用行动赢得尊重

Alan Hou

Notes on coding, AI, and technology.

记录编程、人工智能与技术探索。

Unpacking the Compute Efficiency of Strong-to-Weak OPD Against GRPO

拆解强弱在线策略蒸馏（OPD）对决 GRPO 的算力效率真相

Salesforce AI Research presents the first rigorous, FLOPs-matched comparison between on-policy distillation (OPD) and Group Relative Policy Optimization (GRPO). While external post-trained teachers consistently lose to GRPO across all task difficulty tiers and cause response-length inflation, using the student's own GRPO checkpoint as a 'self-GRPO teacher' changes everything—recovering accuracy with 50x fewer FLOPs. Moreover, multi-teacher OPD consolidates multi-domain specialists into a unified generalist with 42% fewer total FLOPs than joint GRPO.

Nov 8, 2026 Read article

Alan Hou

Unpacking the Compute Efficiency of Strong-to-Weak OPD Against GRPO

Recent Posts / 近期文章

VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding

Benchmarks as Software: Terminal-Bench 2.1 and the Reality of Benchmark Maintenance Debt

Spectral and Subspace Dynamics of LoRA: Why Reinforcement Learning Demands Different Optimization Than SFT

NEO-VLA: Towards Native, Encoder-Free Vision-Language-Action Policies

Visual Token Routing: Learning to Keep Tokens That Matter for Computer-Use Agents

DreamX: Repurposing Video Diffusion Models as Universal Latent Action Learners

You Don't Need Long Context: Semantic State Space Models and the 7x Inference Penalty

Pretraining Data Scale Reverses the Muon-AdamW Fine-Tuning Gap

WebHarbor: Docking Real Websites for Evolving GUI Agent Environments

Series / 系列文章

Nobel Prize in Economics (1969–2025)

Odoo 12 Development Cookbook

Odoo 14 Development Cookbook

TouchDesigner Tutorials

Mastering Python Scripting for System Administrators

Odoo 12 Development Essentials

Hung-yi Lee ML 2025

Dwarkesh Podcast Notes

Andrej Karpathy LLM Notes

freeCodeCamp