论文检索

Huixuan Zhang, Yun Lin, Xiaojun Wan

Large language models (LLMs) are known to be trained on vast amounts of data, which may unintentionally or intentionally include data from commonly used benchmarks. This inclusion can lead to cheatingly high scores on model leaderboards, yet result in disappointing performance in real-world applications. To address this benchmark contamination problem, we first propose a set of requirements that practical contamination detection methods should follow. Following these proposed requirements, we introduce PaCoST, a Paired Confidence Significance Testing to effectively detect benchmark contamination in LLMs. Our method constructs a counterpart for each piece of data with the same distribution, and performs statistical analysis of the corresponding confidence to test whether the model is significantly more confident under the original benchmark. We validate the effectiveness of PaCoST and apply it on popular open-source models and benchmarks. We find that almost all models and benchmarks we tested are suspected contaminated more or less. We finally call for new LLM evaluation methods.

Shaoyang Xu, Weilong Dong, Zishan Guo, Xinwei Wu, Deyi Xiong

Prior research has revealed that certain abstract concepts are linearly represented as directions in the representation space of LLMs, predominantly centered around English. In this paper, we extend this investigation to a multilingual context, with a specific focus on human values-related concepts (i.e., value concepts) due to their significance for AI safety. Through our comprehensive exploration covering 7 types of human values, 16 languages and 3 LLM series with distinct multilinguality (e.g., monolingual, bilingual and multilingual), we first empirically confirm the presence of value concepts within LLMs in a multilingual format. Further analysis on the cross-lingual characteristics of these concepts reveals 3 traits arising from language resource disparities: cross-lingual inconsistency, distorted linguistic relationships, and unidirectional cross-lingual transfer between high- and low-resource languages, all in terms of value concepts. Moreover, we validate the feasibility of cross-lingual control over value alignment capabilities of LLMs, leveraging the dominant language as a source language. Ultimately, recognizing the significant impact of LLMs’ multilinguality on our results, we consolidate our findings and provide prudent suggestions on the composition of multilingual data for LLMs pre-training.

Yubo Wang, Xueguang Ma, Wenhu Chen

Large-scale language models (LLMs) like ChatGPT have demonstrated impressive abilities in generating responses based on human instructions. However, their use in the medical field can be challenging due to their lack of specific, in-depth knowledge. In this study, we present a system called LLMs Augmented with Medical Textbooks (LLM-AMT) designed to enhance the proficiency of LLMs in specialized domains. LLM-AMT integrates authoritative medical textbooks into the LLMs’ framework using plug-and-play modules. These modules include a Query Augmenter, a Hybrid Textbook Retriever, and a Knowledge Self-Refiner. Together, they incorporate authoritative medical knowledge. Additionally, an LLM Reader aids in contextual understanding. Our experimental results on three medical QA tasks demonstrate that LLM-AMT significantly improves response quality, with accuracy gains ranging from 11.6% to 16.6%. Notably, with GPT-4-Turbo as the base model, LLM-AMT outperforms the specialized Med-PaLM 2 model pre-trained on a massive amount of medical corpus by 2-3%. We found that despite being 100 smaller in size, medical textbooks as a retrieval corpus are proven to be a more effective knowledge database than Wikipedia in the medical domain, boosting performance by 7.8%-13.7%.

Chau Minh Pham, Simeng Sun, Mohit Iyyer

Existing research on instruction following largely focuses on tasks with simple instructions and short responses. In this work, we explore multi-constraint instruction following for generating long-form text. We create Suri, a dataset with 20K human-written long-form texts paired with LLM-generated backtranslated instructions that contain multiple complex constraints. Because of prohibitive challenges associated with collecting human preference judgments on long-form texts, preference-tuning algorithms such as DPO are infeasible in our setting; thus, we propose Instructional ORPO (I-ORPO), an alignment method based on the ORPO algorithm. Instead of receiving negative feedback from dispreferred responses, I-ORPO obtains negative feedback from synthetically corrupted instructions generated by an LLM. Using Suri, we perform supervised and I-ORPO fine-tuning on Mistral-7b-Instruct-v0.2. The resulting models, Suri-SFT and Suri-I-ORPO, generate significantly longer texts (5K tokens) than base models without significant quality deterioration. Our human evaluation shows that while both SFT and I-ORPO models satisfy most constraints, Suri-I-ORPO generations are generally preferred for their coherent and informative incorporation of the constraints.

