Generalization in Language Models: Mechanism, Evaluation, and Adaptation
Name
wu-zfw-phd-eecs-2026-thesis.pdf
Size
7.4 MB
Format
Adobe PDF
Checksum (MD5)
6f7c8ac3c4bc17404e2c61874846d8a5
Author(s)
Wu, Zhaofeng
Advisor(s)
Kim, Yoon
Date Issued
May 2026
Publisher
Massachusetts Institute of Technology
Abstract
Modern large-scale language models (LMs) have remarkable capabilities, yet their performance often degrades when tested on distributions that deviate from their curated training data. This dissertation investigates the challenge of LM generalization by framing common LM failure modes—ranging from cross-lingual degradation and formatting brittleness to failures in counterfactual reasoning—as instances of covariate shifts. We argue that a key factor to robust generalization lies in a high-quality representation space learned during pretraining. Specifically, we demonstrate that LMs implicitly develop a semantic hub mechanism, a representation strategy that projects isomorphic inputs across diverse data types—such as languages and modalities—into a shared, functionality-centric latent space. We argue that when this hub functions effectively, it can support zero-shot transfer of post-trained capabilities to unseen input covariates. However, our empirical evaluation reveals that this transfer is fragile when the input covariate shifts into low-resource domains in the representation space. We show that reward models systematically degrade under superficial formatting perturbations. In the domain of code, we demonstrate that task capabilities learned in high-resource programming languages fail to transfer satisfactorily to lower-resource ones. Furthermore, through counterfactual evaluations, we expose a fundamental limitation in LM reasoning: when explicitly asked to operate under counterfactual world models that contradict their pre-training priors, models struggle to faithfully perform many tasks. To address these vulnerabilities, this dissertation introduces algorithmic adaptations to improve LM generalizability. We leverage synthetic data to fill in gaps in training distributions, and align representations of isomorphic inputs either implicitly via structured training data or explicitly using regularization. Together, these contributions provide a comprehensive framework for understanding, evaluating, and algorithmically encouraging generalization in language models.
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