Xinsong Feng (冯欣淞)

I am a first-year Ph.D. student in Computer Science at Georgia Tech.

Previously, I was a research assistant at the Data-Driven Decision Intelligence (D3I) Lab, led by Prof. Haipeng Chen, at William & Mary (2025–2026). I received my M.S. degree in Electrical and Computer Engineering from UCLA (2023–2025), where I was a member of the Wireless Lab at UCLA, supervised by Prof. Ian P. Roberts. Before that, I obtained my B.Eng. in Communication Engineering from Chongqing University (2019–2023).

My research interests include reinforcement learning, generative models, and large language models. Currently, I focus primarily on diffusion language models (DLMs), with particular interests in:

  • discrete DLMs: decoding order and remasking policies
  • continuous DLMs: latent representations, architecture design, efficiency, and parallel generation
  • multimodal DLMs: how semantics emerge during generation, and modality binding
  • training and testing: training objectives, train–test mismatch, and test-time scaling
  • theoretical analysis: learnability and global consistency

Some questions I’m currently exploring:

  • What makes a generative process learnable?
  • How do representations shape learning and generation?

I enjoy exchanging research ideas and sharing what I’ve learned. Feel free to reach out if you’d like to discuss research or explore potential collaborations.

Projects

First-author

When Are Errors Evidence? Deciding Whether to Revise, What to Keep, and How to Commit in Diffusion Language Models

Xinsong Feng, Jinjin He, Peng Du, Zhizhuo Yang, Daniel M. Bikel, Jiayun Wang

Submitted to ACL ARR 2026

We separate post-hoc correction in diffusion language models into three decisions: whether to revise, what to keep, and how to commit. Erasing suspicious tokens is safe when the prompt determines the answer, but when those tokens carry the only evidence, it makes faithful recovery impossible and can launder the error into a consistent but incorrect output.

Rethinking the Roles of Diffusion and Decoding: Predictive-State Flows for Parallel Language Generation

Jinjin He*, Xinsong Feng*, Lucheng Fu, Sinan Wang, Bo Zhu, Jiayun Wang * equal contribution

Submitted to ICLR 2027 Page Paper Code Models

We introduce Predictive-State Flows (PSF), which separates the generative roles of diffusion and decoding through predictive states computed solely from each block's preceding context. The diffusion prior generates high-level predictive information, while the decoder resolves the remaining token-level choices and generates blocks in parallel.

Denoising Blocks, Not Tokens: Efficient Compressed Continuous Diffusion with Branching Token Realization

Xinsong Feng, Peng Du, Zhizhuo Yang, Daniel M. Bikel, Jiayun Wang, Haipeng Chen

Submitted to ICLR 2027 Paper

We introduce Branching Latent Diffusion (BLD), which compresses 1024 tokens into 64 block latents and reconstructs text through parallel autoregressive branches, achieving over 6× higher throughput than both ELF-L and the AR baseline in same-GPU evaluations.

DELA: Learning Controllable Edits for Diffusion Language Models

Xinsong Feng, Chenan Wang, Leshu Tang, Johnny Liu, Dong Li, Da Li, Haipeng Chen

Submitted to ICLR 2027 Paper

We introduce Diffusion Editing Language Adapter (DELA), a plug-in edit controller that separates draft-level control from token-level denoising in frozen diffusion language models. Explicit INSERT, DELETE, REMASK, and STOP actions enable controllable variable-length refinement, with cache-friendly drafting followed by targeted global edits.

Offline Reinforcement Learning with Generative Trajectory Policies

Xinsong Feng, Leshu Tang, Chenan Wang, Haipeng Chen

We propose Generative Trajectory Policies (GTPs), an ODE-based framework that unifies generative policies in offline RL, overcoming the performance–efficiency trade-off and achieving state-of-the-art results on D4RL benchmarks.

Sequential stochastic combinatorial optimization using hierarchal reinforcement learning

Xinsong Feng, Zihan Yu, Yanhai Xiong, Haipeng Chen

ICLR 2025 Paper Poster Code

We propose Wake-Sleep Option (WS-option), a hierarchical reinforcement learning framework for sequential stochastic combinatorial optimization that jointly optimizes budget allocation and node selection in a two-layer MDP.

Co-authored

Keeping Is Not Enough: Separating Selection and Use in Long-Conversation Context Management

Chenan Wang, Xinsong Feng, Yuanhe Li, Johnny Liu, Da Li, Dong Li, Ayan Mukhopadhyay, Haipeng Chen

Submitted to ACL ARR 2026

Steering Is Easy, Routing Is Not: Dual-Stream Provenance Routing for Multi-Turn Constraint Adherence

Chenan Wang, Xinsong Feng, Yuanhe Li, Johnny Liu, Da Li, Dong Li, Ayan Mukhopadhyay, Haipeng Chen

Submitted to AAAI 2027

Solving Constrained Optimization Problems as ODE-based Models Using Reinforcement Learning

Han Meng, Xinsong Feng, Yang Li, Chenan Wang, Kishansingh Rajput, Malachi Schram, Haipeng Chen

TMLR, 2026 Paper Code

Selected Writing