distillation framework that leverages abstract skills as privileged signals to enhance reinforcement learning performance. By employing a dual-context approach with a question-only student and a teacher conditioned on skill cards, SKALD effectively transfers skill advantages into shared model parameters without requiring privileged input during testing. The method significantly outperforms traditional group-relative reward strategies across multiple mathematics benchmarks, demonstrating that abstract skills can provide valuable supervision when conventional rewards fail to yield informative signals.
Abstract skills can transform reinforcement learning by providing dense supervision when traditional reward signals are inadequate, leading to significant performance gains.
Reinforcement learning with verifiable rewards yields no group-relative signal when rollout groups are uniformly correct or uniformly wrong, which account for 63.0-68.0% of groups in our experiments. We propose SKALD (Skill-Anchored Latent Distillation), an on-policy self-distillation framework that uses two context views of the same Qwen3-Base model: a question-only student and a teacher conditioned on an abstract, explicit-answer-filtered skill card. The student is trained on its own prefixes, transferring the skill-induced advantage into shared parameters without privileged input at test time. To stabilize context-induced distribution mismatch, SKALD employs an annealed exponentially tilted objective that downweights teacher-preferred tokens with very low student likelihood; as the tilt vanishes, it converges to teacher cross-entropy and recovers the forward-KL student gradient. An empirical gate activates distillation only when verified rollouts estimate a positive teacher advantage. Across five held-out mathematics benchmarks, SKALD improves overall avg@8 over GRPO by +2.46, +4.85, and +12.01 at 0.6B, 1.7B, and 4B, respectively. At 1.7B, zero-variance-only distillation recovers 84.7% of the full gain, while SKALD remains +4.06 above FLOP-matched GRPO and exceeds contextual skill exposure by +3.77. These results show that abstract skills provide dense supervision where group-relative rewards become uninformative.
