Hy4
Reasoning, coding & long-horizon agents
Visual understanding and computer use for complex coding, research, and real-world workflows.
My contribution RL post-training and multi-agent capabilities, with additional contributions to STEM reasoning.
Senior Researcher, Tencent · Hunyuan Department
I am a Senior Researcher at Tencent Hunyuan and a core contributor to the Hunyuan model family. My work focuses on RL post-training and multi-agent systems for coding, mathematical and scientific reasoning, and long-horizon tasks.
Hunyuan · Selected models
Core contributions to the Hunyuan family, with a current focus on RL post-training and multi-agent capabilities.
Reasoning, coding & long-horizon agents
Visual understanding and computer use for complex coding, research, and real-world workflows.
My contribution RL post-training and multi-agent capabilities, with additional contributions to STEM reasoning.
Multimodal agents for everyday work
12B active parameters and up to 1M-token context, with RL improving the efficiency of end-to-end tasks.
My contribution RL post-training and multi-agent capabilities.
Open-source MoE with stronger reasoning, coding, and agent capabilities.
Post-training for reasoning and instruction following across Think and Instruct.
Foundational work spanning MoE models, reasoning post-training, and adaptive thinking.
The formal Hy4 release adds visual understanding and computer use, with stronger long-horizon agentic capabilities for coding, research, and real-world workflows.
My role: Core contributor to RL post-training and multi-agent capabilities, with additional contributions to STEM reasoning.
An efficient multimodal model with 12B active parameters and up to 1M-token context, built for everyday agentic workflows. Tencent reports that RL reduces average end-to-end task output tokens by 50%.
My role: Core contributor to RL post-training and multi-agent capabilities.
Open-source 770B-parameter MoE flagship with 49B active parameters and 1M context, delivering frontier-level performance across coding, productivity, scientific research, and complex agentic tasks.
My role: Core contributor to post-training, reinforcement learning, and agentic capability optimization.
Open-source 295B-parameter MoE model with 21B active parameters and 256K context, built for reasoning, coding, long-context, and agentic workflows.
My role: Core contributor to post-training, RL recipes, and reasoning optimization across STEM and code.
Next-generation Hunyuan model family for reasoning and instruction-following scenarios.
My role: Core contributor to post-training and evaluation for reasoning and instruction-following models.
Open-source fine-grained MoE model with 80B total parameters, 13B active parameters, hybrid reasoning, long-context support, and strong agent capabilities.
My role: Core author and contributor to Hunyuan-A13B.
My role: Core contributor to reasoning model optimization and adaptive CoT evaluation.
My role: Core contributor to reasoning post-training, code-reasoning optimization, LiveCodeBench SOTA, and code evaluation.
Open-source MoE LLM with long-context support.
My role: Core contributor to model development and evaluation for the Hunyuan open-source LLM family.
Research
My research connects RL for reasoning and action with learning to coordinate agent teams, using code and task-level evaluations to measure progress and guide training.
Improving code and STEM reasoning through RL post-training, critic feedback, and credit assignment.
Learning how agent teams delegate, communicate, and coordinate, and how credit should be assigned across their decisions.
Evaluating how agents build software, use visual feedback, and critique code, then using these insights to guide training and model improvement.
Selected Work
Independent work on multi-agent RL and credit assignment, followed by selected collaborative papers and further research.
Develops an orchestration-trace framework for studying how RL can train teams of LLM agents to spawn, delegate, communicate, aggregate results, and stop. Connects reward design with credit assignment across these decisions, identifies gaps in message-level credit and learned stopping, and releases a tagged literature corpus and a replayable trace schema.
Unifies credit assignment across reasoning steps, tool calls, memory operations, and interacting agents. Combines a literature taxonomy with six diagnostics for identifying when credit estimates are meaningful, analyzes the limits of text-only histories, and introduces a CA-ID Card linking causal claims to evidence and falsification tests.
Introduces a vision-grounded RL framework that trains web-coding agents to generate, visually diagnose, and refine interfaces using a multimodal critic. Render-validity rewards and an improvement-only revision rule stabilize learning; a critic-free self-edit cycle retains most of the gains at inference without repeated critic calls.
Builds a unified benchmark of 2,000 real pull-request tasks spanning eight task types, eight programming scenarios, and ten languages. Evaluations with SWE-Agent and Claude Code expose distinct weaknesses across tasks and languages, connecting agentic coding assessment to the breadth of real developer workflows.
Introduces automated evaluation of executable visual artifacts using screenshots captured before, during, and after interaction, together with code and task-specific checklists. Across 1,825 tasks in nine domains, the framework measures visual quality and behavior; an expert study reports 90.95% pairwise agreement between the multimodal referee and human judgments.
Evaluates web agents across generation, editing, and repair from text, images, and video. Combines checklist-based patch assessment with an Agent-as-a-Judge that runs websites in a browser and creates targeted tests, revealing gaps in visual quality, interaction, and codebase-level reasoning that static code metrics miss.
Introduces a code-critique benchmark covering both code generation and code question answering across difficulty levels. Basic and advanced protocols, including fine-grained checklists, assess the quality of model critiques and constructive feedback, extending evaluation beyond whether a model can generate a correct solution.
Tests whether coding agents can turn a requirements document and an empty workspace into an installable Python library. Execution-based evaluation exposes premature stopping, inconsistent planning, and fragile cross-file dependencies over long development trajectories.
