Curriculum vitae
CV
Web CV · updated October 31, 2026
I build LLM-agent evaluation harnesses and study protocol coordination, critique uptake, task-conditioned memory, and inference quality-cost trade-offs. My industry experience includes experiment-grounded preference modeling and DPO/PPO post-training. At Argonne, I also participate in the AuroraGPT Post-Training Working Group and mentor 4 students.
Resume PDFs
Open the original PDF for clickable paper, code and contact links.
General information
- Full name
- Chih-Hsuan (Bella) Yang
- Current position
- Postdoctoral Researcher, Argonne National Laboratory
- Research identity
- LLM Agents & Multi-Agent Systems · Agentic Evaluation · Preference Learning · AI for Science
- bellayang@anl.gov
Research skills
- Agentic evaluationMatched protocol comparisons, reviewer-to-solver critique uptake, answer-repair diagnostics, held-out tests, confidence calibration, and inference quality-cost trade-offs.
- LLM-agent systemsResearch harnesses, multi-agent workflows, protocol routing, feedback analysis, and task-conditioned memory evaluation.
- Preference learning & post-trainingPreference and reward modeling grounded in experimental outcomes; Direct Preference Optimization (DPO), Proximal Policy Optimization (PPO), and scientist preferences elicited when the preference model is uncertain.
- ImplementationPython, PyTorch, scikit-learn, NumPy/pandas, Hugging Face, pytest, and Git; reproducible Linux/HPC experiments and structured trace/data workflows.
- Open researchReproducible experiments, reasoning-trace analysis, scientific datasets, and linked releases of papers, code, and evaluation resources.
Research experience
- Jul 2025–present
Postdoctoral Researcher
Argonne National Laboratory · Lemont, Illinois
Builds LLM-agent evaluation harnesses for matched protocol comparisons and structured trace analysis. Research examines reviewer-to-solver critique uptake, answer repair, task-conditioned memory utility, and cost-aware routing with held-out tests and confidence calibration. Participates in the AuroraGPT Post-Training Working Group and mentors four students.
- Sep 2024–Jan 2025
Visiting Scientist
Corteva Agriscience · Des Moines, Iowa
Trained preference models from scratch using real experimental outcomes, with rewards grounded in those experiments. Applied DPO and PPO to train multiple language models for domain-specific customization. Incorporated scientists' preference input when the preference model was uncertain.
- Sep–Dec 2023
AI Resident
X, Google LLC (Google X) · Mountain View, California
Developed multi-agent LLM frameworks representing stakeholders with different interests. Studied information flow, communication, coordination, and bidding mechanisms.
- 2022
Deep Learning Engineer Intern
FLX AI · New York, New York
Worked on deep representation learning and forecasting methods for image and time-series applications.
- Jan 2020–Jun 2025
AI Graduate Researcher
Iowa State University · Ames, Iowa
Led and contributed to research in molecular language modeling, materials design, multi-fidelity learning, neural PDE solvers, and multimodal AI for biodiversity.
Education
- 2025
Ph.D. · Mechanical Engineering; co-major in Computer Engineering; minor in Applied Mathematics
Iowa State University
- 2025
M.S. · Computer Science
Iowa State University
- 2019
B.S. · Bioenvironmental Systems Engineering
National Taiwan University
Research areas
- Multi-agent systems & agentic evaluationBuild research harnesses for matched protocol comparisons; analyze critique uptake, answer repair, supervised routing, and inference quality-cost trade-offs.
- Agent memory & skill learningMeasure memory utility through controlled with/without-memory comparisons; study message-level credit assignment and evaluation-aligned training signals.
- Materials & molecular discoveryConnect molecular language models and multi-fidelity design with experiment-grounded preference modeling and DPO/PPO post-training.
- Scientific machine learningBuild multimodal foundation models, scientific datasets, and physics-informed learning methods for real scientific workflows.
