Keping Bi

Shiyu Ni, Keping Bi, Jiafeng Guo et al.

Retrieval effectiveness varies substantially across queries, making it important to estimate ranking quality before relevance judgments are available. Query performance prediction (QPP) addresses this need, but most existing methods rely on external predictors after retrieval or reranking. In this paper, we study \textit{reranker-internal QPP}: can an LLM reranker estimate the quality of the ranking it has just produced? We investigate both training-free and training-based approaches. For training-free estimation, we examine metric-specific self-consistency across sampled rankings and verbalized confidence produced directly by the reranker. Experiments on TREC Deep Learning 2019--2022 with four LLMs show that self-consistency is competitive with the state-of-the-art (SOTA) approach and better calibrated in almost all settings, while direct verbalized confidence is severely overconfident. To improve verbalized confidence, we propose two supervised methods, Verb-Num and Verb-List, which enable LLM rerankers to produce calibrated ranking-quality estimates with only a few additional output tokens.

Yuhan Wang, Shiyu Ni, Zhikai Ding et al.

Confidence calibration is essential for making large language models (LLMs) reliable, yet existing training-free methods have been primarily studied under single-answer question answering. In this paper, we show that these methods break down in the presence of multiple valid answers, where disagreement among equally correct responses leads to systematic underestimation of confidence. To enable a systematic study of this phenomenon, we introduce MACE, a benchmark of 12,000 factual questions spanning six domains with varying numbers of correct answers. Experiments across 15 representative calibration methods and four LLM families (7B-72B) reveal that while accuracy increases with answer cardinality, estimated confidence consistently decreases, causing severe miscalibration for questions with mixed answer counts. To address this issue, we propose Semantic Confidence Aggregation (SCA), which aggregates confidence over multiple high-probability sampled responses. SCA achieves state-of-the-art calibration performance under mixed-answer settings while preserving strong calibration on single-answer questions.

Zhihui Yao, Hengran Zhang, Keping Bi

Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) with external knowledge but remains vulnerable to low-authority sources that can propagate misinformation. We investigate whether LLMs can perceive information authority - a capability extending beyond semantic understanding. To address this, we introduce AuthorityBench, a comprehensive benchmark for evaluating LLM authority perception comprising three datasets: DomainAuth (10K web domains with PageRank-based authority), EntityAuth (22K entities with popularity-based authority), and RAGAuth (120 queries with documents of varying authority for downstream evaluation). We evaluate five LLMs using three judging methods (PointJudge, PairJudge, ListJudge) across multiple output formats. Results show that ListJudge and PairJudge with PointScore output achieve the strongest correlation with ground-truth authority, while ListJudge offers optimal cost-effectiveness. Notably, incorporating webpage text consistently degrades judgment performance, suggesting authority is distinct from textual style. Downstream experiments on RAG demonstrate that authority-guided filtering largely improves answer accuracy, validating the practical importance of authority perception for reliable knowledge retrieval. Code and benchmark are available at: https://github.com/Trustworthy-Information-Access/AuthorityBench.

Sihui Yang, Keping Bi, Wanqing Cui et al.

Non-Factoid (NF) Question Answering (QA) is challenging to evaluate due to diverse potential answers and no objective criterion. The commonly used automatic evaluation metrics like ROUGE or BERTScore cannot accurately measure semantic similarities or answers from different perspectives. Recently, Large Language Models (LLMs) have been resorted to for NFQA evaluation due to their compelling performance on various NLP tasks. Common approaches include pointwise scoring of each candidate answer and pairwise comparisons between answers. Inspired by the evolution from pointwise to pairwise to listwise in learning-to-rank methods, we propose a novel listwise NFQA evaluation approach, that utilizes LLMs to rank candidate answers in a list of reference answers sorted by descending quality. Moreover, for NF questions that do not have multi-grade or any golden answers, we leverage LLMs to generate the reference answer list of various quality to facilitate the listwise evaluation. Extensive experimental results on three NFQA datasets, i.e., ANTIQUE, the TREC-DL-NF, and WebGLM show that our method has significantly higher correlations with human annotations compared to automatic scores and common pointwise and pairwise approaches.

Shiyu Ni, Keping Bi, Lulu Yu et al.

Large language models (LLMs) have been found to produce hallucinations when the question exceeds their internal knowledge boundaries. A reliable model should have a clear perception of its knowledge boundaries, providing correct answers within its scope and refusing to answer when it lacks knowledge. Existing research on LLMs' perception of their knowledge boundaries typically uses either the probability of the generated tokens or the verbalized confidence as the model's confidence in its response. However, these studies overlook the differences and connections between the two. In this paper, we conduct a comprehensive analysis and comparison of LLMs' probabilistic perception and verbalized perception of their factual knowledge boundaries. First, we investigate the pros and cons of these two perceptions. Then, we study how they change under questions of varying frequencies. Finally, we measure the correlation between LLMs' probabilistic confidence and verbalized confidence. Experimental results show that 1) LLMs' probabilistic perception is generally more accurate than verbalized perception but requires an in-domain validation set to adjust the confidence threshold. 2) Both perceptions perform better on less frequent questions. 3) It is challenging for LLMs to accurately express their internal confidence in natural language.

Hengran Zhang, Keping Bi, Jiafeng Guo et al.

General-purpose text embedding models underpin a wide range of NLP and information retrieval applications, and are typically trained on large-scale multi-task corpora to encourage broad generalization. However, it remains unclear how different multi-task training strategies compare in practice, and how to efficiently adapt embedding models as new domains and data types continually emerge. In this work, we present a systematic study of multi-task training for text embeddings from two perspectives: data scheduling and model merging. We compare batch-level shuffling, sequential training variants, two-stage training, and multiple merging granularities, and find that simple batch-level shuffling consistently yields the strongest overall performance, suggesting that task conflicts are limited and training datasets are largely complementary. Despite its effectiveness, batch-level shuffling exhibits two practical limitations: suboptimal out-of-domain (OOD) generalization and poor suitability for incremental learning due to expensive full retraining. To address these issues, we propose Bagging-based rObust mOdel Merging (BOOM), which trains multiple embedding models on sampled subsets and merges them into a single model, improving robustness while retaining single-model inference efficiency. Moreover, BOOM naturally supports efficient incremental updates by training lightweight update models on new data with a small historical subset and merging them into the existing model. Experiments across diverse embedding benchmarks demonstrate that BOOM consistently improves both in-domain and OOD performance over full-corpus batch-level shuffling, while substantially reducing training cost in incremental learning settings.

Minzhu Tu, Shiyu Ni, Keping Bi

Large language models (LLMs) has been widely adopted as a scalable surrogate for human evaluation, yet such judges remain imperfect and susceptible to surface-level biases. One possible reason is that these judges lack sufficient information in assessing answer correctness. With the rise of reasoning-capable models, exposing a generator's reasoning content to the judge provides richer information and is a natural candidate for improving judgment accuracy. However, its actual impact on judge behavior remains understudied. In this paper, we systematically investigate how access to reasoning chains affects LLM-based judgment across factual question answering (QA) and mathematical reasoning benchmarks. We find that weak judges are easily swayed by reasoning presence, frequently accepting incorrect answers accompanied by fluent reasoning, while strong judges can partially leverage reasoning as informative evidence. Nevertheless, even strong judges are misled by seemingly high-quality reasoning chains. Controlled experiments further reveal that both fluency and factuality of reasoning chains are critical signals driving judge decisions. These findings highlight the need for more robust LLM judges that can distinguish genuine reasoning quality from superficial fluency when evaluating modern reasoning models.

Mingkun Zhang, Keping Bi, Wei Chen et al.

In this paper, we aim to build an adversarially robust zero-shot image classifier. We ground our work on CLIP, a vision-language pre-trained encoder model that can perform zero-shot classification by matching an image with text prompts ``a photo of a <class-name>.''. Purification is the path we choose since it does not require adversarial training on specific attack types and thus can cope with any foreseen attacks. We then formulate purification risk as the KL divergence between the joint distributions of the purification process of denoising the adversarial samples and the attack process of adding perturbations to benign samples, through bidirectional Stochastic Differential Equations (SDEs). The final derived results inspire us to explore purification in the multi-modal latent space of CLIP. We propose two variants for our CLIPure approach: CLIPure-Diff which models the likelihood of images' latent vectors with the DiffusionPrior module in DaLLE-2 (modeling the generation process of CLIP's latent vectors), and CLIPure-Cos which models the likelihood with the cosine similarity between the embeddings of an image and ``a photo of a.''. As far as we know, CLIPure is the first purification method in multi-modal latent space and CLIPure-Cos is the first purification method that is not based on generative models, which substantially improves defense efficiency. We conducted extensive experiments on CIFAR-10, ImageNet, and 13 datasets that previous CLIP-based defense methods used for evaluating zero-shot classification robustness. Results show that CLIPure boosts the SOTA robustness by a large margin, e.g., from 71.7% to 91.1% on CIFAR10, from 59.6% to 72.6% on ImageNet, and 108% relative improvements of average robustness on the 13 datasets over previous SOTA. The code is available at https://github.com/TMLResearchGroup-CAS/CLIPure.

Hengran Zhang, Minghao Tang, Keping Bi et al.

Information retrieval systems have traditionally optimized for topical relevance-the degree to which retrieved documents match a query. However, relevance only approximates a deeper goal: utility, namely, whether retrieved information helps accomplish a user's underlying task. The emergence of retrieval-augmented generation (RAG) fundamentally changes this paradigm. Retrieved documents are no longer consumed directly by users but instead serve as evidence for large language models (LLMs) that produce answers. As a result, retrieval effectiveness must be evaluated by its contribution to generation quality rather than by relevance-based ranking metrics alone. This tutorial argues that retrieval objectives are evolving from relevance-centric optimization toward LLM-centric utility. We present a unified framework covering LLM-agnostic versus LLM-specific utility, context-independent versus context-dependent utility, and the connection with LLM information needs and agentic RAG. By synthesizing recent advances, the tutorial provides conceptual foundations and practical guidance for designing retrieval systems aligned with the requirements of LLM-based information access.

Mingkun Zhang, Keping Bi, Wei Chen et al.

Despite ongoing efforts to defend neural classifiers from adversarial attacks, they remain vulnerable, especially to unseen attacks. In contrast, humans are difficult to be cheated by subtle manipulations, since we make judgments only based on essential factors. Inspired by this observation, we attempt to model label generation with essential label-causative factors and incorporate label-non-causative factors to assist data generation. For an adversarial example, we aim to discriminate the perturbations as non-causative factors and make predictions only based on the label-causative factors. Concretely, we propose a casual diffusion model (CausalDiff) that adapts diffusion models for conditional data generation and disentangles the two types of casual factors by learning towards a novel casual information bottleneck objective. Empirically, CausalDiff has significantly outperformed state-of-the-art defense methods on various unseen attacks, achieving an average robustness of 86.39% (+4.01%) on CIFAR-10, 56.25% (+3.13%) on CIFAR-100, and 82.62% (+4.93%) on GTSRB (German Traffic Sign Recognition Benchmark). The code is available at https://github.com/CAS-AISafetyBasicResearchGroup/CausalDiff.

Da Li, Yuxiao Luo, Keping Bi et al.

Vision-language models advance multimodal representation learning by acquiring transferable semantic embeddings, thereby substantially enhancing performance across a range of vision-language tasks, including cross-modal retrieval, clustering, and classification. An effective embedding is expected to comprehensively preserve the semantic content of the input while simultaneously emphasizing features that are discriminative for downstream tasks. Recent approaches demonstrate that VLMs can be adapted into competitive embedding models via large-scale contrastive learning, enabling the simultaneous optimization of two complementary objectives. We argue that the two aforementioned objectives can be decoupled: a comprehensive understanding of the input facilitates the embedding model in achieving superior performance in downstream tasks via contrastive learning. In this paper, we propose CoMa, a compressed pre-training phase, which serves as a warm-up stage for contrastive learning. Experiments demonstrate that with only a small amount of pre-training data, we can transform a VLM into a competitive embedding model. CoMa achieves new state-of-the-art results among VLMs of comparable size on the MMEB, realizing optimization in both efficiency and effectiveness.

Jiahan Chen, Da Li, Hengran Zhang et al.

