Searcharxiv⌕ Search

arXiv subjects

Shengyao Zhuang

Publications and source records attributed to Shengyao Zhuang.

At least 19 recordsLinked to original sources

Tevatron Meets Megatron: Expert-Parallel LLM Reranker Training on an Academic Budget

Training large dense and mixture-of-experts (MoE) rerankers is constrained by GPU memory and distributed-execution efficiency, which has kept academic reranker research largely on dense models. Tevatron trains rerankers through the Hugging Face Trainer with DeepSpeed and PyTorch FSDP1, a stack that stalls on three fronts: DeepSpeed ZeRO-2 has a known gradient-gathering failure on this path, FSDP1 full sharding is memory-safe but throughput-bound, and neither offers the expert parallelism (EP) that MoE training needs. We present Tevatron 3.0, which integrates Megatron-Core as a reranker training backend while preserving Tevatron's data processing, evaluation pipeline, and Hugging Face checkpoint format. For full-parameter dense 8B training, Megatron matches FSDP1 on BEIR-15 NDCG@10 with approximately 4% lower epoch time and twice the peak GPU memory under matched data-parallel and recomputation settings, and approximately 22% lower epoch time in its recommended configuration; it supports both LoRA and full-parameter fine-tuning. Expert parallelism makes it possible to train a 30B-parameter MoE reranker (Qwen3-30B-A3B) on two nodes. A controlled study of architecture, LoRA, and distillation on BEIR-15, with three first-stage retrievers and Hugging Face and vLLM serving, shows that the MoE reranker matches dense-8B quality at less than half the activated parameters and at higher inference throughput. We open-source the framework and trained checkpoints.

cs.IR↗

Tevatron-Elastic: A Unified Abstraction for Training Elastic Retrievers and Rerankers

A single model scale challenges the flexibility of a production retrieval system: some settings need it faster, others need a smaller index, and the right trade-off changes with the workload. In the context of information retrieval (IR), a transformer-based model can be made smaller in three ways --- using fewer layers, passing fewer tokens through the upper layers, or producing a shorter embedding --- and each way saves a different resource. These options have been studied in isolation, each as its own method with its own code and training setup, which makes them hard to combine or adapt to a new model. In this work, we present Tevatron-Elastic to bring all three under one unified abstraction: a single object names any size the model can run at, and a short schedule lists the sizes to train. Training then produces one checkpoint that serves all of those sizes, and at deployment the user has the liberty to pick any of them. The same abstraction covers both retrievers and rerankers and both encoder and decoder models, as it works through interfaces that Hugging Face transformers already expose; adding a new backbone can be a configuration change rather than a model rewrite. Prior methods --- Matryoshka embeddings, early exit, 2D Matryoshka (Starbucks), and layerwise token compression (LTC) --- become special cases of our unified abstraction. With Tevatron-Elastic, we also propose a new elastic algorithm Matryoshka~LTC (MLTC) to jointly train several token-compression ratios in one retriever checkpoint. To validate our framework, we train 20 checkpoints across three backbones and two tasks: the quality curves are smooth, full-point quality remains close to that of models trained for a single size, and a controlled study confirms the wall-clock speedups. We release the framework and all checkpoints to support elastic retrieval research.

cs.CL↗

Beyond Ranking: Are Decision Models Calibrated Relevance Scorers?

Decision models such as Jev are a new kind of relevance scorer. Their typed score primitive takes a question, an ordered set of textual relevance levels and a state, and returns a probability for each level directly, rather than token likelihoods or generated labels. Jev's vendor promises that these probabilities are calibrated. This motivates revisiting 'Beyond Yes and No', which improved zero-shot pointwise LLM reranking using likelihood-derived scores over fine-grained relevance labels but left calibration against human assessors as future work. In this paper, we first reproduce its open-model FLAN-UL2 setting on eight BEIR collections, confirming its ranking trends, and then run Jev and two open-weight decision models with the same protocol. Jev and the 27B pplx-decider rerank ahead of FLAN-UL2 and three cross-encoders on average. We then compare Jev's native probabilities with the human assessors' labels and find them overconfident: of the DL21 pairs it scores above 0.9, only 66% are relevant. Since this is mostly an error of scale, a two-parameter calibration fitted on 2,257 DL19 pairs removes much of it and transfers to DL21. After calibration, Jev has a lower DL21 Brier score than FLAN-UL2 and the three cross-encoders, mainly because it separates relevant from non-relevant pairs better. Finally, with calibrated probabilities, choosing each query's cut-off by expected F1 needs no parameter of its own, beats a fixed top-10 cut and matches a cut-off threshold tuned on DL19.

