This page collects my current papers and preprints. I am especially interested in rigorous evaluation, medical benchmarks, and scientific-intelligence systems for research workflows.
Published in arXiv preprint · ICML 2026 submission (under review), 2026
SciAgentGym benchmarks multi-step scientific tool use for LLM agents with 1,780 tools and long-horizon workflows. It reports systematic failures on extended trajectories and introduces SciForge data synthesis to improve tool-use training. Current manuscript status: ICML 2026 submission (under review).
Published in arXiv preprint, 2026
OpenNovelty builds an evidence-grounded agent pipeline for scholarly novelty assessment. Instead of giving opaque yes/no judgments, it retrieves related literature, compares contribution claims against full texts, and produces verifiable novelty reports with explicit evidence snippets and citations.
Published in ACL 2026 Submission (Under Review), 2025
LLMEval-Fair proposes a dynamic evaluation framework that samples unseen test sets from a large question bank, combines contamination-resistant curation with anti-cheating design, and studies frontier models longitudinally to produce a more reliable picture of progress than static leaderboards. Current manuscript status: ACL 2026 submission (under review).
Published in Findings of EMNLP 2025, 2025
LLMEval-Med is a physician-validated clinical benchmark built from real-world electronic health records and expert-designed scenarios. It targets the weaknesses of existing medical LLM evaluations by moving beyond exam-style questions toward realistic clinical reasoning and checklist-based expert assessment.