| name | triage-role-typed-credit-assignment |
| version | 1.0.0 |
| description | Role-typed credit assignment framework for agentic reinforcement learning that addresses outcome-only credit blind spots |
| tags | ["reinforcement-learning","agentic-ai","credit-assignment","llm-agents","process-reward"] |
| categories | ["reinforcement-learning","agent-research"] |
| source | arXiv 2606.32017 |
| date_collected | 2026-07-02T00:00:00.000Z |
TRIAGE: Role-Typed Credit Assignment for Agentic Reinforcement Learning
Overview
TRIAGE is a role-typed credit assignment framework for agentic reinforcement learning that addresses two critical blind spots in standard GRPO (Group Relative Policy Optimization): punishing useful exploration in failed rollouts and reinforcing redundant actions in successful rollouts.
Problem Statement
Standard GRPO uses final verifier outcomes as uniform advantages over all action tokens. This outcome-only signal is structurally incomplete:
- Failed rollouts: Useful exploration gets punished alongside bad actions
- Successful rollouts: Redundant or regressive actions get reinforced alongside decisive progress
Core Methodology
1. Semantic Role Classification
A structured judge classifies each action segment into four roles:
- Decisive progress: Actions that directly advance toward the goal
- Useful exploration: Actions that gather information but don't directly progress
- No-progress infrastructure: Necessary but non-progressive actions (e.g., navigation)
- Regression: Actions that undo progress or move away from the goal
2. Role-Conditioned Process Rewards
Fixed role-conditioned rules map semantic labels to bounded segment-level process rewards:
- Preserves verifier outcomes as the source of optimization direction
- Adds semantic role axis to outcome credit
- Corrects the two main blind spots of outcome-only credit
3. Theoretical Foundation
Role-conditioned credit is proven to be the optimal segment-level correction expressible from role labels alone:
- It's a projection of per-segment advantage residual onto the role variable
- Fixed role constants reduce advantage estimation error when the judge is reliable
- Connects to lower-variance policy gradients
Key Results
Performance Improvements
- ALFWorld: Improves success rates over GRPO for two policy models
- Search-QA: Outperforms both scalar judge-derived process reward and outcome-supervised value baselines
- WebShop: Reduces environment-facing turns by 14.8% relative to GRPO
Efficiency Gains
- ALFWorld: 10.4% reduction in environment-facing turns on completed rollouts
- WebShop: 14.8% reduction in environment-facing turns on completed rollouts
Ablation Insights
- Regression detection inside successful trajectories is the dominant contributor to gains
- Exploration credit provides consistent secondary gains
- Gains come from role typing, not merely adding dense rewards
Implementation Patterns
Integration with GRPO
for rollout in grpo_rollouts:
segments = segment_trajectory(rollout.actions)
for segment in segments:
role = judge_classify(segment)
process_reward = role_conditioned_reward(role)
segment.advantage = outcome_advantage + process_reward
Judge Design
- Structured judge outputs semantic role labels (not scalar scores)
- Fixed mapping from roles to bounded reward values
- Judge reliability directly impacts correction quality
Advantages Over Alternatives
| Method | Limitation | TRIAGE Solution |
|---|
| Outcome-only (GRPO) | Punishes exploration in failures | Role typing preserves exploration credit |
| Scalar process rewards | No semantic distinction between action types | Four-way role classification |
| Value function baselines | Shared backbone causes cross-domain coupling | Decoupled role-conditioned correction |
When to Use
Apply TRIAGE when:
- Training agentic RL systems with multi-step action sequences
- Standard GRPO shows poor sample efficiency
- You need to distinguish exploration from regression in trajectories
- Actions have varying semantic importance (not just final outcome)
Skip TRIAGE when:
- Actions are homogeneous in importance
- Simple outcome reward is sufficient
- Computational overhead of judge is prohibitive
Activation Keywords
agentic RL, credit assignment, process reward, GRPO, role typing, trajectory analysis, exploration credit, regression detection, multi-step agents, LLM agents
Related Patterns
- [[score-broadcast-decorrelation-credit-assignment]] - Alternative credit assignment via broadcast
- [[ucob-skill-memory-self-distillation]] - Skill memory with bidirectional self-distillation
- [[opcd-on-policy-distillation]] - On-policy distillation for capability integration
References
- Paper: TRIAGE: Role-Typed Credit Assignment for Agentic Reinforcement Learning
- arXiv: 2606.32017
- Date: 2026-06-30
- Categories: cs.LG, cs.AI