Chenglin Li, Qianglong Chen, Zhi Li, Feng Tao, Yicheng Li, Hao Chen, Fei Yu, Yin Zhang

Instruction tuning is a crucial technique for aligning language models with humans’ actual goals in the real world. Extensive research has highlighted the quality of instruction data is essential for the success of this alignment. However, creating high-quality data manually is labor-intensive and time-consuming, which leads researchers to explore using LLMs to synthesize data. Recent studies have focused on using a stronger LLM to iteratively enhance existing instruction data, showing promising results. Nevertheless, previous work often lacks control over the evolution direction, resulting in high uncertainty in the data synthesis process and low-quality instructions. In this paper, we introduce a general and scalable framework, IDEA-MCTS (Instruction Data Enhancement using Monte Carlo Tree Search), a scalable framework for efficiently synthesizing instructions. With tree search and evaluation models, it can efficiently guide each instruction to evolve into a high-quality form, aiding in instruction fine-tuning. Experimental results show that IDEA-MCTS significantly enhances the seed instruction data, raising the average evaluation scores of quality, diversity, and complexity from 2.19 to 3.81. Furthermore, in open-domain benchmarks, experimental results show that IDEA-MCTS improves the accuracy of real-world instruction-following skills in LLMs by an average of 5% in low-resource settings.

Xinyi Chen, Baohao Liao, Jirui Qi, Panagiotis Eustratiadis, Christof Monz, Arianna Bisazza, Maarten de Rijke

Following multiple instructions is a crucial ability for large language models (LLMs). Evaluating this ability comes with significant challenges: (i) limited coherence between multiple instructions, (ii) positional bias where the order of instructions affects model performance, and (iii) a lack of objectively verifiable tasks. To address these issues, we introduce a benchmark designed to evaluate models’ abilities to follow multiple instructions through sequential instruction following (SIFo) tasks. In SIFo, the successful completion of multiple instructions is verifiable by examining only the final instruction. Our benchmark evaluates instruction following using four tasks (text modification, question answering, mathematics, and security rule following), each assessing different aspects of sequential instruction following. Our evaluation of popular LLMs, both closed-source and open-source, shows that more recent and larger models significantly outperform their older and smaller counterparts on the SIFo tasks, validating the benchmark’s effectiveness. All models struggle with following sequences of instructions, hinting at an important lack of robustness of today’s language models.

Chenglin Li, Qianglong Chen, Liangyue Li, Caiyu Wang, Feng Tao, Yicheng Li, Zulong Chen, Yin Zhang

As large language models (LLMs) have demonstrated impressive multiple step-by-step reasoning capabilities in recent natural language processing (NLP) reasoning tasks, many studies are interested in distilling reasoning abilities into smaller language models (SLMs) via fine-tuning. Previous distillation methods usually utilize the capabilities of LLMs to generate chain-of-thought (CoT) samples to teach SLMs. However, this distillation approach performs poorly in certain scenarios due to the limitations of CoT. In this work, we introduce a novel Mixed Distillation (MD) framework, distilling multiple step-by-step reasoning abilities into SLMs. First, we leverage LLMs to generate multiple step-by-step reasoning rationales by sampling automatically. Then, we create high-quality, well-balanced mixed thought data and design a novel multi-task loss to help SLMs better learn and adaptively activate multiple step-by-step reasoning. Our extensive experiments demonstrate that MD enhances both single-path (using either CoT or PoT) and multi-path (using both CoT and PoT) reasoning abilities of SLMs during inference across reasoning tasks. Notably, a single model generated by MD exceeds the comprehensive performance of an ensemble of two individual CoT and PoT distilled models. Mistral-7B using MD can achieve remarkable improvements of 87.5%, 74.0% and 77.1% on SVAMP, GSM8K and ASDIV, respectively, outperforming the teacher model, GPT-3.5-Turbo. We hope our work provides insight into SLMs’ multiple step-by-step reasoning abilities.