Multimodal embedding models, rooted in multimodal large language models (MLLMs), have yielded significant performance improvements across diverse tasks such as retrieval and classification. However, most existing approaches rely heavily on large-scale contrastive learning, with limited exploration of how the architectural and training paradigms of MLLMs affect embedding quality. While effective for generation, the causal attention and next-token prediction paradigm of MLLMs does not explicitly encourage the formation of globally compact representations, limiting their effectiveness as multimodal embedding backbones. To address this, we propose CoCoA, a Content reconstruction pre-training paradigm based on Collaborative Attention for multimodal embedding optimization. Specifically, we restructure the attention flow and introduce an EOS-based reconstruction task, encouraging the model to reconstruct input from the corresponding <EOS> embeddings. This drives the multimodal model to compress the semantic information of the input into the <EOS> token, laying the foundations for subsequent contrastive learning. Extensive experiments on MMEB-V1 demonstrate that CoCoA built upon Qwen2-VL and Qwen2.5-VL significantly improves embedding quality. Results validate that content reconstruction serves as an effective strategy to maximize the value of existing data, enabling multimodal embedding models generate compact and informative representations, raising their performance ceiling.

Lulu Yu, Keping Bi, Shiyu Ni et al.

Unbiased Learning to Rank (ULTR) aims to leverage biased implicit user feedback (e.g., click) to optimize an unbiased ranking model. The effectiveness of the existing ULTR methods has primarily been validated on synthetic datasets. However, their performance on real-world click data remains unclear. Recently, Baidu released a large publicly available dataset of their web search logs. Subsequently, the NTCIR-17 ULTRE-2 task released a subset dataset extracted from it. We conduct experiments on commonly used or effective ULTR methods on this subset to determine whether they maintain their effectiveness. In this paper, we propose a Contextual Dual Learning Algorithm with Listwise Distillation (CDLA-LD) to simultaneously address both position bias and contextual bias. We utilize a listwise-input ranking model to obtain reconstructed feature vectors incorporating local contextual information and employ the Dual Learning Algorithm (DLA) method to jointly train this ranking model and a propensity model to address position bias. As this ranking model learns the interaction information within the documents list of the training set, to enhance the ranking model's generalization ability, we additionally train a pointwise-input ranking model to learn the listwise-input ranking model's capability for relevance judgment in a listwise manner. Extensive experiments and analysis confirm the effectiveness of our approach.

Wanqing Cui, Wei Huang, Yazhi Guo et al.

Visual document retrieval requires understanding heterogeneous and multi-modal content to satisfy information needs. Recent advances use screenshot-based document encoding with fine-grained late interaction, significantly improving retrieval performance. However, retrievers are still trained with coarse global relevance labels, without revealing which regions support the match. As a result, retrievers tend to rely on surface-level cues and struggle to capture implicit semantic connections, hindering their ability to handle non-extractive queries. To alleviate this problem, we propose a \textbf{A}ttention-\textbf{G}rounded \textbf{RE}triever \textbf{E}nhancement (AGREE) framework. AGREE leverages cross-modal attention from multimodal large language models as proxy local supervision to guide the identification of relevant document regions. During training, AGREE combines local signals with the global signals to jointly optimize the retriever, enabling it to learn not only whether documents match, but also which content drives relevance. Experiments on the challenging ViDoRe V2 benchmark show that AGREE significantly outperforms the global-supervision-only baseline. Quantitative and qualitative analyses further demonstrate that AGREE promotes deeper alignment between query terms and document regions, moving beyond surface-level matching toward more accurate and interpretable retrieval. Our code is available at: https://anonymous.4open.science/r/AGREE-2025.

Minghao Tang, Shiyu Ni, Jiafeng Guo et al.

Retrieval-augmented generation (RAG) has been widely adopted to augment large language models (LLMs) with external knowledge for knowledge-intensive tasks. However, its effectiveness is often undermined by the presence of noisy (i.e., low-quality) retrieved passages. Enhancing LLMs' robustness to such noise is critical for improving the reliability of RAG systems. Recent advances have equipped LLMs with strong reasoning and self-reflection capabilities, allowing them to identify and correct errors in their reasoning process. Inspired by this ability, we propose Passage Injection-a simple yet effective method that explicitly incorporates retrieved passages into LLMs' reasoning process, aiming to enhance the model's ability to recognize and resist noisy passages. We validate Passage Injection under general RAG settings using BM25 as the retriever. Experiments on four reasoning-enhanced LLMs across four factual QA datasets demonstrate that Passage Injection significantly improves overall RAG performance. Further analysis on two noisy retrieval settings-random noise, where the model is provided irrelevant passages, and counterfactual noise, where it is given misleading passages-shows that Passage Injection consistently improves robustness. Controlled experiments confirm that Passage Injection can also effectively leverage helpful passages. These findings suggest that incorporating passages in LLMs' reasoning process is a promising direction for building more robust RAG systems. The code can be found \href{here}{https://github.com/Trustworthy-Information-Access/Passage-Injection}.

Yuchen Wen, Keping Bi, Wei Chen et al.

As large language models (LLMs) become an important way of information access, there have been increasing concerns that LLMs may intensify the spread of unethical content, including implicit bias that hurts certain populations without explicit harmful words. In this paper, we conduct a rigorous evaluation of LLMs' implicit bias towards certain demographics by attacking them from a psychometric perspective to elicit agreements to biased viewpoints. Inspired by psychometric principles in cognitive and social psychology, we propose three attack approaches, i.e., Disguise, Deception, and Teaching. Incorporating the corresponding attack instructions, we built two benchmarks: (1) a bilingual dataset with biased statements covering four bias types (2.7K instances) for extensive comparative analysis, and (2) BUMBLE, a larger benchmark spanning nine common bias types (12.7K instances) for comprehensive evaluation. Extensive evaluation of popular commercial and open-source LLMs shows that our methods can elicit LLMs' inner bias more effectively than competitive baselines. Our attack methodology and benchmarks offer an effective means of assessing the ethical risks of LLMs, driving progress toward greater accountability in their development. Our code, data, and benchmarks are available at https://yuchenwen1.github.io/ImplicitBiasEvaluation/.

Shiyu Ni, Keping Bi, Jiafeng Guo et al.

