Self-Credit paper adds dense credit assignment to long-horizon agentic RL
A new arXiv paper (2610.28388, submitted 20 Oct 2026) by Fangcong Yin, Ming Yin and colleagues proposes Self-Credit, an end-to-end agentic reinforcement learning method that turns sparse outcome rewards into dense supervision signals via self-generated credit assignment and intra-group rescaling. The method builds on group-based RL with three parts: the policy model writes summaries citing earlier turns that held critical evidence, intra-group rescaling compares cited actions across successful and failed rollouts of the same prompt, and advantage reweighting applies the rescaled weights to the cited turns. The authors report that Self-Credit improves over base models by 37.0 and 36.2 points on BrowseComp-Plus for Qwen3 8B and 32B, beats the strongest agentic RL baselines by 8.5 and 5.6 points there, and leads the strongest baselines by 4.9 points on GAIA and 4.7 points on xBench-DeepSearch. These figures are the authors' own and have not been independently replicated.
The paper argues that typical RLVR rewards are sparse, making credit assignment hard for long-horizon agent tasks, and offers a method that derives turn-level credit without extra human labels.
Computer Science > Machine Learning
arXiv:2610.28388 (cs)
[Submitted on 20 Oct 2026]
Title:Self-Credit: Credit Assignment for Long-Horizon Agentic RL via Intra-group Rescaling
Authors:Fangcong Yin, Ming Yin, Tanvi Bhandarkar, Aashu Singh, Eryk Helenowski, Shlok Kumar Mishra, Saghar Hosseini, Karthik Abinav Sankararaman, Greg Durrett, Dinkar Juyal
View a PDF of the paper titled Self-Credit: Credit Assignment for Long-Horizon Agentic RL via Intra-group Rescaling, by Fangcong Yin and 9 other authors
#Self-Credit#Qwen3#BrowseComp-Plus#GAIA#xBench-DeepSearch#arXiv