AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Agentic Reinforcement Learning with Implicit Step Rewards

Xiaoqian Liu, Ke Wang, Yuchuan Wu, Fei Huang, Yongbin Li, Jianbin Jiao, Junge Zhang

ICLR 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning · 44 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add implicit step rewards (iStar) to RLOORLOO · WebShopsuccess rate (%): 77.4 -> 86.5Improved
add implicit step rewards (iStar) to RLOORLOO · VisualSokobansuccess rate (%): 86.3 -> 91.7Improved
add iStar to GRPO on SOTOPIA self-chat hardGRPO · SOTOPIA self-chat (Hard), Llama3.1-8Bgoal completion: 7.92 -> 8.06Mixed
add iStar to GRPO with GPT-4o partnerGRPO · SOTOPIA GPT-4o partner (Hard), Llama3.1-8Bgoal completion: 6.68 -> 7.16Improved
combine implicit step rewards at advantage level rather than merging into outcome reward (ablation)iStar · WebShopsuccess rate (%): 81.3 -> 89.1Improved
use step-level instead of token-level implicit rewards (ablation)iStar · WebShopsuccess rate (%): 82.0 -> 89.1Improved
use step-level implicit rewards instead of token-level (ablation)iStar · VisualSokobansuccess rate (%): 89.1 -> 93.0Improved
use implicit step rewards instead of ground-truth step penalties (ablation)iStar · VisualSokobansuccess rate (%): 87.5 -> 93.0Improved
ground-truth process rewards added to RLOO (ablation)RLOO · VisualSokobansuccess rate (%): 85.9 -> 87.5Improved
add iStar to smaller base modelRLOO · WebShop, Qwen2.5-1.5Bsuccess rate (%): 71.9 -> 80.5Improved
add iStar to RLOO (different RL algorithms)RLOO · WebShop and VisualSokobansuccess rate (%): +6.3% on bothImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy