🤖 AI

PEAR: Permutation-Equivariant Adaptive Routing Multi-Agent Debate

The paper introduces PEAR, a permutation-equivariant adaptive routing protocol for multi-agent debates that dynamically reconfigures agent roles and sparse topologies to eliminate positional biases and significantly improve reasoning accuracy across diverse large language models.

Original authors: Yang Feng, Ziwei Xu, Xia Hu, Fengxiang He

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yang Feng, Ziwei Xu, Xia Hu, Fengxiang He

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a group of five experts trying to solve a very difficult puzzle. In the past, researchers tried to improve their answers by having them "debate" each other. However, they usually forced the experts to talk in a rigid, unchanging pattern—like a circle where everyone only talks to the person on their left, or a star where one person talks to everyone else.

The paper argues that these fixed patterns have a big flaw: they create "positional bias." If the person sitting in the "star" spot happens to be wrong, they can drag the whole group down because everyone listens to them. Conversely, if a quiet expert in the "circle" has the right answer, they might never get a chance to speak up because the rules of the conversation don't let them reach the people who need to hear them.

To fix this, the authors created PEAR (Permutation-Equivariant Adaptive Routing Multi-Agent Debate). Think of PEAR not as a rigid meeting room, but as a smart, dynamic seating chart that changes every time the group discusses a new point.

Here is how PEAR works, using simple analogies:

1. The "Smart Seating Chart" (Adaptive Routing)

In a normal debate, the seating is fixed. In PEAR, before every round of discussion, a "router" (an invisible manager) looks at the current state of the group and rearranges who talks to whom.

The Goal: It wants to make sure the right information gets to the right people at the right time.
The Analogy: Imagine a classroom. If a student is confused (low confidence) and a student is very sure of the answer (high confidence), PEAR instantly moves the confident student next to the confused one so they can help. If a student has been talking too much and dominating the conversation, PEAR moves them to a seat where they talk less, giving others a chance to speak.

2. The Three Rules of the Router

The "manager" decides the new seating arrangement based on three simple rules:

Rule A: "Help the Confused" (Targeted Diversity)
If Agent A is very confident and has a different answer than Agent B (who is unsure), the router connects them. It's like a teacher pairing a top student with a struggling student specifically to correct a mistake. It ensures that "dissenting" (different) but confident opinions reach the people who need them most.
Rule B: "Don't Let One Person Dominate" (Influence Balancing)
If an agent has been very influential in previous rounds (meaning many people agreed with them), the router penalizes them. It's like a moderator saying, "You've spoken a lot already; let's hear from someone else." This prevents a single person from accidentally leading the whole group into a wrong conclusion just because they spoke first or loudest.
Rule C: "Filter the Guessers" (Low-Confidence Filtering)
If an agent is guessing and has low confidence, the router stops their critiques from spreading. It's like muting a microphone for someone who says, "I'm not sure, but I think..." so they don't accidentally confuse the others with unreliable information.

3. The "Fairness" Guarantee (Permutation Equivariance)

The paper makes a cool mathematical claim: It doesn't matter who is who.
Imagine you swap the names of the experts (Agent 1 becomes Agent 5, Agent 2 becomes Agent 1, etc.). Because PEAR only cares about how confident they are and what they answered—not their names or labels—the outcome of the debate stays exactly the same. This ensures the system is fair and doesn't accidentally favor a specific person just because of their position in the lineup.

4. The Results: A Better Group Decision

The authors tested this system on four difficult types of problems:

General Knowledge (MMLU-Pro)
Factual Truth (TruthfulQA)
Math Word Problems (GSM8K)
Hard Competition Math (MATH-500)

They tested it on six different AI models (ranging from smaller open-source models to powerful commercial ones).

The Findings:

PEAR consistently won. In almost every test, the group using the dynamic, smart seating chart got the right answer more often than groups using fixed patterns (like circles or stars) or just random pairings.
It fixed errors faster. The system was particularly good at taking a group that started with the wrong answer and guiding them to the right one, whereas fixed groups often got stuck in their initial mistakes.
It was efficient. It didn't need to be a "full mesh" (where everyone talks to everyone) to work well. By being smart about who talks to whom, it saved computing power while getting better results.

Summary

In short, PEAR is a method for making AI groups smarter by stopping them from getting stuck in rigid conversation patterns. Instead of forcing everyone to talk to the same neighbors every time, it dynamically rearranges the conversation so that confident experts can correct confused ones, and no single person can hijack the group's decision. The result is a more accurate, fair, and reliable group answer.

Technical Summary: PEAR (Permutation-Equivariant Adaptive Routing Multi-Agent Debate)

1. Problem Statement

Multi-agent debate (MAD) has emerged as a promising inference-time paradigm for enhancing the reasoning capabilities of Large Language Models (LLMs) by aggregating diverse reasoning paths through iterative peer critiques. However, existing protocols predominantly rely on fixed communication topologies (e.g., chain, star, ring, or clique) that remain static across all debate rounds.

The authors identify two primary failure modes inherent to fixed topologies:

Information Homogenization: Agents repeatedly exchange critiques within predetermined neighborhoods, leading to the reinforcement of errors rather than their challenge. Fixed structures may prevent diverse viewpoints from reaching the agents who would benefit most from them.
Error Cascades: Structural "hubs," early speakers, or privileged agents can exert disproportionate influence. If such an agent provides an incorrect answer, the fixed topology can anchor the entire group to that error, causing persistent downstream misalignment.

