AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Harnessing Uncertainty: Entropy-Modulated Policy Gradients for Long-Horizon LLM Agents

Jiawei Wang, Jiacai Liu, Yuqian Fu, Yingru Li, Xintao Wang, Yuan Lin, Yu Yue, Lin Zhang, Yang Wang, WANG KE

ICML 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning

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 EMPG advantage modulationGRPO Qwen2.5-1.5B-Instruct · ALFWorldsuccess rate (%): 65.6 -> 73.7 (+8.1)Improved
add EMPG advantage modulationDAPO Qwen2.5-1.5B-Instruct · ALFWorldsuccess rate (%): 80.8 -> 88.1 (+7.3)Improved
add EMPG advantage modulationGRPO Qwen2.5-7B-Instruct · ALFWorldsuccess rate (%): 74.8 -> 78.5 (+3.7)Improved
add EMPG advantage modulationDAPO Qwen2.5-7B-Instruct · ALFWorldsuccess rate (%): 90.0 -> 91.6 (+1.6)Improved
add EMPG advantage modulationGRPO Qwen2.5-1.5B-Instruct · WebShopsuccess rate (%): 58.2 -> 60.8 (+2.6)Improved
add EMPG advantage modulationDAPO Qwen2.5-1.5B-Instruct · WebShopsuccess rate (%): 73.2 -> 73.8 (+0.6)No clear effect
add EMPG advantage modulationDAPO Qwen2.5-7B-Instruct · WebShopsuccess rate (%): 79.6 -> 82.7 (+3.1)Improved
add EMPG advantage modulationDAPO Qwen2.5-32B-Instruct · Deep Search overall averagescore: 62.0 -> 65.3 (+3.3)Improved
add EMPG advantage modulationDAPO Qwen2.5-32B-Instruct · Deep Search out-of-domain averagescore: 59.8 -> 63.7 (+3.9)Improved
add self-calibrating gradient scaling only (ablation)DAPO Qwen2.5-32B-Instruct · Deep Search out-of-domain averagescore: 59.8 -> 63.7Improved
add future clarity bonus only (ablation)DAPO Qwen2.5-32B-Instruct · Deep Search in-domain averagescore: 63.5 -> 66.1Improved
add self-calibrating gradient scaling only (ablation)DAPO Qwen2.5-32B-Instruct · Deep Search in-domain averagescore: 63.5 -> 63.7No clear effect

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