CODA: Difficulty-Aware Compute Allocation for Adaptive Reasoning
TL;DR: We propose CODA, which scales reasoning by difficulty, reducing overthinking on easy tasks while promoting deeper reasoning on hard ones.
CV
🎓 About Me
Siye Wu (伍思烨) is a second-year M.S. student in Computer Science at Fudan University, advised by Prof. Yanghua Xiao.
His research interests lie in Natural Language Processing (NLP) and Large Language Models (LLMs), with a focus on:
- Efficient LLMs — Designing computation- and budget-aware LLMs that emphasize adaptive inference, reasoning efficiency, and performance–cost trade-offs across tasks of varying complexity.
- LLM Post-training and Reasoning — Developing post-training methods, such as supervised fine-tuning (SFT) and reinforcement learning (RL), to improve LLMs’ reasoning capabilities.
📖 Educations
- 2024.09 - 2027.06 (Expected) |
M.S. in Computer Science, Fudan University , Shanghai, China - 2020.09 - 2024.06 |
B.S. in Computer Science, Wuhan University , Wuhan, China GPA: 3.97/4.0 (Top 1%)
💻 Internships
- 2025.10 - 2026.04 |
Research Intern on Post-training, StepFun, Post-Train & Agent Group
📝 Selected Publication
See full list on Google Scholar .
Step 3.5 Flash: Open Frontier-Level Intelligence with 11B Active Parameters
TL;DR: Step 3.5 Flash is our most capable open-source foundation model, built for frontier reasoning and agentic tasks with exceptional efficiency.
ARM: Adaptive Reasoning Model
TL;DR: We propose ARM, a reasoning model capable of adaptively selecting appropriate reasoning formats based on the task at hand.
How Easily Do Irrelevant Inputs Skew the Responses of Large Language Models?
TL;DR: We present a comprehensive study of the robustness of LLMs to different types of irrelevant information under various conditions.
