Hamed Zamani

Md Zarif Ul Alam, Alireza Salemi, Hamed Zamani

Agentic search systems iteratively interact with retrieval models to answer complex queries. Despite substantial progress, optimizing retrievers for agentic search remains challenging, often requiring heavy co-training or gold-standard annotations that limit real-world applicability. We propose Critic-R, a framework that explicitly closes the feedback loop between the reasoning agent and the retrieval model during both inference and training. Critic-R introduces a critic model that evaluates the agent's introspective reasoning trace after consuming retrieved evidence to determine whether the retrieved context sufficiently supports the next reasoning step. Critic-R has two complementary mechanisms: Critic-R-Zero, an inference-time query refinement loop that iteratively rewrites queries and retrieval instructions, and Critic-Embed, an optimization approach for retrieval models that leverages successful and failed refinement trajectories as automatic supervision without requiring manual relevance annotation. We evaluate Critic-R on HotpotQA, 2WikiMultihopQA, MuSiQue, and Bamboogle. Results show that Critic-R significantly improves both retrieval quality and downstream answer accuracy.

Heydar Soudani, Hamed Zamani, Faegheh Hasibi

Retrieval-augmented reasoning (RAR) is a recent evolution of retrieval-augmented generation (RAG) that employs multiple reasoning steps for retrieval and generation. While effective for some complex queries, RAR remains vulnerable to errors and misleading outputs. Uncertainty quantification (UQ) offers methods to estimate the confidence of systems' outputs. These methods, however, often handle simple queries with no retrieval or single-step retrieval, without properly handling RAR setup. Accurate estimation of UQ for RAR requires accounting for all sources of uncertainty, including those arising from retrieval and generation. In this paper, we account for all these sources and introduce Retrieval-Augmented Reasoning Consistency (R2C)--a novel UQ method for RAR. The core idea of R2C is to perturb the multi-step reasoning process by applying various actions to reasoning steps. These perturbations alter the retriever's input, which shifts its output and consequently modifies the generator's input at the next step. Through this iterative feedback loop, the retriever and generator continuously reshape one another's inputs, enabling us to capture uncertainty arising from both components. Experiments on five popular RAR systems across diverse QA datasets show that R2C improves AUROC by over 5% on average compared to the state-of-the-art UQ baselines. Extrinsic evaluations using R2C as an external signal further confirm its effectiveness for two downstream tasks: in Abstention, it achieves ~5% gains in both F1Abstain and AccAbstain; in Model Selection, it improves the exact match by ~7% over single models and ~3% over selection methods.

Alireza Salemi, Chang Zeng, Atharva Nijasure et al.

Large Language Model (LLM) search agents have shown strong promise for knowledge-intensive language tasks through multiple rounds of reasoning and information retrieval. Most existing systems access information using a retriever that takes a keyword or natural language query and returns a ranked list of documents using an index of pre-computed document representations. In this work, we explore a complementary perspective in which the search agent treats the corpus itself as the search environment and finds evidence by issuing executable shell commands. We introduce GrepSeek, an optimized direct corpus interaction (DCI) search agent that trains a compact search agent to find, filter, and compose evidence from large text corpora. To address the instability of learning behavior directly with reinforcement learning on large corpora, we propose a two-stage training pipeline. First, we construct a cold-start dataset using an answer-aware Tutor and answer-blind Planner to generate verified, causally grounded search trajectories. Second, we refine the initialized policy with Group Relative Policy Optimization (GRPO), allowing the agent to improve its task-oriented search behavior through direct interaction with the corpus. To make DCI practical at scale, we further use a semantics-preserving sharded-parallel execution engine that accelerates shell-based retrieval by up to $7.6\times$ while preserving byte-exact equivalence with sequential execution of the shell command. Experiments across seven open-domain question answering benchmarks show that GrepSeek achieves the strongest overall token-level $F_1$ and Exact Match. Our analysis also highlights the limitations of purely lexical interaction on queries with substantial surface-form variation, suggesting DCI as a practical and competitive method for search agents that can complement existing retrieval paradigms in the real world.

Ruicheng Xian, Honglei Zhuang, Zhen Qin et al. · deepmind

Domain adaptation aims to transfer the knowledge learned on (data-rich) source domains to (low-resource) target domains, and a popular method is invariant representation learning, which matches and aligns the data distributions on the feature space. Although this method is studied extensively and applied on classification and regression problems, its adoption on ranking problems is sporadic, and the few existing implementations lack theoretical justifications. This paper revisits invariant representation learning for ranking. Upon reviewing prior work, we found that they implement what we call item-level alignment, which aligns the distributions of the items being ranked from all lists in aggregate but ignores their list structure. However, the list structure should be leveraged, because it is intrinsic to ranking problems where the data and the metrics are defined and computed on lists, not the items by themselves. To close this discrepancy, we propose list-level alignment -- learning domain-invariant representations at the higher level of lists. The benefits are twofold: it leads to the first domain adaptation generalization bound for ranking, in turn providing theoretical support for the proposed method, and it achieves better empirical transfer performance for unsupervised domain adaptation on ranking tasks, including passage reranking.

Sebastian Hofstätter, Nick Craswell, Bhaskar Mitra et al. · microsoft-research

Recently, several dense retrieval (DR) models have demonstrated competitive performance to term-based retrieval that are ubiquitous in search systems. In contrast to term-based matching, DR projects queries and documents into a dense vector space and retrieves results via (approximate) nearest neighbor search. Deploying a new system, such as DR, inevitably involves tradeoffs in aspects of its performance. Established retrieval systems running at scale are usually well understood in terms of effectiveness and costs, such as query latency, indexing throughput, or storage requirements. In this work, we propose a framework with a set of criteria that go beyond simple effectiveness measures to thoroughly compare two retrieval systems with the explicit goal of assessing the readiness of one system to replace the other. This includes careful tradeoff considerations between effectiveness and various cost factors. Furthermore, we describe guardrail criteria, since even a system that is better on average may have systematic failures on a minority of queries. The guardrails check for failures on certain query characteristics and novel failure types that are only possible in dense retrieval systems. We demonstrate our decision framework on a Web ranking scenario. In that scenario, state-of-the-art DR models have surprisingly strong results, not only on average performance but passing an extensive set of guardrail tests, showing robustness on different query characteristics, lexical matching, generalization, and number of regressions. It is impossible to predict whether DR will become ubiquitous in the future, but one way this is possible is through repeated applications of decision processes such as the one presented here.

