The advent of LLMs has profoundly reshaped the landscape of Natural Language Processing (NLP) and Artificial Intelligence (AI). While LLMs continue to push the boundaries of generative AI, their inherent general-purpose nature at the time of initial r...
The advent of LLMs has profoundly reshaped the landscape of Natural Language Processing (NLP) and Artificial Intelligence (AI). While LLMs continue to push the boundaries of generative AI, their inherent general-purpose nature at the time of initial release presents significant challenges in controlling and tailoring their outputs for specific downstream tasks, user preferences, or domains. This need for effective control, or "steering,'' is critical for real-world deployments.
Common strategies for steering LLMs are broadly categorized into training-free and post-training-based methods. Training-free approaches modify the LLM without altering the LLM's trainable parameters. This line of works includes prompt engineering, where input prompts are carefully crafted; few-shot learning, which provides in-context demonstrations; controlled generation, involving manipulation of internal hidden states or output logits at inference time; and Retrieval-Augmented Generation (RAG), which grounds responses in external knowledge sources. While computationally inexpensive and straightforward, these methods offer limited control and can be highly sensitive to prompt phrasing and context length limitations.
Conversely, post-training-based methods, such as Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), enable more precise control by modifying the LLM's weights. SFT involves further training a pre-trained LLM on a labeled dataset representative of the desired behavior, using gradient-based optimization with a cross-entropy loss. RL optimizes LLMs based on feedback rather than fixed labels. RL from Human Feedback (RLHF) trains the LLM to maximize a reward signal derived from human preferences, either through human-in-the-loop processes or learned reward models. More recently, RL with verifiable rewards (RLVR), which employs programmatically defined metrics or rules, is being widely adopted; however, its application is largely restricted to tasks where correct and incorrect outputs can be clearly distinguished (e.g., coding or mathematical problem solving). While these post-training approaches offer the highest degree of steerability, their resource-intensiveness presents a significant impediment, particularly in resource-constrained settings. This dissertation addresses these critical challenges by proposing and evaluating novel frameworks for efficiently steering LLMs.
First, DynaInst focuses on circumventing iterative retraining of language models for knowledge accumulation through continual instruction tuning. Like other Deep Neural Networks, language models are susceptible to catastrophic forgetting, where they lose previously acquired capabilities upon learning new tasks. This typically necessitates a joint, computationally prohibitive retraining to integrate additional capabilities. To overcome this, we introduce DynaInst, a novel continual instruction tuning framework that leverages both parameter regularization and experience replay to mitigate catastrophic forgetting and facilitate the continuous accumulation of task-wise generalization. Experimental results demonstrate that the proposed method significantly advances both types of generalization, achieving performance comparable to the upper-bound performance of models trained jointly on all tasks. This work offers a promising path toward building continuously evolving, universal LLMs without the need for iterative re-training on all newly arriving tasks, thereby reducing the computational burden associated with knowledge accumulation in LLMs.
Second, we explore the effective utilization of smaller, open-source LLMs for specific downstream tasks. Although extremely large proprietary LLMs offer unparalleled performance, smaller open-source models, typically with around seven billion parameters, present an efficient and attractive alternative, especially in resource-constrained settings. This paper presents a two-stage supervised fine-tuning framework specifically designed for open-source LLMs for the task of API argument filling in task-oriented conversational systems. Our experimental results demonstrate that when paired with this proposed technique, the argument filling performance of smaller LLMs noticeably improves, outperforming larger proprietary LLMs. This work thus paves a new avenue toward making LLMs more reliable and accurate in a specific, mission-critical task.
Third, HiCUPID tackles the monetary and labor costs associated with collecting data for LLM steering, particularly in the context of personalization and discloses potentials and limitations of existing parameter-efficient fine-tuning (PEFT) strategies. HiCUPID, a synthetically-generated benchmark, includes a novel conversational dataset and an automated evaluation model for developing and assessing LLMs as personalized AI assistants. With HiCUPID, this work conducts extensive experiments to investigate the personalization ability of state-of-the-art closed- and open-source LLMs. The analysis incorporates four widely used inference-time and three train-time parameter-efficient LLM customization methods, addressing the challenge of understanding the unknown potentials and limitations of existing PEFT techniques in the context of LLM personalization. This comprehensive exploration provides valuable insights into how these techniques can be leveraged for effective and resource-efficient personalization.
In summary, this dissertation makes significant contributions to improving the steerability of LLMs in resource-constrained environments by reducing the data, memory, and compute cost associated with post-training. The findings and methodologies presented here pave the way for a new generation of adaptable and steerable LLMs that can thrive even in the most resource-constrained settings.