cs.IR↗

ITER: Interaction-Aware Retrieval for Agentic Search

Deep-research agents answer complex user questions through an iterative sequence of search steps, where the agent autonomously formulates sub-queries to retrieve the evidence needed at each stage. However, existing retriever training typically relies only on the sub-query and its corresponding search results at the current step as training signals, leaving the information accumulated from previous interactions largely underutilized. We introduce ITER, an agent interaction-aware dense retriever trained using agent trajectory learning signals. ITER represents each query by incorporating not only the current sub-query, but also the main question, the agent's pre-search reasoning, and preceding sub-queries, and is trained using trajectory-relative learning signals derived from the agent's interactions. Across six agent backbones from three model families, ITER consistently outperforms the existing agent-trajectory-trained dense retriever, LRAT, achieving an average relative improvement of 6.9% on InfoSeek-Eval and 15.4% on BrowseComp-Plus. At the matched 4B scale, ITER outperforms AgentIR on InfoSeek-Eval for five of six backbones while achieving a higher visit-to-search recall ratio on BrowseComp-Plus across all six backbones. Ablations further show that structured interaction history and pre-search reasoning provide complementary retrieval context, while previously visited and useful documents, used as redundancy negatives in subsequent searches, provide the strongest trajectory-relative supervision.

cs.IR↗

Search, Inspect, Fetch: Exploiting Structure-Aware Boolean Retrieval for Deep-Search Agents

Existing deep-search agents use a Search-Visit workflow that retrieves whole webpages without considering the structure they expose through titles, headings, sections, and metadata. This prevents agents from directly constraining retrieval to parts of a webpage and often carries irrelevant content into their context. We introduce Sieve, a search-inspect-fetch strategy driven by a Boolean Query Language (BQL): it searches webpage fields to filter candidates, uses an interchangeable ranker to order them, presents structure-rich result cards for inspection, and fetches only selected sections. Across three QA collections, Sieve is more accurate than the strongest conventional Search-Visit configuration on each collection while using 20.7-50.6% fewer tokens. Boolean filtering improves every tested ranker, and the accuracy-context advantage persists across retriever choices and agent backbones. Our implementation is included in the SkimSearchAgent library https://github.com/ielab/skim-search-agent.

cs.IR↗

LACONIC: Dense-Level Effectiveness for Scalable Sparse Retrieval via a Two-Phase Training Curriculum

While dense retrieval models have been the standard for state-of-the-art information retrieval, their deployment is often constrained by high memory requirements and reliance on GPU accelerators for vector similarity search at scale. Learned sparse retrieval offers a compelling alternative by enabling efficient search via inverted indices, yet it has historically received less attention than dense approaches. In this paper, we introduce LACONIC, a family of learned sparse retrievers based on the Llama3 architecture (1B, 3B, and 8B). We propose a streamlined two-phase training curriculum consisting of (1) weakly supervised pre-finetuning to adapt causal LLMs for bidirectional contextualization and (2) high-signal finetuning using curated hard negatives. Our results demonstrate that LACONIC effectively bridges the performance gap with dense models: the 8B variant achieves a state-of-the-art 60.2 nDCG@10 on the MTEB Retrieval benchmark, ranking 15th on the leaderboard as of February 5th, 2026, while utilizing 74% less index memory than an equivalent dense model. By delivering high retrieval effectiveness on commodity CPU hardware with a fraction of the compute budget required by competing models, LACONIC provides a scalable and efficient solution for real-world search applications. We fully open source our code implementation and trained checkpoints to facilitate reproducibility.

cs.IR↗

Rethinking On-policy Optimization for Query Augmentation

Recent advances in large language models (LLMs) have led to a surge of interest in query augmentation for information retrieval (IR). Two main approaches have emerged. The first prompts LLMs to generate answers or pseudo-documents that serve as new queries, relying purely on the model's parametric knowledge or contextual information. The second applies reinforcement learning (RL) to fine-tune LLMs for query rewriting, directly optimizing retrieval metrics. While having respective advantages and limitations, the two approaches have not been compared under consistent experimental conditions. In this work, we present the first systematic comparison of prompting-based and RL-based query augmentation across diverse benchmarks, including evidence-seeking, ad hoc, and tool retrieval. Our key finding is that under a compute-aware comparison setting, simple, training-free query augmentation often performs on par with, or even surpasses, more expensive RL-based counterparts, especially when using powerful LLMs. Motivated by this discovery, we introduce a novel hybrid method, On-policy Pseudo-document Query Expansion (OPQE), in which the LLM policy learns to generate a pseudo-document that maximizes retrieval performance, rather than rewriting the query, thus merging the flexibility and generative structure of prompting with the targeted optimization of RL. We show OPQE outperforms both standalone prompting and RL-based rewriting, demonstrating that a synergistic approach yields the best results. We open source our implementation to facilitate reproducibility.

