Traditional retrieval-augmented generation (RAG) benchmarks evaluate systems using heuristic-based metrics, but these require human preferences as the ground truth for reference. In contrast, arena-based benchmarks, where systems compete against each other, require an expensive large language model (LLM) as a judge for a reliable evaluation. We present a simple efficient technique to combine the best of both worlds. The idea is to train a surrogate judge using heuristic metrics as input, to output the LLM as a judge prediction. In our work, we develop MIRAGE-Bench, a synthetic arena-based RAG benchmark for 18 diverse languages on Wikipedia focused on multilingual answer generation evaluation. It extensively couples both heuristic features and LLM as a judge for evaluation. We benchmark 19 multilingual LLMs, and observe a high correlation (Kendall Tau ($\tau$) = 0.909) using our surrogate judge and between GPT-4o as a teacher using the Bradley-Terry framework. Our results show proprietary and large open-source LLMs currently dominate on MIRAGE-Bench. Our code and datasets are made publicly available here: https://github.com/vectara/mirage-bench.
Offline model-based optimization (MBO) aims to identify a design that maximizes a black-box function using only a fixed, pre-collected dataset of designs and their corresponding scores. A common approach in offline MBO is to train a regression-based surrogate model by minimizing mean squared error (MSE) and then find the best design within this surrogate model by different optimizers (e.g., gradient ascent). However, a critical challenge is the risk of out-of-distribution errors, i.e., the surrogate model may typically overestimate the scores and mislead the optimizers into suboptimal regions. Prior works have attempted to address this issue in various ways, such as using regularization techniques and ensemble learning to enhance the robustness of the model, but it still remains. In this paper, we argue that regression models trained with MSE are not well-aligned with the primary goal of offline MBO, which is to select promising designs rather than to predict their scores precisely. Notably, if a surrogate model can maintain the order of candidate designs based on their relative score relationships, it can produce the best designs even without precise predictions. To validate it, we conduct experiments to compare the relationship between the quality of the final designs and MSE, finding that the correlation is really very weak. In contrast, a metric that measures order-maintaining quality shows a significantly stronger correlation. Based on this observation, we propose learning a ranking-based model that leverages learning to rank techniques to prioritize promising designs based on their relative scores. We show that the generalization error on ranking loss can be well bounded. Empirical results across diverse tasks demonstrate the superior performance of our proposed ranking-based models than twenty existing methods.
This paper investigates the design of a unified search engine to serve multiple retrieval-augmented generation (RAG) agents, each with a distinct task, backbone large language model (LLM), and retrieval-augmentation strategy. We introduce an iterative approach where the search engine generates retrieval results for these RAG agents and gathers feedback on the quality of the retrieved documents during an offline phase. This feedback is then used to iteratively optimize the search engine using a novel expectation-maximization algorithm, with the goal of maximizing each agent's utility function. Additionally, we adapt this approach to an online setting, allowing the search engine to refine its behavior based on real-time individual agents feedback to better serve the results for each of them. Experiments on diverse datasets from the Knowledge-Intensive Language Tasks (KILT) benchmark demonstrates that our approach significantly on average outperforms competitive baselines across 18 RAG models. We also demonstrate that our method effectively ``personalizes'' the retrieval process for each RAG agent based on the collected feedback. Finally, we provide a comprehensive ablation study to explore various aspects of our method.
In LLM alignment and many other ML applications, one often faces the Multi-Objective Fine-Tuning (MOFT) problem, i.e., fine-tuning an existing model with datasets labeled w.r.t. different objectives simultaneously. To address the challenge, we propose a Conditioned One-Shot fine-tuning framework (COS-DPO) that extends the Direct Preference Optimization technique, originally developed for efficient LLM alignment with preference data, to accommodate the MOFT settings. By direct conditioning on the weight across auxiliary objectives, our Weight-COS-DPO method enjoys an efficient one-shot training process for profiling the Pareto front and is capable of achieving comprehensive trade-off solutions even in the post-training stage. Based on our theoretical findings on the linear transformation properties of the loss function, we further propose the Temperature-COS-DPO method that augments the temperature parameter to the model input, enhancing the flexibility of post-training control over the trade-offs between the main and auxiliary objectives. We demonstrate the effectiveness and efficiency of the COS-DPO framework through its applications to various tasks, including the Learning-to-Rank (LTR) and LLM alignment tasks, highlighting its viability for large-scale ML deployments.
