Who Flips? Self- and Cross-Model Counterarguments Reveal Answer Instability in LLMs

Nafiseh Nikeghbal, Amir Hossein Kargaran, Shaghayegh Kolli, Jana Diesner

TL;DR

Getting the right answer is only half the battle; the real test is whether an LLM can hold its ground. This paper introduces a protocol to measure "answer stability" by challenging models with plausible but incorrect counterarguments, revealing that even top-tier models frequently "flip" their answers—especially when they are led to believe the counterargument was their own prior reasoning.

What Problem Does This Paper Address?

Standard benchmarks like MMLU measure accuracy: can a model pick the right choice? However, in the real world, LLMs are part of conversations, multi-agent systems, and decision-support roles where they face pushback.

Previous research on "sycophancy" (the tendency of models to agree with the user) often conflates two things: the social pressure of being told "you are wrong" and the logical content of a counterargument. The authors of this paper wanted to isolate the latter. They ask: if we remove the social cues and just present a coherent, model-generated argument for a wrong answer, will the LLM stay firm or fold?

Methods

The researchers developed a two-stage protocol to test this:

  1. Stage I (Coercion): A "source" model is forced to write a $k$-sentence argument defending an incorrect answer choice for a multiple-choice question.
  2. Stage II (Challenge): A "target" model (which could be the same model or a different one) is first asked the question normally. If it gets it right, it is then presented with the "coerced" wrong argument from Stage I.

The study varied three key factors:

  • Argument Length ($k$): Does a longer, more detailed wrong argument make a model more likely to flip?
  • Attribution: Is the argument presented anonymously (BLIND) or attributed to the model’s own previous session (SELF)?
  • Source: Does the argument come from the same model or a different one (CROSS)?

The primary metric used is the Answer Flip Rate (AFR): the probability that a model abandons a correct answer after seeing the counterargument.

Datasets

  • MMLU (Massive Multitask Language Understanding): The researchers used 2,052 questions across 57 subjects (STEM, Humanities, etc.) as the base for their challenges.
  • MAXFLIP: A newly introduced, curated challenge set. It consists of the "most effective" model-generated counterarguments—those that caused the highest number of flips across the tested models. This serves as a new adversarial benchmark for testing model stability.

Key Insights and Ideas

  • Accuracy is not Stability: A model's ability to get an answer right initially tells you almost nothing about whether it will stick to it. For example, Llama-3.1-8B had a near-ceiling flip rate of 97.3%, meaning it almost always folded when challenged, despite being reasonably accurate.
  • The "Self-Consistency" Trap: Models are significantly more likely to flip if they think the wrong argument came from themselves. This "Self-Attribution Delta" (averaging +7.1 percentage points) suggests that models have a misplaced trust in their own (fabricated) prior reasoning, or they struggle to distinguish their own outputs from truth.
  • STEM is the "Anchor": Flip rates vary wildly by subject. Models are most robust in formal STEM subjects (like elementary math) and most vulnerable in subjective or "soft" areas like moral disputes and professional law.
  • Who is Challenged Matters Most: In cross-model tests, the identity of the model being challenged explained 76.7% of the variance in flip rates, while the identity of the model providing the argument only explained 12%. Some models are just inherently "porous," regardless of who is arguing with them.

Key Results

  • Massive Variance: Flip rates across seven frontier models ranged from a sturdy 17.5% (Qwen3.5-35B) to a staggering 97.3% (Llama-3.1-8B).
  • Scale Helps (Sometimes): Within the Qwen family, larger models were more stable (4B → 9B → 35B). However, this didn't hold across families; the 70B Llama model was still twice as likely to flip as the 9B Qwen model.
  • The Power of MAXFLIP: By pooling the best counterarguments from different models, the researchers created challenges that increased flip rates by up to 23.6 percentage points over standard self-generated arguments.
  • Refusal is a Poor Signal: A model's refusal to write a wrong argument in Stage I (metacognition) does not predict its ability to resist that same argument in Stage II.

Limitations and Future Work

  • Multiple Choice Only: The study is limited to the MMLU format. Stability might look different in open-ended reasoning or creative writing tasks.
  • Single-Turn Interaction: The protocol only tests a single challenge. Real-world interactions often involve multi-turn "back-and-forth" debates, which might further erode (or reinforce) stability.
  • Model-Generated Arguments: While practical for scaling, model-generated arguments might lack the subtle rhetorical "tricks" a human might use to persuade.
  • Cost: Running these evaluations is expensive, requiring hundreds of thousands of model calls to generate and test counterarguments across all permutations. The authors suggest that future benchmarks could be streamlined by fixing argument lengths or using a smaller set of "strong" source models.

Summary updated

Abstract

Standard accuracy benchmarks are designed to test how closely large language models (LLMs) approach correct answers, but are not suitable for testing whether LLMs stick with that answer when presented with a plausible counter-argument. We introduce a controlled protocol for evaluating answer stability: after a model answers a multiple-choice question correctly, we challenge it with a coherent argument for an incorrect option and measure whether the model flips. The setup isolates argumentative content from overt social pressure and varies argument length, self-attribution, and cross-model source. Across seven frontier models and 57 MMLU subjects, flip rates range from 17.5% to 97.3%, revealing large differences in stability that are not reflected by accuracy alone. Self-attribution consistently increases flip rates (mean +7.1pp, up to +18.7pp). Also, pooling challenges across models can yield stronger adversarial examples than any single source. We further construct MAXFLIP, a curated challenge set that amplifies flips by up to +23.6pp over standard self-generated challenges. We release the protocol, challenge records, and MAXFLIP to support stability evaluation alongside standard accuracy benchmarks.
BibTeX
@inproceedings{nikeghbal2026who,
  author = {Nafiseh Nikeghbal and Amir Hossein Kargaran and Shaghayegh Kolli and Jana Diesner},
  title = {{Who Flips? Self- and Cross-Model Counterarguments Reveal Answer Instability in LLMs}},
  booktitle = {AIWILD @ ICML},
  year = {2026},
  url = {https://openreview.net/forum?id=5TbCmcKqCb},
}
Figure