Foundation Models & Post-training
Training objectives and inference methods that improve model capability without hiding failure modes behind aggregate scores.
SFTPreference OptimizationDiffusion LMs
Profile
About me
I am a master's student in Software Engineering at Peking University. My research spans reinforcement learning for tool-using agents, LLM evaluation and reward modeling, diffusion language models, and multimodal or embodied intelligence.
My research aims to move language models from generating answers to taking reliable action: enabling autoregressive and diffusion models to learn from verifiable feedback, improve through tools, vision, and real-world interaction, and translate capability gains into measurable value.
At Tencent Youtu Lab (CSIG), I focus on post-training for Code Agents.
Available Jul 2027 LLM Research Scientist · Research Engineer
Mainland China · Hong Kong · Singapore
01
Research directions
From failure modes to training, evaluation, and systems.
Agents & Reinforcement Learning
Credit assignment, tool use, and reward design for agents that must interact with environments rather than only generate text.
Tool UseGRPOCredit Assignment
Evaluation & Reward Modeling
Robust LLM judges, long-form evaluation, uncertainty, and reward models whose scores reflect what users actually care about.
LLM-as-a-JudgePRMLong-form
Multimodal, VLM & Embodied AI
Visual feedback and robot reward models that connect language-model reasoning with perception and action.
VLMVisual FeedbackRobot Rewards
02
Latest news
Research, publication, and career updates.
-
Our paper Length-Adaptive Decoding for Masked Diffusion Machine Translation was accepted to the EMNLP 2026 Main Conference. The paper, code, data, and model weights are now publicly available.
-
I am currently working on Code Agent post-training at Tencent Youtu Lab, CSIG.
-
Our paper RefineSVG was accepted to ACM Multimedia 2026, a CCF-A conference.
-
We released new work on tool-calling reinforcement learning, diffusion-language-model reasoning, and robot reward modeling.
-
I joined Tencent to work on multilingual large-language-model research and agent systems.
-
I started my master’s study at Peking University.
03
Publications & preprints
Filter by topic, then open a paper for details.
Showing all 9 papers.
RefineSVG: Visual Feedback-Driven Reinforcement Learning for Image-to-SVG Generation
A closed-loop visual-feedback framework that lets multimodal language models render, compare, and iteratively correct SVG output, supported by an SVG-oriented vocabulary and agentic reinforcement learning.
ACM Multimedia (ACM MM) 2026
Length-Adaptive Decoding for Masked Diffusion Machine Translation
Entropy-Valley is a training-free method for selecting a masked diffusion model's target canvas before denoising, improving translation adequacy without a learned length predictor.
EMNLP 2026 Main Conference
SLCA-GRPO: Resolving Cross-Segment Credit Misattribution in Tool-Calling RL
Segment-Locked Credit Assignment routes execution rewards to tool tokens and preference rewards to summary tokens, reducing cross-segment credit contamination in tool-calling reinforcement learning.
Submitted to NeurIPS 2026
Why Deterministic PRM Guidance Underperforms in Discrete Diffusion Reasoning
A matched-compute diagnostic showing that deterministic PRM guidance can trail simple outcome-reward reranking because of intermediate-state signal decay, diversity collapse, and readout errors.
Submitted to NeurIPS 2026
EAGLE: Entropy-Adaptive Gating of Pointwise Scores for Pairwise LLM Judges
A selective ranking framework that starts from efficient pointwise scores and invokes pairwise comparison only inside ambiguous near-tie spans.
Submitted to AAAI 2027
AgreeJudge: Predicting Cross-Rubric Support for Frozen LLM Judges
A lightweight predictor of cross-rubric support for a frozen LLM judge's verdict, improving confidence alignment and selective review without changing the verdict.
Submitted to AAAI 2027
Decoupling Localization and Scoring in Long-form Reward Models
A systematic study of 23 aggregation rules showing that error localization and impact scoring are nearly independent capabilities that should be evaluated and modeled separately.
Submitted to AAAI 2027
DECAF: Decoupling Answer Correctness from Reasoning Presentation at the Parameter Level via Counterfactual Adversarial Fine-tuning
Counterfactual masked views and crossed supervision train separate process and answer scorers so misleading reasoning presentation is less likely to override objective answer correctness.
Submitted to AAAI 2027
TrustRoboReward: Preference-Ordered Isotonic Score Editing for Multi-Paradigm Robot Reward Models
A multi-paradigm robot reward-modeling framework whose POISE module calibrates pointwise scores to respect more reliable pairwise preferences.
Submitted to NeurIPS 2026
Under-review manuscripts are shown with their target venues; CCF labels describe venue class only. Classification follows the CCF seventh-edition directory. View directory
04
Industry experience
Research methods applied to real data, models, and systems.
Tencent CSIG · Youtu Lab
Large Language Model Algorithm Research Intern
Beijing, China · On-site
Working on post-training for Code Agents at Tencent Youtu Lab, with a focus on reinforcement learning and tool-using agent behavior.
Code Agentpost-trainingRLtool useCSIGYoutu Lab
Tencent PCG · QQ
Large Language Model Algorithm Research Intern
Shenzhen, China · On-site
Built a one-million-pair multilingual translation corpus for QQ International spanning 13 Eurasian languages, covering data filtering and cleaning, low-resource language-pair distillation, NER entity repair, 4B-class base-model evaluation, and SFT. The resulting model improved in-domain performance by approximately 10 percentage points while matching a substantially larger MoE translation baseline on FLORES-200 and WMT25.
1Mtraining pairs13languages+10ppin-domain
Perfect World
NLP Algorithm Intern
Beijing, China · On-site
Developed an LLM-based advertising-slogan recognition system for player-community moderation. Combined retrieval-augmented generation with LoRA fine-tuning and deployed efficient inference with vLLM. The system reached an F1 score of 85% and supported hourly offline batch recognition.
85%F1 scoreRAG+ LoRAvLLMdeployment
05
Education
Select a school name or mark to visit its official website.
Peking University
ExpectedMaster's in Software Engineering
GPA: 3.71/4.0. Coursework includes machine learning, natural language processing, large language model development, AI practice, and algorithm design.
GPA 3.71 / 4.0Expected Jun 2027
Bachelor's in Digital Media Technology
Overall rank: 2/37, including first place in the 2022–2023 academic year.
Overall rank 2 / 372022–2023 rank 1 / 37
06
Honors & awards
Competitions, scholarships, and student leadership.
First Prize, Beijing Regional Round, Higher Education Main Track
8th China International College Students' 'Internet+' Innovation and Entrepreneurship Competition
Second and Third Prizes, National Finals
National College Student Digital Media Technology Works and Creativity Competition
Second Prize
2021 National Modeling Competition
Academic Excellence Scholarship
Beijing Forestry University
Outstanding Student Scholarship
Beijing Forestry University
Arts and Sports Excellence Scholarship
Beijing Forestry University
Merit Student
Beijing Forestry University
Outstanding Student Organization Officer
Beijing Forestry University