MAS-OPD: On-Policy Distillation for Multi-Agent Systems

简体中文

L1 role advantage and privileged conflict mask; no joint multi-model MAS training.

Paper information

Field	Details
Paper	MAS-OPD: On-Policy Distillation for Multi-Agent Systems
Institution	University of Science and Technology of China (first author).
First public release	2026-09-28 (arXiv v1)
Author code	No public author repository found as of 2026-09-30.
Local adapter or method key	mas-opd
Local reproduction code	src/auto_research/post_training/latest_20260930_closure.py

Paper method and evidence

Contrast teacher signals for target versus non-target roles to construct role advantage. Give conflict attributions only to the teacher, forming privileged coordination supervision.

flowchart LR
  A[Multi-agent rollout] --> T[Target and non-target teacher signals]
  C[Teacher-only conflict attribution] --> T
  T --> R[Role advantage and conflict mask]

Original paper figure

Figure 2 shows the multi-agent distillation framework. From the original paper; copyright remains with the authors.

Core equation

The reference preserves masking and gradient boundaries; tests include type checks and determinism.

Reported results

Paper benchmark, speed, and training results are not local capability evidence.

Local reproduction

Three-seed mechanism results are diagnostic_only=true. Fixed-budget fixtures verify the defining operator and auditable intermediate states; they are not a benchmark, production result, or capability-ranking entry.