Initial Nugget Evaluation Results for the TREC 2024 RAG Track with the AutoNuggetizer Framework

This report provides an initial look at partial results from the TREC 2024 Retrieval-Augmented Generation (RAG) Track. We have identified RAG evaluation as a barrier to continued progress in information access (and more broadly, natural language processing and artificial intelligence), and it is our hope that we can contribute to tackling the many challenges in this space. The central hypothesis we explore in this work is that the nugget evaluation methodology, originally developed for the TREC Question Answering Track in 2003, provides a solid foundation for evaluating RAG systems. As such, our efforts have focused on "refactoring" this methodology, specifically applying large language models to both automatically create nuggets and to automatically assign nuggets to system answers. We call this the AutoNuggetizer framework. Within the TREC setup, we are able to calibrate our fully automatic process against a manual process whereby nuggets are created by human assessors semi-manually and then assigned manually to system answers. Based on initial results across 21 topics from 45 runs, we observe a strong correlation between scores derived from a fully automatic nugget evaluation and a (mostly) manual nugget evaluation by human assessors. This suggests that our fully automatic evaluation process can be used to guide future iterations of RAG systems.

A Large-Scale Study of Relevance Assessments with Large Language Models: An Initial Look

The application of large language models to provide relevance assessments presents exciting opportunities to advance information retrieval, natural language processing, and beyond, but to date many unknowns remain. This paper reports on the results of a large-scale evaluation (the TREC 2024 RAG Track) where four different relevance assessment approaches were deployed in situ: the "standard" fully manual process that NIST has implemented for decades and three different alternatives that take advantage of LLMs to different extents using the open-source UMBRELA tool. This setup allows us to correlate system rankings induced by the different approaches to characterize tradeoffs between cost and quality. We find that in terms of nDCG@20, nDCG@100, and Recall@100, system rankings induced by automatically generated relevance assessments from UMBRELA correlate highly with those induced by fully manual assessments across a diverse set of 77 runs from 19 teams. Our results suggest that automatically generated UMBRELA judgments can replace fully manual judgments to accurately capture run-level effectiveness. Surprisingly, we find that LLM assistance does not appear to increase correlation with fully manual assessments, suggesting that costs associated with human-in-the-loop processes do not bring obvious tangible benefits. Overall, human assessors appear to be stricter than UMBRELA in applying relevance criteria. Our work validates the use of LLMs in academic TREC-style evaluations and provides the foundation for future studies.

An Early FIRST Reproduction and Improvements to Single-Token Decoding for Fast Listwise Reranking

Recent advances have demonstrated that large language models (LLMs) excel as listwise rerankers, but their high computational demands remain a barrier to widespread adoption. Further, the traditional language modeling (LM) objective is not ideally suited for reranking tasks. FIRST is a novel approach that addresses these challenges by integrating a learning-to-rank objective and leveraging the logits of only the first generated token, thereby significantly reducing inference latency compared to traditional LLM rerankers. In this study, we extend the evaluation of FIRST to the TREC Deep Learning datasets (DL19-22), validating its robustness across diverse domains. We investigate the influence of different first-stage retrievers on FIRST rerankers, observing diminishing returns and patterns consistent with traditional LLM rerankers. Through applying the FIRST objective to a broader range of backbone models, we achieve effectiveness surpassing the original implementation. Our experiments confirm that fast reranking with single-token logits does not compromise out-of-domain reranking quality. To better quantify the computational savings in the original study, we measure and compare latency to find a 21%-42% gain across various models and benchmarks. Moreover, while LM training implicitly improves zero-shot single-token reranking, our experiments also raise questions about whether LM pre-training may hinder subsequent fine-tuning with the FIRST objective. These findings pave the way for more efficient and effective listwise reranking in future applications.

