We present the Auto-Judge resource for the meta-evaluation of automated LLM judges, especially judges that evaluate Retrieval-Augmented Generation (RAG) systems that ground their response with citations. The resource couples (i) a data release of topics, pooled RAG responses, and human judgments, with (ii) a standardized protocol and software infrastructure for implementing "LLM-as-a-judge" methods in a reproducible and extensible way, including support for parameter sweeps and variant tracking.

The lack of efficiency in neural information retrieval remains one of the primary obstacles to deploying neural retrieval models as a first-stage retriever at scale. While recent tools have improved the standardized measurement of model efficiency, substantial progress is still needed to enable systematic comparative evaluation, for example, in terms of standards for systems and hardware configurations, cloud-based evaluation, benchmarks, and reproducibility. Beyond measurement, the IR community also needs stronger incentives to move in this direction, such as cost-efficiency as a review criterion or as efficiency and/or effectiveness measures in shared tasks, related teaching materials, efficiency-oriented user studies, and specialized awards for efficiency achievements. In particular, developing more efficient variants of highly effective retrieval algorithms should become an admissible research goal for PhD students if cost-efficiency is to become a first-class design objective in~IR. With ReNeuIR, we have established a recurring forum where these questions and new ideas are discussed and where the community comes together to collaboratively evaluate and improve efficiency benchmarking frameworks---most notably through the organization of a shared task focused on efficiency and reproducibility.

Tip-of-the-Tongue (ToT) retrieval benchmarks have largely focused on English, limiting their applicability to multilingual information access. In this work, we construct multilingual ToT test collections for Chinese, Japanese, Korean, and English, using an LLM-based query simulation framework. We systematically study how prompt language and source document language affect the fidelity of simulated ToT queries, validating synthetic queries through system rank correlation against real user queries. Our results show that effective ToT simulation requires language-aware design choices: non-English language sources are generally important, while English Wikipedia can be beneficial when non-English sources provide insufficient information for query generation. Based on these findings, we release four ToT test collections with 5,000 queries per language across multiple domains. This work provides the first large-scale multilingual ToT benchmark and offers practical guidance for constructing realistic ToT datasets beyond English.

Cranfield-style retrieval evaluations with too few or too many relevant documents or with low inter-assessor agreement on relevance can reduce the reliability of observations. In evaluations with human assessors, information needs are often formalized as retrieval topics to avoid an excessive number of relevant documents while maintaining good agreement. However, emerging evaluation setups that use Large Language Models (LLMs) as relevance assessors often use only queries, potentially decreasing the reliability. To study whether LLM relevance assessors benefit from formalized information needs, we synthetically formalize information needs with LLMs into topics that follow the established structure from previous human relevance assessments (i.e., descriptions and narratives). We compare assessors using synthetically formalized topics against the LLM-default query-only assessor on the~2019/2020~editions of TREC Deep Learning and Robust04. We find that assessors without formalization judge many more documents relevant and have a lower agreement, leading to reduced reliability in retrieval evaluations. Furthermore, we show that the formalized topics improve agreement between human and LLM relevance judgments, even when the topics are not highly similar to their human counterparts. Our findings indicate that LLM relevance assessors should use formalized information needs, as is standard for human assessment, and synthetically formalize topics when no human formalization exists to improve evaluation reliability.

Measuring effectiveness and efficiency in information retrieval has a strong empirical background. While modern retrieval systems substantially improve effectiveness, the community has not yet agreed on how to measure efficiency, making it difficult to contrast effectiveness and efficiency fairly. Efficiency-oriented system comparisons are difficult due to factors such as hardware configurations, software versioning, and experimental settings. Efficiency affects users, researchers, and the environment and can be measured in many dimensions beyond time and space, such as resource consumption, water usage, and sample efficiency. Analyzing the efficiency of algorithms and their trade-off with effectiveness requires revisiting and establishing new standards and principles, from defining relevant concepts to designing new measures and guidelines to assess the findings' significance. ReNeuIR's fourth iteration aims to bring the community together to debate these questions and collaboratively test and improve benchmarking frameworks for efficiency based on discussions and collaborations of its previous iterations, including a shared task focused on efficiency and reproducibility.

