About Me

I am a Ph.D. student (Sep. 2022 – present) at the National University of Defense Technology (NUDT), jointly trained with the Intelligent Game and Decision Lab (IGDL), where I am fortunate to work with Prof. Qingyong Hu. I received my Bachelor's degree from NUDT in 2022. My work is on large language models and their post-training.

My research is about making large models more efficient and more reliable: which data is worth fine-tuning on, why reinforcement learning for reasoning collapses and how to fix it, how to tell when a model is hallucinating, and how well multimodal models actually reason rather than pattern-match. Recently I have been especially interested in distillation and multi-teacher supervision for multi-domain LLMs.

So far I have first-authored papers at NeurIPS, ICML and ICME, and co-authored work at CVPR (Highlight), AAAI (Oral) and IJCAI.

I am always glad to talk about research or collaboration — feel free to reach out by email.

Research Interests

📊 Data-Centric Post-TrainingTACOS

🎯 RL for LLM ReasoningAVSPO

🧪 Knowledge DistillationMT-SDPO

🔍 Trustworthy LLMsEfficient Hallucination Detection, CALRD

👁️ Multimodal ReasoningStemBind, ENC-Bench

🔥 News

2026.09: 🎉🎉 StemBind is accepted by NeurIPS 2026.
2026.09: 🔥🔥 MT-SDPO is released on arXiv, together with the SciKnowEval-L3 data, training code and evaluation protocol.
2026.05: 🔥🔥 StemBind is released on arXiv, with a project page and public leaderboard.
2026.05: 🎉🎉 AVSPO is accepted by ICML 2026.
2026.04: 🎉🎉 CALRD is accepted by IJCAI-ECAI 2026.
2026.02: 🎉🎉 ENC-Bench is accepted by CVPR 2026 as a Highlight.
2025.11: 🎉🎉 Efficient Hallucination Detection is accepted by AAAI 2026 as an Oral.
2025.03: 🎉🎉 TACOS is accepted by ICME 2025.

📝 Publications

Preprints

Preprint

MT-SDPO: answer-verified multi-teacher self-distillation pipeline

Learn from Whoever Is Right: Answer-Verified Multi-Teacher Distillation for Multi-Domain LLMs

🧑‍💻 Xixiang He, Xingming Li, Baiqi Wu, Qiyao Sun, Xuanyu Ji, Ao Cheng, Qingyong Hu# First Author

arXiv preprint, 2026.

[Paper] | [Github]

Accepted Papers

NeurIPS 2026

StemBind overview: nine operations, shared-stem P/R/F probes, and S1-S4 process annotation

StemBind: When MLLMs Get Lost Between Rules and Instances in Abstract Visual Reasoning

🧑‍💻 Xixiang He, Baiqi Wu, Xingming Li, Ao Cheng, Qiyao Sun, Xuanyu Ji, Qingyong Hu# First Author

🏛️ Conference on Neural Information Processing Systems (NeurIPS), Evaluations & Datasets Track, 2026. CCF A.

[Paper] | [Project] | [Github] | [Dataset]

ICML 2026

AVSPO framework: ACR monitor triggers virtual reward injection to restore advantage variance

Advantage Collapse in Group Relative Policy Optimization: Diagnosis and Mitigation

🧑‍💻 Xixiang He, Qiyao Sun, Ao Cheng, Xingming Li, Xuanyu Ji, Hailun Lu, Runke Huang, Qingyong Hu# First Author

🏛️ International Conference on Machine Learning (ICML), 2026. CCF A.

[Paper] | [Project]

IJCAI 2026

CALRD teaser: MLLMs reach the right answer in middle layers and lose it to textual bias in late layers

MLLMs Get It Right, Then Get It Wrong: Tracing and Correcting Late-Layer Textual Bias

🧑‍💻 Xingming Li, Ao Cheng, Qiyao Sun, Xixiang He, Xuanyu Ji, Runke Huang, Qingyong Hu#

🏛️ International Joint Conference on Artificial Intelligence (IJCAI-ECAI), 2026. CCF A.

[Paper] | [Project]

CVPR 2026 Highlight

ENC-Bench overview of electronic navigational chart understanding tasks

ENC-Bench: A Benchmark for Evaluating Multimodal Large Language Models in Electronic Navigational Chart Understanding

🧑‍💻 Ao Cheng, Xingming Li, Xuanyu Ji, Xixiang He, Qiyao Sun, Chunping Qiu, Runke Huang, Qingyong Hu#

🏛️ IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026. CCF A, Highlight.

[Paper] | [Project] | [Github]

AAAI 2026 Oral

Adaptive Bayesian estimation of semantic entropy for hallucination detection

Efficient Hallucination Detection: Adaptive Bayesian Estimation of Semantic Entropy with Guided Semantic Exploration

🧑‍💻 Qiyao Sun, Xingming Li, Xixiang He, Ao Cheng, Xuanyu Ji, Hailun Lu, Runke Huang, Qingyong Hu#

🏛️ AAAI Conference on Artificial Intelligence (AAAI), 2026. CCF A, Oral.

[Paper] | [Project]

ICME 2025

TACOS: open tagging and comparative scoring for instruction data selection

TACOS: Open Tagging and Comparative Scoring for Instruction Fine-Tuning Data Selection

🧑‍💻 Xixiang He, Hao Yu, Qiyao Sun, Ao Cheng, Tailai Zhang, Cong Liu, Shuxuan Guo# First Author

🏛️ IEEE International Conference on Multimedia and Expo (ICME), 2025. CCF B.

[Paper] | [Github]

# Corresponding author. Full list on Google Scholar.

🧰 Open-Source Resources

StemBind Benchmark — a shared-stem diagnostic benchmark for abstract visual reasoning in MLLMs: 2,298 stems and 19,533 perception / rule / full-item tasks, with S1–S4 process annotations and a public leaderboard.
[Dataset] | [Github] | [Leaderboard]
MT-SDPO — training code for answer-verified multi-teacher distillation, with the SciKnowEval-L3 splits, teacher-generated SFT data and the frozen evaluation protocol.
[Github]
TACOS — instruction-tuning data selection code, the curated fine-tuning dataset, and the fine-tuning, inference and evaluation scripts.
[Github]

📖 Educations

2022.09 — Present
Ph.D. Student · Large Language Models & Post-training

National University of Defense Technology

Jointly trained with the Intelligent Game and Decision Lab (IGDL), working with Prof. Qingyong Hu
2018.09 — 2022.06
National University of Defense Technology