arXiv Science⌕ Search

arXiv · 2611.18616

The Judge Is Not Impartial: Self-Preference in Medical LLM Evaluation

Abstract

Large language models (LLMs) are increasingly used to evaluate model-generated medical content, creating a potential circularity: an LLM judge may favor outputs generated by the same model independently of clinical quality. To characterize LLM judge self-preference in medical settings, we measured same-model affinity in single-turn specialist question answering and multi-turn standardized-patient encounters. We evaluated eight models, two from OpenAI, Anthropic, Google, and Qwen, on 620 real point of care questions from clinicians (Real-POCQi). The same eight models also generated and judged 200 standardized patient-clinician scenarios (MedSP1000) up to 8 visible turns. In total, our analysis spans 13,760 LLM judgments. We define three complementary measures of self-preference. Answer Rank Bias, our primary measure, compared the rank assigned to a fixed response by its own-model judge with its average rank from outside judges. Answer Score Bias made the analogous comparison using rubric scores. Own-Answer Win Rate measured how often a judge ranked its own model's response above a competitor. Higher values across all these measures indicate greater degrees of self-preference. In Real-POCQi, pooled Answer Rank Bias was $+1.22$ positions, and all eight judges favored their own answers. Self-preference persisted both with and without a grading rubric. Revealing generator identities modestly weakened self-preference by an average of $-0.094$ positions. In MedSP1000, pooled Answer Rank Bias ranged from $+0.53$ to $+0.62$ positions, with no clear difference by conversation length ($p=0.17$). LLM judges systematically favored responses generated by their own models across multiple formats. Clinical AI deployments should exercise increased vigilance including multiple evaluation measures and judge panels spanning differing model families when deploying LLM-as-a-judge systems.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Zara Ansari, Shlok Natarajan, Vincent Yip, Cally Lin, Avi Udash, Mia Garvey, Aaron Fanous, Roxana Daneshjou. 2026-09-09. The Judge Is Not Impartial: Self-Preference in Medical LLM Evaluation. https://arxiv.science/abs/2611.18616

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Response Attack: Exploiting Contextual Priming to Jailbreak Large Language Models

Contextual priming, where earlier stimuli covertly bias later judgments, offers an unexplored attack surface for large language models (LLMs). We uncover a contextual priming vulnerability in which the previous response in the dialogue can steer its subsequent behavior toward policy-violating content. While existing jailbreak attacks largely rely on single-turn or multi-turn prompt manipulations, or inject static in-context examples, these methods suffer from limited effectiveness, inefficiency, or semantic drift. We introduce Response Attack (RA), a novel framework that strategically leverages intermediate, mildly harmful responses as contextual primers within a dialogue. By reformulating harmful queries and injecting these intermediate responses before issuing a targeted trigger prompt, RA exploits a previously overlooked vulnerability in LLMs. Extensive experiments across eight state-of-the-art LLMs show that RA consistently achieves significantly higher attack success rates than nine leading jailbreak baselines. Our results demonstrate that the success of RA is directly attributable to the strategic use of intermediate responses, which induce models to generate more explicit and relevant harmful content while maintaining stealth, efficiency, and fidelity to the original query. The code and data are available at https://github.com/Dtc7w3PQ/Response-Attack.

cs.CL↗

Stress-Testing Emotional Support Models: Moving from Homogeneous to Diverse Help Seekers

As emotional support chatbots have recently gained significant traction across both research and industry, a common evaluation strategy has emerged: use help-seeker simulators to interact with supporter chatbots. However, current simulators suffer from two critical limitations: (1) they fail to capture the behavioral diversity of real-world seekers, often portraying them as overly cooperative, and (2) they lack the controllability required to simulate specific seeker profiles. To address these challenges, we present a controllable seeker simulator driven by nine psychological and linguistic features that underpin seeker behavior. Using authentic Reddit conversations, we train our model via a Mixture-of-Experts (MoE) architecture, which effectively differentiates diverse seeker behaviors into specialized parameter subspaces, thereby enhancing fine-grained controllability. Our simulator achieves superior profile adherence and behavioral diversity compared to existing approaches. Furthermore, evaluating 7 prominent supporter models with our system uncovers previously obscured performance degradations. These findings underscore the utility of our framework in providing a more faithful and stress-tested evaluation for emotional support chatbots.

cs.CL↗

Thinking Is Not Telling: Information Disclosure in User-Service LLM Agents

User-engaged LLM agents increasingly operate in service scenarios where task success depends on coordination between the agent, the user, and a stateful environment. In such interactions, the agent often has access to rich information that the user does not observe, such as task policies, tool results, and environment states. This makes agent-user communication a central component of task completion. In this work, we study a communication-level failure mode of user-service agents: internal reasoning is counterproductive for the interactive information disclosure. We introduce information disclosure as a measurable communication mechanism, covering state corrections, available options, constraints, consequences, and proactively retrieved environment information. Motivated by recent advances in LLM reasoning, we evaluate and analyse information disclosure under a controlled setting of agents equipped with a "think" tool, which allows agents to reason internally before responding or acting. Across five models and three user-service scenarios, we first show that enforced thinking has unstable utility. We then analyse response scaling under different think efforts and find that more thinking does not reliably translate into richer information disclosure. Using turn-level counterfactual intervention, we show that injecting golden decision-relevant information significantly intervenes agents to correct action trajectories, indicating that missing disclosure can causally affect the agent's immediate downstream behaviour. We further find that failed trajectories most often lack state corrections and constraints, and that a lightweight disclosure-aware prompt effectively mitigates degradation. These results suggest that the key challenge is not simply making agents think more, but ensuring that reasoning and information are externalised as timely information disclosure.

cs.CL↗