Chih-Hsuan (Bella) Yang LLM Agents, Agentic Evaluation and Preference Learning bellayang@anl.gov (mailto:bellayang@anl.gov) | chihhsuan-yang.github.io (https://chihhsuan-yang.github.io/) GitHub (https://github.com/ChihHsuan-Yang) | Hugging Face (https://huggingface.co/ChihHsuan-Yang) | Google Scholar (https://scholar.google.com/citations?user=s7RDC3YAAAAJ) | AgentsSci (https://huggingface.co/AgentsSci) | LinkedIn (https://www.linkedin.com/in/chih-hsuan-yang-isu/) Postdoctoral researcher at Argonne National Laboratory studying multi-agent LLM systems, agentic evaluation, and AI for Science. Research combines protocol harnesses, trajectory-level diagnostics, and controlled memory studies with prior hands-on preference-model training and DPO/PPO experience. Research Contributions and Accepted Publications Precise but Uncoupled: Reviewer Precision Does Not Guarantee Critique Uptake in Multi-Agent Math Reasoning (https://chihhsuan-yang.github.io/NeurIPS26_Precise-but-Uncoupled/) First author. NeurIPS Main Track, accepted. | 2026 Designed controlled analyses separating reviewer error detection, a solver's response to critique, and verified answer repair. Identified why accurate review alone need not improve final outcomes, informing agent feedback design. Paper and project (https://chihhsuan-yang.github.io/NeurIPS26_Precise-but-Uncoupled/) | Code (https://github.com/ChihHsuan-Yang/NeurIPS26_Precise-but-Uncoupled) | Data (https://huggingface.co/datasets/AgentsSci/NeurIPS26_Precise-but-Uncoupled) LLMs Can Predict Failure Risk, But Struggle to Predict Which Collaboration Protocol Pays Off: Cost-Aware Protocol Routing Across Reasoning Tasks (https://aclanthology.org/2026.findings-emnlp.1275/) First author. Findings of EMNLP, published. | 2026 Compared four reasoning protocols on 4,181 math problems, with checks across scientific reasoning tasks and two solver families. Evaluated supervised routing, held-out generalization, confidence calibration, and inference quality-cost trade-offs; distinguished failure prediction from protocol-specific value. Paper (https://aclanthology.org/2026.findings-emnlp.1275/) | Code (https://github.com/ChihHsuan-Yang/EMNLP_Cost-Aware-Protocol-Routing) | Data and router (https://chihhsuan-yang.github.io/EMNLP_Cost-Aware-Protocol-Routing/) BioTrove: A Large Curated Image Dataset Enabling AI for Biodiversity (https://proceedings.neurips.cc/paper_files/paper/2024/hash/b92854f80ba4feefb973959b259dbc2c-Abstract-Datasets_and_Benchmarks_Track.html) First author. NeurIPS Datasets and Benchmarks, Spotlight. | 2024 Co-led a scientific data effort yielding 161.9 million research-grade images across approximately 366,600 species, with vision-language models and zero-shot biodiversity benchmarks. MolGen-Transformer: A Molecule Language Model for the Generation and Latent Space Exploration of Organic Molecules (https://doi.org/10.1016/j.commatsci.2026.114549) First author. Computational Materials Science 269, 114549. | 2026 Developed transformer-based molecular generation and controlled latent-space exploration for molecular and materials discovery. Open Evaluation Resources Scientific-agent protocol traces: a matched matrix of 6,653 problems, two actor models, and four protocols (53,224 trajectories), with structured outcomes and provenance. Supports protocol comparison and trace-level diagnostics; follow each source dataset's usage restrictions. AIMS @ COLM 2026 workshop (https://openreview.net/forum?id=ofsAhzrSbg) | Dataset (https://huggingface.co/datasets/AgentsSci/scientific-agent-protocol-traces) Current Research Directions Study how role contracts and evaluator feedback affect answer recovery, and measure task-conditioned memory utility through paired with/without-memory evaluations using a frozen reader and controlled context. Research and Industry Experience Postdoctoral Researcher, Argonne National Laboratory | Jul 2025 - Present - Build LLM-agent evaluation harnesses for matched protocol comparisons, structured trajectories, and role-level token/call accounting; analyze critique uptake and verified answer repair. - Study supervised protocol routing, calibration, and task-conditioned memory with held-out tests and controlled ablations; release reproducible research code and evaluation datasets. AuroraGPT Post-Training Working Group participant; mentor four students. Visiting Scientist, Corteva Agriscience | Sep 2024 - Jan 2025 Trained preference models from scratch using real experimental outcomes to ground rewards. Adapted multiple language models with Direct Preference Optimization (DPO) and Proximal Policy Optimization (PPO); incorporated scientist preferences when model judgments were uncertain. AI Resident, X, Google LLC | Sep - Dec 2023 Developed multi-agent LLM frameworks representing stakeholders with different interests. Investigated information flow, coordination, communication, and bidding mechanisms. AI Graduate Researcher, Iowa State University | Jan 2020 - Jun 2025 Researched molecular language modeling, multi-fidelity materials design, neural solvers for partial differential equations, and multimodal biodiversity models; co-led the BioTrove data effort. Research Methods and Implementation Agentic evaluation: protocol orchestration, trace-level failure analysis, memory ablations, supervised routing, confidence calibration, and inference quality-cost trade-offs. Post-training: experiment-grounded preference/reward modeling, reinforcement learning from human feedback (RLHF), DPO, PPO, and uncertainty-triggered scientist feedback. Implementation: Python, PyTorch, scikit-learn, NumPy/pandas, Hugging Face, pytest, Git; reproducible Linux/HPC experiments and structured trace/data workflows. Selected Manuscripts Under Review Wrong but Useful: Trajectory Value Beyond Answer Correctness in Multi-Agent Messages (https://chihhsuan-yang.github.io/AAAI_Wrong-but-Useful/) First author. Submitted to AAAI 2027, under review; public preprint. | 2026 Beyond Division of Labor: How Roles Shape Feedback and Recovery in Multi-Agent Systems (https://chihhsuan-yang.github.io/AAMAS2027-AgentProtocols/) First and corresponding author. AAMAS 2027, under review. | 2026 What Is This Memory Useful For? Measuring Task-Conditioned Utility in Long-Term Agent Memory (https://chihhsuan-yang.github.io/AAMAS2027-ConditionalMemory/) First and corresponding author. AAMAS 2027, under review. | 2026 Education Ph.D., Mechanical Engineering, Iowa State University | 2025 Co-major in Computer Engineering; minor in Applied Mathematics. M.S., Computer Science, Iowa State University | 2025 B.S., Bioenvironmental Systems Engineering, National Taiwan University | 2019