$\mathcal{X}$-KD: General Experiential Knowledge Distillation for Large Language Models

Knowledge Distillation (KD) for Large Language Models (LLMs) has become increasingly important as models grow in size and complexity. While existing distillation approaches focus on imitating teacher behavior, they often overlook the original learning environment that shaped the teacher's knowledge. Inspired by the experiential learning theory and inverse reinforcement learning, we propose Experiential Knowledge Distillation ($\mathcal{X}$-KD), a novel and general framework that enables student models to learn in the teacher's original learning environment. $\mathcal{X}$-KD adopts the Approximated Variational Reward Imitation Learning (AVRIL) framework to jointly model the teacher's original reward function and perform policy distillation, encouraging consistency between the student policy and the original reward function. Our derivation demonstrates that $\mathcal{X}$-KD follows the supervised learning framework and applies to both sequence-level and divergence-based distillation methods, underlining the simplicity and flexibility of our approach. Empirical results show that $\mathcal{X}$-KD outperforms the generalized KD and MiniLLM baselines on abstractive summarization, machine translation, and arithmetic reasoning tasks. Additionally, $\mathcal{X}$-KD achieves better performance-diversity trade-off and data efficiency than baseline KD approaches.

Label-Confidence-Aware Uncertainty Estimation in Natural Language Generation

Large Language Models (LLMs) display formidable capabilities in generative tasks but also pose potential risks due to their tendency to generate hallucinatory responses. Uncertainty Quantification (UQ), the evaluation of model output reliability, is crucial for ensuring the safety and robustness of AI systems. Recent studies have concentrated on model uncertainty by analyzing the relationship between output entropy under various sampling conditions and the corresponding labels. However, these methods primarily focus on measuring model entropy with precision to capture response characteristics, often neglecting the uncertainties associated with greedy decoding results-the sources of model labels, which can lead to biased classification outcomes. In this paper, we explore the biases introduced by greedy decoding and propose a label-confidence-aware (LCA) uncertainty estimation based on Kullback-Leibler (KL) divergence bridging between samples and label source, thus enhancing the reliability and stability of uncertainty assessments. Our empirical evaluations across a range of popular LLMs and NLP datasets reveal that different label sources can indeed affect classification, and that our approach can effectively capture differences in sampling results and label sources, demonstrating more effective uncertainty estimation.

Approximated Variational Bayesian Inverse Reinforcement Learning for Large Language Model Alignment

The alignment of large language models (LLMs) is crucial for generating helpful and harmless content. Existing approaches leverage preference-based human feedback data to learn the reward function and align the LLM with the feedback data. However, these approaches focus on modeling the reward difference between the chosen and rejected demonstrations, rather than directly modeling the true reward from each demonstration. Moreover, these approaches assume that the reward is only obtained at the end of the sentence, which overlooks the modeling of intermediate rewards. These issues lead to insufficient use of training signals in the feedback data, limiting the representation and generalization ability of the reward and potentially resulting in reward hacking. In this paper, we formulate LLM alignment as a Bayesian Inverse Reinforcement Learning (BIRL) problem and propose a novel training objective, Approximated Variational Alignment (AVA), to perform LLM alignment through Approximated Variational Reward Imitation Learning (AVRIL). The BIRL formulation facilitates intermediate reward modeling and direct reward modeling on each single demonstration, which enhances the utilization of training signals in the feedback data. Experiments show that AVA outperforms existing LLM alignment approaches in reward modeling, RL fine-tuning, and direct optimization.