Sheshera Mysore, Andrew McCallum, Hamed Zamani

Narrative-driven recommendation (NDR) presents an information access problem where users solicit recommendations with verbose descriptions of their preferences and context, for example, travelers soliciting recommendations for points of interest while describing their likes/dislikes and travel circumstances. These requests are increasingly important with the rise of natural language-based conversational interfaces for search and recommendation systems. However, NDR lacks abundant training data for models, and current platforms commonly do not support these requests. Fortunately, classical user-item interaction datasets contain rich textual data, e.g., reviews, which often describe user preferences and context - this may be used to bootstrap training for NDR models. In this work, we explore using large language models (LLMs) for data augmentation to train NDR models. We use LLMs for authoring synthetic narrative queries from user-item interactions with few-shot prompting and train retrieval models for NDR on synthetic queries and user-item interaction data. Our experiments demonstrate that this is an effective strategy for training small-parameter retrieval models that outperform other retrieval and LLM baselines for narrative-driven recommendation.

Alireza Salemi, Sheshera Mysore, Michael Bendersky et al.

This paper highlights the importance of personalization in large language models and introduces the LaMP benchmark -- a novel benchmark for training and evaluating language models for producing personalized outputs. LaMP offers a comprehensive evaluation framework with diverse language tasks and multiple entries for each user profile. It consists of seven personalized tasks, spanning three text classification and four text generation tasks. We additionally propose two retrieval augmentation approaches that retrieve personal items from each user profile for personalizing language model outputs. To this aim, we study various retrieval models, including term matching, semantic matching, and time-aware methods. Extensive experiments on LaMP for zero-shot and fine-tuned language models demonstrate the efficacy of the proposed retrieval augmentation approach and highlight the impact of personalization in various natural language tasks.

Sheshera Mysore, Mahmood Jasim, Andrew McCallum et al.

Methods for making high-quality recommendations often rely on learning latent representations from interaction data. These methods, while performant, do not provide ready mechanisms for users to control the recommendation they receive. Our work tackles this problem by proposing LACE, a novel concept value bottleneck model for controllable text recommendations. LACE represents each user with a succinct set of human-readable concepts through retrieval given user-interacted documents and learns personalized representations of the concepts based on user documents. This concept based user profile is then leveraged to make recommendations. The design of our model affords control over the recommendations through a number of intuitive interactions with a transparent user profile. We first establish the quality of recommendations obtained from LACE in an offline evaluation on three recommendation tasks spanning six datasets in warm-start, cold-start, and zero-shot setups. Next, we validate the controllability of LACE under simulated user interactions. Finally, we implement LACE in an interactive controllable recommender system and conduct a user study to demonstrate that users are able to improve the quality of recommendations they receive through interactions with an editable user profile.

Haoyang Wen, Honglei Zhuang, Hamed Zamani et al. · deepmind

Knowledge-intensive visual question answering requires models to effectively use external knowledge to help answer visual questions. A typical pipeline includes a knowledge retriever and an answer generator. However, a retriever that utilizes local information, such as an image patch, may not provide reliable question-candidate relevance scores. Besides, the two-tower architecture also limits the relevance score modeling of a retriever to select top candidates for answer generator reasoning. In this paper, we introduce an additional module, a multi-modal reranker, to improve the ranking quality of knowledge candidates for answer generation. Our reranking module takes multi-modal information from both candidates and questions and performs cross-item interaction for better relevance score modeling. Experiments on OK-VQA and A-OKVQA show that multi-modal reranker from distant supervision provides consistent improvements. We also find a training-testing discrepancy with reranking in answer generation, where performance improves if training knowledge candidates are similar to or noisier than those used in testing.

Andrew Drozdov, Shufan Wang, Razieh Rahimi et al.

Retrieval-enhanced language models (LMs), which condition their predictions on text retrieved from large external datastores, have recently shown significant perplexity improvements compared to standard LMs. One such approach, the $k$NN-LM, interpolates any existing LM's predictions with the output of a $k$-nearest neighbors model and requires no additional training. In this paper, we explore the importance of lexical and semantic matching in the context of items retrieved by $k$NN-LM. We find two trends: (1) the presence of large overlapping $n$-grams between the datastore and evaluation set plays an important factor in strong performance, even when the datastore is derived from the training data; and (2) the $k$NN-LM is most beneficial when retrieved items have high semantic similarity with the query. Based on our analysis, we define a new formulation of the $k$NN-LM that uses retrieval quality to assign the interpolation coefficient. We empirically measure the effectiveness of our approach on two English language modeling datasets, Wikitext-103 and PG-19. Our re-formulation of the $k$NN-LM is beneficial in both cases, and leads to nearly 4% improvement in perplexity on the Wikitext-103 test set.

Hamed Zamani, Fernando Diaz, Mostafa Dehghani et al.

Although information access systems have long supported people in accomplishing a wide range of tasks, we propose broadening the scope of users of information access systems to include task-driven machines, such as machine learning models. In this way, the core principles of indexing, representation, retrieval, and ranking can be applied and extended to substantially improve model generalization, scalability, robustness, and interpretability. We describe a generic retrieval-enhanced machine learning (REML) framework, which includes a number of existing models as special cases. REML challenges information retrieval conventions, presenting opportunities for novel advances in core areas, including optimization. The REML research agenda lays a foundation for a new style of information access research and paves a path towards advancing machine learning and artificial intelligence.

Sebastian Hofstätter, Jiecao Chen, Karthik Raman et al.

Retrieval-augmented generation models offer many benefits over standalone language models: besides a textual answer to a given query they provide provenance items retrieved from an updateable knowledge base. However, they are also more complex systems and need to handle long inputs. In this work, we introduce FiD-Light to strongly increase the efficiency of the state-of-the-art retrieval-augmented FiD model, while maintaining the same level of effectiveness. Our FiD-Light model constrains the information flow from the encoder (which encodes passages separately) to the decoder (using concatenated encoded representations). Furthermore, we adapt FiD-Light with re-ranking capabilities through textual source pointers, to improve the top-ranked provenance precision. Our experiments on a diverse set of seven knowledge intensive tasks (KILT) show FiD-Light consistently improves the Pareto frontier between query latency and effectiveness. FiD-Light with source pointing sets substantial new state-of-the-art results on six KILT tasks for combined text generation and provenance retrieval evaluation, while maintaining reasonable efficiency.

