Retrieval-Augmented Generation (RAG) has become a standard paradigm for grounding Large Language Models (LLMs) with external knowledge.However, RAG performance often degrades substantially when faced with noisy, outdated, or conflicting retrieved information.In this work, we empirically demonstrate that Prior-Guided Reasoning-a strategy that explicitly elicits the model's parametric knowledge as prior information to guide reasoning on retrieved documents-effectively mitigates the impact of external conflicts.Building on this, we propose BrPr (Bernoulligated reinforcement learning for Prior-Guided reasoning), a framework that achieves robust performance across varying degrees of external inconsistency.Furthermore, by employing a Bernoulli-gated dropout mechanism during training, BrPr distills the prior-driven reasoning capability into the model parameters, enabling efficient latent reasoning without explicit prior generation.The experimental results demonstrate that BrPr consistently exhibits superior robustness to external conflicts and noise.

Xiaowei Yuan, Zhao Yang, Ziyang Huang, Yequan Wang, Siqi Fan, Yiming Ju, Jun Zhao, Kang Liu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Chain-of-thought Distillation (CoTD) aims at distilling Chain-of-thought (CoT) reasoning ability of large language models (LLMs) to much smaller student models. The core of CoTD is using a large teacher model to generate rationales and fine-tune smaller student models. However, current Chain-of-thought Distillation works have the following limitations: 1) Student models are separately distilled from specific reasoning tasks and lack a collaboration mechanism, hindering the enhancement of reasoning performance through collaboration among various reasoning tasks. 2) The parameter update of student models severely harms the CoT reasoning ability on other unseen reasoning tasks not included in the distillation process. In this work, we introduce a novel CoT Distillation method, MoDE-CoTD, which decouples the CoT reasoning abilities out of the student model by distilling multiple LoRA-Experts and freezing the parameters of the student model. Sequentially, LoRA-Experts are combined and adapted to handle both seen and unseen reasoning tasks, enabling collaboration among diverse reasoning tasks to further enhance CoT reasoning performance. Experimental results on 14 datasets (including 4 unseen datasets) demonstrate the strength of MoDE-CoTD, with an average accuracy gain of 6.3% on seen datasets and 7.8% on unseen datasets.

Chain-of-Thought (CoT) prompting combined with large language models (LLM) has shown great potential in improving performance on challenging reasoning tasks. While understanding why CoT prompting is effective is crucial for the application and improvement of CoT prompting, few studies have addressed this issue. Besides, almost no prior work has conducted theoretical analysis on CoT prompting in the context of black-box models. In this paper, we approach the analysis of CoT prompting in black-box LLMs from an information-theoretic perspective. Specifically, we propose a new metric, EPVI (Estimated Pointwise V-Information), which extends the concept of pointwise V-information to black-box models, quantifying the label-relevant new information introduced by CoT prompting beyond the pre-existing information in the input. Based on this, we conduct a series of experiments at both the task and instance levels to analyze CoT prompting, demonstrating that the effectiveness of CoT prompting can be attributed to its capacity to influence the difficulty of model inference by augmenting or reducing the model-usable information. Furthermore, we show that selecting high-quality demonstrations of CoT reasoning based on EPVI can improve the downstream performance of reasoning tasks.

Deyuan Liu, Zhanyue Qin, Hairu Wang, Zhao Yang, Zecheng Wang, Fangying Rong, Qingbin Liu, Yanchao Hao, Bo Li, Xi Chen, Cunhang Fan, Zhao Lv, Dianhui Chu, Zhiying Tu, Dianbo Sui. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

In the Retrieval-Augmented Generation (RAG) system, advanced Large Language Models (LLMs) have emerged as effective Query Likelihood Models (QLMs) in an unsupervised way, which re-rank documents based on the probability of generating the query given the content of a document.However, directly prompting LLMs to approximate QLMs inherently is biased, where the estimated distribution might diverge from the actual document-specific distribution.In this study, we introduce a novel framework, UR 3 , which leverages Bayesian decision theory to both quantify and mitigate this estimation bias.Specifically, UR 3 reformulates the problem as maximizing the probability of document generation, thereby harmonizing the optimization of query and document generation probabilities under a unified risk minimization objective.Our empirical results indicate that UR 3 significantly enhances re-ranking, particularly in improving the Top-1 accuracy.It benefits the QA tasks by achieving higher accuracy with fewer input documents.

Knowledge distillation is widely used in pre-trained language model compression, which can transfer knowledge from a cumbersome model to a lightweight one. Though knowledge distillation based model compression has achieved promising performance, we observe that explanations between the teacher model and the student model are not consistent. We argue that the student model should study not only the predictions of the teacher model but also the internal reasoning process. To this end, we propose Explanation Guided Knowledge Distillation (EGKD) in this article, which utilizes explanations to represent the thinking process and improve knowledge distillation. To obtain explanations in our distillation framework, we select three typical explanation methods rooted in different mechanisms, namely gradient-based , perturbation-based , and feature selection methods. Then, to improve computational efficiency, we propose different optimization strategies to utilize the explanations obtained by these three different explanation methods, which could provide the student model with better learning guidance. Experimental results on GLUE demonstrate that leveraging explanations can improve the performance of the student model. Moreover, our EGKD could also be applied to model compression with different architectures.

Although In-Context Learning has proven effective across a broad array of tasks, its efficiency is noticeably influenced by the selection of demonstrations.Existing methods tend to select different demonstrations for each test instance, which is time-consuming and poses limitations in practical scenarios.Therefore, this study aims to address the challenge of selecting a representative subset of in-context demonstrations that can effectively prompt different test instances in a specific task.We propose that this representative subset should be of high quality and diversity.Our empirical analyses confirm that demonstrations that meet these criteria can indeed bolster model performance.To satisfy these criteria, this paper further introduces a two-stage Determinantal Point Process (DPP) method designed to incorporate both quality and diversity in the process of demonstration selection, thereby obtaining representative in-context demonstrations.Through comprehensive experimentation, we have confirmed the efficacy of our proposed method, paving the way for more practical and effective In-Context Learning.

Modern deep learning models are notoriously opaque, which has motivated the development of methods for interpreting how deep models predict.This goal is usually approached with attribution method, which assesses the influence of features on model predictions.As an explanation method, the evaluation criteria of attribution methods is how accurately it reflects the actual reasoning process of the model (faithfulness).Meanwhile, since the reasoning process of deep models is inaccessible, researchers design various evaluation methods to demonstrate their arguments.However, some crucial logic traps in these evaluation methods are ignored in most works, causing inaccurate evaluation and unfair comparison.This paper systematically reviews existing methods for evaluating attribution scores and summarizes the logic traps in these methods.We further conduct experiments to demonstrate the existence of each logic trap.Through both theoretical and experimental analysis, we hope to increase attention on the inaccurate evaluation of attribution scores.Moreover, with this paper, we suggest stopping focusing on improving performance under unreliable evaluation systems and starting efforts on reducing the impact of proposed logic traps.

Zhongtao Jiang, Yuanzhe Zhang, Zhao Yang, Jun Zhao, Kang Liu. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.