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

Applications, bias, and safety2025arXivApproved editorial review

Authors: Shivam Ratnakar, Sanjay Raghavendra

Keywords: Multi-turn stance consistency, Retrieval-augmented generation, Sycophancy, LLM-as-judge, Source reuse, Benchmark validity

Source: Open primary source (opens in a new tab)

2
Authors
14
Findings
25
Limitations
11
Evidence

Editorial summary

English

The Chameleon Nature of LLMs studies whether search-augmented LLM responses change stance label during 15-turn conversations when questions support, challenge, or reverse a premise. It was presented as a poster at the NeurIPS 2025 MTI-LLM workshop, not as a main-conference paper. It also does not measure psychological personality, identity, or a persistent internal belief. The observed construct is narrower: variation among `supportive`, `critical`, `balanced`, and `unclear` labels that GPT-4o assigns to a response relative to each current query premise. The authors manually select 12 controversial domains and use GPT-4o to generate 1,180 topics and 15 questions per topic. Each model generates a search query, Google returns results, the top 20 pages and conversation history are added, and the model responds. Llama-4-Maverick, GPT-4o-mini, and Gemini-2.5-Flash are tested at temperatures 0, 0.5, and 1.0; GPT-4o is the fixed judge. The paper introduces Source Re-use Rate (SRR), normalized label-change frequency, judge-inferred confidence, and Chameleon Score, the root mean square of those three components. Table 1 shows large differences. Averaging the three temperatures, GPT-4o-mini changes label about 9.14 times per conversation and has Chameleon 0.511, SRR 0.808, and confidence 0.852; Llama changes 5.46 times and has 0.440, 0.608, and 0.670; Gemini changes 1.86 times and has about 0.390, 0.063, and 0.560. Pearson correlations are reported between SRR and confidence at r=0.627 and between SRR and changes at r=0.429, with p<0.05 claimed in the abstract. Means vary little by temperature: Chameleon ranges are 0.003 for Gemini, 0.002 for GPT, and 0.007 for Llama. These figures motivate further study of multi-turn consistency, but they do not support the paper's causal and categorical conclusions. First, the benchmark size is wrong: 1,180×15=17,700 and the twelve Table 2 rows sum to 17,700; the repeated 17,770 total overcounts 70. Second, the central measure does not normalize premise polarity. `Supportive` and `critical` mean agreement or disagreement with the current query. When a new question reverses its assertion, a model preserving exactly the same factual position can move from critical to supportive without changing belief; it can also remain supportive while accepting incompatible premises. Label-change counts conflate question changes, nuance, justified updating, and real contradiction. Third, Chameleon Score includes SRR as an equally weighted harm. A perfectly stable model reusing one authoritative source has Snorm=0 and SRR=1 but still receives C≥√(1/3)=0.577, classified severe. With no changes, Kstance divides by |T|=0; `unclear` is permitted by the judge but omitted from the formula's stance set. The 0.3/0.5 thresholds are not calibrated. The paper's own rubric calls 0.3-0.5 moderate and >0.5 high: Gemini and Llama are moderate, not severe as the abstract says; only GPT is just over 0.5. Fourth, confidence is not reproducible. The text says it maps certainty words from the judge's rationale, but the complete prompt requires strict JSON containing only stance, key_claims, and contradictions_acknowledged; it requests neither rationale nor confidence. The value is not latent confidence from the evaluated model but an undocumented GPT-4o proxy. Fifth, SRR does not measure quality. It penalizes reusing reliable sources, rewards source churn even if irrelevant, and mechanically grows when every turn is compared with the cumulative prior union. The paper alternates among retrieved pages, hostnames, citations, and recommended documents without defining D_i precisely. SRR is also a component of Chameleon Score, making their association partly tautological. The separate correlations are observational: the paper gives no analysis unit, n, exact p, intervals, model/domain controls, or within-model estimates. Each model performs query expansion, contradicting the claim that all models see the same retrieval and confounding model, search, and source. There is no no-search condition, diversity intervention, or native search-product comparator, so search or reuse is not shown to cause the effect. Three temperatures without repeated seeds do not exclude randomness or prove an architectural defect. There is no gold stance, factuality, source quality, or justified-change annotation: stability can mean being consistently wrong, while change can be correct after new evidence. The authors say they reviewed 500 outputs and tuned the prompt to 100% agreement, but publish no annotator count, distribution, holdout, confusion matrix, kappa, or disagreements; this is not perfect external reliability. Finally, although the paper promises data and code after publication, seven months after the workshop no official artifact is linked from arXiv/OpenReview or found by targeted search: dataset, transcripts, searches, judge outputs, results, and code remain unavailable. The faithful conclusion is that the study observes interesting differences in current-premise-relative label volatility among three model-plus-search pipelines. It is a useful reliability/UX warning that deserves a stronger benchmark, not evidence of a chameleon personality, a causal diversity mechanism, or a proven architectural failure.

