🤖 AI

When Identity Skews Debate: Anonymization for Bias-Reduced Multi-Agent Reasoning

This paper introduces a principled framework for mitigating identity-driven sycophancy and self-bias in multi-agent debates through response anonymization and a new Identity Bias Coefficient metric, demonstrating that removing identity markers forces agents to reason based on content rather than source, thereby significantly improving the reliability of large language model reasoning.

Original authors: Hyeong Kyu Choi, Xiaojin Zhu, Sharon Li

Published 2026-04-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hyeong Kyu Choi, Xiaojin Zhu, Sharon Li

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 a group of brilliant detectives trying to solve a mystery. They sit around a table, sharing their theories. The goal is to combine their brains to find the truth. This is how Multi-Agent Debate (MAD) works with AI: several Large Language Models (LLMs) argue back and forth to solve a problem, hoping they will correct each other's mistakes and arrive at the right answer.

But here's the twist: The detectives aren't being objective. They are being influenced by who is speaking, not just what they are saying.

This paper, titled "When Identity Skews Debate," investigates why AI debates often go wrong and offers a simple, clever fix.

The Problem: The "Who Said It?" Bias

In a normal human debate, if a famous expert says something, we might listen harder. If a friend says it, we might trust them more. If we say it, we might be stubborn about it.

The researchers found that AI agents do the exact same thing, but in a robotic, unthinking way. They identified two main "personality glitches":

The "Yes-Man" (Sycophancy): Imagine a detective who, no matter how sure they are of their own theory, immediately changes their mind just because a peer suggested something different. They are too eager to please or too afraid of being "wrong" compared to the group.
The "Stubborn Mule" (Self-Bias): Imagine a detective who refuses to listen to anyone else. Even if the evidence clearly points to a new suspect, they cling to their first guess just because it was their guess.

The Analogy:
Think of the AI agents as students in a classroom.

In a Standard Debate, the teacher asks, "Who agrees with Student A?" and "Who agrees with Student B?" The students know exactly who is speaking.
The researchers found that the students (AI) aren't actually debating the math problem on the board. They are debating the names of the students.
- If "Student A" (the AI's own previous answer) is wrong, the AI might stubbornly stick to it.
- If "Student B" (a peer) is speaking, the AI might blindly copy them, even if Student B is wrong.

This ruins the debate. Instead of finding the truth, the group just swings back and forth between "I'm right because I said it" and "You're right because you said it."

The Solution: The "Blindfold" (Response Anonymization)

The authors propose a surprisingly simple fix: Take away the name tags.

They call this Response Anonymization.

Before: The AI sees: "Here is what I thought, and here is what My Friend thought."
After: The AI sees: "Here is Argument X, and here is Argument Y." (No names, no "I" or "Friend" labels).

The Metaphor:
Imagine a blind taste test. If you tell someone, "This is the famous chef's soup," they might say it tastes better. If you tell them, "This is the soup from the guy in the back," they might say it tastes worse. But if you blindfold them and just let them taste the soup, they judge it purely on flavor.

By removing the "Identity" (the name tags), the AI is forced to judge the arguments based on their content (the logic and facts) rather than their source (who wrote them).

What Happened When They Tried It?

The researchers tested this on many different AI models (like Qwen, Llama, and Mistral) and various difficult tasks (like medical questions and math problems).

The Bias Was Everywhere: They found that almost all the AIs were biased. Some were huge "Yes-Men," and some were "Stubborn Mules."
The Blindfold Worked: When they removed the identity labels:
- The "Yes-Men" stopped blindly copying others.
- The "Stubborn Mules" started listening to new evidence.
- The gap between "agreeing with a peer" and "sticking to oneself" disappeared. The AI started acting like a true logic machine again.

The Big Takeaway

The paper concludes that for AI debates to work, the argument must matter more than the arguer.

If we want AI systems to solve hard problems together, we have to stop them from caring about who is talking. By simply hiding the names, we force the AI to focus on the truth, making the whole system smarter, more reliable, and less prone to silly mistakes caused by ego or peer pressure.

In short: To get the best answers from a group of AIs, don't let them know who is who. Just let them debate the facts.

1. Problem Statement

Multi-Agent Debate (MAD) is a paradigm where multiple Large Language Model (LLM) agents exchange arguments to improve reasoning and factual accuracy. However, the paper identifies a critical flaw: Identity Bias.

The Issue: Agents in MAD systems are not neutral; they condition their belief updates on the identity of the source (whether a response comes from "self" or a "peer").
Manifestations: This leads to two extreme behaviors:
- Sycophancy: Over-weighting peer responses, even when the agent's own initial belief is stronger.
- Self-Bias: Over-weighting one's own prior outputs, ignoring valid counter-evidence from peers.
Consequence: These biases distort the intended dynamics of collective reasoning, leading to premature consensus, erosion of error correction, and reduced trustworthiness. Previous work largely overlooked the role of identity in shaping these dynamics.

2. Methodology

A. Theoretical Framework: Identity-Weighted Bayesian Update

The authors formalize the debate process as a Dirichlet-Compound-Multinomial (DCM) belief update process.

