LOPD: latent self-distillation beats GRPO using under 30% of its rollout budget
burny_tech · x · 2026-08-16
A paper on Self-Evolving AI proposes LOPD (Latent On-Policy Self-Distillation). The problem it tackles: on-policy self-distillation (OPSD) requires hand-crafting the teacher's privileged context (the "teacher's edge"), which is labor-intensive and doesn't scale. LOPD instead lets the model learn its own privileged context end to end — pulling relevant past experiences, turning them into continuous latent tokens to condition the teacher, and giving the student dense step-by-step supervision.
On coding and tool-use benchmarks, LOPD beats standard RLVR and OPSD while using less than 30% of the rollout budget GRPO needs. Making context learnable in latent space is framed as a major step toward genuinely self-directed agent training.
More from Research
- Study: Capabilities outside training scope hard to recover; K-5 model generalization limited — paulnovosad · 2026-08-16
- Exploring Visual Model Fingerprinting via Tensor Scanning — Glad_Contest_8014 · 2026-08-16
- Ofir Press: The industry is now driven by evals — OfirPress · 2026-08-16
- Study finds forced experience consolidation degrades LLM agent memory — _AndrewZhao · 2026-08-16
- Paper: Gender-Associated Linguistic Bias in LLMs — sbulaev · 2026-08-16
- CUHK's VideoCoCo: executable code as CoT lifts VBench-2.0 average by 25.7 points — 机器之心 · 2026-08-16