To Eun Kim, Alireza Salemi, Andrew Drozdov et al. · cmu

In the field of language modeling, models augmented with retrieval components have emerged as a promising solution to address several challenges faced in the natural language processing (NLP) field, including knowledge grounding, interpretability, and scalability. Despite the primary focus on NLP, we posit that the paradigm of retrieval-enhancement can be extended to a broader spectrum of machine learning (ML) such as computer vision, time series prediction, and computational biology. Therefore, this work introduces a formal framework of this paradigm, Retrieval-Enhanced Machine Learning (REML), by synthesizing the literature in various domains in ML with consistent notations which is missing from the current literature. Also, we found that while a number of studies employ retrieval components to augment their models, there is a lack of integration with foundational Information Retrieval (IR) research. We bridge this gap between the seminal IR research and contemporary REML studies by investigating each component that comprises the REML framework. Ultimately, the goal of this work is to equip researchers across various disciplines with a comprehensive, formally structured framework of retrieval-enhanced models, thereby fostering interdisciplinary future research.

Hansi Zeng, Hamed Zamani, Vishwa Vinay

Recent work has shown that more effective dense retrieval models can be obtained by distilling ranking knowledge from an existing base re-ranking model. In this paper, we propose a generic curriculum learning based optimization framework called CL-DRD that controls the difficulty level of training data produced by the re-ranking (teacher) model. CL-DRD iteratively optimizes the dense retrieval (student) model by increasing the difficulty of the knowledge distillation data made available to it. In more detail, we initially provide the student model coarse-grained preference pairs between documents in the teacher's ranking and progressively move towards finer-grained pairwise document ordering requirements. In our experiments, we apply a simple implementation of the CL-DRD framework to enhance two state-of-the-art dense retrieval models. Experiments on three public passage retrieval datasets demonstrate the effectiveness of our proposed framework.

Alireza Salemi, Juan Altmayer Pizzorno, Hamed Zamani

Knowledge-Intensive Visual Question Answering (KI-VQA) refers to answering a question about an image whose answer does not lie in the image. This paper presents a new pipeline for KI-VQA tasks, consisting of a retriever and a reader. First, we introduce DEDR, a symmetric dual encoding dense retrieval framework in which documents and queries are encoded into a shared embedding space using uni-modal (textual) and multi-modal encoders. We introduce an iterative knowledge distillation approach that bridges the gap between the representation spaces in these two encoders. Extensive evaluation on two well-established KI-VQA datasets, i.e., OK-VQA and FVQA, suggests that DEDR outperforms state-of-the-art baselines by 11.6% and 30.9% on OK-VQA and FVQA, respectively. Utilizing the passages retrieved by DEDR, we further introduce MM-FiD, an encoder-decoder multi-modal fusion-in-decoder model, for generating a textual answer for KI-VQA tasks. MM-FiD encodes the question, the image, and each retrieved passage separately and uses all passages jointly in its decoder. Compared to competitive baselines in the literature, this approach leads to 5.5% and 8.5% improvements in terms of question answering accuracy on OK-VQA and FVQA, respectively.

To Eun Kim, Alireza Salemi, Hamed Zamani et al. · cmu

Modern information access ecosystems consist of mixtures of systems, such as retrieval systems and large language models, and increasingly rely on marketplaces to mediate access to models, tools, and data, making competition between systems inherent to deployment. In such settings, outcomes are shaped not only by benchmark quality but also by competitive pressure, including user switching, routing decisions, and operational constraints. Yet evaluation is still largely conducted on static benchmarks with accuracy-focused measures that assume systems operate in isolation. This mismatch makes it difficult to predict post-deployment success and obscures competitive effects such as early-adoption advantages and market dominance. We introduce Marketplace Evaluation, a simulation-based paradigm that evaluates information access systems as participants in a competitive marketplace. By simulating repeated interactions and evolving user and agent preferences, the framework enables longitudinal evaluation and marketplace-level metrics, such as retention and market share, that complement and can extend beyond traditional accuracy-based metrics. We formalize the framework and outline a research agenda, motivated by business and economics, around marketplace simulation, metrics, optimization, and adoption in evaluation campaigns like TREC.

Alireza Salemi, Mahta Rafiee, Hamed Zamani

This paper studies a category of visual question answering tasks, in which accessing external knowledge is necessary for answering the questions. This category is called outside-knowledge visual question answering (OK-VQA). A major step in developing OK-VQA systems is to retrieve relevant documents for the given multi-modal query. Current state-of-the-art asymmetric dense retrieval model for this task uses an architecture with a multi-modal query encoder and a uni-modal document encoder. Such an architecture requires a large amount of training data for effective performance. We propose an automatic data generation pipeline for pre-training passage retrieval models for OK-VQA tasks. The proposed approach leads to 26.9% Precision@5 improvements compared to the current state-of-the-art asymmetric architecture. Additionally, the proposed pre-training approach exhibits a good ability in zero-shot retrieval scenarios.

Sebastian Hofstätter, Jiecao Chen, Karthik Raman et al.

This paper studies multi-task training of retrieval-augmented generation models for knowledge-intensive tasks. We propose to clean the training set by utilizing a distinct property of knowledge-intensive generation: The connection of query-answer pairs to items in the knowledge base. We filter training examples via a threshold of confidence on the relevance labels, whether a pair is answerable by the knowledge base or not. We train a single Fusion-in-Decoder (FiD) generator on seven combined tasks of the KILT benchmark. The experimental results suggest that our simple yet effective approach substantially improves competitive baselines on two strongly imbalanced tasks; and shows either smaller improvements or no significant regression on the remaining tasks. Furthermore, we demonstrate our multi-task training with relevance label sampling scales well with increased model capacity and achieves state-of-the-art results in five out of seven KILT tasks.

