# TRIAGE: Role-Typed Credit Assignment for Agentic Reinforcement Learning
> Ficha editorial pública de Research IA. Estado: Lectura primaria completa. La interpretación editorial no sustituye la fuente primaria.

- Página canónica: https://luiseduardodemiguel.com/research-ia/papers/triage-role-typed-credit-assignment-for-agentic-reinforcement-learning
- Fuente primaria: https://arxiv.org/abs/2606.32017
- Versión leída: v3
- Fuente comprobada: 2026-08-19 · lectura primaria completa; extracción editorial automatizada, revisión humana pendiente
- Autores: Yuanda Xu, Zhengze Zhou, Hejian Sang, Xiaomin Li, Jiaxin Zhang, Xinchen Du, Sen Na, Zhipeng Wang, Alborz Geramifard
- Fecha del corte: 30 JUNIO 2026.
- Área: AGENTES · RAG

## Tesis y contexto

Mejora RL de agentes asignando crédito por rol semántico de cada segmento: progreso decisivo, exploración útil, infraestructura sin progreso o regresión. Supera GRPO en ALFWorld, Search-QA y WebShop y reduce turnos de entorno en rollouts completados.

- Problema: Usar solo éxito/fallo final castiga exploración útil y premia pasos redundantes.
- Por qué importa: Apunta a agentes más baratos, más estables y con aprendizaje de trayectorias más fino.

## Evidencia reportada

- **reported-result**: With the default Qwen3-8B-thinking judge, TRIAGE improves over GRPO on all three benchmarks for both policies, with the largest gains on ALFWorld and WebShop—the two audited environments with the highest regression mass (48% and 43%; Appendix G ). [localizador](https://arxiv.org/html/2606.32017#S5)
- **reported-result**: The Search-QA gain is smaller but consistent, matching its more exploration-dominated, lower-regression profile. [localizador](https://arxiv.org/html/2606.32017#S5)
- **reported-result**: Substituting the Qwen3-8B no-think judge—which collapses on the R -in-success cell (Table 4 )—drives TRIAGE below the GRPO baseline on ALFWorld and WebShop for both policies, confirming that the gains stem from accurate role typing and not from the extra reward term alone. [localizador](https://arxiv.org/html/2606.32017#S5)
- **reported-result**: Table 5 situates TRIAGE against stronger credit-assignment baselines reproduced under an identical protocol: PPO with a learned critic, GiGPO, which assigns step-level credit by grouping actions from recurring states ( 5 ) , and a shared-backbone value baseline that learns a dense per-segment signal from the same verifier rewards. [localizador](https://arxiv.org/html/2606.32017#S5)

## Lectura y límite

- Método: La lectura de 2 Problem Setup: Segment Credit in Agentic RL describe la intervención y su construcción: Given a task prompt x , GRPO samples G trajectories, scores each with a verifier r_{i}=V(\tau_{i})\in\{0,1\} , and assigns the group-normalized advantage A_{i}^{\mathrm{GRPO}}=(r_{i}-\bar{r})/(\sigma_{r}+\epsilon) uniformly to every token in the trajectory. Some environment logs report raw success rewards on a different scale, such as 10 for success and 0 for failure; throughout training and in all equations, we binarize these raw rewards to r_{i}\in\{0,1\} . An agentic trajectory \tau_{i}=(a_{i,1},o_{i,1},\ldots,a_{i,K_{i}},o_{i,K_{i}}) consists of environment-facing action segments a_{i,k} and their resulting…
- Límite: La lectura primaria permite comprobar método y resultados en el HTML, pero no convierte sus conclusiones en validación independiente. La ficha no demuestra transferencia fuera de los datasets, modelos, herramientas y condiciones descritos en 5 Experiments.
- Confianza editorial: Media
- Limitación: El cierre de la fuente señala: Finally, the discrete four-role label is only the first layer of role-aware judging. On harder tasks or stronger base agents, obvious loops, wrong purchases, and repeated inspections become rare, and the credit problem shifts from detecting coarse failures to estimating how much each segment advances the task or belief state. In that regime the same framework can use a stronger judge to assign finer-grained process rewards within each role rather than a…
- Limitación: La ficha no demuestra transferencia fuera de los datasets, modelos, herramientas y condiciones descritos en 5 Experiments.

## Localizadores de evidencia
- [Fuente primaria · canonical](https://arxiv.org/abs/2606.32017): tipo abstract
- [HTML · lectura completa](https://arxiv.org/html/2606.32017): tipo abstract
- [Método · 2 Problem Setup: Segment Credit in Agentic RL](https://arxiv.org/html/2606.32017#S2): tipo section
- [Evaluación · 5 Experiments](https://arxiv.org/html/2606.32017#S5): tipo section
- [Cierre · 6 Discussion and Limitations](https://arxiv.org/html/2606.32017#S6): tipo section

## Próxima prueba

- ¿La propuesta mejora agentes web frente a la línea base actual?
- Métrica: Comparar la métrica principal de la fuente junto con calidad, coste, latencia y tasa de errores.
- Regla de parada: Parar si no aparece una mejora reproducible o si aumenta el riesgo, la complejidad o el coste sin compensación.

## Recursos reproducibles
- [the following issues](https://github.com/arXiv/html_feedback/issues)
- [list of packages that need conversion](https://github.com/brucemiller/LaTeXML/wiki/Porting-LaTeX-packages-for-LaTeXML)
- [developer contributions](https://github.com/brucemiller/LaTeXML/issues)

## Enlaces relacionados

- [VAKRA](https://luiseduardodemiguel.com/research-ia/markdown/papers/vakra)
- [The Devil Is in the Interface](https://luiseduardodemiguel.com/research-ia/markdown/papers/devil-interface)
- [SkillSentry](https://luiseduardodemiguel.com/research-ia/markdown/papers/skillsentry)