Large Language Models (LLMs) have been found to have difficulty knowing they do not possess certain knowledge and tend to provide specious answers in such cases. Retrieval Augmentation (RA) has been extensively studied to mitigate LLMs' hallucinations. However, due to the extra overhead and unassured quality of retrieval, it may not be optimal to conduct RA all the time. A straightforward idea is to only conduct retrieval when LLMs are uncertain about a question. This motivates us to enhance the LLMs' ability to perceive their knowledge boundaries to help RA. In this paper, we first quantitatively measure LLMs' such ability and confirm their overconfidence. Then, we study how LLMs' certainty about a question correlates with their dependence on external retrieved information. We propose several methods to enhance LLMs' perception of knowledge boundaries and show that they are effective in reducing overconfidence. Additionally, equipped with these methods, LLMs can achieve comparable or even better performance of RA with much fewer retrieval calls.

Shiyu Ni, Keping Bi, Jiafeng Guo et al.

Large language models (LLMs) exhibit impressive performance across diverse tasks but often struggle to accurately gauge their knowledge boundaries, leading to confident yet incorrect responses. This paper explores leveraging LLMs' internal states to enhance their perception of knowledge boundaries from efficiency and risk perspectives. We investigate whether LLMs can estimate their confidence using internal states before response generation, potentially saving computational resources. Our experiments on datasets like Natural Questions, HotpotQA, and MMLU reveal that LLMs demonstrate significant pre-generation perception, which is further refined post-generation, with perception gaps remaining stable across varying conditions. To mitigate risks in critical domains, we introduce Confidence Consistency-based Calibration ($C^3$), which assesses confidence consistency through question reformulation. $C^3$ significantly improves LLMs' ability to recognize their knowledge gaps, enhancing the unknown perception rate by 5.6% on NQ and 4.9% on HotpotQA. Our findings suggest that pre-generation confidence estimation can optimize efficiency, while $C^3$ effectively controls output risks, advancing the reliability of LLMs in practical applications.

Wanqing Cui, Keping Bi, Jiafeng Guo et al.

Since commonsense information has been recorded significantly less frequently than its existence, language models pre-trained by text generation have difficulty to learn sufficient commonsense knowledge. Several studies have leveraged text retrieval to augment the models' commonsense ability. Unlike text, images capture commonsense information inherently but little effort has been paid to effectively utilize them. In this work, we propose a novel Multi-mOdal REtrieval (MORE) augmentation framework, to leverage both text and images to enhance the commonsense ability of language models. Extensive experiments on the Common-Gen task have demonstrated the efficacy of MORE based on the pre-trained models of both single and multiple modalities.

Hengran Zhang, Keping Bi, Jiafeng Guo et al.

Dense retrieval is a crucial task in Information Retrieval (IR), serving as the basis for downstream tasks such as re-ranking and augmenting generation. Recently, large language models (LLMs) have demonstrated impressive semantic understanding capabilities, making them attractive to researchers focusing on dense retrieval. While LLMs, as decoder-style generative models, excel in language generation, they often fall short in modeling global information due to a lack of attention to subsequent tokens. Drawing inspiration from the classical word-based language modeling approach for IR, specifically the query likelihood (QL) model, we aim to leverage the generative strengths of LLMs through QL maximization. Rather than employing QL estimation for document ranking, we propose an auxiliary task of QL maximization to enhance the backbone for subsequent contrastive learning of the retriever. We introduce our model, LLM-QL, which incorporates two key components: Attention Block (AB) and Document Corruption (DC). AB blocks the attention of predictive tokens to the document tokens before the document's ending token, while DC corrupts a document by masking a portion of its tokens during prediction. Evaluations on the in-domain (MS MARCO) and out-of-domain dataset (BEIR) indicate LLM-QL's superiority over other LLM-based retrievers. Furthermore, comprehensive analyses also validate the efficacy of LLM-QL and its components.

Wei Huang, Keping Bi, Yinqiong Cai et al.

Recent studies show that neural retrievers often display source bias, favoring passages generated by LLMs over human-written ones, even when both are semantically similar. This bias has been considered an inherent flaw of retrievers, raising concerns about the fairness and reliability of modern information access systems. Our work challenges this view by showing that source bias stems from supervision in retrieval datasets rather than the models themselves. We found that non-semantic differences, like fluency and term specificity, exist between positive and negative documents, mirroring differences between LLM and human texts. In the embedding space, the bias direction from negatives to positives aligns with the direction from human-written to LLM-generated texts. We theoretically show that retrievers inevitably absorb the artifact imbalances in the training data during contrastive learning, which leads to their preferences over LLM texts. To mitigate the effect, we propose two approaches: 1) reducing artifact differences in training data and 2) adjusting LLM text vectors by removing their projection on the bias vector. Both methods substantially reduce source bias. We hope our study alleviates some concerns regarding LLM-generated texts in information access systems.

Zhikai Ding, Shiyu Ni, Keping Bi

Large vision-language models (LVLMs) demonstrate strong visual question answering (VQA) capabilities but are shown to hallucinate. A reliable model should perceive its knowledge boundaries-knowing what it knows and what it does not. This paper investigates LVLMs' perception of their knowledge boundaries by evaluating three types of confidence signals: probabilistic confidence, answer consistency-based confidence, and verbalized confidence. Experiments on three LVLMs across three VQA datasets show that, although LVLMs possess a reasonable perception level, there is substantial room for improvement. Among the three confidences, probabilistic and consistency-based signals are more reliable indicators, while verbalized confidence often leads to overconfidence. To enhance LVLMs' perception, we adapt several established confidence calibration methods from Large Language Models (LLMs) and propose three effective methods. Additionally, we compare LVLMs with their LLM counterparts, finding that jointly processing visual and textual inputs decreases question-answering performance but reduces confidence, resulting in an improved perception level compared to LLMs.

Shiyu Ni, Keping Bi, Jiafeng Guo et al.

Large language models (LLMs) often fail to recognize their knowledge boundaries, producing confident yet incorrect answers. In this paper, we investigate how knowledge popularity affects LLMs' ability to perceive their knowledge boundaries. Focusing on entity-centric factual question answering (QA), we quantify knowledge popularity from three perspectives: the popularity of entities in the question, the popularity of entities in the answer, and relation popularity, defined as their co-occurrence frequency. Experiments on three representative datasets containing knowledge with varying popularity show that LLMs exhibit better QA performance, higher confidence, and more accurate perception on more popular knowledge, with relation popularity having the strongest correlation. Cause knowledge popularity shows strong correlation with LLMs' QA performance, we propose to leverage these signals for confidence calibration. This improves the accuracy of answer correctness prediction by an average of 5.24% across all models and datasets. Furthermore, we explore prompting LLMs to estimate popularity without external corpora, which yields a viable alternative.

