Accepted · NeurIPS 2026 main track
Precise but Uncoupled
Traces corrective feedback through reviewer detection, the next answer response, and verified repair, showing why accurate review alone need not improve final outcomes.
Researcher · Mentor · CollaboratorPostdoctoral Researcher · Argonne National Laboratory
Chih-Hsuan (Bella) Yang
LLM agents: collaboration, memory, and learning.
I am Chih-Hsuan Yang, also known as Bella Yang, a Postdoctoral Researcher at Argonne National Laboratory. I build research harnesses for LLM agents and multi-agent systems. My agentic evaluation work compares collaboration protocols, traces how solvers use reviewer critiques, and measures memory utility and inference quality-cost trade-offs.
At Corteva, I trained preference models from experimental outcomes and adapted language models with DPO and PPO, incorporating scientist preferences when the model was uncertain. At Google X, I studied information flow, coordination, and bidding among stakeholder agents. These experiences connect my agent research to preference-based post-training and agentic AI for Science, building on earlier work in materials and molecular discovery and BioTrove.
My first-author work includes Precise but Uncoupled (NeurIPS 2026 Main Track) and Cost-Aware Protocol Routing (Findings of EMNLP 2026). I am also a first author of BioTrove (NeurIPS 2024 Datasets and Benchmarks Spotlight).
I mentor 4 students and welcome research collaborations. Team & contact →
Research themes
Questions that connect my work
Selected projects
Recent research, in focus
Published · Findings of EMNLP 2026
Cost-Aware Protocol Routing
Evaluates supervised protocol routing, held-out generalization, and confidence calibration across reasoning tasks. Separates predicting failure from predicting which collaboration protocol is worth its additional inference cost.
Public preprint · AAAI 2027 under review
Wrong but Useful
Uses paired message-ablation replay to measure whether an intermediate message helps or harms downstream team reasoning, distinguishing message correctness from its contribution to the final answer.
Team & collaboration
Research questions worth exploring together.
I currently mentor 4 students. I welcome collaborations on agent evaluation harnesses, multi-agent reasoning, task-conditioned memory, and preference learning for scientific applications.
Research highlights
Publication record →Precise but Uncoupled was accepted to NeurIPS 2026; Cost-Aware Protocol Routing was published in Findings of EMNLP 2026.
The matched scientific-agent trace resource was accepted at AIMS @ COLM 2026, a non-archival workshop.
My first-author manuscripts on role-guided feedback and task-conditioned memory utility are under review at AAMAS 2027; I am the corresponding author of both.