Think Twice Before Assure: Confidence Estimation for Large Language Models through Reflection on Multiple Answers

Confidence estimation aiming to evaluate output trustability is crucial for the application of large language models (LLM), especially the black-box ones. Existing confidence estimation of LLM is typically not calibrated due to the overconfidence of LLM on its generated incorrect answers. Existing approaches addressing the overconfidence issue are hindered by a significant limitation that they merely consider the confidence of one answer generated by LLM. To tackle this limitation, we propose a novel paradigm that thoroughly evaluates the trustability of multiple candidate answers to mitigate the overconfidence on incorrect answers. Building upon this paradigm, we introduce a two-step framework, which firstly instructs LLM to reflect and provide justifications for each answer, and then aggregates the justifications for comprehensive confidence estimation. This framework can be integrated with existing confidence estimation approaches for superior calibration. Experimental results on six datasets of three tasks demonstrate the rationality and effectiveness of the proposed framework.

Gotcha! Don't trick me with unanswerable questions! Self-aligning Large Language Models for Responding to Unknown Questions

Despite the remarkable abilities of Large Language Models (LLMs) to answer questions, they often display a considerable level of overconfidence even when the question does not have a definitive answer. To avoid providing hallucinated answers to these unknown questions, existing studies typically investigate approaches to refusing to answer these questions. In this work, we propose a novel and scalable self-alignment method to utilize the LLM itself to enhance its response-ability to different types of unknown questions, being capable of not only refusing to answer but also providing explanation to the unanswerability of unknown questions. Specifically, the Self-Align method first employ a two-stage class-aware self-augmentation approach to generate a large amount of unknown question-response data. Then we conduct disparity-driven self-curation to select qualified data for fine-tuning the LLM itself for aligning the responses to unknown questions as desired. Experimental results on two datasets across four types of unknown questions validate the superiority of the Self-Align method over existing baselines in terms of three types of task formulation.

TAT-LLM: A Specialized Language Model for Discrete Reasoning over Tabular and Textual Data

In this work, we address question answering (QA) over a hybrid of tabular and textual data that are very common content on the Web (e.g. SEC filings), where discrete reasoning capabilities are often required. Recently, large language models (LLMs) like GPT-4 have demonstrated strong multi-step reasoning capabilities. We then consider harnessing the amazing power of LLMs to solve our task. We abstract a Step-wise Pipeline for tabular and textual QA, which consists of three key steps, including Extractor, Reasoner and Executor, and initially design an instruction to instantiate the pipeline and validate that GPT-4 outperforms all existing methods. However, utilizing an online LLM like GPT-4 holds various challenges in terms of cost, latency, and data security risk, which motivates us to specialize smaller LLMs in this task. We develop a TAT-LLM language model by fine-tuning LLaMA 2 with the training data generated automatically from existing expert-annotated datasets following the Step-wise Pipeline. The experimental results have verified that our TAT-LLM model can outperform all baseline models, including the previous best fine-tuned models and very large-scale LLMs like GPT-4 on FinQA, TAT-QA and TAT-DQA benchmarks. We hope our work can serve as a pioneering example of specializing smaller language models for specific tasks.

Robust Instruction Optimization for Large Language Models with Distribution Shifts

Large Language Models have demonstrated significant ability in accomplishing a wide range of Natural Language Processing (NLP) tasks. However, their performance is highly sensitive to the even minor changes in the phrasing of the task instructions, leading to a line of research in automatic instruction optimization towards better performance for NLP tasks. Unfortunately, existing methods for instruction optimization fail to consider the distribution shift between the seen training data and the unseen test data, where testing on unseen group of data with a different distribution could potentially lead to performance drop. In this paper, we take an initial step of investigating the problem of LLM instruction optimization across data groups with distribution shifts. We find that the optimal instructions do encounter performance drops on LLM under certain distribution shifts. To this end, we propose a framework to derive more robust optimal instructions that improve the performance on the unseen data group without large sacrifice on the seen data group. Experimental results demonstrate the effectiveness of our proposed framework.

Doc2SoarGraph: Discrete Reasoning over Visually-Rich Table-Text Documents with Semantic-Oriented Hierarchical Graphs

Discrete reasoning over table-text documents (e.g., financial reports) gains increasing attention in recent two years. Existing works mostly simplify this challenge by manually selecting and transforming document pages to structured tables and paragraphs, hindering their practical application. In this work, we explore a more realistic problem setting in the form of TAT-DQA, i.e. to answer the question over a visually-rich table-text document. Specifically, we propose a novel Doc2SoarGraph framework with enhanced discrete reasoning capability by harnessing the differences and correlations among different elements (e.g., quantities, dates) of the given question and document with Semantic-oriented hierarchical Graph structures. We conduct extensive experiments on TAT-DQA dataset, and the results show that our proposed framework outperforms the best baseline model by 17.73% and 16.91% in terms of Exact Match (EM) and F1 score respectively on the test set, achieving the new state-of-the-art.