Hengran Zhang, Minghao Tang, Keping Bi et al.

This paper explores the use of large language models (LLMs) for annotating document utility in training retrieval and retrieval-augmented generation (RAG) systems, aiming to reduce dependence on costly human annotations. We address the gap between retrieval relevance and generative utility by employing LLMs to annotate document utility. To effectively utilize multiple positive samples per query, we introduce a novel loss that maximizes their summed marginal likelihood. Using the Qwen-2.5-32B model, we annotate utility on the MS MARCO dataset and conduct retrieval experiments on MS MARCO and BEIR, as well as RAG experiments on MS MARCO QA, NQ, and HotpotQA. Our results show that LLM-generated annotations enhance out-of-domain retrieval performance and improve RAG outcomes compared to models trained solely on human annotations or downstream QA metrics. Furthermore, combining LLM annotations with just 20% of human labels achieves performance comparable to using full human annotations. Our study offers a comprehensive approach to utilizing LLM annotations for initializing QA systems on new corpora.

Shiyu Ni, Keping Bi, Jiafeng Guo et al.

Honesty alignment-the ability of large language models (LLMs) to recognize their knowledge boundaries and express calibrated confidence-is essential for trustworthy deployment. Existing methods either rely on training-free confidence estimation (e.g., token probabilities, self-consistency) or training-based calibration with correctness annotations. While effective, achieving universal honesty alignment with training-based calibration requires costly, large-scale labeling. To support annotation-efficient training, we introduce Elicitation-Then-Calibration (EliCal), a two-stage framework that first elicits internal confidence using inexpensive self-consistency supervision, then calibrates this confidence with a small set of correctness annotations. To support a large-scale study, we release HonestyBench, a benchmark covering ten free-form QA datasets with 560k training and 70k evaluation instances annotated with correctness and self-consistency signals. Experiments show that EliCal achieves near-optimal alignment with only 1k correctness annotations (0.18% of full supervision) and better alignment performance on unseen MMLU tasks than the calibration-only baseline, offering a scalable solution toward universal honesty alignment in LLMs.

Minghao Tang, Shiyu Ni, Jingtong Wu et al.

Retrieval-augmented generation (RAG) enhances large language models (LLMs) by retrieving external documents. As an emerging form of RAG, parametric retrieval-augmented generation (PRAG) encodes documents as model parameters (i.e., LoRA modules) and injects these representations into the model during inference, enabling interaction between the LLM and documents at parametric level. Compared with directly placing documents in the input context, PRAG is more efficient and has the potential to offer deeper model-document interaction. Despite its growing attention, the mechanism underlying parametric injection remains poorly understood. In this work, we present a systematic study of PRAG to clarify the role of parametric injection, showing that parameterized documents capture only partial semantic information of documents, and relying on them alone yields inferior performance compared to interaction at text level. However, these parametric representations encode high-level document information that can enhance the model's understanding of documents within the input context. When combined parameterized documents with textual documents, the model can leverage relevant information more effectively and become more robust to noisy inputs, achieving better performance than either source alone. We recommend jointly using parameterized and textual documents and advocate for increasing the information content of parametric representations to advance PRAG.

Hengran Zhang, Keping Bi, Jiafeng Guo et al.

Retrieval-augmented generation (RAG) enhances large language models (LLMs) by incorporating external knowledge. While traditional retrieval focuses on relevance, RAG's effectiveness depends on the utility of retrieved passages, i.e., the usefulness in facilitating the generation of an accurate and comprehensive answer. Existing studies often treat utility as a generic attribute, ignoring the fact that different LLMs may benefit differently from the same passage due to variations in internal knowledge and comprehension ability. In this work, we introduce and systematically investigate the notion of LLM-specific utility. Through large-scale experiments across multiple datasets and LLMs, we demonstrate that human-annotated passages are not optimal for LLMs and that ground-truth utilitarian passages are not transferable across different LLMs. These findings highlight the necessity of adopting the LLM-specific utility in RAG research. Our findings indicate that some human-annotated passages are not ground-truth utilitarian passages for specific LLMs, partially due to the varying readability of queries and passages for LLMs, a tendency for which perplexity is a key metric. Based on these findings, we propose a benchmarking procedure for LLM-specific utility judgments. We evaluate existing utility judgment methods on six datasets and find that while verbalized methods using pseudo-answers perform robustly, LLMs struggle to assess utility effectively-failing to reject all passages for known queries and to select truly useful ones for unknown queries.

Wei Huang, Keping Bi, Yinqiong Cai et al.

As more content generated by large language models (LLMs) floods into the Internet, information retrieval (IR) systems now face the challenge of distinguishing and handling a blend of human-authored and machine-generated texts. Recent studies suggest that neural retrievers may exhibit a preferential inclination toward LLM-generated content, while classic term-based retrievers like BM25 tend to favor human-written documents. This paper investigates the influence of LLM-generated content on term-based retrieval models, which are valued for their efficiency and robust generalization across domains. Our linguistic analysis reveals that LLM-generated texts exhibit smoother high-frequency and steeper low-frequency Zipf slopes, higher term specificity, and greater document-level diversity. These traits are aligned with LLMs being trained to optimize reader experience through diverse and precise expressions. Our study further explores whether term-based retrieval models demonstrate source bias, concluding that these models prioritize documents whose term distributions closely correspond to those of the queries, rather than displaying an inherent source bias. This work provides a foundation for understanding and addressing potential biases in term-based IR systems managing mixed-source content.

Hengran Zhang, Keping Bi, Jiafeng Guo et al.