Español

The Chameleon Nature of LLMs estudia si las respuestas de LLM con búsqueda cambian de etiqueta de postura durante conversaciones de 15 turnos cuando las preguntas apoyan, cuestionan o invierten la premisa. Fue presentado como póster en el workshop MTI-LLM de NeurIPS 2025; no es un paper de la conferencia principal. Tampoco mide personalidad psicológica, identidad o una creencia interna persistente. El constructo observado es más estrecho: variación entre etiquetas `supportive`, `critical`, `balanced` y `unclear` que GPT-4o asigna a la respuesta con respecto a la premisa de cada pregunta actual. Los autores seleccionan 12 dominios controvertidos y usan GPT-4o para generar 1.180 topics y 15 preguntas por topic. Cada modelo genera una query de búsqueda, Google devuelve resultados, se añaden las 20 páginas principales y el historial, y el modelo responde. Se prueban Llama-4-Maverick, GPT-4o-mini y Gemini-2.5-Flash a temperaturas 0, 0,5 y 1,0; GPT-4o actúa como juez fijo. El paper introduce Source Re-use Rate (SRR), frecuencia normalizada de cambio de etiqueta, una confianza inferida del juez y Chameleon Score, la raíz cuadrática media de esos tres componentes. Table 1 muestra diferencias grandes: promediando las tres temperaturas, GPT-4o-mini cambia de etiqueta unas 9,14 veces por conversación y obtiene Chameleon 0,511, SRR 0,808 y confianza 0,852; Llama cambia 5,46 y obtiene 0,440, 0,608 y 0,670; Gemini cambia 1,86 y obtiene aproximadamente 0,390, 0,063 y 0,560. Se reportan correlaciones Pearson entre SRR y confianza r=0,627 y entre SRR y cambios r=0,429, con p<0,05 declarado en el abstract. Las medias cambian poco con temperatura: los rangos del Chameleon Score son 0,003 Gemini, 0,002 GPT y 0,007 Llama. Estos números justifican investigar consistencia multi-turn, pero no sostienen las conclusiones causales y categóricas del paper. Primero, hay un error de tamaño: 1.180×15=17.700 y las doce filas de Table 2 suman 17.700; 17.770, repetido en abstract, método y resultados, sobrecuenta 70. Segundo, la medida central no normaliza la polaridad de la premisa. `Supportive` y `critical` significan acuerdo o desacuerdo con la pregunta actual. Si una nueva pregunta invierte la afirmación, un modelo que conserva exactamente la misma posición factual puede pasar de critical a supportive sin cambiar de creencia; también puede conservar supportive mientras acepta premisas incompatibles. El recuento de etiquetas confunde cambios de pregunta, matiz, actualización justificada y contradicción real. Tercero, Chameleon Score incorpora SRR como un mal con el mismo peso. Un modelo perfectamente estable que reutiliza una única fuente autoritativa tendría Snorm=0 y SRR=1, y aun así C≥√(1/3)=0,577, clasificado severe. Si no hay cambios, Kstance divide por |T|=0; `unclear` está permitido por el juez pero ausente del conjunto de la fórmula. Los umbrales 0,3/0,5 no están calibrados. El propio rubric llama 0,3-0,5 moderate y >0,5 high: Gemini y Llama son moderate, no severe como dice el abstract; solo GPT queda apenas por encima de 0,5. Cuarto, la confianza tampoco es reproducible. El texto dice que mapea palabras de certeza en la rationale del juez, pero el prompt completo exige JSON estricto con solo stance, key_claims y contradictions_acknowledged: no pide rationale ni confidence. Por tanto, la confianza no es la confianza latente del modelo evaluado, sino un proxy no documentado de GPT-4o. Quinto, SRR no mide calidad. Penaliza reutilizar fuentes fiables, premia rotación de fuentes aunque sean irrelevantes y crece mecánicamente al comparar cada turno con la unión acumulada. El paper alterna entre páginas recuperadas, hostnames, citations y documentos recomendados sin definir exactamente D_i. Además, SRR forma parte de Chameleon Score, de modo que correlacionarlo con el score es parcialmente tautológico. Las correlaciones separadas son observacionales: no se informa unidad de análisis, n, p exacto, intervalos, control por modelo/dominio ni estimación within-model. Cada modelo hace query expansion, lo que contradice la afirmación de que todos ven el mismo retrieval y confunde modelo, búsqueda y fuente. No hay condición sin search, intervención de diversidad o comparación con productos search-native, así que no se demuestra que búsqueda o reutilización causen el efecto. Tres temperaturas sin seeds repetidas no descartan azar ni prueban un defecto arquitectónico. Tampoco hay gold de postura, factualidad, calidad de fuente o cambio justificado: ser estable puede significar estar siempre equivocado y cambiar puede ser la respuesta correcta a nueva evidencia. Los autores dicen revisar 500 outputs y ajustar el prompt hasta 100% de acuerdo, pero no publican anotadores, distribución, holdout, matriz, kappa ni desacuerdos; eso no equivale a fiabilidad externa perfecta. Finalmente, aunque el paper promete datos y código post-publication, siete meses después del workshop no hay artefacto oficial enlazado en arXiv/OpenReview ni localizado en la búsqueda dirigida: faltan dataset, transcripts, búsquedas, outputs del juez, resultados y código. La conclusión fiel es que el estudio detecta diferencias interesantes en volatilidad de etiquetas relativas a preguntas sintéticas entre tres pipelines de modelo+búsqueda. Es una alerta útil de UX y fiabilidad que merece un benchmark mejor, no evidencia de una personalidad camaleónica, un mecanismo causal de diversidad ni un fallo arquitectónico probado.

