Advancing Language Models for Reasoning and Beyond Text
Name
zha-kzha-phd-eecs-2026-thesis.pdf
Size
36.65 MB
Format
Adobe PDF
Checksum (MD5)
4c8512f0be124c3d7d615a8e9ee52e47
Author(s)
Zha, Kaiwen
Advisor(s)
Katabi, Dina
Date Issued
May 2026
Publisher
Massachusetts Institute of Technology
Abstract
Modern language models have achieved remarkable success in natural language processing and increasingly serve as a foundation for generative modeling. However, two fundamental challenges remain: current training paradigms still struggle with complex multi-step and long-context reasoning, and language modeling remains difficult to effectively extend beyond text to other modalities and domains. This thesis addresses these challenges through a unified study of language models for reasoning and multimodal sequence modeling. The first part of the thesis focuses on improving the reasoning capabilities of language models by advancing their training paradigms. We first present RL Tango, a co-evolutionary reinforcement learning framework that co-trains the policy model with the judge to improve reasoning capabilities through process-level supervision for multi-step reasoning. We then introduce RLKV, a method that improves long-context reasoning by extending RL sampling beyond the response space into the context space through KV sampling, allowing models to adaptively process and compress long inputs. The second part of the thesis extends language modeling to new modalities and develops more effective tokenization schemes for continuous and high-dimensional data. We introduce B2B, a waveform-conditional generative framework that models physiological time series as a tokenized language, translating nocturnal respiration into sleep electroencephalography (EEG) by bridging continuous signals and discrete token representations. We then propose TexTok, a language-guided image tokenization framework that incorporates semantic information from text captions into the tokenization process, enabling more efficient and expressive representations for image generation. Together, these works advance language models along two key directions: improving their reasoning abilities and broadening their applicability beyond text. By addressing both challenges, this thesis demonstrates how language models can be made more capable of handling complex reasoning tasks and more broadly applicable across modalities, moving toward more general-purpose sequence modeling systems.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
