SLIC и Semantic Shapley Value: как оценить вклад агентов в LLM-системах, −93% расчётов

In July 2026, a preprint appeared on arXiv describing SLIC — an algorithm that estimates each agent's contribution in a multi-agent LLM system in a single pass and cuts computational cost by 93.3% compared to the baseline Monte Carlo Shapley method.

Why agent contribution is hard to compute

Classical attribution methods require repeatedly rerunning the system — removing individual agents and comparing how the final result changes. In language workflows, each such run means new model calls: an exact Shapley-based contribution calculation requires enumerating all 2^n subsets of agents, which produces high-variance estimates and quickly becomes intractable as the number of agents grows.

The authors point to a second problem: counterfactual methods don't see the intermediate semantic states — how agents generate, preserve, and transform task-relevant information as messages are exchanged. Only the system's input and output are evaluated, not the actual logic of meaning transfer between roles.

How Semantic Shapley Value works

SLIC builds a semantic hypergraph from the realized language flow and computes the Semantic Shapley Value (SSV) without rerunning agent subsets. The algorithm recovers minimal semantic supports, applies Boolean absorption, and distributes contribution according to the logic of semantic support over a single execution trajectory — hence "single-trajectory" in the method's description.

The authors prove that SSV reduces to the classical Shapley value under standard conditions: multilinear representation, full observability, and order-independence. In other words, the new method is not a heuristic but a correct generalization of well-known cooperative game theory to language processes, where step order and intermediate meaning matter.

*Method: Semantic Cooperative Games (SCG) + Semantic Shapley Value (SSV)

*Algorithm: SLIC — single-pass computation, no counterfactual reruns

*Savings: −93.3% computational cost on a medical benchmark

*Consistency: results close to the baseline Monte Carlo Shapley

*Publication: preprint arXiv:2607.18255, July 2026

What the tests showed

On a medical benchmark satisfying the theorem's conditions, SLIC cut computational cost by 93.3% while remaining highly consistent with the baseline Monte Carlo Shapley method. This is the paper's key result: essentially the same contribution estimates, but almost 15 times cheaper to compute.

In more general multi-role workflows, the SSV value matches the profiles of score degradation under perturbations and reveals cases where an agent's semantic contribution and its impact on failure diverge — that is, an agent can carry a lot of meaning without being a point of failure, and vice versa.

"SLIC is a fast, counterfactual-free, and interpretable attribution

method for complex multi-agent LLM systems," the paper's abstract on arXiv states.

What this means

As products are assembled from multiple interacting LLM agents, the question of "who exactly delivered value, and who broke the result" becomes an engineering question rather than a philosophical one. SLIC offers a way to answer it cheaply and without a black box — which matters for debugging, optimization, and fairly evaluating agents in multi-agent systems.

Frequently asked questions

What is contribution attribution in multi-agent LLM systems?

It is the assessment of which of several agents influenced the final result, and by how much, when the answer emerges through a chain of messages and workflow dependencies. SLIC solves this through a semantic hypergraph and Semantic Shapley Value instead of expensive counterfactual reruns.

How much faster is SLIC than previous methods?

On a medical benchmark, SLIC cut computational cost by 93.3% relative to the baseline Monte Carlo Shapley, while keeping results close to it and without rerunning agent subsets.