analiza aiqbench
This research introduces RePro, a training framework that leverages retrospective progress assessment to enhance long-horizon LLM agent performance without requiring separate process reward models or continuous teacher annotation. By combining forward action estimation with hindsight-aligned reflection on completed trajectories, the method achieves significant success rate improvements across diverse environments like WebShop and ALFWorld. The findings highlight the potential of self-generated supervision signals to overcome limited explicit feedback in complex task completion scenarios.
Retrospective Progress-Aware Self-Refinement for LLM Agent Training
arXiv:2606.14302v2 Announce Type: replace Abstract: Long-horizon LLM-based agents receive rich environmental observations during interaction, yet outcome rewards provide limited explicit supervision about how individual actions advance task completion. We investigate whether agents can turn this interaction evidence into useful training signals through retrospective progress assessment. A WebShop pilot shows that direct progress prompting reduces task success, whereas hindsight-annotated demonstrations improve it. We introduce RePro, Retrospective Progress-Aware Training, with a forward-then-reflect rollout: the agent estimates progress while acting, then reassesses each step using the completed trajectory and outcome. After warmup with externally generated demonstrations, policy optimization combines self-generated progress differences, online-retrospective alignment, and format rewards with environment feedback, requiring neither a separate process reward model nor ongoing teacher annotation. Experiments on WebShop, ALFWorld, and Sokoban show that RePro enhances the Qwen family's performance, with up to 11.57% success rate gains.
Abstract:Long-horizon LLM-based agents receive rich environmental observations during interaction, yet outcome rewards provide limited explicit supervision about how individual actions advance task completion. We investigate whether agents can turn this interaction evidence into useful training signals through retrospective progress assessment. A WebShop pilot shows that direct progress prompting reduces task success, whereas hindsight-annotated demonstrations improve it. We introduce RePro, Retrospective Progress-Aware Training, with a forward-then-reflect rollout: the agent estimates progress while acting, then reassesses each step using the completed trajectory and outcome. After warmup with externally generated demonstrations, policy optimization combines self-generated progress differences, online-retrospective alignment, and format rewards with environment feedback, requiring neither a separate process reward model nor ongoing teacher annotation. Experiments on WebShop, ALFWorld, and Sokoban show that RePro enhances the Qwen family's performance, with up to 11.57% success rate gains.
Computation and Language (cs.CL)
arXiv:2606.14302 [cs.CL]
(or arXiv:2606.14302v2 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2606.14302 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 227 cuvinte
Subiecte conexe
Retrospective Progress-Aware Self-Refinement for LLM Agent Training
arXiv · 15 iun. 2026
The Continual Constitutional Convention: How language models revise the documents that govern them
AI Alignment Forum · 13 oct. 2026
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · 17 sept. 2026
[AINews] not much happened today
Latent Space · 10 sept. 2026