Tomato, Tomahto, Tomate: Measuring the Role of Shared Semantics among Subwords in Multilingual Language Models

Human understanding of language is robust to different word choices as far as they represent similar semantic concepts. To what extent does our human intuition transfer to language models, which represent all subwords as distinct embeddings? In this work, we take an initial step on measuring the role of shared semantics among subwords in the encoder-only multilingual language models (mLMs). To this end, we form "semantic tokens" by merging the semantically similar subwords and their embeddings, and evaluate the updated mLMs on 5 heterogeneous multilingual downstream tasks. Results show that the general shared semantics could get the models a long way in making the predictions on mLMs with different tokenizers and model sizes. Inspections on the grouped subwords show that they exhibit a wide range of semantic similarities, including synonyms and translations across many languages and scripts. Lastly, we found the zero-shot results with semantic tokens are on par or even better than the original models on certain classification tasks, suggesting that the shared subword-level semantics may serve as the anchors for cross-lingual transferring.

MM-Embed: Universal Multimodal Retrieval with Multimodal LLMs

State-of-the-art retrieval models typically address a straightforward search scenario, where retrieval tasks are fixed (e.g., finding a passage to answer a specific question) and only a single modality is supported for both queries and retrieved results. This paper introduces techniques for advancing information retrieval with multimodal large language models (MLLMs), enabling a broader search scenario, termed universal multimodal retrieval, where multiple modalities and diverse retrieval tasks are accommodated. To this end, we first study fine-tuning an MLLM as a bi-encoder retriever on 10 datasets with 16 retrieval tasks. Our empirical results show that the fine-tuned MLLM retriever is capable of understanding challenging queries, composed of both text and image, but underperforms a smaller CLIP retriever in cross-modal retrieval tasks due to modality bias from MLLMs. To address the issue, we propose modality-aware hard negative mining to mitigate the modality bias exhibited by MLLM retrievers. Second, we propose to continually fine-tune the universal multimodal retriever to enhance its text retrieval capability while maintaining multimodal retrieval capability. As a result, our model, MM-Embed, achieves state-of-the-art performance on the multimodal retrieval benchmark M-BEIR, which spans multiple domains and tasks, while also surpassing the state-of-the-art text retrieval model, NV-Embed-v1, on MTEB retrieval benchmark. Finally, we explore to prompt the off-the-shelf MLLMs as the zero-shot rerankers to refine the ranking of the candidates from the multimodal retriever. We find that through prompt-and-reranking, MLLMs can further improve multimodal retrieval when the user queries (e.g., text-image composed queries) are more complex and challenging to understand. These findings also pave the way to advance universal multimodal retrieval in the future.

MIRAGE-Bench: Automatic Multilingual Benchmark Arena for Retrieval-Augmented Generation Systems

Traditional Retrieval-Augmented Generation (RAG) benchmarks rely on different heuristic-based metrics for evaluation, but these require human preferences as ground truth for reference. In contrast, arena-based benchmarks, where two models compete each other, require an expensive Large Language Model (LLM) as a judge for a reliable evaluation. We present an easy and efficient technique to get the best of both worlds. The idea is to train a learning to rank model as a "surrogate" judge using RAG-based evaluation heuristics as input, to produce a synthetic arena-based leaderboard. Using this idea, We develop MIRAGE-Bench, a standardized arena-based multilingual RAG benchmark for 18 diverse languages on Wikipedia. The benchmark is constructed using MIRACL, a retrieval dataset, and extended for multilingual generation evaluation. MIRAGE-Bench evaluates RAG extensively coupling both heuristic features and LLM as a judge evaluator. In our work, we benchmark 19 diverse multilingual-focused LLMs, and achieve a high correlation (Kendall Tau ($\tau$) = 0.909) using our surrogate judge learned using heuristic features with pairwise evaluations and between GPT-4o as a teacher on the MIRAGE-Bench leaderboard using the Bradley-Terry framework. We observe proprietary and large open-source LLMs currently dominate in multilingual RAG. MIRAGE-Bench is available at: https://github.com/vectara/mirage-bench.