Hamed Zamani, Michael Bendersky

Dense retrieval models use bi-encoder network architectures for learning query and document representations. These representations are often in the form of a vector representation and their similarities are often computed using the dot product function. In this paper, we propose a new representation learning framework for dense retrieval. Instead of learning a vector for each query and document, our framework learns a multivariate distribution and uses negative multivariate KL divergence to compute the similarity between distributions. For simplicity and efficiency reasons, we assume that the distributions are multivariate normals and then train large language models to produce mean and variance vectors for these distributions. We provide a theoretical foundation for the proposed framework and show that it can be seamlessly integrated into the existing approximate nearest neighbor algorithms to perform retrieval efficiently. We conduct an extensive suite of experiments on a wide range of datasets, and demonstrate significant improvements compared to competitive dense retrieval models.

Bowen Jin, Hansi Zeng, Zhenrui Yue et al.

Efficiently acquiring external knowledge and up-to-date information is essential for effective reasoning and text generation in large language models (LLMs). Prompting advanced LLMs with reasoning capabilities to use search engines during inference is often suboptimal, as the LLM might not fully possess the capability on how to interact optimally with the search engine. This paper introduces Search-R1, an extension of reinforcement learning (RL) for reasoning frameworks where the LLM learns to autonomously generate (multiple) search queries during step-by-step reasoning with real-time retrieval. Search-R1 optimizes LLM reasoning trajectories with multi-turn search interactions, leveraging retrieved token masking for stable RL training and a simple outcome-based reward function. Experiments on seven question-answering datasets show that Search-R1 improves performance by 41% (Qwen2.5-7B) and 20% (Qwen2.5-3B) over various RAG baselines under the same setting. This paper further provides empirical insights into RL optimization methods, LLM choices, and response length dynamics in retrieval-augmented reasoning. The code and model checkpoints are available at https://github.com/PeterGriffinJin/Search-R1.

Mingchen Li, Hansi Zeng, Zhuo Qian et al.

Retrieval is increasingly moving from one-shot matching toward interactive reasoning, where language agents iteratively inspect evidence, reformulate queries, and search again. Training such agents raises a credit-assignment challenge: executable actions such as queries or summaries can be directly evaluated by the retriever, while latent reasoning steps are not directly observable and only affect future executable actions. This asymmetry makes outcome-level reward assignment unreliable, as the same final reward may credit reasoning steps that did not actually shape retrieval success. We propose RICE-PO, a critic-free policy optimization framework that converts retrieval interactions into localized learning signals. RICE-PO selects high-uncertainty executable actions as anchors, evaluates local counterfactual branches using retrieval metrics, and propagates credit to latent reasoning steps only when reasoning-to-action influence is strong and future residual effects are stable. On BRIGHT and BEIR, RICE-PO consistently outperforms prompt-based agents and group-based RL baselines under the same retriever setting. These results show that the structure of agent-environment interaction itself can provide useful supervision for training reasoning-based retrieval agents.

Alireza Salemi, Hamed Zamani

Despite its substantial impact on various search, recommendation, and question answering tasks, privacy-preserving methods for personalizing large language models (LLMs) have received relatively limited exploration. There is one primary approach in this area through retrieval-augmented generation (RAG), which generates personalized outputs by enriching the input prompt with information retrieved from the user's personal data. This paper studies an orthogonal approach to RAG that involves learning user-dependent LLM parameters through parameter-efficient fine-tuning (PEFT). This paper presents the first systematic study for exploration of PEFT for LLM personalization and provides an extensive comparisons between RAG- and PEFT-based solutions, across a broad set of seven diverse datasets from the LaMP benchmark. Our results demonstrate that, on average, both RAG- and PEFT-based personalization methods yield 14.92% and 1.07% improvements over non-personalized LLMs, respectively. When combining RAG with PEFT, we observe a further improvement of 15.98%, highlighting the effectiveness of their integration in enhancing personalized text generation. Additionally, we identify a positive correlation between the amount of user data available and the effectiveness of PEFT. This finding suggests that RAG is particularly beneficial for cold-start users -- users with limited personal data -- while PEFT performs better when more user-specific data is available.

Yaxin Zhu, Hamed Zamani

This paper focuses on the task of Extreme Multi-Label Classification (XMC) whose goal is to predict multiple labels for each instance from an extremely large label space. While existing research has primarily focused on fully supervised XMC, real-world scenarios often lack supervision signals, highlighting the importance of zero-shot settings. Given the large label space, utilizing in-context learning approaches is not trivial. We address this issue by introducing In-Context Extreme Multilabel Learning (ICXML), a two-stage framework that cuts down the search space by generating a set of candidate labels through incontext learning and then reranks them. Extensive experiments suggest that ICXML advances the state of the art on two diverse public benchmarks.

Mahta Rafiee, Heydar Soudani, Zahra Abbasiantaeb et al.

Deep research agents have emerged as LLM-based systems designed to perform multi-step information seeking and reasoning over large, open-domain sources to answer complex questions by synthesizing information from multiple information sources. Given the complexity of the task and despite various recent efforts, evaluation of deep research agents remains fundamentally challenging. This paper identifies a list of requirements and optional properties for evaluating deep research agents. We observe that existing benchmarks do not satisfy all identified requirements. Inspired by prior research on TREC Total Recall Tracks, we introduce the task of Total Recall Question Answering and develop a framework for deep research agents evaluation that satisfies the identified criteria. Our framework constructs single-answer, total recall queries with precise evaluation and relevance judgments derived from a structured knowledge base paired with a text corpus, enabling large-scale data construction. Using this framework, we build TRQA, a deep research benchmark constructed from Wikidata-Wikipedia as a real-world source and a synthetically generated e-commerce knowledge base and corpus to mitigate the effects of data contamination. We benchmark the collection with representative retriever and deep research models and establish baseline retrieval and end-to-end results for future comparative evaluation.

Rahul Seetharaman, Aman Bansal, Hamed Zamani et al.

Scaling laws are well studied for language models and first-stage retrieval, but not for reranking. We present the first systematic study of scaling laws for cross-encoder rerankers across pointwise, pairwise, and listwise objectives. Across model size and training exposure, ranking quality follows predictable power laws, enabling larger rerankers to be forecast from smaller runs. Using models up to 150M parameters, we forecast 400M and 1B rerankers on MSMARCO-dev and TREC DL. Beyond forecasting, we derive compute-allocation rules from the fitted joint scaling law and compare them with equal-compute checkpoints, showing that retrieval metrics often favor data-heavy scaling, though the recommendation depends on the training objective. The forecasts are accurate and typically conservative, making them useful for planning expensive large-model training. These results provide practical scaling principles for industrial reranking systems, and we will release code and evaluation protocols.

