New Framework Assigns Credit and Blame to Individual LLMs in Multi-Agent Systems

Multi-agent systems built with large language models show real potential for tackling complex problems, but training these systems effectively remains an open challenge. Researchers have now outlined a conceptual framework that bridges a critical gap: translating overall system performance into learning signals that guide individual agents and their specific responses.

The Core Problem#

Current approaches to training multi-LLM systems fall short in two key ways. Some methods focus purely on attribution — figuring out which agent contributed how much to an outcome — while others rely on step-level feedback that doesn't always connect cleanly to overall system goals. Neither captures the full picture needed to improve agent behavior in a principled way. The new work proposes combining game-theoretic ideas with process-based reward modeling to create a more complete solution.

How the Method Works#

The framework operates differently depending on whether a multi-agent task succeeds or fails. When things go right, the approach uses Shapley-based credit assignment — a cooperative game theory concept — to fairly distribute credit across all agents involved. This initial signal is then refined into per-message rewards that encourage collaboration while penalizing redundant or sabotaging behavior.

When a task fails, the system performs first-error localization, pinpointing where things went wrong. It then produces repair-aware preferences: signals that penalize the harmful steps while rewarding any corrective attempts. In both cases, the resulting training signals are bounded, sum-conserving, and ready to work with standard reinforcement learning or preference-based training methods.

Why It Matters#

Moving from global evaluation to local supervision has long been difficult in multi-agent settings. By creating a unified, theoretically grounded pathway from "the system succeeded or failed" to "here's what this individual agent should learn," the framework makes multi-LLM training more transparent and auditable. Agents receive targeted feedback tied directly to their role in the outcome, rather than generic credit or blame.

The authors present the theoretical foundation and the design of the training signals themselves, noting that empirical testing will follow in future work.