Danning Xie
42 posts
Danning Xie
@danning_x
Research Scientist @Meta | Ex @PyTorch @aws | CS PhD @PurdueCS | GenAI | LLM4Code | Software Engineering Researcher
Menlo Park, CA Katılım Mayıs 2022
274 Takip Edilen607 Takipçiler
I’m at #NeurIPS2025 this week! Come say hi 👋
I’ll be presenting our ✨Spotlight✨ paper 🥑CoRe on #LLM code reasoning at poster session:
📅 Fri, Dec 5
⏰ 4:30–7:30 PM
📍 Exhibit Hall C/D/E
Benchmark + details: corebench.github.io
#AI #Reasoning #StaticAnalysis #LLM4code
Lin Tan@Lin0Tan
At #NeurIPS? Come by our CoRe (✨ Spotlight) poster on #LLM code-#reasoning capabilities and meet my student Danning (@danning_x). She‘ll be in Exhibit Hall C/D/E on Fri, Dec 5, 4:30–7:30 p.m. PST. Stop in if you’re around! @PurdueCS @NeurIPSConf @cerias
Coronado, CA 🇺🇸 English
Glad to be one of the contributors. Check it out
PyTorch@PyTorch
Today Meta announced torchforge, a brand-new PyTorch-native library that makes it easy to use reinforcement learning (RL) to train AI agents. Forge provides high-performance building blocks and ready-to-use examples, so you can focus on what’s novel about your use case rather than reinventing infrastructure. Built by the PyTorch team at Meta, Forge integrates seamlessly with environments, trainers, and deployment systems—bridging the gap between RL research and scalable, production-ready agentic development. Read our blog to learn more: hubs.la/Q03PCXV_0 #Forge #PyTorch #OpenSourceAI
English
Thank you, Lin! I’m deeply grateful for the journey with @PurdueCS. Excited to share that I’ll be joining Meta Superintelligence Lab (MSL), working on PyTorch open-source reinforcement learning infrastructure.
Lin Tan@Lin0Tan
🎓 Congratulations, Dr. Danning Xie (@danning_x) and Dr. Jiannan Wang (@JiannanWang51)! Hooding time is the best time—so proud to see your growth! @PurdueCS
English
🎓 Congratulations, Dr. Danning Xie (@danning_x) and Dr. Jiannan Wang (@JiannanWang51)! Hooding time is the best time—so proud to see your growth! @PurdueCS
English
CoRe highlights the gap between code generation and code reasoning.
Bridging this gap is key for building trustworthy, analysis-aware AI systems for software engineering.
Co-authors: @MingweiZheng2, @Shell97, @JiannanWang51, @Chasen86341870, @Lin0Tan, and Xiangyu Zhang
English
LLMs can generate and repair code — but do they reason about it?
Our work CoRe🥑 (accepted to NeurIPS 2025 Spotlight✨) takes a deep look at this question through static analysis–based reasoning tasks.
🔗corebench.github.io
#NeurIPS2025 #LLMs
English
We evaluated 10 major LLMs:
GPT-o3, Gemini 2.5 Pro, DeepSeek-R1/V3, Claude 3.7, Llama 3.1, Qwen 3, and more.
✅ Good at identifying dependencies
⚠️ Still struggle with multi-step reasoning and complex control structures.
English
- It contains 12,553 human-verified task instances from 180 programs.
- Each tests fine-grained reasoning beyond surface patterns — e.g., tracking value propagation, branch logic, and semantic dependencies.
English
CoRe benchmarks whether LLMs truly understand program semantics, not just generate syntactically correct code.
It evaluates reasoning across:
• Data dependency
• Control dependency
• Information flow
in C/C++, Java, and Python.
English
We evaluated 10 major LLMs:
GPT-o3, Gemini 2.5 Pro, DeepSeek-R1/V3, Claude 3.7, Llama 3.1, Qwen 3, and more.
✅ Good at identifying dependencies
⚠️ Still struggle with multi-step reasoning and complex control structures.
English
- It contains 12,553 human-verified task instances from 180 programs.
- Each tests fine-grained reasoning beyond surface patterns — e.g., tracking value propagation, branch logic, and semantic dependencies.
English
CoRe benchmarks whether LLMs truly understand program semantics, not just generate syntactically correct code.
It evaluates reasoning across:
• Data dependency
• Control dependency
• Information flow
in C/C++, Java, and Python.
English
🎉Excited to announce our paper: “ParDiff: Practical Static Differential Analysis of Network Protocol Parsers” won #OOPSLA2024 Distinguished Paper Award! I will present tomorrow morning @splashcon ! Hope to see you in Pasadena! #SoftwareTesting #StaticAnalysis #ProtocolTesting
English
Danning Xie retweetledi
Introducing Nova, a series of foundation models for binary/assembly code.
We have also released fine-tuned models for binary code decompilation. #LLM #LLM4Code #BinaryAnalysis #Security @NanJiang719, Chengxiao Wang, Kevin Liu, @XiangzheX, Xiangyu Zhang, Petr Babkin.
English
Danning Xie retweetledi
English