cs.CL↗

Improving Long-Context Retrieval with Multi-Prefix Embedding

Long-context retrieval exposes a tension: single-vector embeddings lose fine-grained detail, while token-level multi-vector methods incur prohibitive storage. We propose Multi-Prefix Embedding (MPE), which partitions a document into chunks separated by EOS tokens, encodes the full sequence in a single causal forward pass, and extracts one embedding at each prefix boundary. MPE retains cross-chunk context, enables chunk-level MaxSim matching, and trains with only document-level relevance labels. Experiments on MLDR-en, BrowseComp-Plus, and LongEmbed show that MPE is competitive with or outperforms single-vector, independent-chunk, and multi-vector baselines, while providing a natural source attribution mechanism for locating evidence chunks.

cs.IR↗

Towards Retrieving Interaction Spaces for Agentic Search

Retrieval for search agents is still inherited from non-agentic information retrieval: a retriever ranks the corpus and the agent reads a small set of returned documents. Recent direct corpus interaction (DCI) work shows that agents can instead interact with the raw corpus through shell tools such as grep and file reads. But unbounded interaction does not scale: every broad shell command is a scan over the whole corpus, and latency degrades sharply as the corpus grows. We argue that the role of retrieval for agentic search is not just to select documents that fit in the LLM context window, but to construct an interaction space: a bounded subset of the corpus the agent can explore with associated tools. Two design consequences follow. The space needs a boundary supplied by retrieval, and the objects within it should be processed for interaction. As a proof of concept, we propose RISE (Retrieving Interaction SpacE): we use BM25 to construct the interaction space; meanwhile, its documents are processed during indexing for shell-style navigation. On BrowseComp-Plus, RISE matches the pure-shell DCI baseline at 78% accuracy with gpt-5.4-mini at roughly one quarter of the per-query cost. At 1M documents, RISE-BM25 reaches 81% on gpt-5.4-mini, whereas DCI on gpt-5.4-nano degrades to 60% with 33 of 100 wall-clock failures.

cs.IR↗

DiffRetriever: Parallel Representative Tokens for Retrieval with Diffusion Language Models

This paper shows how diffusion language models (DLMs) can be used as effective and efficient retrievers. Existing DLM-based retrievers (e.g., DiffEmbed) follow BERT-style encoding, representing each query or passage as a single mean-pooled vector. This ignores how DLMs are trained to generate responses through masked-position prediction under bidirectional attention, a capability that can provide stronger retrieval signals. We propose DiffRetriever, which uses the DLM's native masked-position prediction directly for retrieval. For each query or passage, DiffRetriever appends one or more masked positions, using the outputs as retrieval representations in a single forward pass. With one masked position, single-representation DiffRetriever already improves over DiffEmbed on the same backbones. DiffRetriever also naturally extends to multi-representation retrieval: DLMs process multiple masked positions jointly, enabling ColBERT-style fine-grained matching with little additional encoding latency. In autoregressive LLM retrievers, the same multi-representation strategy requires sequential decoding and therefore incurs much higher latency. DiffRetriever obtains the strongest aggregate effectiveness within our matched comparison, outperforming DiffEmbed, PromptReps, and RepLLaMA. Masked-position counts selected on training data transfer well across datasets, while per-query variation suggests headroom for adaptive allocation. Code is available at https://github.com/ielab/diffretriever.

cs.IR↗

Layer-wise Token Compression for Efficient Document Reranking

Transformer-based document cross-encoder rerankers are a central component of modern information retrieval systems. Despite their success, these models suffer from high computational costs due to processing long query-document sequences at inference time. A known approach to improve efficiency is token compression, which consists of aggregating groups of tokens together in the initial embedding layer, reducing the effective number of tokens, and making the computation faster. While token compression has proven to be successful for bi-encoder retrievers, we empirically observed that this approach may be ineffective for cross-encoder rerankers. In this paper, we propose Layer-wise Token Compression (LTC), which applies adaptive token pooling at intermediate transformer layers. Through extensive ablation studies on MS MARCO passage and document ranking tasks, we demonstrate that compression at middle layers preserves ranking quality while increasing inference QPS by up to 25% for passage ranking and up to 116% for document ranking. We also extend LTC to listwise LLM rerankers and show that the same approach can be easily applied to long-context listwise reranking, where the QPS improvements are even greater. More surprisingly, when applying rerankers trained on short passages to long-document ranking tasks, models trained with compression outperform their uncompressed counterparts, suggesting that compression may act as a beneficial regularizer that encourages length-invariant representations.