At Expedia, learning-to-rank (LTR) models plays a key role on our website in sorting and presenting information more relevant to users, such as search filters, property rooms, amenities, and images. A major challenge in deploying these models is ensuring consistent feature scaling between training and production data, as discrepancies can lead to unreliable rankings when deployed. Normalization techniques like feature standardization and batch normalization could address these issues but are impractical in production due to latency impacts and the difficulty of distributed real-time inference. To address consistent feature scaling issue, we introduce a scale-invariant LTR framework which combines a deep and a wide neural network to mathematically guarantee scale-invariance in the model at both training and prediction time. We evaluate our framework in simulated real-world scenarios with injected feature scale issues by perturbing the test set at prediction time, and show that even with inconsistent train-test scaling, using framework achieves better performance than without.
Information Retrieval (IR) systems used in search and recommendation platforms frequently employ Learning-to-Rank (LTR) models to rank items in response to user queries. These models heavily rely on features derived from user interactions, such as clicks and engagement data. This dependence introduces cold start issues for items lacking user engagement and poses challenges in adapting to non-stationary shifts in user behavior over time. We address both challenges holistically as an online learning problem and propose BayesCNS, a Bayesian approach designed to handle cold start and non-stationary distribution shifts in search systems at scale. BayesCNS achieves this by estimating prior distributions for user-item interactions, which are continuously updated with new user interactions gathered online. This online learning procedure is guided by a ranker model, enabling efficient exploration of relevant items using contextual information provided by the ranker. We successfully deployed BayesCNS in a large-scale search system and demonstrated its efficacy through comprehensive offline and online experiments. Notably, an online A/B experiment showed a 10.60% increase in new item interactions and a 1.05% improvement in overall success metrics over the existing production baseline.
Supervisory signals are a critical resource for training learning to rank models. In many real-world search and retrieval scenarios, these signals may not be readily available or could be costly to obtain for some queries. The examples include domains where labeling requires professional expertise, applications with strong privacy constraints, and user engagement information that are too scarce. We refer to these scenarios as sparsely supervised queries which pose significant challenges to traditional learning to rank models. In this work, we address sparsely supervised queries by proposing a novel meta learning to rank framework which leverages fast learning and adaption capability of meta-learning. The proposed approach accounts for the fact that different queries have different optimal parameters for their rankers, in contrast to traditional learning to rank models which only learn a global ranking model applied to all the queries. In consequence, the proposed method would yield significant advantages especially when new queries are of different characteristics with the training queries. Moreover, the proposed meta learning to rank framework is generic and flexible. We conduct a set of comprehensive experiments on both public datasets and a real-world e-commerce dataset. The results demonstrate that the proposed meta-learning approach can significantly enhance the performance of learning to rank models with sparsely labeled queries.
Training Learning-to-Rank models for e-commerce product search ranking can be challenging due to the lack of a gold standard of ranking relevance. In this paper, we decompose ranking relevance into content-based and engagement-based aspects, and we propose to leverage Large Language Models (LLMs) for both label and feature generation in model training, primarily aiming to improve the model's predictive capability for content-based relevance. Additionally, we introduce different sigmoid transformations on the LLM outputs to polarize relevance scores in labeling, enhancing the model's ability to balance content-based and engagement-based relevances and thus prioritize highly relevant items overall. Comprehensive online tests and offline evaluations are also conducted for the proposed design. Our work sheds light on advanced strategies for integrating LLMs into e-commerce product search ranking model training, offering a pathway to more effective and balanced models with improved ranking relevance.
News recommendation is a challenging task that involves personalization based on the interaction history and preferences of each user. Recent works have leveraged the power of pretrained language models (PLMs) to directly rank news items by using inference approaches that predominately fall into three categories: pointwise, pairwise, and listwise learning-to-rank. While pointwise methods offer linear inference complexity, they fail to capture crucial comparative information between items that is more effective for ranking tasks. Conversely, pairwise and listwise approaches excel at incorporating these comparisons but suffer from practical limitations: pairwise approaches are either computationally expensive or lack theoretical guarantees, and listwise methods often perform poorly in practice. In this paper, we propose a novel framework for PLM-based news recommendation that integrates both pointwise relevance prediction and pairwise comparisons in a scalable manner. We present a rigorous theoretical analysis of our framework, establishing conditions under which our approach guarantees improved performance. Extensive experiments show that our approach outperforms the state-of-the-art methods on the MIND and Adressa news recommendation datasets.