Retrieval-augmented generation (RAG) enhances large language models (LLMs) by incorporating retrieved information. Standard retrieval process prioritized relevance, focusing on topical alignment between queries and passages. In contrast, in RAG, the emphasis has shifted to utility, which considers the usefulness of passages for generating accurate answers. Despite empirical evidence showing the benefits of utility-based retrieval in RAG, the high computational cost of using LLMs for utility judgments limits the number of passages evaluated. This restriction is problematic for complex queries requiring extensive information. To address this, we propose a method to distill the utility judgment capabilities of LLMs into smaller, more efficient models. Our approach focuses on utility-based selection rather than ranking, enabling dynamic passage selection tailored to specific queries without the need for fixed thresholds. We train student models to learn pseudo-answer generation and utility judgments from teacher LLMs, using a sliding window method that dynamically selects useful passages. Our experiments demonstrate that utility-based selection provides a flexible and cost-effective solution for RAG, significantly reducing computational costs while improving answer quality. We present the distillation results using Qwen3-32B as the teacher model for both relevance ranking and utility-based selection, distilled into RankQwen1.7B and UtilityQwen1.7B. Our findings indicate that for complex questions, utility-based selection is more effective than relevance ranking in enhancing answer generation performance. We will release the relevance ranking and utility-based selection annotations for the MS MARCO dataset, supporting further research in this area.

Hengran Zhang, Keping Bi, Jiafeng Guo

While large language models (LLMs) are increasingly deployed as dense retrievers, the impact of their domain-specific specialization on retrieval effectiveness remains underexplored. This investigation systematically examines how task-specific adaptations in LLMs influence their retrieval capabilities, an essential step toward developing unified retrievers capable of handling text, code, images, and multimodal content. We conduct extensive experiments with eight Qwen2.5 7B LLMs, including base, instruction-tuned, code/math-specialized, long reasoning, and vision-language models across zero-shot retrieval settings and the supervised setting. For the zero-shot retrieval settings, we consider text retrieval from the BEIR benchmark and code retrieval from the CoIR benchmark. Further, to evaluate supervised performance, all LLMs are fine-tuned on the MS MARCO dataset. We find that mathematical specialization and the long reasoning capability cause consistent degradation in three settings, indicating conflicts between mathematical reasoning and semantic matching. The vision-language model and code-specialized LLMs demonstrate superior zero-shot performance compared to other LLMs, even surpassing BM25 on the code retrieval task, and maintain comparable performance to base LLMs in supervised settings. These findings suggest promising directions for the unified retrieval task leveraging cross-domain and cross-modal fusion.

Wanqing Cui, Keping Bi, Jiafeng Guo et al.

Commonsense plausibility estimation is critical for evaluating language models (LMs), yet existing generative approaches--reliant on likelihoods or verbalized judgments--struggle with fine-grained discrimination. In this paper, we propose ComPaSS, a novel discriminative framework that quantifies commonsense plausibility by measuring semantic shifts when augmenting sentences with commonsense-related information. Plausible augmentations induce minimal shifts in semantics, while implausible ones result in substantial deviations. Evaluations on two types of fine-grained commonsense plausibility estimation tasks across different backbones, including LLMs and vision-language models (VLMs), show that ComPaSS consistently outperforms baselines. It demonstrates the advantage of discriminative approaches over generative methods in fine-grained commonsense plausibility evaluation. Experiments also show that (1) VLMs yield superior performance to LMs, when integrated with ComPaSS, on vision-grounded commonsense tasks. (2) contrastive pre-training sharpens backbone models' ability to capture semantic nuances, thereby further enhancing ComPaSS.

Hengran Zhang, Keping Bi, Jiafeng Guo et al.

Relevance and utility are two frequently used measures to evaluate the effectiveness of an information retrieval (IR) system. Relevance emphasizes the aboutness of a result to a query, while utility refers to the result's usefulness or value to an information seeker. In Retrieval-Augmented Generation (RAG), high-utility results should be prioritized to feed to LLMs due to their limited input bandwidth. Re-examining RAG's three core components -- relevance ranking derived from retrieval models, utility judgments, and answer generation -- aligns with Schutz's philosophical system of relevances, which encompasses three types of relevance representing different levels of human cognition that enhance each other. These three RAG components also reflect three cognitive levels for LLMs in question-answering. Therefore, we propose an Iterative utiliTy judgmEnt fraMework (ITEM) to promote each step in RAG. We conducted extensive experiments on retrieval (TREC DL, WebAP), utility judgment task (GTI-NQ), and factoid question-answering (NQ) datasets. Experimental results demonstrate significant improvements of ITEM in utility judgments, ranking, and answer generation upon representative baselines.

Keping Bi, Qingyao Ai, W. Bruce Croft

Users often need to look through multiple search result pages or reformulate queries when they have complex information-seeking needs. Conversational search systems make it possible to improve user satisfaction by asking questions to clarify users' search intents. This, however, can take significant effort to answer a series of questions starting with "what/why/how". To quickly identify user intent and reduce effort during interactions, we propose an intent clarification task based on yes/no questions where the system needs to ask the correct question about intents within the fewest conversation turns. In this task, it is essential to use negative feedback about the previous questions in the conversation history. To this end, we propose a Maximum-Marginal-Relevance (MMR) based BERT model (MMR-BERT) to leverage negative feedback based on the MMR principle for the next clarifying question selection. Experiments on the Qulac dataset show that MMR-BERT outperforms state-of-the-art baselines significantly on the intent identification task and the selected questions also achieve significantly better performance in the associated document retrieval tasks.

Keping Bi, Pavel Metrikov, Chunyuan Li et al.

An effective email search engine can facilitate users' search tasks and improve their communication efficiency. Users could have varied preferences on various ranking signals of an email, such as relevance and recency based on their tasks at hand and even their jobs. Thus a uniform matching pattern is not optimal for all users. Instead, an effective email ranker should conduct personalized ranking by taking users' characteristics into account. Existing studies have explored user characteristics from various angles to make email search results personalized. However, little attention has been given to users' search history for characterizing users. Although users' historical behaviors have been shown to be beneficial as context in Web search, their effect in email search has not been studied and remains unknown. Given these observations, we propose to leverage user search history as query context to characterize users and build a context-aware ranking model for email search. In contrast to previous context-dependent ranking techniques that are based on raw texts, we use ranking features in the search history. This frees us from potential privacy leakage while giving a better generalization power to unseen users. Accordingly, we propose a context-dependent neural ranking model (CNRM) that encodes the ranking features in users' search history as query context and show that it can significantly outperform the baseline neural model without using the context. We also investigate the benefit of the query context vectors obtained from CNRM on the state-of-the-art learning-to-rank model LambdaMart by clustering the vectors and incorporating the cluster information. Experimental results show that significantly better results can be achieved on LambdaMart as well, indicating that the query clusters can characterize different users and effectively turn the ranking model personalized.

