AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

GLARE: Scalable Neuro-Symbolic Reward Shaping for LLM Agents via Group-Level Automata

Jingyuan Yan, Qingchen Liu, Qichao Ma, Jiahu Qin

ICML 2026 · Reinforcement learning › RL from feedback · approach: Scaling RL with verifiable rewards

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 GLARE automaton-based dense reward to GRPOGRPO Qwen2.5-1.5B · ALFWorldsuccess rate (%): 72.4 -> 84.5Improved
add GLARE automaton-based dense reward to GRPOGRPO Qwen2.5-7B · ALFWorldsuccess rate (%): 75.7 -> 86.7Improved
use GLARE rewards instead of trace-level LLM judge rewardGRPO Qwen2.5-1.5B · ALFWorldsuccess rate (%): 76.4 -> 84.5Improved
use GLARE rewards instead of trace-level LLM judge rewardGRPO Qwen2.5-7B · ALFWorldsuccess rate (%): 82.4 -> 86.7Improved
add trace-level LLM judge dense reward to GRPOGRPO Qwen2.5-1.5B · ALFWorldsuccess rate (%): 72.4 -> 76.4Improved
add trace-level LLM judge dense reward to GRPOGRPO Qwen2.5-7B · ALFWorldsuccess rate (%): 75.7 -> 82.4Improved
use GLARE automata instead of trace-level LLM scoring for token costGLARE · ALFWorldcompletion token usage: 15% of trace-level usageImproved
use negative-only LTL rewards (remove positive milestone rewards) (ablation)GLARE Qwen2.5-1.5B · ALFWorldsuccess rate (%)Improved
add negative LTL penalties on top of positive-only rewards (ablation)GLARE Qwen2.5-1.5B · ALFWorldsuccess rate (%)Improved
use GiGPO instead of GLARE under original observationsGLARE · ALFWorldsuccess rate (%): 86.7 -> 83.3Hurt
perturb observations with distractor descriptions (GiGPO)GiGPO · ALFWorldsuccess rate (%): 86.7 -> 71.4Hurt
use GLARE instead of GiGPO on perturbed observationsGiGPO · ALFWorld perturbedsuccess rate (%): 71.4 -> 83.3Improved

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