Learning to rank (LTR) is widely employed in web searches to prioritize pertinent webpages from retrieved content based on input queries. However, traditional LTR models encounter two principal obstacles that lead to suboptimal performance: (1) the lack of well-annotated query-webpage pairs with ranking scores covering a diverse range of search query popularities, which hampers their ability to address queries across the popularity spectrum, and (2) inadequately trained models that fail to induce generalized representations for LTR, resulting in overfitting. To address these challenges, we propose a \emph{\uline{G}enerative \uline{S}emi-\uline{S}upervised \uline{P}re-trained} (GS2P) LTR model. We conduct extensive offline experiments on both a publicly available dataset and a real-world dataset collected from a large-scale search engine. Furthermore, we deploy GS2P in a large-scale web search engine with realistic traffic, where we observe significant improvements in the real-world application.
Both Transformer and Graph Neural Networks (GNNs) have been employed in the domain of learning to rank (LTR). However, these approaches adhere to two distinct yet complementary problem formulations: ranking score regression based on query-webpage pairs, and link prediction within query-webpage bipartite graphs, respectively. While it is possible to pre-train GNNs or Transformers on source datasets and subsequently fine-tune them on sparsely annotated LTR datasets, the distributional shifts between the pair-based and bipartite graph domains present significant challenges in integrating these heterogeneous models into a unified LTR framework at web scale. To address this, we introduce the novel MPGraf model, which leverages a modular and capsule-based pre-training strategy, aiming to cohesively integrate the regression capabilities of Transformers with the link prediction strengths of GNNs. We conduct extensive offline and online experiments to rigorously evaluate the performance of MPGraf.
Contrastive image-text pre-trained models such as CLIP have shown remarkable adaptability to downstream tasks. However, they face challenges due to the high computational requirements of the Vision Transformer (ViT) backbone. Current strategies to boost ViT efficiency focus on pruning patch tokens but fall short in addressing the multimodal nature of CLIP and identifying the optimal subset of tokens for maximum performance. To address this, we propose greedy search methods to establish a "Golden Ranking" and introduce a lightweight predictor specifically trained to approximate this Ranking. To compensate for any performance degradation resulting from token pruning, we incorporate learnable visual tokens that aid in restoring and potentially enhancing the model's performance. Our work presents a comprehensive and systematic investigation of pruning tokens within the ViT backbone of CLIP models. Through our framework, we successfully reduced 40% of patch tokens in CLIP's ViT while only suffering a minimal average accuracy loss of 0.3 across seven datasets. Our study lays the groundwork for building more computationally efficient multimodal models without sacrificing their performance, addressing a key challenge in the application of advanced vision-language models.
Learning to rank is a rare technology compared with other techniques such as deep neural networks. The number of experts in the field is roughly 1/6 of the number of professionals in deep learning. Being an effective ranking methodology, learning to rank has been widely used in the field of information retrieval. However, in recent years, learning to rank as a recommendation approach has been on decline. In this paper, we take full advantage of order statistic approximation and power law distribution to design a zeroshot listwise learning to rank algorithm for recommendation. We prove in the experiment section that our approach is both accurate and fair.
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.
This paper addresses the problem of ranking Content Providers for Content Recommendation System. Content Providers are the sources of news and other types of content, such as lifestyle, travel, gardening. We propose a framework that leverages explicit user feedback, such as clicks and reactions, and content-based features, such as writing style and frequency of publishing, to rank Content Providers for a given topic. We also use language models to engineer prompts that help us create a ground truth dataset for the previous unsupervised ranking problem. Using this ground truth, we expand with a self-attention based network to train on Learning to Rank ListWise task. We evaluate our framework using online experiments and show that it can improve the quality, credibility, and diversity of the content recommended to users.
Despite the popularity of the two-tower model for unbiased learning to rank (ULTR) tasks, recent work suggests that it suffers from a major limitation that could lead to its collapse in industry applications: the problem of logging policy confounding. Several potential solutions have even been proposed; however, the evaluation of these methods was mostly conducted using semi-synthetic simulation experiments. This paper bridges the gap between theory and practice by investigating the confounding problem on the largest real-world dataset, Baidu-ULTR. Our main contributions are threefold: 1) we show that the conditions for the confounding problem are given on Baidu-ULTR, 2) the confounding problem bears no significant effect on the two-tower model, and 3) we point to a potential mismatch between expert annotations, the golden standard in ULTR, and user click behavior.