cs.IR↗

Where Relevance Emerges: A Layer-Wise Study of Internal Attention for Zero-Shot Re-Ranking

Zero-shot document re-ranking with Large Language Models (LLMs) has evolved from Pointwise methods to Listwise and Setwise approaches that optimize computational efficiency. Despite their success, these methods predominantly rely on generative scoring or output logits, which face bottlenecks in inference latency and result consistency. In-Context Re-ranking (ICR) has recently been proposed as an O(1) alternative method. ICR extracts internal attention signals directly, avoiding the overhead of text generation. However, existing ICR methods simply aggregate signals across all layers; layer-wise contributions and their consistency across architectures have been left unexplored. Furthermore, no unified study has compared internal attention with traditional generative and likelihood-based mechanisms across diverse ranking frameworks under consistent conditions. In this paper, we conduct an orthogonal evaluation of generation, likelihood, and internal attention mechanisms across multiple ranking frameworks. We further identify a universal "bell-curve" distribution of relevance signals across transformer layers, which motivates the proposed Selective-ICR strategy that reduces inference latency by 30%-50% without compromising effectiveness. Finally, evaluation on the reasoning-intensive BRIGHT benchmark shows that precisely capturing high-quality in-context attention signals fundamentally reduces the need for model scaling and reinforcement learning: a zero-shot 8B model matches the performance of 14B reinforcement-learned re-rankers, while even a 0.6B model outperforms state-of-the-art generation-based approaches. These findings redefine the efficiency-effectiveness frontier for LLM-based re-ranking and highlight the latent potential of internal signals for complex reasoning ranking tasks. Our code and results are publicly available at https://github.com/ielab/Selective-ICR.

cs.IR↗

AgentIR: Reasoning-Aware Retrieval for Deep Research Agents

Deep Research agents are rapidly emerging as primary consumers of modern retrieval systems. Unlike human users who issue and refine queries without documenting their intermediate thought processes, Deep Research agents generate explicit natural language reasoning before each search call, revealing rich intent and contextual information that existing retrievers entirely ignore. To exploit this overlooked signal, we introduce: (1) Reasoning-Aware Retrieval, a retrieval paradigm that jointly embeds the agent's reasoning trace alongside its query; and (2) DR-Synth, a data synthesis method that generates Deep Research retriever training data from standard QA datasets. We demonstrate that both components are independently effective, and their combination yields a trained embedding model, AgentIR-4B, with substantial gains. On the challenging BrowseComp-Plus benchmark, AgentIR-4B achieves 68\% accuracy with the open-weight agent Tongyi-DeepResearch, compared to 50\% with conventional embedding models twice its size, and 37\% with BM25. Code and data are available at: https://texttron.github.io/AgentIR/.

cs.CL↗

Leveraging Reference Documents for Zero-Shot Ranking via Large Language Models

Large Language Models (LLMs) have demonstrated exceptional performance in the task of text ranking for information retrieval. While Pointwise ranking approaches offer computational efficiency by scoring documents independently, they often yield biased relevance estimates due to the lack of inter-document comparisons. In contrast, Pairwise methods improve ranking accuracy by explicitly comparing document pairs, but suffer from substantial computational overhead with quadratic complexity ($O(n^2)$). To address this tradeoff, we propose \textbf{RefRank}, a simple and effective comparative ranking method based on a fixed reference document. Instead of comparing all document pairs, RefRank prompts the LLM to evaluate each candidate relative to a shared reference anchor. By selecting the reference anchor that encapsulates the core query intent, RefRank implicitly captures relevance cues, enabling indirect comparison between documents via this common anchor. This reduces computational cost to linear time ($O(n)$) while importantly, preserving the advantages of comparative evaluation. To further enhance robustness, we aggregate multiple RefRank outputs using a weighted averaging scheme across different reference choices. Experiments on several benchmark datasets and with various LLMs show that RefRank significantly outperforms Pointwise baselines and could achieve performance at least on par with Pairwise approaches with a significantly lower computational cost.

