The Language Intelligence and Knowledge Engineering Lab (LIKE Lab) is based at the Institute of Computational Linguistics, School of Computer Science, Peking University, within the State Key Laboratory of Multimedia Information Processing. Led by Professor Zhifang Sui, we pursue frontier research in artificial intelligence, with a focus on natural language processing, large language models, and multimodal intelligence.
Our recent work spans the foundations, training, and applications of large models. We have made a series of advances in natural language processing and multimodal learning, with papers published at leading international conferences including ICLR, ICML, NeurIPS, ACL, CVPR, AAAI and EMNLP. Our alumni have joined leading LLM research and development teams at DeepSeek, Google DeepMind, Xiaomi MiMo, Tencent, Alibaba, ByteDance and so on, contributing to continued progress in AI research and its real-world applications.
We also work closely with industry, collaborating with ByteDance, Xiaomi, Tencent, Huawei and Baidu on LLM technologies and intelligent applications. Several of our students have contributed directly to the training and optimization of the DeepSeek, MiMo and Hunyuan model families, helping advance the next generation of large language models.
From understanding language to building intelligent systems.
We study how large language models learn, reason, and act. Our research spans pre-training, post-training, reinforcement learning, and efficient inference, with the goal of building more capable and reliable language intelligence.
Explore the researchWe connect language with perception to build models that understand and reason across modalities. Our work explores multimodal learning, visual grounding, and the understanding of images and narratives, alongside evaluation in interactive environments.
Explore the researchWe investigate how knowledge is acquired, represented, and used by intelligent systems. Our research spans knowledge graphs, entity linking, and knowledge in language models, with an emphasis on interpretability, factual accuracy, and reliable reasoning.
Explore the researchAn end-to-end evaluation suite for autonomous agents that measures task completion, safety, and robustness using execution traces, audit logs, and environment snapshots.
Read paperRollout Routing Replay reuses inference-time expert routing decisions during training. Aligning the two routers reduces training–inference mismatch and prevents MoE reinforcement learning collapse without compromising training speed.
Read paperTrains agents to navigate spreadsheets, PDFs, documents, and slides in a persistent sandbox. Stateful execution feedback helps models plan across multiple steps, recover from errors, and reason over real-world files.
Read paperUses sparse autoencoder features to trace how ideas evolve across concepts, corpora, and time. Shared, interpretable representations support comparative analysis of conceptual history, including ideas expressed implicitly.
Read paperIdentifies shifts in expert routing as a source of unstable MoE reinforcement learning. Router-Shift Policy Optimization accounts for this drift to improve training stability and performance on mathematics and code.
Read paperLearns to estimate and allocate reasoning tokens according to problem difficulty. Budget-guided reinforcement learning reduces average response length by 61% on math tasks while maintaining accuracy and supporting user-defined budgets.
Read paperMulti-Teacher On-Policy Distillation integrates the capabilities of specialized teachers into a single language model, distilling their guidance on the student’s own rollouts during post-training.
Read paperEvaluates visual grounding across discrimination, spatial relations, limited visibility, and rejection of ungroundable queries. A 1,005-example benchmark reveals substantial gaps in how multimodal models connect language to visual evidence.
Read paperA systematic overview of how language models learn from examples in context, covering training, prompt design, analysis, applications, and open challenges.
Read paperAdvancing expert specialization through fine-grained experts and shared expert isolation, enabling more efficient mixture-of-experts language models.
Read paperUncovering positional bias in LLM-based evaluation and introducing a calibration framework that brings model judgments closer to human assessments.
Read paperLearns to verify and reward each step of mathematical reasoning using automatically generated supervision. The process reward model supports both answer selection and reinforcement learning, without human step-by-step annotations.
Read paperShared meals, adventures, and discoveries — in the lab and beyond.
We welcome conversations about language intelligence, knowledge engineering, and opportunities for research collaboration.
LIKE Lab