Counterfactual learning to rank (CLTR) can be risky and, in various circumstances, can produce sub-optimal models that hurt performance when deployed. Safe CLTR was introduced to mitigate these risks when using inverse propensity scoring to correct for position bias. However, the existing safety measure for CLTR is not applicable to state-of-the-art CLTR methods, cannot handle trust bias, and relies on specific assumptions about user behavior. We propose a novel approach, proximal ranking policy optimization (PRPO), that provides safety in deployment without assumptions about user behavior. PRPO removes incentives for learning ranking behavior that is too dissimilar to a safe ranking model. Thereby, PRPO imposes a limit on how much learned models can degrade performance metrics, without relying on any specific user assumptions. Our experiments show that PRPO provides higher performance than the existing safe inverse propensity scoring approach. PRPO always maintains safety, even in maximally adversarial situations. By avoiding assumptions, PRPO is the first method with unconditional safety in deployment that translates to robust safety for real-world applications.
Automatic image-based severity estimation is an important task in computer-aided diagnosis. Severity estimation by deep learning requires a large amount of training data to achieve a high performance. In general, severity estimation uses training data annotated with discrete (i.e., quantized) severity labels. Annotating discrete labels is often difficult in images with ambiguous severity, and the annotation cost is high. In contrast, relative annotation, in which the severity between a pair of images is compared, can avoid quantizing severity and thus makes it easier. We can estimate relative disease severity using a learning-to-rank framework with relative annotations, but relative annotation has the problem of the enormous number of pairs that can be annotated. Therefore, the selection of appropriate pairs is essential for relative annotation. In this paper, we propose a deep Bayesian active learning-to-rank that automatically selects appropriate pairs for relative annotation. Our method preferentially annotates unlabeled pairs with high learning efficiency from the model uncertainty of the samples. We prove the theoretical basis for adapting Bayesian neural networks to pairwise learning-to-rank and demonstrate the efficiency of our method through experiments on endoscopic images of ulcerative colitis on both private and public datasets. We also show that our method achieves a high performance under conditions of significant class imbalance because it automatically selects samples from the minority classes.
In Large Language Model (LLM) inference, the output length of an LLM request is typically regarded as not known a priori. Consequently, most LLM serving systems employ a simple First-come-first-serve (FCFS) scheduling strategy, leading to Head-Of-Line (HOL) blocking and reduced throughput and service quality. In this paper, we reexamine this assumption -- we show that, although predicting the exact generation length of each request is infeasible, it is possible to predict the relative ranks of output lengths in a batch of requests, using learning to rank. The ranking information offers valuable guidance for scheduling requests. Building on this insight, we develop a novel scheduler for LLM inference and serving that can approximate the shortest-job-first (SJF) schedule better than existing approaches. We integrate this scheduler with the state-of-the-art LLM serving system and show significant performance improvement in several important applications: 2.8x lower latency in chatbot serving and 6.5x higher throughput in synthetic data generation. Our code is available at https://github.com/hao-ai-lab/vllm-ltr.git
Active learning aims to alleviate the amount of labor involved in data labeling by automating the selection of unlabeled samples via an acquisition function. For example, variational adversarial active learning (VAAL) leverages an adversarial network to discriminate unlabeled samples from labeled ones using latent space information. However, VAAL has the following shortcomings: (i) it does not exploit target task information, and (ii) unlabeled data is only used for sample selection rather than model training. To address these limitations, we introduce novel techniques that significantly improve the use of abundant unlabeled data during training and take into account the task information. Concretely, we propose an improved pseudo-labeling algorithm that leverages information from all unlabeled data in a semi-supervised manner, thus allowing a model to explore a richer data space. In addition, we develop a ranking-based loss prediction module that converts predicted relative ranking information into a differentiable ranking loss. This loss can be embedded as a rank variable into the latent space of a variational autoencoder and then trained with a discriminator in an adversarial fashion for sample selection. We demonstrate the superior performance of our approach over the state of the art on various image classification and segmentation benchmark datasets.