New LLM learning framework uses trajectory quality for improved agent performance

Researchers have introduced TQG-Exp, a novel framework for experience-based learning in large language model (LLM) agents. This method goes beyond traditional outcome feedback by incorporating process feedback through a Trajectory Quality Loss (TQL) signal. TQL evaluates both task completion and the quality of the execution process, guiding experience optimization through paired rollouts. Experiments on ALFWorld, WebShop, and ScienceWorld demonstrated that TQG-Exp significantly improves performance, outperforming existing methods on DeepSeek-V4-Flash and Qwen3.5-Flash models. AI

IMPACT This new learning framework could enhance the efficiency and effectiveness of LLM agents in complex tasks by optimizing not just the final outcome but the entire execution process.

RANK_REASON The cluster contains a research paper detailing a new framework for LLM agents. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• ALFWorld
• arXiv
• DeepSeek-V4 Flash
• Hugging Face
• LLM
• Qwen3.5 Flash
• TQG-Exp
• Trajectory Quality Loss
• WebShop

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →