ByteDance Research: Filtering Spurious Signals in LLM Distillation

ByteDance · hf · 2026-08-06

On-Policy distillation (OPD) is commonly used to transfer large model capabilities to student models. However, ByteDance researchers point out that token-level judgments are often driven by input-agnostic language priors or formatting conventions, generating spurious signals. These signals create large gradients but contribute little to actual task improvement.

To address this, the team proposed SA-OPD (Spurious-Signal-Aware On-Policy Distillation). The framework introduces a lightweight proxy to estimate whether a token-level distillation signal truly depends on the input. It filters tokens exhibiting both "low input-groundedness" and "extreme distillation divergence," removing high-impact spurious updates. Experiments show this method consistently outperforms Vanilla OPD in both LLM and VLM settings.

Original post →

More from Research

Research channel →