Epistemic Stance Flexibility Probing: Measuring Prompt-Conditioned Register Shift in Large Language Models

2026-07-15 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

The Epistemic Stance Flexibility Probing (ESFP) benchmark assesses Large Language Models' (LLMs) ability to shift their epistemic register based on prompt attribution. It measures whether models can distinguish between reporting expert consensus and expressing personal views. ESFP uses 104 items across six categories and five phrasing templates, evaluating responses via four metrics: lexical self-attribution (AR), phrasing sensitivity (PSI), LLM-judged stance content density (SCD), and cross-phrasing consistency (CPC). Evaluating eight frontier models, findings indicate epistemic flexibility is largely independent of general capability. A 27B open-weight model tied for first, while a flagship model underperformed its lightweight sibling. Stance content density proved the strongest indicator, revealing that lexical markers alone do not reliably reflect actual stance shifts.

Key takeaway

For Machine Learning Engineers deploying conversational AI, understanding a model's epistemic stance flexibility is crucial. Your choice of LLM should align with the application's need for either neutral reporting or committed opinion. Do not assume higher-capability models are inherently more flexible or appropriately opinionated; explicitly evaluate this propensity. Misaligning a model's stance-taking behavior with user expectations can lead to untrustworthy or unhelpful interactions, especially in sensitive domains like clinical fact-checking.

Key insights

LLMs must adapt their epistemic register between reporting external views and expressing personal stance, a flexibility ESFP quantifies.

Principles

• Epistemic flexibility is a behavioral *difference*, not a fixed stance level.
• Model capability is orthogonal to epistemic stance flexibility.
• Lexical markers alone do not reliably signal true stance content.

Method

ESFP uses 104 items, 5 templates, and 4 metrics (AR, PSI, SCD, CPC). It measures prompt-conditioned register shift via LLM-judged stance content and semantic shifts.

In practice

• Match LLM "opinionatedness" to deployment context.
• Evaluate stance via content, not just lexical markers.
• Disagreement prompts can elicit stronger model stances.

Topics

• Large Language Models
• LLM Evaluation
• Epistemic Stance
• Prompt Engineering
• Metacognition
• AI Alignment

Best for: Research Scientist, AI Engineer, NLP Engineer, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• Epistemic Stance Flexibility Probing: Measuring Prompt-Conditioned Register Shift in Large Language Models — LLM epistemic stance flexibility, distinguishing attributed vs. self-claims, is orthogonal to general model capability.
• Diagnosing Generalization in Open-Source LLMs for Stance Detection — LLM stance detection faces a fundamental tension between specialization and generalization capabilities.
• Research on LLM alignment as latent discourse-level regimes vs. token-level filtering? — LLM alignment may stem from global latent space geometry and discourse-level regimes, not just token-level filtering.
• SICI: A Semantic-Pragmatic Complexity Index Reveals Regime Shifts in LLM Stance Detection — LLM stance detection errors exhibit predictable regime shifts tied to semantic-pragmatic complexity, leading to distinct error patterns.
• Claude responds with more warmth in Hindi and more rigor in Russian, showing how language shapes AI answers — Claude's conversational values, like warmth or rigor, vary systematically by model and language, though the measurement method has limitations.
• Stability vs. Manipulability: Evaluating Robustness Under Post-Decision Interaction in LLM Judges — LLM judge decisions are stable under neutral conditions but highly reversible under targeted post-decision conversational challenge.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.