ReGE turns recorded agent experience into guidance for a retry
Original title: ReGE: Stronger Agents through Experience Synthesis
ReGE-Syn raises accuracy 9.9 points for GPT and 8.8 for Claude while cutting guidance-construction tokens versus adapted AggAgent.
ReGE converts recorded interactive attempts into guidance for a fresh run from the initial environment state. ReGE-Syn uses one coach call on the same base model, with no parameter updates or stronger teacher. Across four benchmarks it improves GPT and Claude versus unguided execution, and on BrowseComp-Plus it shortens guidance relative to adapted AggAgent while reducing combined tokens.
- Guidance-construction tokens fall 98.2% on GPT and 95.5% on Claude versus adapted AggAgent.
- Combined construction and execution tokens drop 15.7% and 35.4%, excluding source generation.
- Joint synthesis yields shorter guidance than Direct history or Independent summaries.
- Process information is associated with fewer student steps on BrowseComp-Plus ablations.
From the original article
1 min read at huggingface.co
Repeated attempts at interactive tasks uncover facts, useful operations, and failures that can inform subsequent execution. We introduce ReGE, a framework for turning this recorded experience into guidance for a fresh attempt.
By Jiarong Liang, Yubo Wang, Zijie Liu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.