Relevance judgments can differ between assessors, but previous work has shown that such disagreements have little impact on the effectiveness rankings of retrieval systems. This applies to disagreements between humans as well as between human and large language model (LLM) assessors. However, the agreement between different LLM~assessors has not yet been systematically investigated. To close this gap, we compare eight LLM~assessors on the TREC DL tracks and the retrieval task of the RAG track with each other and with human assessors. We find that the agreement between LLM~assessors is higher than between LLMs and humans and, importantly, that LLM~assessors favor retrieval systems that use LLMs in their ranking decisions: our analyses with 30-50 retrieval systems show that the system rankings obtained by LLM~assessors overestimate LLM-based re-rankers by 9~to 17~positions on average.

How good are humans at writing and judging responses in retrieval-augmented generation (RAG) scenarios? To answer this question, we investigate the efficacy of crowdsourcing for RAG through two complementary studies: response writing and response utility judgment. Our new Webis Crowd RAG Corpus 2025 (Webis-CrowdRAG-25) consists of 903 human-written and 903 LLM-generated responses for the 301 topics of the TREC 2024 RAG~track, with each response composed according to one of the three discourse styles 'bullet list', 'essay', or 'news'. For a selection of 65 topics, the corpus further contains 47,320 pairwise human judgments and 10,556 pairwise LLM judgments across seven utility dimensions (e.g., coverage and coherence). Our analyses give insights into human writing behavior for RAG and the viability of crowdsourcing for RAG evaluation. We find that human pairwise judgments provide reliable and cost-effective results. This is much less the case for LLM-based pairwise and human/LLM-based pointwise judgments, nor for automated comparisons with human-written reference responses. All our data and tools are freely available.

The reproducibility and transparency of retrieval experiments depends on the availability of information about the experimental setup. However, the manual collection of experiment metadata can be tedious, error-prone, and inconsistent, which calls for an automated systematic collection. Expanding ir_metadata, we present the TIREx tracker, a tool that records hardware configurations, power/CPU/RAM/GPU usage, and experiment/system versions. Implemented as a lightweight platform-independent C binary, the TIREx tracker integrates seamlessly into Python, Java, or C/C++ workflows and can be easily integrated into shard task submissions, as we demonstrate for the TIRA/TIREx platform. Code, binaries, and documentation of the TIREx tracker are publicly available at https://github.com/tira-io/tirex-tracker.

The fundamental property of Cranfield-style evaluations, that system rankings are stable even when assessors disagree on individual relevance decisions, was validated on traditional test collections. However, the paradigm shift towards neural retrieval models affected the characteristics of modern test collections, e.g., documents are short, judged with four grades of relevance, and information needs have no descriptions or narratives. Under these changes, it is unclear whether assessor disagreement remains negligible for system comparisons. We investigate this aspect under the additional condition that the few modern test collections are heavily re-used. Given more possible query interpretations due to less formalized information needs, an ''expiration date'' for test collections might be needed if top-effectiveness requires overfitting to a single interpretation of relevance. We run a reproducibility study and re-annotate the relevance judgments of the 2019~TREC Deep Learning track. We can reproduce prior work in the neural retrieval setting, showing that assessor disagreement does not affect system rankings. However, we observe that some models substantially degrade with our new relevance judgments, and some have already reached the effectiveness of humans as rankers, providing evidence that test collections can expire.

With the emergence of voice assistants and large language models, conversational interaction with information has become part of everyday life. The eighth edition of the search-oriented conversational AI (SCAI) workshop brings together practitioners and researchers from various disciplines to discuss challenges and advances in conversational search systems. This year’s edition focuses on evaluations beyond relevance and accuracy and looks at conversational search from the user’s perspective. The workshop features a shared task on user-centered evaluation datasets and metrics, challenging participants to develop new and innovative ways to evaluate conversational search systems while accounting for the needs and preferences of users.