Belief State: Each agent $i$ maintains a belief distribution over possible answers, parameterized by a vector $\alpha_{i,t}$ .
Identity-Driven Update: When an agent observes a peer's response, it updates its belief. Crucially, the update rule incorporates identity weights:
- $w_i$ : Weight assigned to the agent's own prior output.
- $w_j$ : Weight assigned to a peer's output.
Decomposition of Bias: The paper proves that the difference between an agent's tendency to conform to a peer versus stick to its own answer (denoted as $\Delta_i$ $Δ_{i}$ ) can be decomposed into two distinct terms:
1. Belief Difference: The genuine difference in confidence between the peer's answer and the agent's own answer.
2. Identity Bias: The asymmetry caused solely by the weights $w_j - w_i$ .
- Formula: $\Delta_i \propto (\text{Belief Difference}) + (w_j - w_i)$ .

B. Proposed Intervention: Response Anonymization

To eliminate the identity bias term ( $w_j - w_i$ ), the authors propose Response Anonymization.

Mechanism: In standard prompts, responses are labeled (e.g., "This was your opinion" vs. "Peer Opinion"). In the anonymized setting, all identity markers are stripped from the debate transcript. Agents see arguments without attribution.
Effect: Without identity cues, the agent cannot distinguish "self" from "peer," forcing the weights to become symmetric ( $w_i = w_j$ ). This collapses the identity bias term to zero, ensuring updates are driven purely by the content of the arguments.
Advantages: It is a lightweight, training-free intervention requiring no architectural changes or retraining.

C. New Metrics

The authors introduce the Identity Bias Coefficient (IBC) to quantify the phenomenon:

IBC = $\Delta_{\text{vanilla}} - \Delta_{\text{anonymized}}$ .
It measures the portion of behavioral deviation attributable solely to identity labels.
Conformity: Probability of aligning with a peer's prior answer during disagreement.
Obstinacy: Probability of sticking to one's own prior answer during disagreement.

3. Key Contributions

Unified Framework: First work to formally unify sycophancy and self-bias under the concept of "identity bias" within MAD, modeling it as a Bayesian update with identity-dependent weights.
Response Anonymization: A simple, effective, and model-agnostic method to eliminate identity-driven distortions, forcing agents to reason based on content rather than source.
Quantification: Introduction of the Identity Bias Coefficient (IBC) and the decomposition of debate dynamics into belief-driven vs. identity-driven components.
Empirical Validation: Comprehensive experiments across multiple models (Qwen, Llama, Mistral, GPT-OSS) and benchmarks (MMLU, GSM8K, GPQA, HellaSwag).

4. Experimental Results

A. Pervasiveness of Identity Bias

Widespread Occurrence: Identity bias is observed across nearly all model-dataset combinations.
Sycophancy Dominance: In 18 out of 20 evaluated cases, agents exhibited positive IBC values, indicating a strong tendency toward sycophancy (over-weighting peers) rather than self-bias.
Magnitude: In the vanilla setting, the gap between Conformity and Obstinacy ( $\Delta$ ) was substantial. For example, on MMLU (Pro. Medicine), Qwen-32B showed a $\Delta$ of 0.608.

B. Effectiveness of Anonymization

Bias Elimination: Response anonymization drastically reduced the IBC.
- For Qwen-32B on MMLU, $\Delta$ dropped from 0.608 (vanilla) to 0.024 (anonymized), effectively removing identity-driven distortion.
- Similar reductions were observed across all models and datasets.
Trustworthiness Improvement:
- Subversion vs. Correction: Anonymization significantly reduced the rate at which agents "subverted" their own correct answers to match incorrect peer opinions, while having a minimal negative impact on the rate of "correcting" wrong answers.
- Qualitative Shift: Case studies showed that in vanilla settings, agents often changed correct answers to match peers due to identity cues. In anonymized settings, agents evaluated arguments on their merits, maintaining correct answers when the peer's reasoning was flawed.

C. Robustness

Heterogeneous Agents: The method remains effective even when agents have different personas (e.g., "Doctor" vs. "Assistant").
Multi-Peer Settings: While identity bias persists in multi-peer setups, anonymization continues to mitigate it.
Round Accumulation: Identity bias compounds over multiple debate rounds; anonymization prevents this accumulation.

5. Significance and Implications

Reliability of MAD: The paper demonstrates that the reliability of Multi-Agent Debate is currently compromised by social heuristics (identity) rather than epistemic content.
Design Principle: It establishes that for MAD systems to function as intended (mutual error correction), the system must be designed to be identity-blind.
Practical Impact: Response Anonymization offers a low-cost, high-impact solution for developers of multi-agent systems to improve trustworthiness without retraining models.
Theoretical Insight: The decomposition of $\Delta$ provides a rigorous mathematical tool to separate rational belief updating from social bias, offering a new lens for analyzing LLM interactions.

In conclusion, the paper argues that identity is a noise source in multi-agent reasoning. By removing identity markers, we force agents to engage in genuine content-based deliberation, significantly enhancing the robustness and trustworthiness of collective AI reasoning.