The Interaction Tax: Unchecked Multi-Agent Communication Erases Diversity

Does multi-agent LLM interaction help or hurt? A new ICML 2026 paper from Ann, Liu, and Tan resolves the contradiction with a single crisp finding: when agents read each other's complete outputs, their proposals converge within one round, erasing the diversity that motivated using multiple models. They call this the interaction tax.

---

The Problem: Multi-Agent Is Not a Free Lunch

The pitch for multi-agent systems is intuitive: different models have different strengths, so letting them talk should produce better results. But a growing body of work under matched budgets finds that independent sampling often matches or beats multi-agent interaction. The contradiction suggests something structural is wrong with the default communication pattern.

The Method: 11 Tasks, Matched Budgets, Three Communication Modes

The authors test 11 verifier-scored optimization tasks where ground truth is unambiguous (e.g., code correctness, constraint satisfaction, puzzle solving). Each task has a verifier that scores solutions deterministically, removing the ambiguity of human evaluation. Three communication modes are compared under matched token budgets:

• Independent generation — each agent produces a solution without seeing others' outputs.
• Full-solution interaction — agents read each other's complete solutions before responding.
• Conditional critique — agents are shown only specific parts of another agent's solution, typically parts that violate known constraints.

graph TB
  subgraph "Independent (No Collapse)"
    A1[Agent 1] --> O1[Solution A]
    A2[Agent 2] --> O2[Solution B]
    A3[Agent 3] --> O3[Solution C]
    O1 --> V[Verifier selects best]
    O2 --> V
    O3 --> V
  end
  subgraph "Full-Solution (Collapse)"
    B1[Agent 1] --> S1[Solution A]
    B2[Agent 2] -->|Reads A| S2["Solution A' ≈ A"]
    B3[Agent 3] -->|Reads A'| S3["Solution A'' ≈ A"]
    S2 --> V2[All nearly identical]
    S3 --> V2
  end
  style S2 fill:#3b1f3b,stroke:#d8b4fe
  style S3 fill:#3b1f3b,stroke:#d8b4fe

The Results: Convergence in One Round

The results are stark. Full-solution interaction collapses diversity within a single round. Different model families find structurally different solutions when working independently. But when they see each other's complete outputs, they converge within one exchange to whichever solution appeared first.

Key numbers from the study:

• Independent generation consistently matches or beats full-solution interaction across all 11 tasks under matched token budgets.
• Critique helps only in narrow conditions: when the violated rule is easy for the LLM to both find and fix — what the authors call a "simple constraint violation." For complex or subtle errors, critique is no better than independent resampling.
• Full-solution interaction makes agents stay close to the first solution they see rather than trying different approaches. The convergence is driven by anchoring, not by genuine improvement.

Why This Happens: Proposal-Level Anchoring

The paper identifies a specific mechanism: proposal-level anchoring. When an agent reads a complete solution from another agent, that solution anchors the search space. Even when instructed to improve or critique, the anchor dominates — the new solution is a perturbation of the original, not an independent attempt. The authors show that interaction helps only when agents share specific constraints, partial results, or targeted feedback — not complete solutions.

Limitations

• Verifier-scored tasks only — the study uses tasks with deterministic ground truth. In open-ended domains, dynamics may differ. Collapse might be less harmful where consensus is desirable.
• Single-round interaction — multi-turn debate may produce different dynamics, though one-round collapse suggests diminishing returns.
• Frontier models only — tested on GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro. Smaller models may behave differently.
• No formal bounds — the paper identifies the empirical pattern but doesn't derive theoretical limits on when interaction helps versus hurts.

What It Means for Multi-Agent Builders

This paper should change how you wire agents together:

• Independent sampling is your baseline. Before adding interaction protocols, compare against the same budget of independent samples from the best single model.
• Share constraints, not solutions. Partial information preserves diversity. Comprehensive sharing destroys it.
• Critique needs guardrails. Critique only works when the error is easy to detect and fix. For complex errors, it's expensive noise.
• Measure diversity explicitly. If your system converges without improving quality, you're paying the interaction tax.

Multi-agent performance depends less on the number of agents than on the information they exchange. The right architecture isn't "more agents" — it's "less shared."

What's Next

The paper's core contribution is identifying the problem precisely. The open questions are harder — when does partial information sharing help, what information should be shared at which point in a multi-turn protocol, and can we design interaction graphs that preserve diversity while enabling useful collaboration? These are the right questions to be asking, and this paper is a strong step toward empirical rigor around interaction patterns.

---