Operational Advice for Dense and Sparse Retrievers: HNSW, Flat, or Inverted Indexes?

Practitioners working on dense retrieval today face a bewildering number of choices. Beyond selecting the embedding model, another consequential choice is the actual implementation of nearest-neighbor vector search. While best practices recommend HNSW indexes, flat vector indexes with brute-force search represent another viable option, particularly for smaller corpora and for rapid prototyping. In this paper, we provide experimental results on the BEIR dataset using the open-source Lucene search library that explicate the tradeoffs between HNSW and flat indexes (including quantized variants) from the perspectives of indexing time, query evaluation performance, and retrieval quality. With additional comparisons between dense and sparse retrievers, our results provide guidance for today's search practitioner in understanding the design space of dense and sparse retrievers. To our knowledge, we are the first to provide operational advice supported by empirical experiments in this regard.

ConvKGYarn: Spinning Configurable and Scalable Conversational Knowledge Graph QA datasets with Large Language Models

The rapid advancement of Large Language Models (LLMs) and conversational assistants necessitates dynamic, scalable, and configurable conversational datasets for training and evaluation. These datasets must accommodate diverse user interaction modes, including text and voice, each presenting unique modeling challenges. Knowledge Graphs (KGs), with their structured and evolving nature, offer an ideal foundation for current and precise knowledge. Although human-curated KG-based conversational datasets exist, they struggle to keep pace with the rapidly changing user information needs. We present ConvKGYarn, a scalable method for generating up-to-date and configurable conversational KGQA datasets. Qualitative psychometric analyses confirm our method can generate high-quality datasets rivaling a popular conversational KGQA dataset while offering it at scale and covering a wide range of human-interaction configurations. We showcase its utility by testing LLMs on diverse conversations - exploring model behavior on conversational KGQA sets with different configurations grounded in the same KG fact set. Our results highlight the ability of ConvKGYarn to improve KGQA foundations and evaluate parametric knowledge of LLMs, thus offering a robust solution to the constantly evolving landscape of conversational assistants.

Toward Automatic Relevance Judgment using Vision--Language Models for Image--Text Retrieval Evaluation

Vision--Language Models (VLMs) have demonstrated success across diverse applications, yet their potential to assist in relevance judgments remains uncertain. This paper assesses the relevance estimation capabilities of VLMs, including CLIP, LLaVA, and GPT-4V, within a large-scale \textit{ad hoc} retrieval task tailored for multimedia content creation in a zero-shot fashion. Preliminary experiments reveal the following: (1) Both LLaVA and GPT-4V, encompassing open-source and closed-source visual-instruction-tuned Large Language Models (LLMs), achieve notable Kendall's $\tau \sim 0.4$ when compared to human relevance judgments, surpassing the CLIPScore metric. (2) While CLIPScore is strongly preferred, LLMs are less biased towards CLIP-based retrieval systems. (3) GPT-4V's score distribution aligns more closely with human judgments than other models, achieving a Cohen's $\kappa$ value of around 0.08, which outperforms CLIPScore at approximately -0.096. These findings underscore the potential of LLM-powered VLMs in enhancing relevance judgments.

Systematic Evaluation of Neural Retrieval Models on the Touché 2020 Argument Retrieval Subset of BEIR

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 Touch\'e 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 Touch\'e 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 Touch\'e 2020 data, and we also find that quite a few of the neural models' results are unjudged in the Touch\'e 2020 data. As many of the short Touch\'e passages are not argumentative and thus non-relevant per se, and as the missing judgments complicate fair comparison, we denoise the Touch\'e 2020 data by excluding very short passages (less than 20 words) and by augmenting the unjudged data with post-hoc judgments following the Touch\'e 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 Touch\'e 2020 dataset are available at \url{https://github.com/castorini/touche-error-analysis}.