Publications

Papers on retrieval-augmented generation, LLM personalization, agentic systems, and long-term memory. For citation counts, see Google Scholar.

2026

  1. Critic-R: Improving Agentic Search using Instruction-tuned Retrievers with Natural Language Introspective Feedback

    Md Zarif Ul Alam, Alireza Salemi, Hamed Zamani

    Findings of EMNLP 2026

    Paper

    Proposes Critic-R, a framework that improves agentic search through retriever optimization. Combines Critic-R-Zero, an inference-time loop that iteratively rewrites queries using introspective feedback, with Critic-Embed, which trains retrievers from automatically supervised refinement trajectories. Improves retrieval quality and answer accuracy on HotpotQA, 2WikiMultihopQA, MuSiQue, and Bamboogle.

  2. GrepSeek: Training Search Agents for Direct Corpus Interaction

    Alireza Salemi, Chang Zeng, Atharva Nijasure, Jui-Hui Chung, Razieh Rahimi, Fernando Diaz, Hamed Zamani

    arXiv

    PaperCode

    Introduces GrepSeek, where LLM agents search corpora by executing shell commands rather than relying on keyword-based retrieval. Uses a two-stage training approach (verified trajectory generation followed by policy optimization) and a parallel execution engine for fast shell-based retrieval, showing strong results across seven QA benchmarks.

  3. Evaluation of Agents under Simulated AI Marketplace Dynamics

    To Eun Kim, Alireza Salemi, Hamed Zamani, Fernando Diaz

    SIGIR 2026

    Paper

    Introduces a simulation-based evaluation framework for information access systems as participants in competitive AI marketplaces. Measures marketplace-level outcomes like retention and market share, capturing early-adoption advantages and competitive dynamics overlooked by static benchmarks.

  4. Learning to Reason for Multi-Step Retrieval of Personal Context in Personalized Question Answering

    Maryam Amirizaniani, Alireza Salemi, Hamed Zamani

    SIGIR 2026

    PaperCode

    Introduces PR2, a reinforcement learning framework that trains LLMs to decide when and what to retrieve from user profiles across multiple reasoning steps for personalized QA. Achieves 8.8–12% relative improvement over existing approaches on LaMP-QA.

  5. Improving User Privacy in Personalized Generation: Client-Side Retrieval-Augmented Modification of Server-Side Generated Speculations

    Alireza Salemi, Hamed Zamani

    arXiv

    PaperCode

    Proposes P³, a system where a large server-side model generates draft tokens from the query alone while a small client-side model refines them using private user profiles. Recovers 90–96% of full-personalization quality while limiting private data exposure.

  6. Beyond a Million Tokens: Benchmarking and Enhancing Long-Term Memory in LLMs

    Mohammad Tavakoli, Alireza Salemi, Carrie Ye, Mohamed Abdalla, Hamed Zamani, J Ross Mitchell

    ICLR 2026

    Paper

    Introduces BEAM, a benchmark with conversations up to 10M tokens, and LIGHT, a cognitive-inspired framework equipping LLMs with episodic, working, and scratchpad memory. Shows even 1M-context models struggle with long dialogues, while LIGHT improves recall by 3.5–12.7%.

  7. Pathways of Thoughts: Multi-Directional Thinking for Long-form Personalized Question Answering

    Alireza Salemi, Cheng Li, Mingyang Zhang, Qiaozhu Mei, Zhuowan Li, Spurthi Amba Hombaiah, Weize Kong, Tao Chen, Hamed Zamani, Michael Bendersky

    WWW 2026

    Paper

    Proposes PoT, an inference-time technique that generates multiple reasoning trajectories by dynamically selecting cognitive operations (reason, revise, personalize, clarify), then combines candidate responses weighted by inferred user preferences. Achieves up to 10.8% relative improvement on LaMP-QA.

  8. LLM-based Multi-Agent Blackboard System for Information Discovery in Data Science

    Alireza Salemi, Mihir Parmar, Palash Goyal, Yiwen Song, Jinsung Yoon, Hamed Zamani, Hamid Palangi, Tomas Pfister

    NeurIPS 2026

    Paper

    Introduces a blackboard-inspired multi-agent architecture where a central agent broadcasts tasks and specialized agents volunteer based on their expertise, eliminating rigid hierarchical coordination. Achieves 13–57% relative improvements on data science benchmarks over the best baseline.

  9. Learning from Natural Language Feedback for Personalized Question Answering

    Alireza Salemi, Hamed Zamani

    SIGIR 2026

    PaperCode

    Proposes VAC, which replaces scalar reward signals with natural language feedback derived from user profiles to train personalized response models via iterative self-training. The final model requires no feedback at inference time and achieves state-of-the-art results on LaMP-QA.

  10. Plan-and-Refine: Diverse and Comprehensive Retrieval-Augmented Generation

    Alireza Salemi, Chris Samarinas, Hamed Zamani

    ICTIR 2026

    PaperCode

    Proposes a two-phase RAG framework that first generates a diverse set of query-aspect plans, then refines candidate responses using those plans and selects the best output via a reward model. Achieves up to 15.4% improvement on TREC datasets.