Keping Bi, Qingyao Ai, W. Bruce Croft

Product search is an important way for people to browse and purchase items on E-commerce platforms. While customers tend to make choices based on their personal tastes and preferences, analysis of commercial product search logs has shown that personalization does not always improve product search quality. Most existing product search techniques, however, conduct undifferentiated personalization across search sessions. They either use a fixed coefficient to control the influence of personalization or let personalization take effect all the time with an attention mechanism. The only notable exception is the recently proposed zero-attention model (ZAM) that can adaptively adjust the effect of personalization by allowing the query to attend to a zero vector. Nonetheless, in ZAM, personalization can act at most as equally important as the query and the representations of items are static across the collection regardless of the items co-occurring in the user's historical purchases. Aware of these limitations, we propose a transformer-based embedding model (TEM) for personalized product search, which could dynamically control the influence of personalization by encoding the sequence of query and user's purchase history with a transformer architecture. Personalization could have a dominant impact when necessary and interactions between items can be taken into consideration when computing attention weights. Experimental results show that TEM outperforms state-of-the-art personalization product retrieval models significantly.

Rahul Jha, Keping Bi, Yang Li et al.

We describe Artemis (Annotation methodology for Rich, Tractable, Extractive, Multi-domain, Indicative Summarization), a novel hierarchical annotation process that produces indicative summaries for documents from multiple domains. Current summarization evaluation datasets are single-domain and focused on a few domains for which naturally occurring summaries can be easily found, such as news and scientific articles. These are not sufficient for training and evaluation of summarization models for use in document management and information retrieval systems, which need to deal with documents from multiple domains. Compared to other annotation methods such as Relative Utility and Pyramid, Artemis is more tractable because judges don't need to look at all the sentences in a document when making an importance judgment for one of the sentences, while providing similarly rich sentence importance annotations. We describe the annotation process in detail and compare it with other similar evaluation systems. We also present analysis and experimental results over a sample set of 532 annotated documents.

Keping Bi, Qingyao Ai, W. Bruce Croft

Product search has been a crucial entry point to serve people shopping online. Most existing personalized product models follow the paradigm of representing and matching user intents and items in the semantic space, where finer-grained matching is totally discarded and the ranking of an item cannot be explained further than just user/item level similarity. In addition, while some models in existing studies have created dynamic user representations based on search context, their representations for items are static across all search sessions. This makes every piece of information about the item always equally important in representing the item during matching with various user intents. Aware of the above limitations, we propose a review-based transformer model (RTM) for personalized product search, which encodes the sequence of query, user reviews, and item reviews with a transformer architecture. RTM conducts review-level matching between the user and item, where each review has a dynamic effect according to the context in the sequence. This makes it possible to identify useful reviews to explain the scoring. Experimental results show that RTM significantly outperforms state-of-the-art personalized product search baselines.

Keping Bi, Rahul Jha, W. Bruce Croft et al.

Redundancy-aware extractive summarization systems score the redundancy of the sentences to be included in a summary either jointly with their salience information or separately as an additional sentence scoring step. Previous work shows the efficacy of jointly scoring and selecting sentences with neural sequence generation models. It is, however, not well-understood if the gain is due to better encoding techniques or better redundancy reduction approaches. Similarly, the contribution of salience versus diversity components on the created summary is not studied well. Building on the state-of-the-art encoding methods for summarization, we present two adaptive learning models: AREDSUM-SEQ that jointly considers salience and novelty during sentence selection; and a two-step AREDSUM-CTX that scores salience first, then learns to balance salience and redundancy, enabling the measurement of the impact of each aspect. Empirical results on CNN/DailyMail and NYT50 datasets show that by modeling diversity explicitly in a separate step, AREDSUM-CTX achieves significantly better performance than AREDSUM-SEQ as well as state-of-the-art extractive summarization baselines.

Qingyao Ai, Yongfeng Zhang, Keping Bi et al.

Product search is one of the most popular methods for customers to discover products online. Most existing studies on product search focus on developing effective retrieval models that rank items by their likelihood to be purchased. They, however, ignore the problem that there is a gap between how systems and customers perceive the relevance of items. Without explanations, users may not understand why product search engines retrieve certain items for them, which consequentially leads to imperfect user experience and suboptimal system performance in practice. In this work, we tackle this problem by constructing explainable retrieval models for product search. Specifically, we propose to model the "search and purchase" behavior as a dynamic relation between users and items, and create a dynamic knowledge graph based on both the multi-relational product data and the context of the search session. Ranking is conducted based on the relationship between users and items in the latent space, and explanations are generated with logic inferences and entity soft matching on the knowledge graph. Empirical experiments show that our model, which we refer to as the Dynamic Relation Embedding Model (DREM), significantly outperforms the state-of-the-art baselines and has the ability to produce reasonable explanations for search results.

Keping Bi, Choon Hui Teo, Yesh Dattatreya et al.

In product search, users tend to browse results on multiple search result pages (SERPs) (e.g., for queries on clothing and shoes) before deciding which item to purchase. Users' clicks can be considered as implicit feedback which indicates their preferences and used to re-rank subsequent SERPs. Relevance feedback (RF) techniques are usually involved to deal with such scenarios. However, these methods are designed for document retrieval, where relevance is the most important criterion. In contrast, product search engines need to retrieve items that are not only relevant but also satisfactory in terms of customers' preferences. Personalization based on users' purchase history has been shown to be effective in product search. However, this method captures users' long-term interest, which does not always align with their short-term interest, and does not benefit customers with little or no purchase history. In this paper, we study RF techniques based on both long-term and short-term context dependencies in multi-page product search. We also propose an end-to-end context-aware embedding model which can capture both types of context. Our experimental results show that short-term context leads to much better performance compared with long-term and no context. Moreover, our proposed model is more effective than state-of-art word-based RF models.

Keping Bi, Qingyao Ai, Yongfeng Zhang et al.