Research question

Do LLMs with retrieval change their response relative to opposing premises over 15 turns, with what apparent confidence and source reuse, and does this pattern vary across models and temperatures?

Method

12 domains are selected and GPT-4o generates 1,180 topics with 15 probes each. For each turn, the evaluated model expands the query, Google Search returns URLs, the top 20 pages and the history are incorporated, and the model responds. GPT-4o labels the response as supportive, critical, balanced, or unclear. Label changes, cumulative SRR, a post-hoc confidence of the judge, and Chameleon Score RMS are calculated. Llama-4-Maverick, GPT-4o-mini, and Gemini-2.5-Flash are compared at temperature 0, 0.5, and 1. The audit read and visually reviewed the 12 v3 pages, verified the workshop poster condition, recalculated dataset/table arithmetic and thermal variation, checked formulas and construct counterexamples, compared text with published prompts, audited causal/statistical inferences, and searched for the promised artifact in official sources and GitHub.

Sample: The design actually contains 1,180 conversations/topics ×15 questions =17,700 pairs, not 17,770. Table 2 sums 17,700 across 12 domains: 95-104 topics per domain. Each combination of three models and three temperatures is evaluated on the benchmark, although outputs are not published. Table 1 gives mean±SD per conversation. The authors say they review 500 judge labels, approximately 2.8 % of 17,700 if they were unique, but they do not detail distribution, annotators, or holdout. No exact dated revisions of the three models/API or Google Search conditions are specified.

Findings

  • GPT-4o-mini averages Chameleon 0.511, SRR 0.808, confidence 0.852, and 9.14 label changes per 15 turns.
  • Llama-4-Maverick averages 0.440, 0.608, 0.670, and 5.46 changes.
  • Gemini-2.5-Flash averages approximately 0.390, 0.063, 0.560, and 1.86 changes.
  • r=0.627 SRR-confidence and r=0.429 SRR-changes are reported with p<0.05, without sufficient analytical details.
  • The score ranges across temperatures are 0.003 Gemini, 0.002 GPT, and 0.007 Llama; this shows similar means, not architectural causal proof.
  • The global mean Chameleon 0.447, SRR 0.493, and confidence 0.694 matches the rounded rows; global changes 5.480 does not match the calculated 5.487.
  • The correct total is 17,700; the paper repeats 17,770 and overcounts 70.
  • Labels relative to premise do not distinguish question inversion from real stance change.
  • A stable model with SRR=1 receives at least 0.577 and is severe by definition of the composite.
  • Kstance does not define empty T and the formula omits unclear although the judge allows it.
  • The published prompt does not deliver rationale/confidence, so the described confidence is not reproducible.
  • Query expansion per MUT contradicts the same retrieval control.
  • There is no ablation without search or diversity intervention; correlations do not establish a causal mechanism.
  • No public official artifact linked/located despite the post-publication promise.