Young Kyun Jang, Junmo Kang, Yong Jae Lee, Donghyun Kim

While advancements in Vision Language Models (VLMs) have significantly improved the alignment of visual and textual data, these models primarily focus on aligning images with short descriptive captions. This focus limits their ability to handle complex text interactions, particularly with longer texts such as lengthy captions or documents, which have not been extensively explored yet. In this paper, we introduce Meet At The Embedding (MATE), a novel approach that combines the capabilities of VLMs with Large Language Models (LLMs) to overcome this challenge without the need for additional image-long text pairs. Specifically, we replace the text encoder of the VLM with a pretrained LLM-based encoder that excels in understanding long texts. To bridge the gap between VLM and LLM, MATE incorporates a projection module that is trained in a multi-stage manner. It starts by aligning the embeddings from the VLM text encoder with those from the LLM using extensive text pairs. This module is then employed to seamlessly align image embeddings closely with LLM embeddings. We propose two new cross-modal retrieval benchmarks to assess the task of connecting images with long texts (lengthy captions / documents). Extensive experimental results demonstrate that MATE effectively connects images with long texts, uncovering diverse semantic relationships.

Enrica Troiano, Sofie Labat, Marco Antonio Stranisci, Rossana Damiano, Viviana Patti, Roman Klinger

There is a mismatch between psychological and computational studies on emotions. Psychological research aims at explaining and documenting internal mechanisms of these phenomena, while computational work often simplifies them into labels. Many emotion fundamentals remain under-explored in natural language processing, particularly how emotions develop and how people cope with them. To help reduce this gap, we follow theories on coping, and treat emotions as strategies to cope with salient situations (i.e., how people deal with emotion-eliciting events). This approach allows us to investigate the link between emotions and behavior, which also emerges in language. We introduce the task of coping identification, together with a corpus to do so, constructed via role-playing. We find that coping strategies realize in text even though they are challenging to recognize, both for humans and automatic systems trained and prompted on the same task. We thus open up a promising research direction to enhance the capability of models to better capture emotion mechanisms from text.

Cennet Oguz, Pascal Denis, Simon Ostermann, Emmanuel Vincent, Natalia Skachkova, Josef van Genabith

Multilingual anaphora resolution identifies referring expressions and implicit arguments in texts and links to antecedents that cover several languages. In the most challenging setting, cross-lingual anaphora resolution, training data, and test data are in different languages. As knowledge needs to be transferred across languages, this task is challenging, both in the multilingual and cross-lingual setting. We hypothesize that one way to alleviate some of the difficulty of the task is to include multimodal information in the form of images (i.e. frames extracted from instructional videos). Such visual inputs are by nature language agnostic, therefore cross- and multilingual anaphora resolution should benefit from visual information. In this paper, we provide the first multilingual and multimodal dataset annotated with anaphoric relations and present experimental results for end-to-end multimodal and multilingual anaphora resolution. Given gold mentions, multimodal features improve anaphora resolution results by ~10 % for unseen languages.

Yuting Wei, Yuanxing Xu, Xinru Wei, Simin Yang, Yangfu Zhu, Yuqing Li, Di Liu, Bin Wu

Given the importance of ancient Chinese in capturing the essence of rich historical and cultural heritage, the rapid advancements in Large Language Models (LLMs) necessitate benchmarks that can effectively evaluate their understanding of ancient contexts. To meet this need, we present AC-EVAL, an innovative benchmark designed to assess the advanced knowledge and reasoning capabilities of LLMs within the context of ancient Chinese. AC-EVAL is structured across three levels of difficulty reflecting different facets of language comprehension: general historical knowledge, short text understanding, and long text comprehension. The benchmark comprises 13 tasks, spanning historical facts, geography, social customs, art, philosophy, classical poetry and prose, providing a comprehensive assessment framework. Our extensive evaluation of top-performing LLMs, tailored for both English and Chinese, reveals a substantial potential for enhancing ancient text comprehension. By highlighting the strengths and weaknesses of LLMs, AC-EVAL aims to promote their development and application forward in the realms of ancient Chinese language education and scholarly research.

Yekun Chai, Yewei Fang, Qiwei Peng, Xuhong Li

Language models typically tokenize raw text into sequences of subword identifiers from a predefined vocabulary, a process inherently sensitive to typographical errors, length variations, and largely oblivious to the internal structure of tokens—issues we term *the curse of tokenization*. In this study, we delve into these drawbacks and demonstrate that large language models (LLMs) remain susceptible to these problems. This study systematically investigates these challenges and their impact on LLMs through three critical research questions: (1) complex problem solving, (2) token structure probing, and (3) resilience to typographical variation. Our findings reveal that scaling model parameters can mitigate the issue of tokenization; however, LLMs still suffer from biases induced by typos and other text format variations. Our experiments show that subword regularization such as BPE-dropout can mitigate this issue. We release our evaluation code and data at https://github.com/FloatAI/TKEval.

