AgentHER framework boosts LLM agent training with failed trajectory relabeling

Researchers have developed AgentHER, a new framework designed to improve the training of LLM agents by repurposing failed trajectories. The system adapts Hindsight Experience Replay to natural language, identifying alternative achievable goals within failed attempts. This method converts discarded data into valuable training material, significantly boosting agent performance and data efficiency across various model sizes. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Enhances LLM agent training efficiency by leveraging failed trajectories, potentially improving performance on complex real-world tasks.

RANK_REASON Academic paper introducing a novel framework for LLM agent training.

Read on arXiv cs.CL →

• AgentHER
• LLM
• GPT-4o
• WebArena
• ToolBench
• Zhou et al.
• Qin et al.
• Andrychowicz et al.
• Qwen2.5-72B
• Qwen2.5-7B
• LLaMA-3.1-8B
• ShareGPT
• DPO
• SFT

• paper
• model release