Julian Killingback, Ofer Meshi, Henry Li et al.

Traditional Retrieval-Augmented Generation (RAG) approaches generally assume that retrieval and generation occur on powerful servers removed from the end user. While this reduces local hardware constraints, it introduces significant drawbacks: privacy concerns regarding data access, recurring maintenance and storage costs, increased latency, and the necessity of an internet connection. On-device RAG addresses these challenges by executing the entire pipeline locally, making it ideal for querying sensitive personal information such as financial documents, contact details, and medical history. However, on-device deployment necessitates a delicate balance between limited memory and disk space. Specifically, the context size provided to the generative model must be restricted to manage KV cache and attention memory usage, while the size of stored embeddings must be minimized to preserve disk space. In this work, we propose a unified model that compresses the RAG context and utilizes the same representations for retrieval. This approach minimizes disk utilization compared to using separate representations, while significantly reducing the context size required for generation. With an average of 1/10 of the context, our model matches the performance of a traditional RAG reader without increasing storage requirements compared to a multi-vector retrieval model. This approach represents the first model to unify retrieval and context compression using a shared model and representation. We believe this work will inspire further consolidation of distinct models to optimize on-device performance.

Junda Wang, Zonghai Tao, Hansi Zeng et al.

Complex clinical decision making often fails not because a model lacks facts, but because it cannot reliably select and apply the right procedural knowledge and the right prior example at the right reasoning step. We frame clinical question answering as an agent problem with two explicit, retrievable resources: skills, reusable clinical procedures such as guidelines, protocols, and pharmacologic mechanisms; and experience, verified reasoning trajectories from previously solved cases (e.g., chain-of-thought solutions and their step-level decompositions). At test time, the agent retrieves both relevant skills and experiences from curated libraries and performs lightweight test-time adaptation to align the language model's intermediate reasoning with clinically valid logic. Concretely, we build (i) a skills library from guideline-style documents organized as executable decision rules, (ii) an experience library of exemplar clinical reasoning chains indexed by step-level transitions, and (iii) a step-aware retriever that selects the most useful skill and experience items for the current case. We then adapt the model on the retrieved items to reduce instance-step misalignment and to prevent reasoning from drifting toward unsupported shortcuts. Experiments on medical question-answering benchmarks show consistent gains over strong medical RAG baselines and prompting-only reasoning methods. Our results suggest that explicitly separating and retrieving clinical skills and experience, and then aligning the model at test time, is a practical approach to more reliable medical agents.

Hansi Zeng, Liam Collins, Bhuvesh Kumar et al.

Agentic search -- the task of training agents that iteratively reason, issue queries, and synthesize retrieved information to answer complex questions -- has achieved remarkable progress through reinforcement learning (RL). However, existing approaches such as Search-R1, treat the retrieval system as a fixed tool, optimizing only the reasoning agent while the retrieval component remains unchanged. A preliminary experiment reveals that the gap between an oracle and a fixed retrieval system reaches up to +26.8% relative F1 improvement across seven QA benchmarks, suggesting that the retrieval system is a key bottleneck in scaling agentic search performance. Motivated by this finding, we propose CoSearch, a framework that jointly trains a multi-step reasoning agent and a generative document ranking model via Group Relative Policy Optimization (GRPO). To enable effective GRPO training for the ranker -- whose inputs vary across reasoning trajectories -- we introduce a semantic grouping strategy that clusters sub-queries by token-level similarity, forming valid optimization groups without additional rollouts. We further design a composite reward combining ranking quality signals with trajectory-level outcome feedback, providing the ranker with both immediate and long-term learning signals. Experiments on seven single-hop and multi-hop QA benchmarks demonstrate consistent improvements over strong baselines, with ablation studies validating each design choice. Our results show that joint training of the reasoning agent and retrieval system is both feasible and strongly performant, pointing to a key ingredient for future search agents.

Maryam Amirizaniani, Alireza Salemi, Hamed Zamani

Personalization in Question Answering (QA) requires answers that are both accurate and aligned with users' background, preferences, and historical context. Existing state-of-the-art methods primarily rely on retrieval-augmented generation (RAG) solutions that construct personal context by retrieving relevant items from the user's profile. Existing methods use the user's query directly to retrieve personal documents, and such strategies often lead to surface-level personalization. We propose PR2 (Personalized Retrieval-Augmented Reasoning), a reinforcement learning framework that integrates reasoning and retrieval from personal context for personalization. PR2 learns adaptive retrieval-reasoning policies, determining when to retrieve, what evidence to retrieve from user profiles, and how to incorporate it into intermediate reasoning steps. By optimizing multi-turn reasoning trajectories under a personalized reward function, the framework reinforces reasoning paths that better align with user-specific preferences and contextual signals reflected by the reward model. Extensive experiments on the LaMP-QA benchmark using three LLMs show that PR2 consistently outperforms strong baselines, achieving an average relative improvement of 8.8%-12% in personalized QA.

Mohammad Tavakoli, Alireza Salemi, Carrie Ye et al.

Evaluating the abilities of large language models (LLMs) for tasks that require long-term memory and thus long-context reasoning, for example in conversational settings, is hampered by the existing benchmarks, which often lack narrative coherence, cover narrow domains, and only test simple recall-oriented tasks. This paper introduces a comprehensive solution to these challenges. First, we present a novel framework for automatically generating long (up to 10M tokens), coherent, and topically diverse conversations, accompanied by probing questions targeting a wide range of memory abilities. From this, we construct BEAM, a new benchmark comprising 100 conversations and 2,000 validated questions. Second, to enhance model performance, we propose LIGHT-a framework inspired by human cognition that equips LLMs with three complementary memory systems: a long-term episodic memory, a short-term working memory, and a scratchpad for accumulating salient facts. Our experiments on BEAM reveal that even LLMs with 1M token context windows (with and without retrieval-augmentation) struggle as dialogues lengthen. In contrast, LIGHT consistently improves performance across various models, achieving an average improvement of 3.5%-12.69% over the strongest baselines, depending on the backbone LLM. An ablation study further confirms the contribution of each memory component.