The first International Workshop on Open Web Search (WOWS) was held on Thursday, March 28th, at ECIR 2024 in Glasgow, UK. The full-day workshop had two calls for contributions: the first call aimed at scientific contributions to building, operating, and evaluating search engines cooperatively and the cooperative use of the web as a resource for researchers and innovators. The second call for implementations of retrieval components aimed to gain practical experience with joint, cooperative evaluation of search engines and their components. In total, 2~papers were accepted for the first call, and 11~software components were submitted for the second. The workshop ended with breakout sessions on how the OpenWebSearch.eu project can incorporate collaborative evaluations and a hub of search engines.

Recent advances in large language models have enabled the development of viable generative retrieval systems. Instead of a traditional document ranking, generative retrieval systems often directly return a grounded generated text as a response to a query. Quantifying the utility of the textual responses is essential for appropriately evaluating such generative ad hoc retrieval. Yet, the established evaluation methodology for ranking-based ad hoc retrieval is not suited for the reliable and reproducible evaluation of generated responses. To lay a foundation for developing new evaluation methods for generative retrieval systems, we survey the relevant literature from the fields of information retrieval and natural language processing, identify search tasks and system architectures in generative retrieval, develop a new user model, and study its operationalization.

The zero-shot effectiveness of neural retrieval models is often evaluated on the BEIR benchmark---a combination of different IR evaluation datasets. Interestingly, previous studies found that particularly on the BEIR~subset Touché 2020, an argument retrieval task, neural retrieval models are considerably less effective than BM25. Still, so far, no further investigation has been conducted on what makes argument retrieval so "special''. To more deeply analyze the respective potential limits of neural retrieval models, we run a reproducibility study on the Touché 2020 data. In our study, we focus on two experiments: (i) a black-box evaluation (i.e., no model retraining), incorporating a theoretical exploration using retrieval axioms, and (ii) a data denoising evaluation involving post-hoc relevance judgments. Our black-box evaluation reveals an inherent bias of neural models towards retrieving short passages from the Touché 2020 data, and we also find that quite a few of the neural models' results are unjudged in the Touché 2020 data. As many of the short Touché passages are not argumentative and thus non-relevant per se, and as the missing judgments complicate fair comparison, we denoise the Touché 2020 data by excluding very short passages (less than 20 words) and by augmenting the unjudged data with post-hoc judgments following the Touché guidelines. On the denoised data, the effectiveness of the neural models improves by up to 0.52 in nDCG@10, but BM25 is still more effective. Our code and the augmented Touché 2020 dataset are available at https://github.com/castorini/touche-error-analysis.

Conditional generative models such as DALL-E and Stable Diffusion generate images based on a user-defined text, the prompt. Finding and refining prompts that produce a desired image has become the art of prompt engineering. Generative models do not provide a built-in retrieval model for a user’s information need expressed through prompts. In light of an extensive literature review, we reframe prompt engineering for generative models as interactive text-based retrieval on a novel kind of “infinite index”. We apply these insights for the first time in a case study on image generation for game design with an expert. Finally, we envision how active learning may help to guide the retrieval of generated images.

Given a text query on a controversial topic, the task of Image Retrieval for Argumentation is to rank images according to how well they can be used to support a discussion on the topic. An important subtask therein is to determine the stance of the retrieved images, i.e., whether an image supports the pro or con side of the topic. In this paper, we conduct a comprehensive reproducibility study of the state of the art as represented by the CLEF'22 Touché lab and an in-house extension of it. Based on the submitted approaches, we developed a unified and modular retrieval process and reimplemented the submitted approaches according to this process. Through this unified reproduction (which also includes models not previously considered), we achieve an effectiveness improvement in argumentative image detection of up to 0.832 [email protected] However, despite this reproduction success, our study also revealed a previously unknown negative result: for stance detection, none of the reproduced or new approaches can convincingly beat a random baseline. To understand the apparent challenges inherent to image stance detection, we conduct a thorough error analysis and provide insight into potential new ways to approach this task.

We integrate irdatasets, ir_measures, and PyTerrier with TIRA in the Information Retrieval Experiment Platform (TIREx) to promote more standardized, reproducible, scalable, and even blinded retrieval experiments. Standardization is achieved when a retrieval approach implements PyTerrier's interfaces and the input and output of an experiment are compatible with ir_datasets and ir_measures. However, none of this is a must for reproducibility and scalability, as TIRA can run any dockerized software locally or remotely in a cloud-native execution environment. Version control and caching ensure efficient (re)execution. TIRA allows for blind evaluation when an experiment runs on a remote server or cloud not under the control of the experimenter. The test data and ground truth are then hidden from public access, and the retrieval software has to process them in a sandbox that prevents data leaks.

