Research
My research centers on generative models, agentic AI, and AI for science. Below are my papers and manuscripts in progress; code lives on GitHub.
Conference Papers
Accepted at Conference on Empirical Methods in Natural Language Processing (EMNLP 2026), 2026
Social deduction games are a popular testbed for reasoning and deception in LLM agents, but scoring only win rates cannot tell whether an agent’s words are grounded in what it actually saw and did. QUACK audits agents at three levels — outcomes, trajectories, and utterance-level consistency — and finds even the strongest VLM hallucinates 15.1% of its verifiable spatial claims.
Recommended citation: Ye Yuan, Rui Song, Weien Li, et al. (including Yonghan Yang), and Xue Liu. (2026). "QUACK: Questioning, Understanding, and Auditing Communicated Knowledge in Multimodal Social Deduction Agents." EMNLP 2026. arXiv:2605.27068.
Download Paper
Accepted at Annual Meeting of the Association for Computational Linguistics (ACL 2027), 2026
Herculean evaluates AI agents on practical financial workflows across four domains — Trading, Hedging, Market Insights, and Auditing — each a standardized environment with its own tools and success criteria. Frontier agents do well on Trading and Market Insights but struggle badly with Hedging and Auditing, where long-horizon coordination, state consistency, and structured verification matter.
Recommended citation: Xueqing Peng, Zhuohan Xie, Yupeng Cao, et al. (including Yonghan Yang). (2026). "Herculean: An Agentic Benchmark for Financial Intelligence." ACL 2027. arXiv:2605.14355.
Download Paper
Accepted at Conference on Neural Information Processing Systems (NeurIPS 2026), 2026
A biomedical ML coding benchmark testing whether agents can write task-specific model-building code for heterogeneous, often multi-modal biomedical data. 76 end-to-end tasks across 9 domains, each with hidden labels, held-out graders, and biology-aware metrics on a common 0–1 scale — agents must write runnable code, train models, and submit predictions on private test samples.
Recommended citation: Loka Li, Duzhen Zhang, Xingbo Du, et al. (including Yonghan Yang), Bin Zhang, and Le Song. (2026). "BioXArena: Benchmarking LLM Agents on Multi-Modal Biomedical Machine Learning Tasks." NeurIPS 2026. arXiv:2605.15766.
Download Paper
Published in International Conference on Machine Learning (ICML), 2026
SPADE casts forward surrogate modeling as a calibrated conditional diffusion problem and adds a kNN support-proximity prior, so an offline optimizer stays expressive without exploiting unsupported, out-of-distribution regions. We prove the regularizer is equivalent to Bayesian inference under a valid design prior, and it tops Design-Bench and LLM-optimization tasks.
Recommended citation: Yonghan Yang*, Ye Yuan*, Zipeng Sun, Linfeng Du, Bowei He, Haolun Wu, Can Chen, and Xue Liu. (2026). "Support-Proximity Augmented Diffusion Estimation for Offline Black-Box Optimization." Proceedings of the 43rd International Conference on Machine Learning (ICML 2026), PMLR 306, Seoul, South Korea. (* equal contribution)
Download Paper | Download Slides | Download Bibtex
Preprints & Works in Progress
arXiv preprint · under review at TMLR, 2026
A survey arguing that in discrete diffusion, tokenization is not preprocessing but the primary design axis: it fixes the geometry of corruption, the difficulty of denoising, controllability, and cost. Transition-matrix, masked/absorbing-state, and score/ratio-based models all emerge as instances of one four-component framework.
Recommended citation: Ye Yuan, Weien Li, Rui Song, et al. (including Yonghan Yang), Dawn Song, Philip S. Yu, and Xue Liu. (2026). "Discrete Diffusion Models: A Unified Framework from Tokenization to Generation." arXiv:2607.13431. Under review at TMLR.
Download Paper
Submitted to the NeurIPS 2026 Workshop on Dynamic Alignment in Human–AI Coupled Systems (BiAlign), 2026
Optimizing what an autonomous agent recalls: we train memory retrieval offline with a learned surrogate that scores which past experiences most improve downstream task success, rather than relying on raw semantic similarity.
Recommended citation: Yonghan Yang*, Ye Yuan*, et al. Surrogate-Guided Memory Retrieval for Autonomous Agents. Submitted to the NeurIPS 2026 BiAlign Workshop, 2026. (* equal contribution)
Manuscript in preparation, 2026
Predicting and analyzing higher-order drug-combination-disease relationships from a comprehensive dataset, using semi-supervised learning to exploit the large space of unlabeled combinations.
Recommended citation: Yonghan Yang, et al. Semi-Supervised High-Order Relation Learning for Drug–Combination–Disease Prediction. Manuscript in preparation, 2026.
