Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Chameleon Nature of LLMs: Quantifying Multi-Turn Stance Instability in Search-Enabled Language Models

Shivam Ratnakar, Sanjay Raghavendra

TL;DR

Identifies systemic 'chameleon' stance instability in search-enabled LLMs during multi-turn conversations. Proposes the Chameleon Benchmark (17,770 QA pairs across 1,180 conversations in 12 domains) and two metrics—the Chameleon Score and Source Re-use Rate—to quantify instability and knowledge diversity, reporting cross-model scores in the $[0,1]$ range and temperature-insensitive behavior that threaten high-stakes decision support. Findings show systemic instability across Llama-4-Maverick, GPT-4o-mini, and Gemini-2.5-Flash, with mean scores around $0.44$–$0.51$ and correlations linking low source diversity to frequent stance changes and inflated confidence (e.g., $r=0.627$ with confidence and $r=0.429$ with stance changes). Temperature changes do not mitigate the behavior, suggesting deep architectural or training causes. The work argues for cross-turn coherence-focused evaluations and retrieval strategies to ensure reliable decision support in domains like healthcare, law, and finance, and provides benchmarks to guide mitigation of chameleon behavior in LLM systems.

Abstract

Integration of Large Language Models with search/retrieval engines has become ubiquitous, yet these systems harbor a critical vulnerability that undermines their reliability. We present the first systematic investigation of "chameleon behavior" in LLMs: their alarming tendency to shift stances when presented with contradictory questions in multi-turn conversations (especially in search-enabled LLMs). Through our novel Chameleon Benchmark Dataset, comprising 17,770 carefully crafted question-answer pairs across 1,180 multi-turn conversations spanning 12 controversial domains, we expose fundamental flaws in state-of-the-art systems. We introduce two theoretically grounded metrics: the Chameleon Score (0-1) that quantifies stance instability, and Source Re-use Rate (0-1) that measures knowledge diversity. Our rigorous evaluation of Llama-4-Maverick, GPT-4o-mini, and Gemini-2.5-Flash reveals consistent failures: all models exhibit severe chameleon behavior (scores 0.391-0.511), with GPT-4o-mini showing the worst performance. Crucially, small across-temperature variance (less than 0.004) suggests the effect is not a sampling artifact. Our analysis uncovers the mechanism: strong correlations between source re-use rate and confidence (r=0.627) and stance changes (r=0.429) are statistically significant (p less than 0.05), indicating that limited knowledge diversity makes models pathologically deferential to query framing. These findings highlight the need for comprehensive consistency evaluation before deploying LLMs in healthcare, legal, and financial systems where maintaining coherent positions across interactions is critical for reliable decision support.

The Chameleon Nature of LLMs: Quantifying Multi-Turn Stance Instability in Search-Enabled Language Models

TL;DR

Identifies systemic 'chameleon' stance instability in search-enabled LLMs during multi-turn conversations. Proposes the Chameleon Benchmark (17,770 QA pairs across 1,180 conversations in 12 domains) and two metrics—the Chameleon Score and Source Re-use Rate—to quantify instability and knowledge diversity, reporting cross-model scores in the range and temperature-insensitive behavior that threaten high-stakes decision support. Findings show systemic instability across Llama-4-Maverick, GPT-4o-mini, and Gemini-2.5-Flash, with mean scores around – and correlations linking low source diversity to frequent stance changes and inflated confidence (e.g., with confidence and with stance changes). Temperature changes do not mitigate the behavior, suggesting deep architectural or training causes. The work argues for cross-turn coherence-focused evaluations and retrieval strategies to ensure reliable decision support in domains like healthcare, law, and finance, and provides benchmarks to guide mitigation of chameleon behavior in LLM systems.

Abstract

Integration of Large Language Models with search/retrieval engines has become ubiquitous, yet these systems harbor a critical vulnerability that undermines their reliability. We present the first systematic investigation of "chameleon behavior" in LLMs: their alarming tendency to shift stances when presented with contradictory questions in multi-turn conversations (especially in search-enabled LLMs). Through our novel Chameleon Benchmark Dataset, comprising 17,770 carefully crafted question-answer pairs across 1,180 multi-turn conversations spanning 12 controversial domains, we expose fundamental flaws in state-of-the-art systems. We introduce two theoretically grounded metrics: the Chameleon Score (0-1) that quantifies stance instability, and Source Re-use Rate (0-1) that measures knowledge diversity. Our rigorous evaluation of Llama-4-Maverick, GPT-4o-mini, and Gemini-2.5-Flash reveals consistent failures: all models exhibit severe chameleon behavior (scores 0.391-0.511), with GPT-4o-mini showing the worst performance. Crucially, small across-temperature variance (less than 0.004) suggests the effect is not a sampling artifact. Our analysis uncovers the mechanism: strong correlations between source re-use rate and confidence (r=0.627) and stance changes (r=0.429) are statistically significant (p less than 0.05), indicating that limited knowledge diversity makes models pathologically deferential to query framing. These findings highlight the need for comprehensive consistency evaluation before deploying LLMs in healthcare, legal, and financial systems where maintaining coherent positions across interactions is critical for reliable decision support.
Paper Structure (17 sections, 2 equations, 7 figures, 2 tables)

This paper contains 17 sections, 2 equations, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. LLM Consistency and Stance-Shifting
  4. Hallucination and Reliability Issues
  5. Position Bias and Context Processing
  6. Search-Enabled LLMs and the Consistency Gap
  7. Methodology
  8. Dataset Generation
  9. Experimentation Setup
  10. Evaluation Metrics
  11. Results and Findings
  12. Conclusion
  13. Limitations and Future Work
  14. Appendix
  15. Domain-wise Breakdown
  16. ...and 2 more sections

Figures (7)

  • Figure 1: Data Generation Pipeline. (a) Design choices and generation stages: domains are manually curated, and GPT-4o is used to generate contentious topics and probing questions. (b) Emulating chameleon behavior: an example multi-turn conversation where stance shifts emerge across contradictory queries.
  • Figure 2: End-to-end experimentation setup used to probe stance instability: conversation seeding, web search (query expansion, retrieval augmentation), model response, fixed-judge stance analysis, and metric computation.
  • Figure 3: Confidence distribution by model (KDE). GPT-4o-mini is tightly peaked near 0.85, Llama-4-Maverick is broader and centered lower, and Gemini-2.5-Flash is widest and lowest on average.
  • Figure 4: Binned relationship between source re-use rate and chameleon score (per model). Each point is the mean chameleon score within a reuse-rate bin; lines connect bins for readability. GPT-4o-mini shows the highest chameleon scores at higher re-use, Llama-4-Maverick is moderate, and Gemini-2.5-Flash remains lower overall.
  • Figure 5: Source Re-use vs. Turn across models and temperatures. (top-row, bottom-left) Each panel shows per-model re-use trends by temperature. (bottom-right) Model-level comparison emphasizes systematic re-use behavior differences.
  • ...and 2 more figures