The Archive Query Log (AQL) is a previously unused, comprehensive query log collected at the Internet Archive over the last 25 years. Its first version includes 356 million queries, 137 million search result pages, and 1.4 billion search results across 550 search providers. Although many query logs have been studied in the literature, the search providers that own them generally do not publish their logs to protect user privacy and vital business data. Of the few query logs publicly available, none combines size, scope, and diversity. The AQL is the first to do so, enabling research on new retrieval models and (diachronic) search engine analyses. Provided in a privacy-preserving manner, it promotes open research as well as more transparency and accountability in the search industry.

Axiomatic approaches to information retrieval have played a key role in determining basic constraints that characterize good retrieval models. Beyond their importance in retrieval theory, axioms have been operationalized to improve an initial ranking, to "guide" retrieval, or to explain some model's rankings. However, recent open-source retrieval frameworks like PyTerrier and Pyserini, which made it easy to experiment with sparse and dense retrieval models, have not included any retrieval axiom support so far.

The amount of near-duplicates in web crawls like the ClueWeb or Common Crawl demands from their users either to develop a preprocessing pipeline for deduplication, which is costly both computationally and in person hours, or accepting the undesired effects that near-duplicates have on reliability and validity of experiments. We introduce ChatNoir-CopyCat-21, which simplifies deduplication significantly. It comes in two parts: (1) A compilation of near-duplicate documents within the ClueWeb09, the ClueWeb12, and two Common Crawl snapshots, as well as between selections of these crawls, and (2) a software library that implements the deduplication of arbitrary document sets. Our analysis shows that 14--52, of the documents within a crawl and around~0.7--2.5, between the crawls are near-duplicates. Two showcases demonstrate the application and usefulness of our resource.

We present the IR Anthology, a corpus of information retrieval publications accessible via a metadata browser and a full-text search engine. Following the example of the well-known ACL Anthology, the IR Anthology serves as a hub for researchers interested in information retrieval. Our search engine ChatNoir indexes the publications' full texts, enabling a focused search and linking users to the respective publisher's site for personal access. Listing more than 40,000 publications at the time of writing, the IR Anthology can be freely accessed at https://IR.webis.de.

NDCG is one of the most commonly used measures to quantify system performance in retrieval experiments. Though originally not considered, graded relevance judgments nowadays frequently include negative labels. Negative relevance labels cause NDCG to be unbounded. This is probably why widely used implementations of NDCG map negative relevance labels to zero, thus ensuring the resulting scores to originate from the [0,1] range. But zeroing negative labels discards valuable relevance information, e.g., by treating spam documents the same as unjudged ones, which are assigned the relevance label of zero by default. We show that, instead of zeroing negative labels, a min-max-normalization of NDCG retains its statistical power while improving its reliability and stability.

Learning to rank~(LTR) is the de facto standard for web search, improving upon classical retrieval models by exploiting (in)direct relevance feedback from user judgments, interaction logs, etc. We investigate for the first time the effect of a sampling bias on LTR~models due to the potential presence of near-duplicate web pages in the training data, and how (in)consistent relevance feedback of duplicates influences an LTR~model's decisions. To examine this bias, we construct a series of specialized LTR~datasets based on the ClueWeb09 corpus with varying amounts of near-duplicates. We devise worst-case and average-case train/test splits that are evaluated on popular pointwise, pairwise, and listwise LTR~models. Our experiments demonstrate that duplication causes overfitting and thus less effective models, making a strong case for the benefits of systematic deduplication before training and model evaluation.

\beginabstract We analyze comparative questions, i.e., questions asking to compare different items, that were submitted to Yandex in 2012. Responses to such questions might be quite different from the simple "ten blue links'' and could, for example, aggregate pros and cons of the different options as direct answers. However, changing the result presentation is an intricate decision such that the classification of comparative questions forms a highly precision-oriented task.