proomt

Search

Search posts, papers, and topics

All posts

Hugging Face Daily PapersXiuYu Zhang, Wei Chow, Junfeng Fang1 min readpaperadvanced

What Does Privileged Information Add to On-Policy Self-Distillation?

Summary

The paper isolates the contribution of privileged teacher information in on‑policy self‑distillation (OPSD) using a 5k‑problem math suite. Reference‑free distillation explains most gains, with only modest extra benefit from full solution traces.

  • Reference‑free OPSD accounts for the bulk of Qwen3‑1.7B's improvement; privileged references add only modest gains.
  • Polished solution references give small benefit, while full reasoning traces add ~2 % accuracy for SmolLM3‑3B at step 50.
  • Switching from short direct‑response rollouts to long thinking‑enabled rollouts can turn gains into losses, even with the same data.
  • Altering token‑level teacher supervision (teacher profiles) has little impact on the student model's behavior.

LLM researchers and engineers designing self‑distillation pipelines need to know that privileged references yield limited extra performance, guiding data and teacher design choices.

6/10

Related reading

  1. When EOS Tokens Disagree: Understanding Length Inflation in On-Policy Distillation

    On-policy distillation (OPD) can lead to excessively long student responses, a phenomenon called length inflation. This paper identifies "termination-token mismatch" between base students and post-trained teachers as a key source, where models place stopping probability on different EOS tokens. Treating functionally equivalent EOS tokens as a shared semantic stopping action substantially mitigate…

    Hugging Face Daily Papersarxiv.org1 minpaper
  2. RetireOPD: Self-Retiring On-Policy Distillation for Agentic Reinforcement Learning

    RetireOPD introduces a self‑retiring on‑policy distillation framework for multi‑turn RL agents. A skill‑conditioned teacher is first trained with environment rewards, then a skill‑free student learns jointly via RL and token‑level distillation. The student automatically drops the teacher once its performance gap stops shrinking and it reaches a target success‑rate fraction, after which training c…

    Hugging Face Daily Papersarxiv.org1 minpaper
  3. Trajectory as the Teacher: Few-Step Discrete Flow Matching via Energy-Navigated Distillation

    Apple researchers propose Trajectory‑Shaped Discrete Flow Matching (TS‑DFM), a training‑time distillation method that replaces blind stochastic jumps in discrete flow‑matching with an energy‑based compass to select higher‑quality intermediate tokens. On a 170 M‑parameter language model, the 8‑step student outperforms the 1 024‑step teacher by 32 % perplexity while being 128× faster, beating basel…

    Apple Machine Learning Researchapple.com1 minpaper
  4. Article: Your Next DSL Author Is a Language Model

    Typed Domain Grounding (TDG) embeds a DSL inside a mainstream language the LLM already knows (e.g., Kotlin) and uses the host compiler as an oracle. The author describes five building blocks—embedding, choosing a host language with high training‑data frequency, compiler‑driven type safety, a generate‑compile‑repair loop, and an on‑demand teaching tool—and shows measured results from kUML, a Kotli…

    InfoQinfoq.com18 min