Selected work

Representative Papers

How to post-train on a surrogate: Envelope sampling mitigates reward hacking

Oct 08, 2026

This study addresses the vulnerability of large language models to reward hacking during post-training with cheap proxy rewards, a problem exacerbated by calibration bias. To mitigate this, the authors propose Envelope Sampling, a theoretically grounded method that recalibrates LLM judges using a small number of ground-truth annotations. This approach integrates rejection sampling with a regret minimization algorithm under an L2-ball constraint to optimize the modified reward function, thereby overcoming the limitations of conventional heuristics and effectively handling misjudgments on rare outputs. Empirical evaluations on clinical note generation and sycophancy tasks demonstrate that the proposed method significantly alleviates reward hacking, achieving superior performance compared to baseline sampling-based recalibration schemes.

0 citationsRead paper

What AI Benchmarks Actually Measure: Adapting Convergent and Discriminant Validity to Interrogate Fifty-Six AI Benchmarks

Sep 08, 2026

该研究通过适应收敛和区分效度的方法，评估了56个AI基准测试是否准确衡量其声称的概念，如推理、拒绝等，并发现许多基准测试在概念一致性上存在问题。

0 citationsRead paper

From Generation to Collaboration: Using LLMs to Edit for Empathy in Healthcare

Jan 22, 2026

This work proposes a collaborative framework in which large language models (LLMs) serve as “empathy editors” to refine clinician-authored messages, enhancing empathetic expression while preserving medical accuracy. Recognizing that clinicians under pressure often struggle to balance factual precision with emotional attunement—leading to perceived empathy deficits in patient communication—the approach emphasizes human–AI collaboration rather than end-to-end generation. To evaluate performance, the study introduces two novel metrics: the Empathy Ranking Score, which assesses perceived empathy, and the MedFactChecking Score, which verifies clinical fidelity. Experimental results demonstrate that LLM-edited texts significantly improve patients’ perception of empathy without compromising medical correctness, outperforming fully LLM-generated responses.

0 citationsRead paper

From Feedback to Checklists: Grounded Evaluation of AI-Generated Clinical Notes

Jul 23, 2025

Existing AI-based clinical note evaluation methods suffer from misalignment between automated metrics and physician preferences, alongside the subjectivity and scalability limitations of expert reviews. Method: We propose Feedback2Checklist—a novel framework that distills large-scale, de-identified real-world clinical feedback into structured, interpretable, and actionable evaluation checklists, and builds an LLM-powered automated evaluator. Contribution/Results: Our approach significantly improves agreement between automated scores and physician preferences (+32.7% Spearman correlation), demonstrating high coverage, diversity, and robustness to quality degradation. Offline experiments show superior performance in identifying low-quality clinical notes compared to baselines, while ensuring strong clinical alignment and practical deployability.

0 citationsRead paper

The Curious Language Model: Strategic Test-Time Information Acquisition

Jun 10, 2025

Large language models (LLMs) struggle to balance information value against acquisition cost in data-scarce scenarios, leading to suboptimal decision-making. Method: This paper proposes a test-time zero-shot active information acquisition framework. Its core is CuriosiTree—a novel, cost-aware greedy tree search strategy that integrates heuristic search, zero-shot information gain estimation, and fusion of heterogeneous multi-source information—requiring no fine-tuning. Crucially, it formalizes information acquisition as a dynamic sequential decision problem, optimizing query actions in real time within a single inference pass. Contribution/Results: Empirical evaluation on clinical diagnosis simulation demonstrates that CuriosiTree achieves significantly higher diagnostic accuracy at lower total acquisition cost, outperforming baselines including random sampling and naive greedy strategies across all metrics.

0 citationsRead paper

Recent publications

Latest Papers

How to post-train on a surrogate: Envelope sampling mitigates reward hacking

Oct 08, 2026

This study addresses the vulnerability of large language models to reward hacking during post-training with cheap proxy rewards, a problem exacerbated by calibration bias. To mitigate this, the authors propose Envelope Sampling, a theoretically grounded method that recalibrates LLM judges using a small number of ground-truth annotations. This approach integrates rejection sampling with a regret minimization algorithm under an L2-ball constraint to optimize the modified reward function, thereby overcoming the limitations of conventional heuristics and effectively handling misjudgments on rare outputs. Empirical evaluations on clinical note generation and sycophancy tasks demonstrate that the proposed method significantly alleviates reward hacking, achieving superior performance compared to baseline sampling-based recalibration schemes.

0 citationsRead paper

What AI Benchmarks Actually Measure: Adapting Convergent and Discriminant Validity to Interrogate Fifty-Six AI Benchmarks

Sep 08, 2026

该研究通过适应收敛和区分效度的方法，评估了56个AI基准测试是否准确衡量其声称的概念，如推理、拒绝等，并发现许多基准测试在概念一致性上存在问题。

0 citationsRead paper

From Generation to Collaboration: Using LLMs to Edit for Empathy in Healthcare

Jan 22, 2026

This work proposes a collaborative framework in which large language models (LLMs) serve as “empathy editors” to refine clinician-authored messages, enhancing empathetic expression while preserving medical accuracy. Recognizing that clinicians under pressure often struggle to balance factual precision with emotional attunement—leading to perceived empathy deficits in patient communication—the approach emphasizes human–AI collaboration rather than end-to-end generation. To evaluate performance, the study introduces two novel metrics: the Empathy Ranking Score, which assesses perceived empathy, and the MedFactChecking Score, which verifies clinical fidelity. Experimental results demonstrate that LLM-edited texts significantly improve patients’ perception of empathy without compromising medical correctness, outperforming fully LLM-generated responses.

0 citationsRead paper

From Feedback to Checklists: Grounded Evaluation of AI-Generated Clinical Notes

Jul 23, 2025

Existing AI-based clinical note evaluation methods suffer from misalignment between automated metrics and physician preferences, alongside the subjectivity and scalability limitations of expert reviews. Method: We propose Feedback2Checklist—a novel framework that distills large-scale, de-identified real-world clinical feedback into structured, interpretable, and actionable evaluation checklists, and builds an LLM-powered automated evaluator. Contribution/Results: Our approach significantly improves agreement between automated scores and physician preferences (+32.7% Spearman correlation), demonstrating high coverage, diversity, and robustness to quality degradation. Offline experiments show superior performance in identifying low-quality clinical notes compared to baselines, while ensuring strong clinical alignment and practical deployability.

0 citationsRead paper

The Curious Language Model: Strategic Test-Time Information Acquisition

Jun 10, 2025

Large language models (LLMs) struggle to balance information value against acquisition cost in data-scarce scenarios, leading to suboptimal decision-making. Method: This paper proposes a test-time zero-shot active information acquisition framework. Its core is CuriosiTree—a novel, cost-aware greedy tree search strategy that integrates heuristic search, zero-shot information gain estimation, and fusion of heterogeneous multi-source information—requiring no fine-tuning. Crucially, it formalizes information acquisition as a dynamic sequential decision problem, optimizing query actions in real time within a single inference pass. Contribution/Results: Empirical evaluation on clinical diagnosis simulation demonstrates that CuriosiTree achieves significantly higher diagnostic accuracy at lower total acquisition cost, outperforming baselines including random sampling and naive greedy strategies across all metrics.

0 citationsRead paper