Theses Doctoral

Towards Data-Efficient and Explainable Large Language Models

Chen, Yanda

Data-efficient learning is crucial for building language models that can adapt to a wide variety of tasks with minimal annotations of labeled examples. Recently, the advent of large language models (LLMs) has given rise to a new ability called in-context learning (ICL), where LLMs can learn and perform a new task via inference on a prompt that consists of a few input-output pairs, all while keeping their parameters frozen.

While ICL excels on academic benchmarks, it faces several challenges in real-world deployment, including sensitivity to prompt artifacts, poor calibration of model confidence, and inefficiency due to large model sizes. We conduct a systematic study of ICL sensitivity and find a negative correlation between ICL sensitivity and accuracy. To improve ICL calibration, we propose a sensitivity-based method that assigns the negative value of sensitivity as a confidence score, and demonstrate that our approach outperforms baselines in selective prediction tasks. Additionally, we propose to enhancing the efficiency of ICL through a new method called in-context tuning, which involves fine-tuning small language models on ICL prompts. We further augment the ICL capabilities of small LMs by incorporating distillation from larger LLMs into the in-context tuning process.

Besides proposing new strategies to improve the reliability, accuracy, and efficiency of ICL, we also present a study on understanding how ICL emerges. The emergence of ICL is mysterious, as ICL prompts consisting of input-output concatenations are rare in natural text, yet pre-training on natural text alone is sufficient for ICL to emerge. We identify a structure called parallel structures, which capture pairs of phrases sampled from the same distribution, and verify through ablation experiments that these structures are a major source of ICL.

Finally, we investigate the effectiveness of LLMs in explaining themselves when prompted with ICL demonstration examples. We propose a new metric called counterfactual simulatability, which measures whether humans can use LLM-generated explanations to construct precise and generalized mental models of the LLMs. Our results demonstrate that LLMs’ capacity to provide faithful explanations is significantly lower than that of humans, even with ICL examples. To address this, we propose explanation-precision fine-tuning, which uses data augmentation to generate synthetic fine-tuning data with explanations that are consistent with answers on relevant inputs.

Our contributions advance the accuracy, reliability, efficiency and understanding of ICL of LLMs, offering methods to mitigate sensitivity, improve calibration, enhance efficiency, and strengthen the self-explaining power of LLMs. This work paves the way for more data-efficient, reliable and explainable language models for diverse real-world applications.

Files

  • thumnail for Chen_columbia_0054D_18976.pdf Chen_columbia_0054D_18976.pdf application/pdf 1.72 MB Download File

More About This Work

Academic Units
Computer Science
Thesis Advisors
McKeown, Kathleen
Yu, Zhou
Degree
Ph.D., Columbia University
Published Here
January 15, 2025