AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

Hierarchy-of-Groups Policy Optimization for Long-Horizon Agentic Tasks

Shuo He, Lang Feng, Qi Wei, Xin Cheng, Lei Feng, Bo An

ICLR 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants · 33 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 hierarchical groups (G1:K) beyond step-level group (ablation)HGPO K=2 (Qwen2.5-1.5B) · ALFWorld in-distributionsuccess rate (%): 90.88 -> 92.77Improved
add hierarchical groups (G1:K) beyond step-level group (ablation)HGPO K=4 (Qwen2.5-1.5B) · ALFWorld in-distributionsuccess rate (%): 91.14 -> 94.85Improved
add hierarchical groups to WebShop (ablation)HGPO K=2 (Qwen2.5-1.5B) · WebShoptask success rate (%): 73.24 -> 71.54Hurt
use adaptive weights (alpha=1) instead of uniform weights (ablation)HGPO K=4 (Qwen2.5-1.5B) · ALFWorld in-distributionsuccess rate (%): 92.12 -> 94.85Improved
use adaptive weights instead of uniform weights (ablation)HGPO K=2 (Qwen2.5-1.5B) · ALFWorld in-distributionsuccess rate (%): 94.33 -> 92.77Hurt
use alpha=2 instead of alpha=1 (ablation)HGPO K=4 (Qwen2.5-1.5B) · ALFWorld in-distributionsuccess rate (%): 94.85 -> 93.55Hurt
add trajectory-level advantage to final estimate (ablation)HGPO K=4 (Qwen2.5-1.5B) · ALFWorld in-distributionsuccess rate (%): 94.85 -> 91.27Hurt
HGPO vs GiGPOHGPO K=4 (Qwen2.5-1.5B) · WebShoptask success rate (%): 73.24 -> 78.12Improved
HGPO vs GiGPO on larger modelHGPO K=4 (Qwen2.5-7B) · ALFWorld in-distributionsuccess rate (%): 95.63 -> 95.96No clear effect
increase memory length K from 2 to 4HGPO (Qwen2.5-1.5B) · ALFWorld in-distributionsuccess rate (%): 92.77 -> 94.85Improved
use only Oracle steps for policy optimization (ablation)Stepwise GRPO · ALFWorld/WebShopsuccess rateHurt

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