Selected publications
Complete publication list →First-author papers include Precise but Uncoupled (NeurIPS 2026 Main Track) and Cost-Aware Protocol Routing (Findings of EMNLP 2026). I am a first author of BioTrove (NeurIPS 2024 Datasets and Benchmarks Spotlight). Additional accepted journal, conference, and workshop contributions are listed below.
- 2026Measuring When Multi-Agent Collaboration Pays: A Matched Four-Protocol Trace Resource for Scientific ReasoningAI Measurement Science Workshop at COLM 2026Accepted · non-archival workshop
- 2026LLMs Can Predict Failure Risk, But Struggle to Predict Which Collaboration Protocol Pays Off: Cost-Aware Protocol Routing Across Reasoning TasksFindings of EMNLP 2026Published · Findings of EMNLP 2026
- 2026Precise but Uncoupled: Reviewer Precision Does Not Guarantee Critique Uptake in Multi-Agent Math ReasoningNeurIPS 2026, main track, posterAccepted · main track
- 2026MolGen-Transformer: A Molecule Language Model for the Generation and Latent Space Exploration of Organic MoleculesComputational Materials Science 269, 114549, 2026
- 2025Evaluating Molecular Similarity Measures: Do Similarity Measures Reflect Electronic Structure Properties?Journal of Chemical Information and Modeling 65(9), 4311–4319, 2025
- 2024BioTrove: A Large Curated Image Dataset Enabling AI for BiodiversityAdvances in Neural Information Processing Systems 37, Datasets and Benchmarks Track, Spotlight, 2024Published · spotlightEqual first authors: Chih-Hsuan Yang and Benjamin Feuer.
- 2022Multi-Fidelity Machine Learning Models for Structure–Property Mapping of Organic ElectronicsComputational Materials Science 213, 111599, 2022
- 2021Fast Inverse Design of Microstructures via Generative Invariance NetworksNature Computational Science 1, 229–238, 2021
Preprints & manuscripts under review
I am a first author and corresponding author of the AAMAS 2027 submissions Beyond Division of Labor and What Is This Memory Useful For?. Both are under review. Submission venues are listed with the current status of each manuscript.
- 2026Beyond Division of Labor: How Roles Shape Feedback and Recovery in Multi-Agent SystemsSubmitted to AAMAS 2027Under reviewEqual first authors: Chih-Hsuan Yang and Xinyang Li. Corresponding author: Chih-Hsuan Yang.
- 2026What Is This Memory Useful For? Measuring Task-Conditioned Utility in Long-Term Agent MemorySubmitted to AAMAS 2027Under reviewEqual first authors: Chih-Hsuan Yang and Anjir Ahmed Chowdhury. Corresponding author: Chih-Hsuan Yang.
- 2026
Safe to Omit, Worth Omitting? Keeping the Answer and Cutting Latency When an LLM Agent Skips Another's Message
Submitted to ICLR 2027Under reviewCorresponding author: Chih-Hsuan Yang. - 2026
From Practice to Model Weights: Consolidating the Skill Library of a Self-Improving LLM Agent into LoRA
Submitted to ICLR 2027Under review - 2026
Which Messages, at What Budget? A Complete-Subset Audit of Learned Message Sharing
Submitted to ICLR 2027Under review - 2026LAM: Efficient Lossy Agent Memory Framework With A Retrieval-Score Error BoundarXiv preprint, 2026Preprint
- 2026What Confidence Routing Is Actually Doing: Auditing Routing, Calibration, and Commitment in Multi-Agent DeliberationSubmitted to AAAI 2027; public preprintUnder review
- 2026Measure Before You Manage: Evaluating Agent Working Memory in Coding AgentsarXiv preprint, 2026Preprint
- 2026Wrong but Useful: Trajectory Value Beyond Answer Correctness in Multi-Agent MessagesSubmitted to AAAI 2027; public preprintUnder review
Mentoring & collaboration
Alongside hands-on research, I mentor 4 students working on collaborative AI and AI for Science. I am open to research collaborations across multi-agent systems, scientific foundation models, and AI-enabled scientific discovery. See Team & collaboration for mentoring information and how to get in touch.