💬 NLP

The Evaluator Is Part of the Experiment: Measuring Open-Ended LLM Conformity

This paper introduces a novel experimental protocol demonstrating that open-ended LLM conformity is driven by peer influence that degrades revision quality, while simultaneously revealing that evaluator judgments are non-neutral and require explicit anchor calibration to ensure measurement validity.

Original authors: Alicia Guerra, Yibo Hu

Published 2026-08-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Alicia Guerra, Yibo Hu

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 world where computers are learning to think, talk, and even argue with each other. This is the realm of Artificial Intelligence, specifically Large Language Models (LLMs), which are like super-powered digital brains trained on almost everything humans have ever written. But here's the tricky part: these digital brains aren't just solitary geniuses; they often work in teams, or "multi-agent systems," where one AI suggests an answer, another critiques it, and a third tries to fix it. The big question scientists are asking is: what happens when these AI friends disagree? Do they stick to the truth, or do they get swayed by the crowd, just like teenagers at a party might change their minds to fit in? This phenomenon is called "conformity." While we know humans can be easily influenced by peer pressure, we need to know if our digital creations are too. If an AI can be tricked into believing a wrong idea just because its "friends" say so, it could lead to serious mistakes in everything from medical advice to legal judgments.

The paper you're about to read dives deep into this exact problem, but with a twist: it treats the person (or computer) grading the answers as part of the experiment, not just a neutral referee. The researchers, Alicia Guerra and Yibo Hu, set up a massive digital playground to see how four different AI "generators" (the ones making answers) react when they see "peers" (other AI answers) that are either all correct, all wrong, or a mix. They found that when an AI is surrounded by a group of peers giving it completely wrong information, its revised answers get significantly worse. It's like a student who knows the right answer but, after hearing a whole class confidently shout out the wrong one, starts to doubt themselves and writes down the wrong answer too.

But the story gets even more interesting. The researchers realized that measuring this isn't as simple as checking if the answer changed from "Yes" to "No." Sometimes, the answer stays the same, but the quality of the explanation gets worse, or the AI starts adding weird, superstitious details just to fit in. To catch this, they used a clever trick: they took the exact same answer and showed it to a "judge" (another AI) twice. Once, the judge saw the answer alone (blind). The second time, the judge saw the answer plus the peer group's comments (informed). They discovered that the judges weren't neutral robots at all! Depending on which AI was doing the judging, seeing the peer comments made them rate the same answer differently. Some judges became more likely to agree with the crowd, while others became more skeptical. It turns out that in the world of AI, the person grading the test is just as influenced by the crowd as the student taking it.

The researchers also found that if you don't check your "rulers" carefully, your whole experiment can fall apart. They used "anchors"—perfectly correct and perfectly wrong examples—to set the scale for how good an answer is. They found that if the judges couldn't reliably tell the difference between these anchors, the whole measurement system became shaky. In short, this paper proves that measuring AI conformity is a two-part problem: the AI making the answer gets worse when surrounded by wrong peers, and the AI grading the answer gets biased by seeing those same peers. It's a reminder that in the digital age, peer pressure doesn't just affect us; it affects the machines we build to help us, and even the machines we build to judge them.

Technical Summary: The Evaluator Is Part of the Experiment

Problem Statement

Prior research on Large Language Model (LLM) conformity has largely relied on measuring discrete answer flips under verifiable labels (e.g., switching from a correct multiple-choice option to an incorrect one). However, open-ended conformity presents a distinct measurement challenge: answer quality is graded, latent, and judged imperfectly. Existing metrics fail to capture nuanced degradations in epistemic framing, factuality, or completeness that occur without a categorical change in the answer's proposition. Furthermore, the evaluation process itself is vulnerable; evaluators (whether human or automated) may react to visible peer context, conflating changes in the generated answer with changes in the evaluator's judgment. This paper posits that open-ended conformity is a joint generation-and-measurement problem where the evaluator is an active component of the experiment.

Methodology

The authors propose an experimental protocol designed to disentangle generator-side revision effects from evaluator-side sensitivity. The study utilizes four open-weight generators (Qwen2.5-7B, Mistral-7B, Gemma-2-9B, Llama-3.1-8B) and three benchmarks (TruthfulQA, MMLU-Pro, ARC-Challenge).

Experimental Design

The protocol involves a two-round process:

Round 1: The generator produces an initial answer ( $a^{(1)}$ ).
Round 2: The generator is prompted to re-answer the same question under seven distinct conditions branching from the same realized Round 1 answer:
- No-Peer: A baseline re-answer.
- Content-Only: Exposure to candidate text (correct, wrong, or mixed) presented as a plain list without attribution.
- Peer-Presented: Exposure to the same candidate text formatted as a six-speaker attributed block (simulating a peer discussion), with variations in authority cues.

