This dissertation categorizes the generative capabilities of deep learning-based language models from low-level to high-level dimensions. First, the model must produce human-level, fluent text. Second, it should generate content that faithfully adhere...
This dissertation categorizes the generative capabilities of deep learning-based language models from low-level to high-level dimensions. First, the model must produce human-level, fluent text. Second, it should generate content that faithfully adheres to user-specified conditions while maintaining human-level fluency, which is a requirement central to controlled text generation. Third, its answers to user queries must be trustworthy and free from internal biases. Finally, the model should effectively utilize complex contextual information to generate reasoning grounded in context. For each of these generative capabilities, this dissertation studies the corresponding forms of model misbehavior and proposes methods that directly address them. We refer to these techniques, which guide a model toward more appropriate behavior by correcting undesirable tendencies, as behavioral adaptation.
To improve fundamental linguistic fluency, we address the well-known text degeneration phenomenon, in which a model produces repetitive and unnatural sequences. We link this misbehavior to deficiencies in the token unembedding process and propose Adaptive Gradient Gating (AGG), a method that enables the model to learn a more expressive unembedding matrix. By enriching the semantic structure of token representations and facilitating more discriminative unembedding, AGG reduces repetitive outputs and promotes diverse, fluent language generation.
For controlled text generation, we address the misbehavior wherein a model, when required to satisfy a specific condition, tends to produce less fluent output. We introduce the Score-based Progressive Editor (ScoPE), which guides the model toward controllable yet fluent text by externally intervening during generation. Trained using a score reflecting the target condition, the editor adjusts the model’s intermediate outputs such that the final text naturally complies with the intended constraint. This yields an efficient adaptation approach that enhances the controllability of diverse language models.
Regarding trustworthiness in question-answering settings, we study the overproduction of negative responses in binary decisions, a misbehavior we term the negative bias problem. We present Negative Attention Score Alignment (NASA), an attention-head–level framework that mitigates this bias. By identifying attention heads causing negative bias and applying parameter-efficient tuning specifically to them, NASA reduces unjustified negative responses and improves the reliability of model outputs.
Finally, for reasoning grounded in complex contexts, we address the misbehavior in which a model fails to utilize the provided information and instead produces incorrect inferences. We observe that certain orderings of contextual information hinder the model’s ability to recognize relevant details properly. To resolve this issue, we propose Context Repetition (CoRe), which reinforces the model’s grounding by repeatedly presenting the context in a structured manner. CoRe effectively targets and mitigates failures in context utilization, leading to more accurate reasoning.
Collectively, this dissertation demonstrates adaptation techniques that address major forms of misbehavior arising during text generation in deep learning–based language models. The adapted models generate fluent natural language, produce condition-aligned output, make more trustworthy decisions, and perform reasoning grounded in all relevant contextual information. Ultimately, this work provides a foundation for enabling language models to engage in meaningful discussions within specialized domains. It also serves as a stepping stone toward future directions, such as test-time scaling for deep reasoning and more effective agentic behaviors in language models.