Combines a 13.1M-sample multilingual multimodal instruction dataset, a trained code model, and a dedicated benchmark to translate visual designs into programs. The resulting 7B model matches much larger models on the reported evaluation.
Evaluates long-horizon GUI agents in specialized professional software. Diagnoses omitted workflow stages, cascading errors, and goal drift, showing why success on short, general-purpose computer tasks does not establish competence on complete professional workflows.
Introduces AutoCodeGen, a pipeline that combines LLM-generated inputs, sandbox execution, reverse problem generation, and filtering to construct challenging coding tasks. Releases a benchmark with balanced coverage of 20 languages, alongside verifiable training data and a multilingual sandbox.
Releases an open code-model training recipe with model weights, reproducible data pipelines, training protocols, and ablations. Identifies the roles of cleaning, deduplication, code-related text, and synthetic data in building competitive code LLMs.
Trains a critic to judge whether Lean 4 formalizations preserve the meaning of natural-language mathematics. Combines critic-guided RL, a semantic-fidelity benchmark, and a corpus of over 509K problems to improve reliability beyond compilation success.
Derives RL rewards directly from general text through a next-segment reasoning objective, reducing dependence on manually annotated post-training tasks. Demonstrates gains in continual pre-training and a stronger starting point for subsequent RL post-training.
Co-trains parallel reasoning proposers and a central synthesizer under a shared RL reward, explicitly optimizing the information they exchange. On mathematical reasoning benchmarks, the method improves accuracy while reducing latency under matched token budgets.
Releases a bilingual 7B model trained on 4.5T tokens together with the cleaned corpus, data pipeline, checkpoints, and training infrastructure. Makes the full model-building process available for reproducible research in language, reasoning, and code.
Introduces domain-aware knowledge distillation that smoothly adapts the training-data mixture to performance gaps between teacher and student models. Directs learning toward domains where the student needs more help, improving transfer over fixed-mixture distillation and continued pre-training baselines.
Models how domain-specific continual pre-training performance changes with data mixture, model size, and training volume. Uses small-scale experiments to select mixtures and transfer scaling predictions across domains, reducing the cost of exhaustive training searches.
My contribution: Core author; wrote most of the underlying codebase for domain-specific continual pre-training and scaling-law experiments.
Extends usable context lengths through a single training run on short sequences. Scale and position augmentations of rotary embeddings improve robustness to longer contexts, avoiding the need to collect long-context training data for each target window.
Introduces 400 questions requiring agents to retrieve and combine evidence from web text, images, and video. Verified evidence checklists support analysis of multimodal dependencies and reasoning paths, exposing limitations hidden by text-only browsing evaluations.
Introduces Automated Weakness Exploration to generate and verify variations near a model’s apparent capability boundary. Evaluates 24 models on 1,142 variants and identifies 18 recurring failure modes, showing how strong aggregate scores can mask brittle reasoning across mathematics, code, science, and knowledge.
Tests joint audio-visual reasoning with 1,000 manually verified questions across 13 task types, supported by step-by-step reasoning annotations. Emphasizes complementary evidence from sound and video to distinguish multimodal understanding from single-modality shortcuts.
Evaluates tool use across five settings, covering single- and multi-turn interaction, one or multiple tools, and out-of-distribution tasks. Uses ground-truth-based metrics without model judges and provides MTU-Instruct data to improve tool-use abilities.
Organizes English and Chinese mathematics problems into a hierarchy of concepts, revealing weaknesses obscured by average benchmark scores. Couples fine-grained diagnosis with targeted fine-tuning to improve underperforming mathematical concepts.
Measures empathy across 97 group identities in 11 dimensions. Distinguishes understanding a group’s perspective from spontaneously responding in alignment with it, revealing uneven behavior across identities and sensitivity to explicit perspective-taking prompts.
Synthesizes long-context research across data, architectures, workflows, training and inference infrastructure, and evaluation. Connects methods for extending context with analyses of whether models reliably use it, alongside a maintained literature repository.
Updates
Selected paper releases, conference updates, and open-source project milestones.
Hy4 and Hy4 flash released. Core contributions to RL and multi-agent capabilities, plus STEM reasoning in Hy4.
Hy4 preview released and open-sourced.
Four papers were accepted to EMNLP 2026: ArtifactsBench, WebCompass, and MM-BrowseComp to the Main Conference, and JaggedBench to Findings.
Hy3 formally released and open-sourced.
ICML 2026: SWE-Compass, NL2Repo-Bench, and From Diagrams to Code.
ACL 2026: RLPT was selected as a SAC Highlight Paper; CriticLean, ReLook, and Rhombus were also accepted.
Hy3 Preview open-sourced and reached #1 on OpenRouter daily global API usage.
ICLR 2026: AutoCodeBench and OmniVideoBench.
Core contributor to the HY2.0 model family.
Hunyuan-A13B released and open-sourced.
Hunyuan-TurboS technical report released.
ACL 2025: OpenCoder.
Core contributor to reasoning-focused Hunyuan-T1.
ICLR 2025: MTU-Bench.
Hunyuan-Large weights and technical report released.
ACL 2024: E2-LLM and ConceptMath.
Background
Research, AI infrastructure, and engineering roles across large-scale language model systems.
Hunyuan Department · Senior Researcher
Algorithm Engineer · AI Infrastructure & Research, with early work focused on AI infra
Algorithm Engineer Intern · AI Infrastructure & research systems
Algorithm Engineer Intern