Xin Li李鑫

03 / 03

Orchestrate

How should models, tools, and verifiers work together to complete tasks reliably?

The system around a model at inference time: harnesses and tools, verifier choice, and protocols for several agents working together — including when that system breaks answers that were already right.

Work on this thread4 papers

NeurIPS 2026

Measuring Collapse and Correction in Homogeneous-Panel LLM Debate

Xin Li*, Mengbing Liu*, Chau Yuen

Debate can turn a wrong answer right, and it can also talk a correct majority into a wrong one. DebateLedger counts the two apart, and finds that stopping collapse also stops correction: across 6,925 logged MMLU-Pro debates, 253 collapses turned a correct majority wrong, and a probe-gated freeze prevents 29 of them but gives up 108 corrections — net −79 under equal weights. 58.9% of collapses begin in the first debate round.

ProjectPaperCodeData

Preprint 2026

Share the Symbol, Not the Vector: Training-Free Fact Communication Across Language Models

Xin Li, Mengbing Liu, Hao Jiang, Xin Gao, Annan Wang, Xingwei Qu, Yichi Zhang, Chau Yuen

Models with different architectures pass facts without fine-tuning or a learned translator: the sender sends word indices from a shared list, and each receiver writes them with its own word directions. With three senders and nine newer receivers, four-answer success on new scenes reaches 72.0–98.3%, and the two largest receivers match plain text.

ProjectPaperCodeDataModels

EMNLP 2026 Findings

Target-Local Verifier Choice in Best-of-K Reasoning Selection

Xin Li, Hao Jiang, Weisi Lin

On a 34-generator, 7-verifier Best-of-K math panel, one strong process reward model is the best fixed verifier overall — and still not the best verifier for every generator. Label-free candidate statistics predict which verifier class wins.

ProjectCode

NeurIPS 2025 Workshop (AI4NextG)

LACP: LLM Agent Communication Protocol Requires Urgent Standardization

Xin Li, Mengbing Liu, Chau Yuen

A position paper arguing that LLM agents need one telecom-style communication standard before today's ad-hoc protocols entrench, and proposing LACP, a three-layer protocol (semantic, transactional, transport) in which authentication and transactional integrity are part of the core rather than add-ons.

ProjectPaper