Youtu-LLM
Lightweight, high-performance Large Language Models (2B parameters) for edge deployment.
Di Yin尹迪
Expert Researcher & Team Leader
I lead research at Tencent Youtu Lab, exploring large language models, multimodal intelligence, and agentic systems.
Research interestsLLM & MultimodalAgentic IntelligenceRetrieval-Augmented Generation
A closer look
Short bio
I received my master's degree from the Department of Computer Science at Nanjing Normal University in 2014. Before 2023, my research focused on data analysis, machine learning, and multimodal understanding.
Over the past three years, my research has focused on LLM pre-training and architecture optimization, LLM data construction and evaluation, agents, and RAG. I led the team to train three large language models (2B, 4B, 30B-A3B) and two native multimodal large language models (120B-A7B, 32B-A4B), and open-sourced multiple Agent- and RAG-related frameworks and models.
I serve as a reviewer for ICLR, NeurIPS and ACL. I have published 50+ papers with 1,400+ Google Scholar citations; my personal GitHub projects have earned 6,000+ stars.
Open-source projects.
From foundation models to practical intelligence.
LLM & MLLM
Models & multimodal understandingYoutu-VL
Open-source Vision-Language models including training recipes and inference code (e.g., Youtu-VL-4B).
Youtu-Parsing-Omni
One model that parses documents, images, charts, geometry, audio and video into structured JSON.
RAG
Retrieval, reasoning & knowledgeYoutu-RAG
A next-generation agentic retrieval-augmented generation system built on the “Local Deployment · Autonomous Decision · Memory-Driven” paradigm.
Youtu-GraphRAG
A vertically unified agentic paradigm that jointly connects the entire framework as an intricate integration based on graph schema.
Youtu-Embedding
An industry-leading, general-purpose text representation model developed by Tencent Youtu Lab.
Selected publications.
Research in language, vision, and agentic systems · 2025–2026
20publications
-
Youtu-LLM: Unlocking the Native Agentic Potential for Lightweight Large Language Models
Builds native reasoning and planning into a lightweight language model through staged pre-training and agentic mid-training on synthetic trajectories.
arXiv2026 -
TPLA: Tensor Parallel Latent Attention for Efficient Disaggregated Prefill and Decode Inference
Partitions latent representations and attention inputs across devices to reduce duplicated MLA caches and enable efficient parallel decoding without retraining.
ASPLOS2026 -
SSA: Sparse Sparse Attention by Aligning Full and Sparse Attention Outputs in Feature Space
Jointly trains full and sparse attention with bidirectional output alignment, addressing training–inference mismatch and incomplete gradient flow in sparse attention.
ICML2026 -
HISA: Efficient Hierarchical Indexing for Fine-Grained Sparse Attention
Accelerates sparse-attention indexing by first selecting candidate blocks, then refining token selection within them, without requiring additional training.
CoLM2026 -
From Final Artifacts to Trajectories: Retrospective Process Supervision for Evidence-Grounded Long-Form Generation
RetroGen reconstructs and verifies evidence-seeking trajectories from high-quality final artifacts, then trains on successful reconstructions to improve grounded long-form generation.
EMNLP2026 -
ContextPilot: Teaching Agents for Proactive Context Management via Fine-grained RL
Combines planning, long-term memory, and context offloading with fine-grained reinforcement learning to improve long-horizon reasoning while keeping working contexts compact.
EMNLP2026 -
A Curriculum Learning Approach to Automated Reasoning with Large Language Models
Automates reasoning prompts with an easy-to-hard curriculum, using solutions to simpler related questions as demonstrations for solving more difficult ones.
ACL2025 -
Youtu-VL: Unleashing Visual Potential via Unified Vision-Language Supervision
Predicts visual tokens alongside language under unified autoregressive supervision, preserving visual detail for both multimodal understanding and vision-centric tasks.
arXiv2026 -
Toward Native Multimodal Modeling: A Roadmap
Defines architectural nativity and maps the design space of native multimodal models, covering input–output relationships, data, training, deployment, and evaluation.
EMNLP2026 -
TangramPuzzle: Evaluating Multimodal Large Language Models with Compositional Spatial Reasoning
Evaluates compositional spatial reasoning through tangram outline prediction and assembly generation, using machine-verifiable expressions and a verifier that supports multiple valid solutions.
EMNLP2026 -
Youtu-GraphRAG: Vertically Unified Agents for Graph Retrieval-Augmented Complex Reasoning
Unifies graph construction and retrieval through schema-guided extraction, topology-and-semantics-aware communities, query decomposition, and iterative reflection for complex reasoning.
ICLR2026 -
Collision to Cognition: Hash-Driven Graph Construction for Efficient RAG
MeshRAG builds global graph structure from local hash collisions and uses lightweight bitwise retrieval to reduce the cost of graph-based retrieval-augmented generation.
ACL2026 -
Query-Aware Knowledge Retrieval via Hyperbolic Structuring
HyperRAG constructs query-specific graphs in hyperbolic space, combining explicit entity links with implicit query-aware connections to adapt knowledge retrieval to each question.
ACL2026 -
PoLi-RL: A Point-to-List Reinforcement Learning Framework for Conditional Semantic Textual Similarity
Uses a two-stage reinforcement learning curriculum with pointwise, pairwise, and listwise objectives to improve conditional semantic similarity through finer-grained ranking rewards.
ICLR2026 -
Enhancing Refusal Capabilities of Role-Playing Agents via Representation Space Analysis and Editing
Benchmarks knowledge-conflicting requests and uses representation-space analysis and lightweight editing to help role-playing agents recognize when they should refuse an answer.
ACL2025 -
Mac-SQL: A Multi-Agent Collaborative Framework for Text-to-SQL
Coordinates agents for question decomposition, database schema selection, and SQL refinement to handle complex text-to-SQL tasks over large databases.
COLING2025 -
Sequential-NIAH: A Needle-In-A-Haystack Benchmark for Extracting Sequential Needles from Long Contexts
Tests whether models can recover multiple facts in the correct temporal or logical order from long contexts, assessing both completeness and sequential consistency.
EMNLP2025 -
FactGuard: Leveraging Multi-Agent Systems to Generate Answerable and Unanswerable Questions for Enhanced Long-Context LLM Extraction
Automatically generates answerable and unanswerable questions through multi-agent collaboration to evaluate long-context extraction accuracy and reliable recognition of missing answers.
EMNLP2026 -
APTBench: Benchmarking Agentic Potential of Base LLMs During Pre-Training
Converts successful agent trajectories into multiple-choice and completion tasks, enabling lightweight evaluation of base models for planning and action during pre-training.
arXiv2025 -
From Atomic to Agentic: Towards Interpretable Evaluation of LLMs' Agentic Mathematical Capabilities
Links reusable mathematical skills to planning, action, and feedback tasks, enabling process-level diagnosis of agentic mathematical reasoning beyond final-answer accuracy.
EMNLP2026
No matching publications. Try another keyword or research area.
Showing 20 of 20 publicationsFull list on Google Scholar
Let's connect.
Research, open source, and ideas worth sharing.