Correctness, efficiency, and code quality
CLEARTRANS and SWIFTTRANS treat translated code as an engineering artifact: it should run correctly, run efficiently, and avoid structural debt.
HIT Shenzhen / KLC Lab / Code LLMs
Chenhao Hu
I study how language models learn from executable feedback: code generation with denser rewards, translation systems that preserve both correctness and quality, and reasoning workflows that know when evidence is sufficient.
- Conference papers
- ICML 2026 x2
- Education
- HIT SZ 2027
Research Axis
Clean signals for models that write code.
The work is practical at its core: feedback should be observable, reasoning should be controllable, and translated code should remain useful after the benchmark is over.
Reasoning with a stopping rule
SuCo and safety data construction explore sufficiency-guided reasoning, benchmark validation, and human-AI collaboration for trustworthy evaluation.
Harbin Institute of Technology, Shenzhen
Computer Science and Technology training across systems, programming languages, mathematics, and engineering practice.
Undergraduate Researcher, KLC Lab
Researching code LLMs, reasoning strategies, translation quality, and safety evaluation through executable feedback.
Research pipeline across RL, translation, and reasoning
Current work spans CLEARTRANS, SWIFTTRANS, SuCo, and a Shenzhen Natural Science Foundation safety testing project.
Selected Work
Projects framed by the behavior they change.
The page stays quiet visually, but the cards respond to motion: hover, tilt, and soft focus reveal how each project connects reward, reasoning, and evaluation.
Clean Code: Eliminating Code Smells in Code Translation with Anchor-to-Thought CoT Synthesis
Introduces CLEARTRANS, a reasoning LLM framework for smell-free code translation using Anchor-to-Thought CoT synthesis and dual-reward reinforcement learning.
translated code -> maintainable artifactBridging Functional Correctness and Runtime Efficiency Gaps in LLM-Based Code Translation
Implements SWIFTTRANS, a two-stage LLM code translation framework that improves functional correctness and runtime efficiency across three major benchmarks.
correctness + efficiency, evaluated togetherSuCo: Sufficiency-guided Continuous Adaptive Reasoning
Led experimental validation across benchmark datasets and improved response accuracy by about 3%, demonstrating the value of sufficiency-guided adaptive reasoning.
reason only as long as evidence requiresSafety and Trustworthiness Testing of Large AI Models
Devised a human-AI collaborative annotation framework using LLM auto-generation and classification to synthesize cross-modal security data for real-world applications.
safety data -> auditable workflowSignal
Small details, real traction.
I like work that is simple to explain after it is hard to build: clear feedback, sharp experiments, and enough taste to make the system feel inevitable.
Technical stack
Python
Java
C++
Linux
PyTorch
vLLM
VERL
RL
Code LLMs
Award
First Prize, Provincial Level, Chinese Mathematics Competitions for University Students, 2024.
Outside the lab
Feet on the trail, lens on the world.
Volunteer at Shenzhen Marathon and technical department member of the university table tennis club.
Contact
Open to research conversations and collaboration.
Based in Shenzhen, Guangdong, China. Best reached by email for code LLMs, reasoning, or safety evaluation discussions.