2025

  1. The Second Tutorial on Retrieval-Enhanced Machine Learning: Synthesis and Opportunities

    Fernando Diaz, Andrew Drozdov, To Eun Kim, Alireza Salemi, Hamed Zamani

    SIGIR 2025

    Paper

    Tutorial synthesizing retrieval-enhanced machine learning paradigms across NLP, vision, and computational biology, presenting a unified formal framework that bridges foundational IR research with modern RAG systems.

  2. CIIR@LiveRAG 2025: Optimizing Multi-Agent Retrieval Augmented Generation through Self-Training

    Alireza Salemi, Mukta Maddipatla, Hamed Zamani

    LiveRAG @ SIGIR 2025

    PaperCode

    Presents mRAG, a multi-agent RAG system with specialized agents for planning, searching, and reasoning, trained via reward-guided self-training over sampled trajectories. Demonstrated strong performance in the SIGIR 2025 LiveRAG competition.

  3. LaMP-QA: A Benchmark for Personalized Long-form Question Answering

    Alireza Salemi, Hamed Zamani

    EMNLP 2025

    Introduces LaMP-QA, a benchmark for personalized long-form QA spanning 45+ subcategories across three broad domains. Shows that systems incorporating personalized context achieve up to 39% performance improvement over non-personalized baselines.

  4. Personalized Generation In Large Model Era: A Survey

    Yiyan Xu, Jinghao Zhang, Alireza Salemi, Xinting Hu, Wenjie Wang, Fuli Feng, Hamed Zamani, Xiangnan He, Tat-Seng Chua

    ACL 2025

    Paper

    The first comprehensive survey on personalized generation with large models, establishing a unified framework and multi-level taxonomy covering technical advances, datasets, and evaluation across modalities and tasks. Discusses open challenges and future directions.

  5. ExPerT: Effective and Explainable Evaluation of Personalized Long-Form Text Generation

    Alireza Salemi, Julian Killingback, Hamed Zamani

    Findings of ACL 2025

    PaperCode

    Introduces ExPerT, an LLM-based evaluation framework that extracts atomic aspects and evidence from generated and reference texts, then measures alignment on content and writing style with detailed explanations. Outperforms existing metrics by 7.2% in alignment with human judgments.

  6. Reasoning-Enhanced Self-Training for Long-Form Personalized Text Generation

    Alireza Salemi, Cheng Li, Mingyang Zhang, Qiaozhu Mei, Weize Kong, Tao Chen, Zhuowan Li, Michael Bendersky, Hamed Zamani

    arXiv

    Paper

    Introduces REST-PG, training LLMs to reason through personal data before generating long-form responses via an EM-based reinforced self-training loop that progressively refines on the model's own high-quality outputs. Achieves 14.5% average relative improvement on LongLaMP.

  7. Beyond Factual Accuracy: Evaluating Coverage of Diverse Factual Information in Long-form Text Generation

    Chris Samarinas, Alexander Krubner, Alireza Salemi, Youngwoo Kim, Hamed Zamani

    Findings of ACL 2025

    PaperCode

    Introduces ICAT, a modular evaluation framework that decomposes LLM-generated text into atomic claims, validates them against retrieved knowledge sources, and measures coverage of expected factual aspects. Shows strong agreement with human evaluations on TREC Web Track data.

  8. Accelerating Retrieval-Augmented Generation

    Derrick Quinn, Mohammad Nouri, Neel Patel, John Salihu, Alireza Salemi, Sukhan Lee, Hamed Zamani, Mohammad Alian

    ASPLOS 2025

    Paper

    Finds that exact nearest-neighbor retrieval, despite higher cost, can reduce overall RAG inference time by improving document quality. Introduces IKS, a near-memory hardware accelerator achieving 13–28× faster exact retrieval on 512GB databases.

  9. Learning to Rank for Multiple Retrieval-Augmented Models through Iterative Utility Maximization

    Alireza Salemi, Hamed Zamani

    ICTIR 2025

    PaperCode

    Proposes a unified search engine for serving multiple heterogeneous RAG models, using iterative expectation-maximization to maximize each agent's utility based on offline document quality feedback. Significantly outperforms baselines across 18 RAG models on KILT.

  10. Comparing Retrieval-Augmentation and Parameter-Efficient Fine-Tuning for Privacy-Preserving Personalization of Large Language Models

    Alireza Salemi, Hamed Zamani

    ICTIR 2025

    PaperCode

    First systematic comparison of RAG vs. PEFT as privacy-preserving personalization strategies across seven LaMP tasks. Finds RAG excels for data-scarce users (+14.9%) while PEFT wins with abundant user data, and their combination reaches +16.0%.

2024

  1. Tutorial on Retrieval-Enhanced Machine Learning: Synthesis and Opportunities

    Fernando Diaz, Andrew Drozdov, To Eun Kim, Alireza Salemi, Hamed Zamani

    SIGIR-AP 2024

    Paper

    Tutorial at SIGIR-AP 2024 presenting a formal unified framework for retrieval-enhanced machine learning, extending RAG paradigms beyond NLP to computer vision, time series prediction, and computational biology.