Furthermore, fixed topologies introduce sensitivity to arbitrary role assignments, where the outcome depends heavily on which agent occupies a specific structural position (e.g., the center of a star graph) rather than the agent's actual reasoning capability or state.

2. Methodology: PEAR

The paper introduces Permutation-Equivariant Adaptive Routing Multi-Agent Debate (PEAR), an inference-time framework that replaces fixed communication graphs with state-aware adaptive routing. Instead of a static graph, PEAR dynamically reconfigures the sparse communication topology at every debate round based on the current state of all agents.

Core Mechanism

PEAR operates in four phases, with the latter three repeating for $R$ rounds:

Initialization: Agents generate independent initial responses (answer, reasoning, self-reported confidence). A base role graph $G_0$ (a sparse $k$ -regular template) is defined, but the assignment of specific agents to roles is permuted dynamically.
Adaptive Routing: A router selects a specific permutation $\pi$ $π$ of agents to roles for the current round. This selection is governed by a composite scoring function $S(E | s_r)$ $S (E ∣ s_{r})$ that evaluates candidate edge sets based on three state-aware components:
- Targeted Diversity: Rewards edges connecting high-confidence agents with differing answers (sources) to low-confidence agents (targets). This ensures reliable dissenting viewpoints reach agents most in need of correction.
- Influence Balancing: Penalizes routing structures that allocate additional out-degree to agents who have already accumulated disproportionate historical influence (based on how often their past critiques were adopted). This prevents any single agent from dominating the information flow regardless of answer correctness.
- Low-Confidence Filtering: Suppresses critique edges originating from agents with low self-reported confidence, reducing the propagation of unreliable signals.
Critique & Revision: Based on the selected topology, source agents generate targeted critiques. Target agents revise their answers and reasoning, explicitly accepting or rejecting incoming critiques.
Influence Update: The router updates the accumulated influence statistic for each agent based on the adoption rate of their critiques in the current round, closing the loop for the next routing decision.

Theoretical Foundation

The authors characterize PEAR as an agent-permutation equivariant router.

Equivariance: Relabeling agents (permuting their identities) results in a relabeled communication graph but leaves the final answer distribution unchanged. This is achieved because the routing scores depend only on label-invariant features (disagreement, confidence, accumulated influence) and the search is over sparse relabelings of a fixed base graph.
Generalization: The paper proves that equivariant routing preserves expected accuracy under exchangeable agent populations and does not increase the covering complexity of the routing class, providing a uniform generalization guarantee.

3. Key Contributions

Novel Routing Protocol: PEAR is the first framework to realize equivariant communication in multi-agent debate, dynamically adapting sparse topologies based on agent states rather than using fixed structures.
Theoretical Guarantees: The work provides formal proofs demonstrating that PEAR is permutation-equivariant and that such equivariant routing preserves expected accuracy while reducing routing complexity.
Composite Objective: The introduction of a three-component routing objective (Targeted Diversity, Influence Balancing, Low-Confidence Filtering) that actively promotes "wrong-to-right" corrections while attenuating error cascades.
Comprehensive Evaluation: Extensive empirical validation across four diverse benchmarks (MMLU-Pro, TruthfulQA, GSM8K, MATH-500) and six LLM backbones (ranging from 8B to 30B parameters, including open-source and closed-source models).

4. Experimental Results

PEAR was evaluated against single-agent baselines (CoT, CoT-SC) and various fixed-topology debate variants (Clique, Star, Chain, Ring, Random).

Accuracy Gains: PEAR consistently outperformed all baselines across all settings. It achieved an average accuracy of 0.701, compared to 0.620 for the strongest fixed baseline (Fixed Clique).
Specific Improvements:
- Gains of up to 9.0 percentage points over the strongest baseline on specific model-dataset combinations.
- On weaker backbones (e.g., Llama-3.1-8B), PEAR showed significant improvements (8.0–9.5 points), suggesting that adaptive routing is particularly crucial when individual agent capabilities are limited.
- Improvements were observed across all four benchmarks: MMLU-Pro (+6.0), TruthfulQA (+3.9), GSM8K (+4.8), and MATH-500 (+5.6) over the best fixed topology.
Trajectory Diagnostics:
- Correction Quality: PEAR achieved the highest net correction rate (0.243), indicating beneficial updates substantially outweighed harmful flips.
- Diversity: It exhibited a significantly higher cross-answer routing rate (0.676 vs. 0.43–0.50 for baselines), confirming that adaptive routing encourages information exchange across diverse reasoning trajectories.
- Influence Distribution: PEAR achieved the highest influence entropy (0.979), reflecting a more balanced distribution of influence compared to fixed topologies where hubs often dominate.
Efficiency: PEAR operates on the cost-accuracy Pareto frontier. It uses approximately 32.3k tokens per example, which is comparable to a Random baseline with the same edge budget and 12% fewer tokens than the densest Fixed Clique baseline, while delivering significantly higher accuracy.

5. Significance and Claims

The paper claims that PEAR significantly improves the performance of LLM reasoning by addressing the structural limitations of fixed-topology debates. By combining adaptivity with the natural symmetries of multi-agent reasoning (via permutation equivariance), PEAR prevents agents from permanently occupying privileged network positions and distributes influence more evenly.

The authors emphasize that the performance gains are driven by improved correction dynamics and more diverse communication, rather than simply increasing computational cost. The framework is presented as a principled approach to mitigating information homogenization and error cascades, offering a robust inference-time protocol for complex reasoning tasks. The code is made available to ensure reproducibility.