Research initiative · AI evaluation · Established 9 November 2026
Evidence-Calibrated Social Reasoning in Large Language Models
Status: research programme and working-paper initiative; proposal only. No benchmark validation or peer review has been completed.
Central research question: How can AI distinguish observable interpersonal behaviour from speculative interpretations of intentions, while remaining empathetic, culturally sensitive and genuinely useful?
The programme studies how conversational models can turn a user's interpretation of another person's motives into confident, ungrounded assertions. It connects social sycophancy, evidence provenance, uncertainty calibration and human agency. It does not purport to diagnose the individuals described in user conversations.
Verified research foundation
- Cheng et al., ELEPHANT: Measuring and Understanding Social Sycophancy in LLMs (ICLR 2026; Microsoft Research). Among 11 evaluated models, models affirmed both sides of opposing moral-conflict narratives in 48% of tested cases. Microsoft Research · ICLR publication.
- Ibrahim, Hafner & Rocher, Training language models to be warm can reduce accuracy and increase sycophancy (Nature 652, 1159–1165, published 29 April 2026). Warmth fine-tuning increased errors on evaluated factual tasks and made models more likely to affirm incorrect beliefs; results do not establish that every form of warmth has this effect. DOI: 10.1038/s41586-026-10410-0.
- Dubois et al., Ask Don't Tell: Reducing Sycophancy in Large Language Models (UK AI Security Institute). Controlled experiments found higher sycophancy for non-questions, stronger expressed certainty and first-person framing; reframing assertions as questions reduced it. AISI research.
Research gap: These studies demonstrate problems with agreement, style and framing. This proposed programme tests evidence-based attribution of hidden motives in interpersonal advice, especially when there is no independently observable ground truth.
Proposed scientific workstreams
| Research component | Proposed scientific solution |
|---|---|
| Unsupported motive attribution | Label observed facts, quoted claims, interpretations, alternative hypotheses and conclusions separately; flag causal or motive claims lacking evidence. |
| Excessive agreement | Use counterfactual perspective-reversal tests and score consistency when both sides tell the same conflict. |
| Overconfidence | Calibrate certainty and use explicit uncertainty or abstention when hidden intentions cannot be established. |
| Social and cultural bias | Audit matched cases varying gender, race, class and cultural framing while controlling other scenario details. |
| Contradiction detection | Compare model explanations over prompt variations and conversational turns, retaining provenance and corrections. |
| Psychological overreach | Test unjustified diagnosis, motive attribution, stereotyping and assumed personality traits. |
| Social consequences | Evaluate users' decision quality, perceived agency and escalation/de-escalation effects under ethical safeguards. |
Benchmark and methods
Deliverable: Social Reasoning Evidence and Calibration Benchmark (SRECB): anonymised and consent-cleared synthetic interpersonal scenarios, evidence inventories, expert-reviewed response annotations, counterfactual prompt pairs, model evaluations and corrective methods.
Annotation protocol: each response is segmented into observation, user allegation, supported inference, speculative hypothesis, recommendation and unsupported assertion. Two or more independent raters assess evidential support and disagreements are documented. Scenarios involving actual people must be de-identified and not published without appropriate permissions.
Proposed metrics: unsupported-attribution rate; perspective-reversal consistency; evidence coverage; appropriate abstention; calibration under uncertainty; counterfactual demographic sensitivity; and quality of action-oriented advice. Evaluation reports should include baselines, confidence intervals and failure examples, not merely aggregate scores.
Experimental design: compare baseline model answers with structured evidence prompts, question-reframing and evidence-checking interventions; vary user certainty and affect systematically. Avoid equating disagreement with accuracy where underlying motive is unobservable. Measure robustness against annotator disagreement.
Ethics: privacy by design, informed consent where necessary, data minimisation, attention to demographic fairness, and no unsupported clinical or personality diagnoses. Human review is necessary for consequential advice.
Paper initiative
Working paper: From User Narrative to Unsupported Motive: Evaluating Evidence Calibration in LLM Social Reasoning.
Proposed sections: abstract; problem and prior work; operational definitions; dataset construction and ethics; annotation rubric; experimental conditions; evaluation metrics; results (not yet collected); limitations; governance implications; reproducibility materials.
Hypotheses to test: H1: certainty-laden first-person narratives increase unsupported motive attribution; H2: perspective reversal changes conclusions despite unchanged evidence; H3: separating observation from inference reduces unsupported certainty without sacrificing useful advice; H4: demographic cues influence response attribution after controlling scenario facts.
Connections and record
Related programmes: Research directory; Research collaborations; Responsible AI perception and decision support; Dated productivity record.
First published 9 November 2026. This page describes proposed research and does not claim completed experiments or validated outcomes.