  2. Retrieval-Enhanced Machine Learning: Synthesis and Opportunities

    To Eun Kim, Alireza Salemi, Andrew Drozdov, Fernando Diaz, Hamed Zamani

    arXiv

    Paper

    Introduces a formal framework for retrieval-enhanced machine learning with consistent notation across ML domains, bridging foundational IR research with contemporary REML literature across NLP, vision, and computational biology.

  3. LongLaMP: A Benchmark for Personalized Long-form Text Generation

    Ishita Kumar, Snigdha Viswanathan, Sushrita Yerra, Alireza Salemi, Ryan Rossi, Franck Dernoncourt, Hanieh Deilamsalehy, Xiang Chen, Ruiyi Zhang, Shubham Agarwal, Nedim Lipka, Hamed Zamani

    arXiv

    Paper

    Introduces LongLaMP, the first benchmark focused on personalized long-form text generation, covering tasks such as email composition and review writing. Demonstrates that personalization consistently benefits long-form generation across all tasks.

  4. LaMP: When Large Language Models Meet Personalization

    Alireza Salemi, Sheshera Mysore, Michael Bendersky, Hamed Zamani

    ACL 2024

    Introduces LaMP, a benchmark of seven personalization tasks (three classification, four generation), and proposes retrieval-augmented methods using term-based, semantic, and time-aware strategies to personalize LLM outputs from user profiles.

  5. Towards a Search Engine for Machines: Unified Ranking for Multiple Retrieval-Augmented Large Language Models

    Alireza Salemi, Hamed Zamani

    SIGIR 2024

    PaperCode

    Introduces uRAG, a unified retrieval engine that serves 36 diverse downstream RAG systems spanning QA, fact verification, entity linking, and relation extraction. Establishes training guidelines for machine-oriented search engines optimized for downstream task utility.

  6. Evaluating Retrieval Quality in Retrieval-Augmented Generation

    Alireza Salemi, Hamed Zamani

    SIGIR 2024Best Short Paper Award

    PaperCode

    Proposes eRAG, which evaluates each retrieved document individually by measuring its contribution to downstream task performance, achieving Kendall's τ improvements of 0.168–0.494 over traditional methods at up to 50× lower GPU cost.

  7. Optimization Methods for Personalizing Large Language Models through Retrieval Augmentation

    Alireza Salemi, Surya Kallumadi, Hamed Zamani

    SIGIR 2024

    PaperCode

    First work to optimize retrieval models that deliver personal documents to LLMs for customized output, developing RL-based and knowledge distillation approaches. Achieves statistically significant improvements on 6 of 7 LaMP benchmark tasks.

2023

  1. Pre-Training Multi-Modal Dense Retrievers for Outside-Knowledge Visual Question Answering

    Alireza Salemi, Mahta Rafiee, Hamed Zamani

    ICTIR 2023

    PaperCode

    Proposes an automatic pipeline for generating pre-training data for outside-knowledge VQA retrieval models. Achieves 26.9% Precision@5 improvement over asymmetric architectures and strong zero-shot generalization with less training data.

  2. A Symmetric Dual Encoding Dense Retrieval Framework for Knowledge-Intensive Visual Question Answering

    Alireza Salemi, Juan Altmayer Pizzorno, Hamed Zamani

    SIGIR 2023Best Student Paper Award

    PaperCode

    Introduces DEDR, a dense retrieval framework for knowledge-intensive VQA that encodes queries and documents into a shared space via uni-modal and multi-modal encoders with iterative distillation. Improves retrieval by 11.6–30.9% and enables state-of-the-art end-to-end VQA via MM-FiD.

  3. PEACH: Pre-Training Sequence-to-Sequence Multilingual Models for Translation with Semi-Supervised Pseudo-Parallel Document Generation

    Alireza Salemi, Amirhossein Abaskohi, Sara Tavakoli, Yadollah Yaghoobzadeh, Azadeh Shakery

    LoResMT 2023

    PaperCode

    Proposes semi-supervised pseudo-parallel document generation for pre-training multilingual seq2seq models in low-resource translation settings, reducing dependence on large parallel corpora.

2021

  1. ARMAN: Pre-training with Semantically Selecting and Reordering of Sentences for Persian Abstractive Summarization

    Alireza Salemi, Emad Kebriaei, Ghazal Neisi Minaei, Azadeh Shakery

    EMNLP 2021

    PaperCode

    Introduces ARMAN, a Transformer encoder-decoder pre-trained with objectives that select semantically salient sentences and reorder them to match human writing conventions. Achieves state-of-the-art results across six Persian summarization benchmarks.

  2. UTNLP at SemEval-2021 Task 5: A Comparative Analysis of Toxic Span Detection using Attention-based, Named Entity Recognition, and Ensemble Models

    Alireza Salemi, Nazanin Sabri, Emad Kebriaei, Behnam Bahrak, Azadeh Shakery

    SemEval 2021

    PaperCode

    Comparative analysis of attention-based, NER-based, and ensemble approaches for toxic span detection at SemEval-2021, evaluating strengths and trade-offs of each paradigm on the shared task dataset.