Treating the rollout policy as an optimization target rather than a fixed choice lifted on-policy distillation accuracy 6.91 points and cut generated length by 41 percent
AspireAI community · arXiv
On-policy distillation trains a student on its own sampled trajectories under teacher supervision, which normally fixes the rollout policy at the current student. This work shows that choice compounds: because the teacher must supervise whatever the student samples, including trajectories the teacher itself would never reach, supervision on those can turn noisy or misleading and push the student into patterns like repetition that then persist across rounds. MPOPD instead derives a student-anchored, teacher-guided rollout distribution from a local trust-region objective over relative teacher support, and approximates that geometric mixture of student and teacher policies with a proposal policy that generates rollouts in a single forward pass. Across 16 math, knowledge and medical reasoning benchmarks the method reports consistent gains over on-policy distillation and recent variants, with average accuracy on math reasoning up 6.91 points over the OPD baseline on Qwen3-4B-Base while average generation length fell 41 percent. The caveat is that the trusted-region objective and the single-pass mixture approximation still need tuning per setup, and the paper does not report results on long-horizon agentic tasks.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.