Research papers
Research on collaborative AI, agent communication and memory, scientific machine learning, and materials and molecular discovery. My name is highlighted in each author list.
Updated October 31, 2026
Accepted conference papers and journal articles.
2026
Findings of EMNLP 2026.
Published · Findings of EMNLP 2026
Evaluates cost-aware routing across reasoning protocols and the gap between failure prediction and protocol selection.
NeurIPS 2026, main track, poster.
Accepted · main track
Traces corrective feedback through reviewer detection, the next answer response, and verified repair, showing why accurate review alone need not improve final outcomes.
Computational Materials Science 269, 114549, 2026.
Introduces a transformer trained on 198 million organic molecules, with complete reconstruction and controllable latent-space generation for molecular discovery.
2025
Journal of Chemical Information and Modeling 65(9), 4311–4319, 2025.
Evaluates molecular fingerprints and distance metrics against electronic, redox, and optical properties across more than 350 million molecular pairs.
2024
Advances in Neural Information Processing Systems 37, Datasets and Benchmarks Track, Spotlight, 2024.
Published · spotlight
Equal first authors: Chih-Hsuan Yang and Benjamin Feuer.
Presents 161.9 million research-grade images across approximately 366,600 species, together with multimodal models and biodiversity benchmarks.
Computer-Aided Design 172, 103709, 2024.
Develops a physics-informed neural framework for solving PDEs across irregular geometries and boundary conditions.
Frontiers in Veterinary Science 11, 1425394, 2024.
Studies whether online search behavior can provide an early signal of African swine fever outbreaks in Southeast Asia.
Journal of Chemical Education 101(11), 5060–5067, 2024.
Describes an interdisciplinary workshop joining computational chemistry, electrochemistry, and data science.
2022
Computational Materials Science 213, 111599, 2022.
Combines inexpensive low-fidelity and scarce high-fidelity simulation data for efficient organic photovoltaic structure–property prediction.
2021
Nature Computational Science 1, 229–238, 2021.
Uses generative invariance networks and multi-fidelity surrogates for rapid, target-conditioned photovoltaic microstructure design.
Workshop contributions, with their venue and publication status noted below.
2026
AI Measurement Science Workshop at COLM 2026.
Accepted · non-archival workshop
Presents matched four-protocol scientific reasoning traces with aligned outcomes and inference costs to measure when additional collaboration pays.
2025
LAW 2025 Workshop at NeurIPS 2025; revised 2026.
Workshop version · revised preprint
Develops evaluation-aligned agent- and message-level training signals from system-level outcomes.
The LAW 2025 workshop version is titled “Who Gets the Reward & Who Gets the Blame? Evaluation-Aligned Post-Training for Multi-LLM Agents” and has a different author list; the authors shown here follow the revised arXiv version.
2023
AAAI Workshop on AI to Accelerate Science and Engineering, 2023.
Explores reinforcement learning in a continuous molecular latent space for feedback-guided molecule design.
Current manuscripts and public preprints. Submission venues indicate work under review; they do not imply acceptance.
2026
Submitted to AAMAS 2027.
Under review
Equal first authors: Chih-Hsuan Yang and Xinyang Li. Corresponding author: Chih-Hsuan Yang.
Studies how role specialization, evaluator feedback, and revision behavior affect recovery after an initially incorrect answer.
Submitted to AAMAS 2027.
Under review
Equal first authors: Chih-Hsuan Yang and Anjir Ahmed Chowdhury. Corresponding author: Chih-Hsuan Yang.
Measures how the usefulness of a stored memory depends on the future task, its representation, and the reader model.
Submitted to ICLR 2027.
Under review
Corresponding author: Chih-Hsuan Yang.
Studies when a receiving agent can skip another agent's message while measuring answer changes and actual latency savings.
Submitted to ICLR 2027.
Under review
Studies when a self-improving agent should consolidate its retrieved skill library into LoRA weights, accounting for context use and task performance.
Submitted to ICLR 2027.
Under review
Audits message combinations, selector decisions, and outcome stability to explain where learned message sharing gains are lost.
arXiv preprint, 2026.
Preprint
Combines deterministic memory deduplication, cached-prefix preservation, and a performance model to study efficient compaction of agent working memory.
Submitted to AAAI 2027; public preprint.
Under review
Audits confidence-driven agent selection by separately measuring routing discrimination, probability calibration, and public answer commitment.
arXiv preprint, 2026.
Preprint
Evaluates coding-agent memory through stored state, delivered context, management work, and task outcomes.