Intelligent assistants change the way people interact with computers and make it possible for people to search for products through conversations when they have purchase needs. During the interactions, the system could ask questions on certain aspects of the ideal products to clarify the users' needs. For example, previous work proposed to ask users the exact characteristics of their ideal items before showing results. However, users may not have clear ideas about what an ideal item looks like, especially when they have not seen any item. So it is more feasible to facilitate the conversational search by showing example items and asking for feedback instead. In addition, when the users provide negative feedback for the presented items, it is easier to collect their detailed feedback on certain properties (aspect-value pairs) of the non-relevant items. By breaking down the item-level negative feedback to fine-grained feedback on aspect-value pairs, more information is available to help clarify users' intents. So in this paper, we propose a conversational paradigm for product search driven by non-relevant items, based on which fine-grained feedback is collected and utilized to show better results in the next iteration. We then propose an aspect-value likelihood model to incorporate both positive and negative feedback on fine-grained aspect-value pairs of the non-relevant items. Experimental results show that our model is significantly better than state-of-the-art product search baselines without using feedback and those baselines using item-level negative feedback.

Keping Bi, Choon Hui Teo, Yesh Dattatreya et al.

Product search serves as an important entry point for online shopping. In contrast to web search, the retrieved results in product search not only need to be relevant but also should satisfy customers' preferences in order to elicit purchases. Previous work has shown the efficacy of purchase history in personalized product search. However, customers with little or no purchase history do not benefit from personalized product search. Furthermore, preferences extracted from a customer's purchase history are usually long-term and may not always align with her short-term interests. Hence, in this paper, we leverage clicks within a query session, as implicit feedback, to represent users' hidden intents, which further act as the basis for re-ranking subsequent result pages for the query. It has been studied extensively to model user preference with implicit feedback in recommendation tasks. However, there has been little research on modeling users' short-term interest in product search. We study whether short-term context could help promote users' ideal item in the following result pages for a query. Furthermore, we propose an end-to-end context-aware embedding model which can capture long-term and short-term context dependencies. Our experimental results on the datasets collected from the search log of a commercial product search engine show that short-term context leads to much better performance compared with long-term and no context. Our results also show that our proposed model is more effective than word-based context-aware models.

Keping Bi, Qingyao Ai, W. Bruce Croft

Relevance feedback techniques assume that users provide relevance judgments for the top k (usually 10) documents and then re-rank using a new query model based on those judgments. Even though this is effective, there has been little research recently on this topic because requiring users to provide substantial feedback on a result list is impractical in a typical web search scenario. In new environments such as voice-based search with smart home devices, however, feedback about result quality can potentially be obtained during users' interactions with the system. Since there are severe limitations on the length and number of results that can be presented in a single interaction in this environment, the focus should move from browsing result lists to iterative retrieval and from retrieving documents to retrieving answers. In this paper, we study iterative relevance feedback techniques with a focus on retrieving answer passages. We first show that iterative feedback is more effective than the top-k approach for answer retrieval. Then we propose an iterative feedback model based on passage-level semantic match and show that it can produce significant improvements compared to both word-based iterative feedback models and those based on term-level semantic similarity.

Keping Bi, Qingyao Ai, W. Bruce Croft

As more and more search traffic comes from mobile phones, intelligent assistants, and smart-home devices, new challenges (e.g., limited presentation space) and opportunities come up in information retrieval. Previously, an effective technique, relevance feedback (RF), has rarely been used in real search scenarios due to the overhead of collecting users' relevance judgments. However, since users tend to interact more with the search results shown on the new interfaces, it becomes feasible to obtain users' assessments on a few results during each interaction. This makes iterative relevance feedback (IRF) techniques look promising today. IRF has not been studied systematically in the new search scenarios and its effectiveness is mostly unknown. In this paper, we re-visit IRF and extend it with RF models proposed in recent years. We conduct extensive experiments to analyze and compare IRF with the standard top-k RF framework on document and passage retrieval. Experimental results show that IRF is at least as effective as the standard top-k RF framework for documents and much more effective for passages. This indicates that IRF for passage retrieval has huge potential.

Qingyao Ai, Keping Bi, Cheng Luo et al.

Learning to rank with biased click data is a well-known challenge. A variety of methods has been explored to debias click data for learning to rank such as click models, result interleaving and, more recently, the unbiased learning-to-rank framework based on inverse propensity weighting. Despite their differences, most existing studies separate the estimation of click bias (namely the \textit{propensity model}) from the learning of ranking algorithms. To estimate click propensities, they either conduct online result randomization, which can negatively affect the user experience, or offline parameter estimation, which has special requirements for click data and is optimized for objectives (e.g. click likelihood) that are not directly related to the ranking performance of the system. In this work, we address those problems by unifying the learning of propensity models and ranking models. We find that the problem of estimating a propensity model from click data is a dual problem of unbiased learning to rank. Based on this observation, we propose a Dual Learning Algorithm (DLA) that jointly learns an unbiased ranker and an \textit{unbiased propensity model}. DLA is an automatic unbiased learning-to-rank framework as it directly learns unbiased ranking models from biased click data without any preprocessing. It can adapt to the change of bias distributions and is applicable to online learning. Our empirical experiments with synthetic and real-world data show that the models trained with DLA significantly outperformed the unbiased learning-to-rank algorithms based on result randomization and the models trained with relevance signals extracted by click models.

Qingyao Ai, Keping Bi, Jiafeng Guo et al.

Learning to rank has been intensively studied and widely applied in information retrieval. Typically, a global ranking function is learned from a set of labeled data, which can achieve good performance on average but may be suboptimal for individual queries by ignoring the fact that relevant documents for different queries may have different distributions in the feature space. Inspired by the idea of pseudo relevance feedback where top ranked documents, which we refer as the \textit{local ranking context}, can provide important information about the query's characteristics, we propose to use the inherent feature distributions of the top results to learn a Deep Listwise Context Model that helps us fine tune the initial ranked list. Specifically, we employ a recurrent neural network to sequentially encode the top results using their feature vectors, learn a local context model and use it to re-rank the top results. There are three merits with our model: (1) Our model can capture the local ranking context based on the complex interactions between top results using a deep neural network; (2) Our model can be built upon existing learning-to-rank methods by directly using their extracted feature vectors; (3) Our model is trained with an attention-based loss function, which is more effective and efficient than many existing listwise methods. Experimental results show that the proposed model can significantly improve the state-of-the-art learning to rank methods on benchmark retrieval corpora.