cs.IR↗

Distillation versus Contrastive Learning: How to Train Your Rerankers

Training effective text rerankers is crucial for information retrieval. Two strategies are widely used: contrastive learning (optimizing directly on ground-truth labels) and knowledge distillation (transferring knowledge from a larger reranker). While both have been studied extensively, a clear comparison of their effectiveness for training cross-encoder rerankers under practical conditions is needed. This paper empirically compares these strategies by training rerankers of different sizes (0.5B, 1.5B, 3B, 7B) and architectures (Transformer, Recurrent) using both methods on the same data, with a strong contrastive learning model acting as the distillation teacher. Our results show that knowledge distillation generally yields better in-domain and out-of-domain ranking performance than contrastive learning when distilling from a more performant teacher model. This finding is consistent across student model sizes and architectures. However, distilling from a teacher of the same capacity does not provide the same advantage, particularly for out-of-domain tasks. These findings offer practical guidance for choosing a training strategy based on available teacher models. We recommend using knowledge distillation to train smaller rerankers if a larger, more performant teacher is accessible; in its absence, contrastive learning remains a robust baseline. Our code implementation is made available to facilitate reproducbility.

cs.CL↗

BrowseComp-Plus: A More Fair and Transparent Evaluation Benchmark of Deep-Research Agent

Deep-Research agents, which integrate large language models (LLMs) with search tools, have shown success in improving the effectiveness of handling complex queries that require iterative search planning and reasoning over search results. Evaluations on current benchmarks like BrowseComp relies on black-box live web search APIs, have notable limitations in (1) fairness: dynamic and opaque web APIs hinder fair comparisons and reproducibility of deep research methods; (2) transparency: lack of control over the document corpus makes it difficult to isolate retriever contributions. In other words, the current evaluations may compare a complete deep research system at a given time, but they do not foster well-controlled experiments to provide insights into the capability of underlying deep research LLMs. To address these challenges, we introduce BrowseComp-Plus, a benchmark derived from BrowseComp, employing a fixed, carefully curated corpus. Each query in BrowseComp-Plus includes human-verified supporting documents and mined challenging negatives, enabling controlled experimentation. The benchmark is shown to be effective in distinguishing the performance of deep research systems. For instance, the open-source model Search-R1, when paired with the BM25 retriever, achieves 3.86% accuracy, whereas the GPT-5 achieves 55.9%. Integrating the GPT-5 with the Qwen3-Embedding-8B retriever further enhances its accuracy to 70.1% with fewer search calls. This benchmark allows comprehensive evaluation and disentangled analysis of deep research agents and retrieval methods, fostering insights into retrieval effectiveness, citation accuracy, and context engineering in Deep-Research system.

cs.CL↗

An Investigation of Prompt Variations for Zero-shot LLM-based Rankers

We provide a systematic understanding of the impact of specific components and wordings used in prompts on the effectiveness of rankers based on zero-shot Large Language Models (LLMs). Several zero-shot ranking methods based on LLMs have recently been proposed. Among many aspects, methods differ across (1) the ranking algorithm they implement, e.g., pointwise vs. listwise, (2) the backbone LLMs used, e.g., GPT3.5 vs. FLAN-T5, (3) the components and wording used in prompts, e.g., the use or not of role-definition (role-playing) and the actual words used to express this. It is currently unclear whether performance differences are due to the underlying ranking algorithm, or because of spurious factors such as better choice of words used in prompts. This confusion risks to undermine future research. Through our large-scale experimentation and analysis, we find that ranking algorithms do contribute to differences between methods for zero-shot LLM ranking. However, so do the LLM backbones -- but even more importantly, the choice of prompt components and wordings affect the ranking. In fact, in our experiments, we find that, at times, these latter elements have more impact on the ranker's effectiveness than the actual ranking algorithms, and that differences among ranking methods become more blurred when prompt variations are considered.

cs.IR↗

MAGMaR Shared Task System Description: Video Retrieval with OmniEmbed

Effective video retrieval remains challenging due to the complexity of integrating visual, auditory, and textual modalities. In this paper, we explore unified retrieval methods using OmniEmbed, a powerful multimodal embedding model from the Tevatron 2.0 toolkit, in the context of the MAGMaR shared task. Evaluated on the comprehensive MultiVENT 2.0 dataset, OmniEmbed generates unified embeddings for text, images, audio, and video, enabling robust multimodal retrieval. By finetuning OmniEmbed with the combined multimodal data--visual frames, audio tracks, and textual descriptions provided in MultiVENT 2.0, we achieve substantial improvements in complex, multilingual video retrieval tasks. Our submission achieved the highest score on the MAGMaR shared task leaderboard among public submissions as of May 20th, 2025, highlighting the practical effectiveness of our unified multimodal retrieval approach. Model checkpoint in this work is opensourced.

cs.IR↗