Chris Samarinas, Haw-Shiuan Chang, Hamed Zamani

Training large language models to reason with search engines via reinforcement learning is hindered by a fundamental credit assignment problem: existing methods such as Search-R1 provide only a sparse outcome reward after an entire multi-step trajectory, making it infeasible to attribute success or failure to individual reasoning and retrieval decisions. Process-reward methods like StepSearch alleviate this by introducing step-level supervision, but rely on heuristic rewards such as TF-IDF overlap with gold documents, and still sample k complete trajectories per example, retaining high gradient variance. We propose SLATE, a framework built on two complementary ideas: (1) truncated step-level sampling, which generates k trajectories that share a common prefix and differ only at the next step, and (2) dense LLM-as-judge rewards, which replace heuristic scoring with a capable LLM evaluator that assesses the quality of each reasoning step, search query, and answer, providing richer and more reliable supervision. We theoretically prove that under the same dense reward structure, truncated sampling reduces the variance of advantage estimates by up to a factor of T compared to full-trajectory sampling for T-step trajectories, yielding lower-variance, better-targeted policy gradients. Experiments on seven QA benchmarks confirm that SLATE consistently outperforms both sparse-reward and process-reward baselines, with the largest gains on harder multi-hop tasks and smaller models.

Alireza Salemi, Hamed Zamani

Personalization is essential for question answering systems that are user-centric. Despite its importance, personalization in answer generation has been relatively underexplored. This is mainly due to lack of resources for training and evaluating personalized question answering systems. We address this gap by introducing LaMP-QA -- a benchmark designed for evaluating personalized long-form answer generation. The benchmark covers questions from three major categories: (1) Arts & Entertainment, (2) Lifestyle & Personal Development, and (3) Society & Culture, encompassing over 45 subcategories in total. To assess the quality and potential impact of the LaMP-QA benchmark for personalized question answering, we conduct comprehensive human and automatic evaluations, to compare multiple evaluation strategies for evaluating generated personalized responses and measure their alignment with human preferences. Furthermore, we benchmark a number of non-personalized and personalized approaches based on open-source and proprietary large language models. Our results show that incorporating the personalized context provided leads to up to 39% performance improvements. The benchmark is publicly released to support future research in this area.

Alireza Salemi, Mihir Parmar, Palash Goyal et al.

The rapid advancement of Large Language Models (LLMs) has opened new opportunities in data science, yet their practical deployment is often constrained by the challenge of discovering relevant data within large heterogeneous data lakes. Existing methods struggle with this: single-agent systems are quickly overwhelmed by large, heterogeneous files in the large data lakes, while multi-agent systems designed based on a master-slave paradigm depend on a rigid central controller for task allocation that requires precise knowledge of each sub-agent's capabilities. To address these limitations, we propose a novel multi-agent communication paradigm inspired by the blackboard architecture for traditional AI models. In this framework, a central agent posts requests to a shared blackboard, and autonomous subordinate agents -- either responsible for a partition of the data lake or general information retrieval -- volunteer to respond based on their capabilities. This design improves scalability and flexibility by eliminating the need for a central coordinator to have prior knowledge of all sub-agents' expertise. We evaluate our method on three benchmarks that require explicit data discovery: KramaBench and modified versions of DS-Bench and DA-Code to incorporate data discovery. Experimental results demonstrate that the blackboard architecture substantially outperforms baselines, including RAG and the master-slave multi-agent paradigm, achieving between 13% to 57% relative improvement in end-to-end task success and up to a 9% relative gain in F1 score for data discovery over the best-performing baselines across both proprietary and open-source LLMs. Our findings establish the blackboard paradigm as a scalable and generalizable communication framework for multi-agent systems.

Alireza Salemi, Hamed Zamani

Evaluating retrieval-augmented generation (RAG) presents challenges, particularly for retrieval models within these systems. Traditional end-to-end evaluation methods are computationally expensive. Furthermore, evaluation of the retrieval model's performance based on query-document relevance labels shows a small correlation with the RAG system's downstream performance. We propose a novel evaluation approach, eRAG, where each document in the retrieval list is individually utilized by the large language model within the RAG system. The output generated for each document is then evaluated based on the downstream task ground truth labels. In this manner, the downstream performance for each document serves as its relevance label. We employ various downstream task metrics to obtain document-level annotations and aggregate them using set-based or ranking metrics. Extensive experiments on a wide range of datasets demonstrate that eRAG achieves a higher correlation with downstream RAG performance compared to baseline methods, with improvements in Kendall's $τ$ correlation ranging from 0.168 to 0.494. Additionally, eRAG offers significant computational advantages, improving runtime and consuming up to 50 times less GPU memory than end-to-end evaluation.

Sheshera Mysore, Tim O'Gorman, Andrew McCallum et al.

Query by Example is a well-known information retrieval task in which a document is chosen by the user as the search query and the goal is to retrieve relevant documents from a large collection. However, a document often covers multiple aspects of a topic. To address this scenario we introduce the task of faceted Query by Example in which users can also specify a finer grained aspect in addition to the input query document. We focus on the application of this task in scientific literature search. We envision models which are able to retrieve scientific papers analogous to a query scientific paper along specifically chosen rhetorical structure elements as one solution to this problem. In this work, the rhetorical structure elements, which we refer to as facets, indicate objectives, methods, or results of a scientific paper. We introduce and describe an expert annotated test collection to evaluate models trained to perform this task. Our test collection consists of a diverse set of 50 query documents in English, drawn from computational linguistics and machine learning venues. We carefully follow the annotation guideline used by TREC for depth-k pooling (k = 100 or 250) and the resulting data collection consists of graded relevance scores with high annotation agreement. State of the art models evaluated on our dataset show a significant gap to be closed in further work. Our dataset may be accessed here: https://github.com/iesl/CSFCube

Hamed Zamani, Nick Craswell

Conversational information seeking (CIS) has been recognized as a major emerging research area in information retrieval. Such research will require data and tools, to allow the implementation and study of conversational systems. This paper introduces Macaw, an open-source framework with a modular architecture for CIS research. Macaw supports multi-turn, multi-modal, and mixed-initiative interactions, and enables research for tasks such as document retrieval, question answering, recommendation, and structured data exploration. It has a modular design to encourage the study of new CIS algorithms, which can be evaluated in batch mode. It can also integrate with a user interface, which allows user studies and data collection in an interactive mode, where the back end can be fully algorithmic or a wizard of oz setup. Macaw is distributed under the MIT License.