Lihu Chen, Alexandre Perez-Lebel, Fabian M. Suchanek, Gaël Varoquaux

Large Language Models (LLMs), such as GPT and LLaMA, are susceptible to generating hallucinated answers in a confident tone. While previous efforts to elicit and calibrate confidence scores have shown some success, they often overlook biases towards certain groups, such as specific nationalities. Existing calibration methods typically focus on average performance, failing to address this disparity. In our study, we demonstrate that the concept of grouping loss is an effective metric for understanding and correcting the heterogeneity in confidence levels. We introduce a novel evaluation dataset, derived from a knowledge base, specifically designed to assess the confidence scores of LLM responses across different groups. Our experimental results highlight significant variations in confidence, which are accurately captured by grouping loss. To tackle this issue, we propose a new method to calibrate the confidence scores of LLMs by considering different groups, a process we term reconfidencing. Our findings indicate that this approach effectively mitigates biases against minority groups, contributing to the development of fairer LLMs.

Miaoyu Li, Haoxin Li, Zilin Du, Boyang Li

Knowledge-based Visual Qustion-answering (K-VQA) often requires the use of background knowledge beyond the image. However, we discover that a single knowledge generation strategy is often insuffcient for all K-VQA questions. To this end, we propose Diversifcation, Evidence Truncation, and Combination for Knowledge-based Elucidation (DietCoke), which utilizes a bundle of complementary question-answering tactics and aggregates their answers using textual rationales. DietCoke comprises of three stages: diversifcation, rationalization, and ensemble. The diversification stage generates three distinctive decision contexts, each leading to its own answer candidate. The rationalization stage generates two rationales, the automatic rationale and the mechanistic rationale, for each answer candidate using decorrelated techniques. Finally, in the ensemble stage, an LLM informed by the rationales selects one answer from the three candidates. Experiments show that DietCoke significantly outperforms state-of-the-art LLM-based baselines by 2.8% on OK-VOA and 4.7% on A-OKVOA and that the strategies in the ensembles are highly complementary.

Yibin Yan, Weidi Xie

Knowledge-based Visual Question Answering (KVQA) tasks require answering questions about images using extensive background knowledge. Despite significant advancements, generative models often struggle with these tasks due to the limited integration of external knowledge. In this paper, we introduce **EchoSight**, a novel multimodal Retrieval-Augmented Generation (RAG) framework that enables large language models (LLMs) to answer visual questions requiring fine-grained encyclopedic knowledge. To strive for high-performing retrieval, EchoSight first searches wiki articles by using visual-only information, subsequently, these candidate articles are further reranked according to their relevance to the combined text-image query. This approach significantly improves the integration of multimodal knowledge, leading to enhanced retrieval outcomes and more accurate VQA responses. Our experimental results on the E-VQA and InfoSeek datasets demonstrate that EchoSight establishes new state-of-the-art results in knowledge-based VQA, achieving an accuracy of 41.8% on E-VQA and 31.3% on InfoSeek.

Bozhong Tian, Xiaozhuan Liang, Siyuan Cheng, Qingbin Liu, Mengru Wang, Dianbo Sui, Xi Chen, Huajun Chen, Ningyu Zhang

Large Language Models (LLMs) trained on extensive corpora inevitably retain sensitive data, such as personal privacy information and copyrighted material. Recent advancements in knowledge unlearning involve updating LLM parameters to erase specific knowledge. However, current unlearning paradigms are mired in vague forgetting boundaries, often erasing knowledge indiscriminately. In this work, we introduce KnowUnDo, a benchmark containing copyrighted content and user privacy domains to evaluate if the unlearning process inadvertently erases essential knowledge. Our findings indicate that existing unlearning methods often suffer from excessive unlearning. To address this, we propose a simple yet effective method, MemFlex, which utilizes gradient information to precisely target and unlearn sensitive parameters. Experimental results show that MemFlex is superior to existing methods in both precise knowledge unlearning and general knowledge retaining of LLMs.

