ExRole: Learning Executable Roles for Multi-Agent Language Models
A recent paper on arXiv (2608.11949) presents a novel framework named ExRole, which reconceptualizes roles in multi-agent language models as dynamic control variables instead of fixed prompt labels. This framework learns role prototypes that are aware of future contexts from prefix-local team traces, transforming them into understandable instructions and token-aligned role markers. Additionally, it can route shared LoRA rank slots with turn-aligned credit. In testing on the MuSiQue and 2WikiMultiHopQA benchmarks, ExRole outperforms single-agent searches by 15.0/14.4 and 13.5/16.1 EM/F1 points, respectively. Compared to the best non-ExRole controls, improvements are still notable at 11.5/11.6 and 7.7/9.7 points. The authors contend that roles should encapsulate behaviors that predict future utility, guide future interactions, and pinpoint trainable capacities linked to those behaviors, contributing to the advancement of more interpretable and adaptive multi-agent AI systems.
Key facts
- ExRole is a trajectory-to-role framework for multi-agent language models.
- Roles are treated as executable control variables, not just prompt labels.
- It learns future-aware role prototypes from prefix-local team traces.
- It resolves roles into readable instructions and token-aligned role markers.
- It optionally routes shared LoRA rank slots with turn-aligned credit.
- On MuSiQue, ExRole improves over single-agent search by 15.0 EM and 14.4 F1.
- On 2WikiMultiHopQA, improvements are 13.5 EM and 16.1 F1.
- Against strongest non-ExRole controls, gains are 11.5/11.6 and 7.7/9.7 points.
- The paper is available on arXiv with ID 2608.11949.
Entities
Institutions
- arXiv