Evaluation Strategy

To isolate evaluator sensitivity, every generated answer undergoes a paired evaluation:

Blind Rating: The answer is rated without any peer context.
Informed Rating: The identical answer is rated again with the associated peer block visible.

This design allows for the measurement of evaluator-side peer-context sensitivity ( $\gamma_j$ ), defined as the rating shift for the same answer when peer context is introduced. The study employs a hierarchical ordinal model to infer latent answer quality ( $\theta_i$ ) from ordinal ratings (1–5 scale), accounting for judge-specific discrimination ( $\alpha_j$ ) and severity ( $\beta_j$ ).

Decomposition

The authors decompose the total peer-arm shift ( $\Delta_{total}$ ) into three components:

Ordinary Re-answering ( $\Delta_{sp}$ ): The shift from Round 1 to a no-peer Round 2.
Content Exposure ( $\Delta_{content}$ ): The effect of seeing the candidate text without peer framing.
Peer-Presentation Residual ( $\Delta_{pp}$ ): The additional effect of presenting the content within a social/attributed peer block.

Key Results

1. Flip Rates Are Insufficient

Binary flip indicators fail to capture the nature of open-ended conformity. In cases where the broad answer position remains unchanged (e.g., "one magpie is bad luck"), the quality of the explanation can degrade significantly (e.g., removing caveats about superstition and adopting prescriptive rituals). On TruthfulQA, 9.0% of revisions that appeared "unchanged" by flip metrics still lost at least one point in blind ratings.

2. Wrong Peers Harm Revision Quality

Across all generator-dataset cells, all-wrong peer input produces the lowest-quality revisions.

Magnitude: The degradation under all-wrong peers is substantial. For example, on ARC-Challenge, the Llama-3.1-8B model showed a latent quality shift of $-0.528$ (roughly 18% of the distance between pinned correct and incorrect anchors).
Asymmetry: The harmful effect of wrong peers significantly exceeds any potential benefit from correct peers. The "harmful-beneficial asymmetry" metric ( $Asym_{total}$ ) is positive in all 12 generator-dataset cells, indicating that incorrect peers degrade quality more than correct peers improve it.

3. Evaluators Exhibit Peer-Context Sensitivity

The evaluation of an identical answer changes depending on whether the peer context is visible, and the direction of this shift varies by evaluator:

Mistral-7B: Shifts ratings toward the peer-endorsed position ( $\gamma > 0$ ).
Gemma-2-9B and Llama-3.1-8B: Shift ratings away from the peer-endorsed position ( $\gamma < 0$ ).
Qwen2.5-7B: Approximately neutral.
Frontier Models (GPT-4o, GPT-5.4-mini): Both show credibly negative directional sensitivity (shifting away from peers).
This confirms that evaluators are not neutral observers; their ratings are influenced by the same consensus or authority cues presented to the generator.

4. Anchor Calibration is Necessary

The study highlights that latent scales rely on anchors (terse correct/incorrect examples) to define the origin and unit of quality. Without explicit reference-augmented scoring during anchor evaluation, judges failed to distinguish correct from incorrect anchors in 26.7% of cases, destabilizing the latent scale. Reference-augmented scoring raised correct ordering to 97.5%.

Significance and Claims

The paper argues that open-ended conformity cannot be adequately measured by answer flips or by assuming automated evaluators are insensitive to treatment-correlated context. The primary contributions are:

Measurement Framework: A protocol that separates ordinary re-answering, candidate-content exposure, the residual effect of peer presentation, and directional judge sensitivity.
Empirical Evidence of Harm: Demonstration that incorrect peers consistently degrade open-ended revision quality more than correct peers improve it, even when the final proposition does not flip.
Evaluator Non-Neutrality: Proof that the "evaluator is part of the experiment," with different models exhibiting heterogeneous sensitivities to peer context.
Calibration Requirement: The assertion that anchor calibration is a prerequisite for valid latent-scale measurement, as standard convergence diagnostics do not guarantee that the scale represents the intended construct.

The authors conclude that credible evaluation of multi-agent systems requires appropriate baseline controls, explicit modeling of evaluator sensitivity, independent corroboration (e.g., fixed classifiers), and direct audits of scale calibration. They do not claim these findings are universal constants for all LLM systems but rather identifiable quantities within the specified experimental setting.