Zhenran Xu, Senbao Shi, Baotian Hu, Longyue Wang, Min Zhang

We propose MultiSkill, an evaluation protocol that assesses large multimodal models (LMMs) across multiple fine-grained skills for alignment with human values. Recent LMMs have shown various intriguing abilities, such as solving graph theory problems and explaining visual jokes. However, existing multimodal benchmarks have mainly focused on coarse-grained evaluation (e.g., accuracy), without considering the skill composition required by specific instructions. To this end, we present MultiSkill, designed to decompose coarse-level scoring to a fine-grained skill set-level scoring tailored to each instruction. MultiSkill defines five core vision-language capabilities and divides into 12 skills that are necessary to align with user instructions. For evaluation metrics on specific skills, we propose an LMM-based evaluator for open-ended outputs. Based on the diverse instructions collected from 66 datasets spanning 10 domains, we compare multiple representative open-source and proprietary LMMs and find a high correlation between model-based and human-based evaluations. Our experiments underscore the importance of fine-grained evaluation in providing a holistic view of model performance and enhancing the reliability of the evaluation.

Li Yang, Qifan Wang, Jianfeng Chi, Jiahao Liu, Jingang Wang, Fuli Feng, Zenglin Xu, Yi Fang, Lifu Huang, Dongfang Liu

Product attribute value extraction involves identifying the specific values associated with various attributes from a product profile. While existing methods often prioritize the development of effective models to improve extraction performance, there has been limited emphasis on extraction efficiency. However, in real-world scenarios, products are typically associated with multiple attributes, necessitating multiple extractions to obtain all corresponding values. In this work, we propose an Efficient product Attribute Value Extraction (EAVE) approach via lightweight sparse-layer interaction. Specifically, we employ a heavy encoder to separately encode the product context and attribute. The resulting non-interacting heavy representations of the context can be cached and reused for all attributes. Additionally, we introduce a light encoder to jointly encode the context and the attribute, facilitating lightweight interactions between them. To enrich the interaction within the lightweight encoder, we design a sparse-layer interaction module to fuse the non-interacting heavy representation into the lightweight encoder. Comprehensive evaluation on two benchmarks demonstrate that our method achieves significant efficiency gains with neutral or marginal loss in performance when the context is long and number of attributes is large. Our code is available at: https://anonymous.4open.science/r/EAVE-EA18.

Tongxin Yuan, Zhiwei He, Lingzhong Dong, Yiming Wang, Ruijie Zhao, Tian Xia, Lizhen Xu, Binglin Zhou, Fangqi Li, Zhuosheng Zhang 等

Large language models (LLMs) have exhibited great potential in autonomously completing tasks across real-world applications. Despite this, these LLM agents introduce unexpected safety risks when operating in interactive environments. Instead of centering on the harmlessness of LLM-generated content in most prior studies, this work addresses the imperative need for benchmarking the behavioral safety of LLM agents within diverse environments. We introduce R-Judge, a benchmark crafted to evaluate the proficiency of LLMs in judging and identifying safety risks given agent interaction records. R-Judge comprises 569 records of multi-turn agent interaction, encompassing 27 key risk scenarios among 5 application categories and 10 risk types. It is of high-quality curation with annotated safety labels and risk descriptions. Evaluation of 11 LLMs on R-Judge shows considerable room for enhancing the risk awareness of LLMs: The best-performing model, GPT-4o, achieves 74.42% while no other models significantly exceed the random. Moreover, we reveal that risk awareness in open agent scenarios is a multi-dimensional capability involving knowledge and reasoning, thus challenging for LLMs. With further experiments, we find that fine-tuning on safety judgment significantly improve model performance while straightforward prompting mechanisms fail. R-Judge is publicly available at Annoymous.

Hyeonbin Hwang, Doyoung Kim, Seungone Kim, Seonghyeon Ye, Minjoon Seo

Training on large amounts of rationales (i.e., CoT Fine-tuning) has been found effective for improving mathematical reasoning of large language models (LLMs). However, acquiring human-authored solutions or augmenting rationales from proprietary models is costly and not scalable. In this paper, we study the problem of whether LLMs could self-improve mathematical reasoning capabilities. To this end, we propose Self-Explore, where the LLM is tasked to explore the first wrong step (i.e., the first pit) within the rationale and use such signals as fine-grained rewards for further improvement. On the GSM8K and MATH test set, Self-Explore achieves 11.57% and 2.89% improvement on average across three LLMs compared to supervised fine-tuning (SFT). Our code is available here]9.