Alireza Salemi, Surya Kallumadi, Hamed Zamani

This paper studies retrieval-augmented approaches for personalizing large language models (LLMs), which potentially have a substantial impact on various applications and domains. We propose the first attempt to optimize the retrieval models that deliver a limited number of personal documents to large language models for the purpose of personalized generation. We develop two optimization algorithms that solicit feedback from the downstream personalized generation tasks for retrieval optimization -- one based on reinforcement learning whose reward function is defined using any arbitrary metric for personalized generation and another based on knowledge distillation from the downstream LLM to the retrieval model. This paper also introduces a pre- and post-generation retriever selection model that decides what retriever to choose for each LLM input. Extensive experiments on diverse tasks from the language model personalization (LaMP) benchmark reveal statistically significant improvements in six out of seven datasets.

Zhehao Zhang, Ryan A. Rossi, Branislav Kveton et al.

Personalization of Large Language Models (LLMs) has recently become increasingly important with a wide range of applications. Despite the importance and recent progress, most existing works on personalized LLMs have focused either entirely on (a) personalized text generation or (b) leveraging LLMs for personalization-related downstream applications, such as recommendation systems. In this work, we bridge the gap between these two separate main directions for the first time by introducing a taxonomy for personalized LLM usage and summarizing the key differences and challenges. We provide a formalization of the foundations of personalized LLMs that consolidates and expands notions of personalization of LLMs, defining and discussing novel facets of personalization, usage, and desiderata of personalized LLMs. We then unify the literature across these diverse fields and usage scenarios by proposing systematic taxonomies for the granularity of personalization, personalization techniques, datasets, evaluation methods, and applications of personalized LLMs. Finally, we highlight challenges and important open problems that remain to be addressed. By unifying and surveying recent research using the proposed taxonomies, we aim to provide a clear guide to the existing literature and different facets of personalization in LLMs, empowering both researchers and practitioners.

Hamed Zamani, Michael Bendersky

This paper introduces Stochastic RAG--a novel approach for end-to-end optimization of retrieval-augmented generation (RAG) models that relaxes the simplifying assumptions of marginalization and document independence, made in most prior work. Stochastic RAG casts the retrieval process in RAG as a stochastic sampling without replacement process. Through this formulation, we employ straight-through Gumbel-top-k that provides a differentiable approximation for sampling without replacement and enables effective end-to-end optimization for RAG. We conduct extensive experiments on seven diverse datasets on a wide range of tasks, from open-domain question answering to fact verification to slot-filling for relation extraction and to dialogue systems. By applying this optimization method to a recent and effective RAG model, we advance state-of-the-art results on six out of seven datasets.

Hansi Zeng, Chen Luo, Hamed Zamani

This paper introduces PAG-a novel optimization and decoding approach that guides autoregressive generation of document identifiers in generative retrieval models through simultaneous decoding. To this aim, PAG constructs a set-based and sequential identifier for each document. Motivated by the bag-of-words assumption in information retrieval, the set-based identifier is built on lexical tokens. The sequential identifier, on the other hand, is obtained via quantizing relevance-based representations of documents. Extensive experiments on MSMARCO and TREC Deep Learning Track data reveal that PAG outperforms the state-of-the-art generative retrieval model by a large margin (e.g., 15.6% MRR improvements on MS MARCO), while achieving 22x speed up in terms of query latency.

Derrick Quinn, Mohammad Nouri, Neel Patel et al.

An evolving solution to address hallucination and enhance accuracy in large language models (LLMs) is Retrieval-Augmented Generation (RAG), which involves augmenting LLMs with information retrieved from an external knowledge source, such as the web. This paper profiles several RAG execution pipelines and demystifies the complex interplay between their retrieval and generation phases. We demonstrate that while exact retrieval schemes are expensive, they can reduce inference time compared to approximate retrieval variants because an exact retrieval model can send a smaller but more accurate list of documents to the generative model while maintaining the same end-to-end accuracy. This observation motivates the acceleration of the exact nearest neighbor search for RAG. In this work, we design Intelligent Knowledge Store (IKS), a type-2 CXL device that implements a scale-out near-memory acceleration architecture with a novel cache-coherent interface between the host CPU and near-memory accelerators. IKS offers 13.4-27.9x faster exact nearest neighbor search over a 512GB vector database compared with executing the search on Intel Sapphire Rapids CPUs. This higher search performance translates to 1.7-26.3x lower end-to-end inference time for representative RAG applications. IKS is inherently a memory expander; its internal DRAM can be disaggregated and used for other applications running on the server to prevent DRAM, which is the most expensive component in today's servers, from being stranded.

Alireza Salemi, Hamed Zamani

This paper introduces uRAG--a framework with a unified retrieval engine that serves multiple downstream retrieval-augmented generation (RAG) systems. Each RAG system consumes the retrieval results for a unique purpose, such as open-domain question answering, fact verification, entity linking, and relation extraction. We introduce a generic training guideline that standardizes the communication between the search engine and the downstream RAG systems that engage in optimizing the retrieval model. This lays the groundwork for us to build a large-scale experimentation ecosystem consisting of 18 RAG systems that engage in training and 18 unknown RAG systems that use the uRAG as the new users of the search engine. Using this experimentation ecosystem, we answer a number of fundamental research questions that improve our understanding of promises and challenges in developing search engines for machines.

Alireza Salemi, Cheng Li, Mingyang Zhang et al.

Personalized text generation requires a unique ability of large language models (LLMs) to learn from context that they often do not encounter during their standard training. One way to encourage LLMs to better use personalized context for generating outputs that better align with the user's expectations is to instruct them to reason over the user's past preferences, background knowledge, or writing style. To achieve this, we propose Reasoning-Enhanced Self-Training for Personalized Text Generation (REST-PG), a framework that trains LLMs to reason over personal data during response generation. REST-PG first generates reasoning paths to train the LLM's reasoning abilities and then employs Expectation-Maximization Reinforced Self-Training to iteratively train the LLM based on its own high-reward outputs. We evaluate REST-PG on the LongLaMP benchmark, consisting of four diverse personalized long-form text generation tasks. Our experiments demonstrate that REST-PG achieves significant improvements over state-of-the-art baselines, with an average relative performance gain of 14.5% on the benchmark.