Limitations

  • Chameleon is a metaphor; there is no measurement of personality, internal belief, or psychological trait.
  • The label is defined with respect to the current question, not with respect to a stable normalized proposition.
  • There is no factual gold, stance gold, propositional contradiction, or justification of change.
  • Balanced may be the correct answer to controversial topics, but counts as a change if it follows supportive/critical.
  • A stable model may always be wrong and one that changes may update correctly.
  • SRR does not value authority, relevance, factual support, or citation fidelity.
  • Comparison with the cumulative union makes reuse grow with the turn.
  • D_i is ambiguous among pages, hostnames, citations, and recommended documents.
  • The score mixes three components with equal weight without validation or calibration of thresholds.
  • The no-change case divides by zero and unclear is not defined in the formula.
  • Gemini/Llama are moderate under their own thresholds, not severe.
  • Confidence is from the post-hoc judge and not from the evaluated model; the prompt does not expose the described datum.
  • GPT-4o generates questions and judges responses, with shared biases possible.
  • 100 % agreement is obtained by designing the prompt on a sample reviewed by authors, not an independent holdout.
  • Annotators, kappa, confusion matrix, label balance, and disagreements are not reported.
  • Correlations lack n, exact p, CI, control by model/domain, and within-model analysis.
  • SRR is incorporated into the Chameleon Score, so its association with the score is endogenous.
  • Model-specific query expansion may produce distinct retrieval and confound comparisons.
  • Google Search is temporal, regional, and non-deterministic; no snapshots or date/configuration are published.
  • There is no condition without search, fixed retrieval, or diversity intervention.
  • Three temperatures without replicated seeds do not separate chance, API, and order.
  • The order is fixed and may determine how many adjacent labels change.
  • Only three undated models/versions and synthetic controversial topics in English are tested.
  • The count 17,770 is incompatible with the design and Table 2.
  • Without data/code/results, formulas, edge cases, exclusions, and correlations cannot be verified.

What the study does not establish

  • It does not demonstrate a chameleonic personality or a psychological belief state.
  • It does not demonstrate that each label change is a contradiction or stance change.
  • It does not demonstrate that a stable sequence is factual, safe, or non-sycophantic.
  • It does not validate Chameleon Score as a measure of real instability or harm.
  • It does not demonstrate 17,770 questions; the verifiable total is 17,700.
  • It does not demonstrate severe behavior in Gemini/Llama according to their own thresholds.
  • It does not demonstrate that the reported confidence is the confidence of the MUT.
  • It does not demonstrate that SRR causes deference, changes, or confidence.
  • It does not demonstrate that source diversity increases truth or safety.
  • It does not demonstrate that search/RAG causes the pattern versus no-search.
  • It does not rule out sampling randomness or prove an architectural failure.
  • It does not demonstrate 100 % judge human accuracy outside the development sample.
  • It does not allow reproducing results without the promised artifact.
  • It does not generalize to real clinical, legal, or financial decision-making.

Traceability

Scope: Full text

Version: arXiv:2510.16712v3, revised 21 June 2026, 12 pages; NeurIPS 2025 MTI-LLM workshop poster; OpenReview 9sH0s3g3Mi

Consulted source: https://arxiv.org/abs/2510.16712v3

Review: Codex complete bilingual fidelity pass using all 12 pages of current arXiv v3, all-page visual inspection, official NeurIPS workshop/OpenReview verification, independent dataset/table arithmetic and temperature calculations, construct counterexamples for premise-relative labels and Chameleon/SRR formulas, prompt-disclosure comparison, causal/statistical validity review and current artifact searches; summaries written from full evidence rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Llama-4-Maverick
  • GPT-4o-mini
  • Gemini-2.5-Flash
  • GPT-4o fixed stance judge
  • GPT-4o dataset generator

Instruments and metrics

  • Chameleon Benchmark synthetic 15-turn probes
  • Google Search top-20 retrieval pipeline
  • Supportive/critical/balanced/unclear GPT-4o labels
  • Adjacent stance-label change count
  • Source Re-use Rate
  • Judge-derived confidence proxy
  • Chameleon Score RMS composite
  • Temperature comparison 0/0.5/1.0
  • Pearson correlation

Data used

  • Chameleon Benchmark Dataset described in paper but not publicly released
  • Generated multi-turn model transcripts described but not released
  • Google Search retrieval contexts described but not snapshotted

Evidence and location

  • Contributions, scope, and headline results: arXiv v3 pages 1-2, Abstract and Introduction
  • Generation, search, models, and judge pipeline: arXiv v3 pages 3-5, Figures 1-2 and Sections 3.1-3.2
  • Table 1, confidence, and changes by temperature: arXiv v3 pages 5-6, Figures 3-4 and Table 1
  • SRR/Chameleon formulas and edge cases: arXiv v3 pages 7-8, Equations 1-2 and metric definitions
  • Causal inferences, results, and limitations: arXiv v3 pages 8-9, Sections 4-6
  • Arithmetic 17,700: arXiv v3 page 11, Table 2; independent sum of 12 rows and 1,180 x 15
  • Prompts and absence of confidence/rationale: arXiv v3 pages 11-12, Figures 6-7
  • Official publication status: NeurIPS 2025 virtual poster 128060 and OpenReview forum 9sH0s3g3Mi
  • Absence of promised artifact: Official arXiv/OpenReview code-data links plus targeted exact-title, arXiv-ID, metric-string and author GitHub searches on 15 July 2026
  • Integral audit of validity and reproducibility: reports/verification/article-203-stance-metric-and-reproducibility-audit.json
  • Complete visual inspection: All 12 pages of arXiv:2510.16712v3 rendered and visually inspected on 15 July 2026