CLM Author Adds Acknowledgments: RL Built on Polar, Data from OpenResearcher Offline Pipeline

RulinShao · x · 2026-10-02

After a self-review, CLM paper author Rulin Shao added acknowledgments for two important citations that were previously missed. First, @billxbf's Polar was used as the infrastructure for CLM's reinforcement learning training—its harness-agnostic design allows training language models with any harness. Second, @DongfuJiang's OpenResearcher was used as DR training data—that work proposes a fully offline pipeline for synthesizing deep-research trajectories (GPT-OSS-120B + a local retriever + a 10T-token corpus, requiring no search APIs and fully reproducible). The original SFT recipe boosted Nemotron-3-Nano-30B-A3B from 20.8% to 54.8% (+34.0) on BrowseComp-Plus, and CLM borrowed its QA setup and local data storage. The author welcomes pointers to further related work.

Related event: Meta's Context Language Models: Let Models Edit Their Own Context, Cutting Compute and Lifting Agent Scores(12 posts)→

Original post →

More from Research

Research channel →