Alireza Salemi, Julian Killingback, Hamed Zamani

Evaluating personalized text generated by large language models (LLMs) is challenging, as only the LLM user, i.e., prompt author, can reliably assess the output, but re-engaging the same individuals across studies is infeasible. This paper addresses the challenge of evaluating personalized text generation by introducing ExPerT, an explainable reference-based evaluation framework. ExPerT leverages an LLM to extract atomic aspects and their evidence from the generated and reference texts, match the aspects, and evaluate their alignment based on content and writing style -- two key attributes in personalized text generation. Additionally, ExPerT generates detailed, fine-grained explanations for every step of the evaluation process, enhancing transparency and interpretability. Our experiments demonstrate that ExPerT achieves a 7.2% relative improvement in alignment with human judgments compared to the state-of-the-art text generation evaluation methods. Furthermore, human evaluators rated the usability of ExPerT's explanations at 4.7 out of 5, highlighting its effectiveness in making evaluation decisions more interpretable.

Alireza Salemi, Hamed Zamani

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 RAG strategy. We introduce an iterative approach where the search engine generates retrieval results for the 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 an expectation-maximization algorithm, with the goal of maximizing each agent's utility function. Additionally, we adapt this 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 datasets from the Knowledge-Intensive Language Tasks (KILT) benchmark demonstrates that our approach significantly on average outperforms baselines across 18 RAG models. We demonstrate that our method effectively ``personalizes'' the retrieval for each RAG agent based on the collected feedback. Finally, we provide a comprehensive ablation study to explore various aspects of our method.

Chris Samarinas, Alexander Krubner, Alireza Salemi et al.

This paper presents ICAT, an evaluation framework for measuring coverage of diverse factual information in long-form text generation. ICAT breaks down a long output text into a list of atomic claims and not only verifies each claim through retrieval from a (reliable) knowledge source, but also computes the alignment between the atomic factual claims and various aspects expected to be presented in the output. We study three implementations of the ICAT framework, each with a different assumption on the availability of aspects and alignment method. By adopting data from the diversification task in the TREC Web Track and the ClueWeb corpus, we evaluate the ICAT framework. We demonstrate strong correlation with human judgments and provide comprehensive evaluation across multiple state-of-the-art LLMs. Our framework further offers interpretable and fine-grained analysis of diversity and coverage. Its modular design allows for easy adaptation to different domains and datasets, making it a valuable tool for evaluating the qualitative aspects of long-form responses produced by LLMs.

Julian Killingback, Hansi Zeng, Hamed Zamani

Existing information retrieval systems are largely constrained by their reliance on vector inner products to assess query-document relevance, which naturally limits the expressiveness of the relevance score they can produce. We propose a new paradigm; instead of representing a query as a vector, we use a small neural network that acts as a learned query-specific relevance function. This small neural network takes a document representation as input (in this work we use a single vector) and produces a scalar relevance score. To produce the small neural network we use a hypernetwork, a network that produces the weights of other networks, as our query encoder. We name this category of encoder models Hypencoders. Experiments on in-domain search tasks show that Hypencoders significantly outperform strong dense retrieval models and even surpass reranking models and retrieval models with an order of magnitude more parameters. To assess the extent of Hypencoders' capabilities, we evaluate on a set of hard retrieval tasks including tip-of-the-tongue and instruction-following retrieval tasks. On harder tasks, we find that the performance gap widens substantially compared to standard retrieval tasks. Furthermore, to demonstrate the practicality of our method, we implement an approximate search algorithm and show that our model is able to retrieve from a corpus of 8.8M documents in under 60 milliseconds.

Chris Samarinas, Hamed Zamani

We present a novel approach for training small language models for reasoning-intensive document ranking that combines knowledge distillation with reinforcement learning optimization. While existing methods often rely on expensive human annotations or large black-box language models, our methodology leverages web data and a teacher LLM to automatically generate high-quality training examples with relevance explanations. By framing document ranking as a reinforcement learning problem and incentivizing explicit reasoning capabilities, we train a compact 3B parameter language model that achieves state-of-the-art performance on the BRIGHT benchmark. Our model ranks third on the leaderboard while using substantially fewer parameters than other approaches, outperforming models that are over 20 times larger. Through extensive experiments, we demonstrate that generating explanations during inference, rather than directly predicting relevance scores, enables more effective reasoning with smaller language models. The self-supervised nature of our method offers a scalable and interpretable solution for modern information retrieval systems.

James Allan, Eunsol Choi, Daniel P. Lopresti et al.

In the fast-evolving field of information retrieval (IR), the integration of generative AI technologies such as large language models (LLMs) is transforming how users search for and interact with information. Recognizing this paradigm shift at the intersection of IR and generative AI (IR-GenAI), a visioning workshop supported by the Computing Community Consortium (CCC) was held in July 2024 to discuss the future of IR in the age of generative AI. This workshop convened 44 experts in information retrieval, natural language processing, human-computer interaction, and artificial intelligence from academia, industry, and government to explore how generative AI can enhance IR and vice versa, and to identify the major challenges and opportunities in this rapidly advancing field. This report contains a summary of discussions as potentially important research topics and contains a list of recommendations for academics, industry practitioners, institutions, evaluation campaigns, and funding agencies.

Alireza Salemi, Hamed Zamani

Personalization is crucial for enhancing both the effectiveness and user satisfaction of language technologies, particularly in information-seeking tasks like question answering. Current approaches for personalizing large language models (LLMs) often rely on retrieval-augmented generation (RAG), followed by reinforcement learning with scalar reward signals to teach models how to use retrieved personal context. We believe that these scalar rewards sometimes provide weak, non-instructive feedback, limiting learning efficiency and personalization quality. We introduce VAC, a novel framework for personalized response generation that replaces scalar rewards with natural language feedback (NLF) that are generated conditioned on the user profiles and the question narratives. NLF serves as a rich and actionable supervision signal, allowing the policy model to iteratively refine its outputs and internalize effective personalization strategies. Training alternates between optimizing the feedback model and fine-tuning the policy model on the improved responses, resulting in a policy model that no longer requires feedback at inference. Evaluation on the LaMP-QA benchmark that consists of three diverse domains demonstrates consistent and significant improvements over the state-of-the-art results. Human evaluations further confirm the superior quality of the generated responses. These results demonstrate that NLF provides more effective signals for optimizing personalized question answering.