New Information Retrieval (cs.IR) papers, Thursday 22 October 2026
25 papers announced on arXiv. Get them ranked for you, free →
RESEARCHHUB: A Personalized Literature Discovery System for Biomedical Preprints
AI summaryResearchHub is an interactive system that personalizes biomedical preprint discovery by retrieving relevant papers and encouraging novelty, diversity, and curriculum ordering. Its bi‑encoder retriever, prism, achieves 12.4 % higher Recall@10 than SPECTER2‑Search on a 44k MedRxiv dataset.
TAGR: Task-Adaptive Geo-Semantic Retrieval across Heterogeneous Urban Delivery Tasks
AI summaryTAGR builds an LLM‑derived geo‑semantic prior and task‑specific retrieval banks to share regional knowledge across demand forecasting, ETA prediction, and next‑stop prediction, using a task‑adaptive augmentation module that reduces mismatch and negative transfer. It outperforms baselines on three city datasets.
Can AI Chat Replace Traditional Search? Evidence from a RAG-based Live Events Experiment
AI summaryThe study evaluates a RAG‑based chat interface versus a traditional filter and a hybrid in an online live‑events search experiment. It finds the chat alone performs worst, mainly due to echo‑chamber effects and prompting fatigue, while the filter or hybrid match each other.
Transaction Foundation Models for Retrieval-Augmented Generation over Financial Transaction Data: A Comparative Evaluation with Text Embeddings
AI summaryThey benchmark a 29M‑parameter Transaction Foundation Model against generic and domain‑text embeddings for RAG on a synthetic AML dataset. TFM beats generic models in retrieval and accuracy, but a hand‑crafted summary with a generic encoder matches or outperforms it on many categories.
Controllable Generative Recommendation via Guided Token Refinement
AI summaryReGen changes autoregressive recommendation from left‑to‑right decoding into parallel mask‑filling and iterative token refinement, enabling real‑time control of accuracy and diversity without retraining. It uses a topology‑constrained tree to enforce valid Semantic IDs, achieving a better accuracy‑diversity trade‑off.
Building Knowledge Graphs for Theses: A Survey
AI summaryThe paper reviews 46 studies (2021‑2025) on constructing knowledge graphs from academic theses, classifying methods from NLP extraction to ontology engineering, semantic web, embeddings, LOD, and LLMs. It summarizes current practices and outlines future research paths.
FutureLens: Longitudinal Topic Patterns in Scholarly Text and Future-Work References
Future-work statements offer a record of the questions researchers choose to leave open. Using them to describe a field, however, requires separating thematic change from changes in document coverage, topic assignment, and scholarly writing. FutureLens examines these measurement…
Constrained Tagging for Span Labeling
AI summaryConstrained tagging forces a generative LLM to output input text with inline span labels, ensuring every labeled span appears exactly in the input. The method achieves near‑perfect coverage across six datasets, matching or exceeding prompting accuracy with negligible latency.
Scriptorium: System Design and Implementation of a Multi-Human, Multi-AI Collaborative Digital Library Search Platform
AI summaryScriptorium is a digital library search interface that enables multiple users and seven specialized generative AI agents to collaborate transparently. The system’s Manager–Worker architecture, detailed prompts, and guardrails aim to boost serendipitous cross‑pollination between collaborators’ research threads.
Our Own Jev: Typed Decisions in One Forward Pass, from an In-House Judge to Open LoRA Models
AI summaryTyped decisions predict option probabilities in one forward pass by reading logits or hidden states, replacing expensive text generation. The paper shows a LoRA‑adapted 35B model reaches 79.3% on JevBench, matching commercial performance, while a 1B system triages 148K coding sessions cost‑effectively.
Progressive Learning from Content to Relevance to Personalization for CTR Prediction
AI summaryFineSID tokenizes multimodal content into fine‑grained Semantic IDs, pretrained with query‑item relevance and fine‑tuned for user clicks. It outperforms ID‑only models on CTR and improves creative discrimination, showing strong semantic generalization.
Reasoning-Search-Augmented LLMs: An Operational Taxonomy, Benchmark Analysis and Controlled Evidence Synthesis
AI summaryThis paper proposes an operational taxonomy for reasoning-search-augmented LLMs based on four architectural axes and learning signals. It analyzes benchmark demands and synthesizes evidence on the quality-cost tradeoffs of these interactive retrieval systems.
Preference Is Not Permission: Personalizing the Auto-Purchase Line in Agentic Commerce
AI summaryThe paper defines a personalized auto‑purchase line with a three‑state decision (don’t buy, confirm, buy) based on warrant, reversibility, price, and urgency, and shows that shallow models trained on a few dozen personal labels can predict this boundary with 0.79 accuracy, outperforming all prompt‑model baselines.
Learning Marginal Viewpoint Utility for Budgeted Evidence Selection in RAG
AI summaryLMVU selects evidence for retrieval‑augmented generation by scoring candidates based on their marginal viewpoint coverage. Across four benchmarks at a 1,024‑token budget, it improves RVC$_1$ by 0.047 and RVC$_2$ by 0.052 over the best non‑oracle baseline, reaching the exact viewpoint‑breadth ceiling on two datasets.
Extending FAIR Principles with Observability for Trustworthy Federated Learning
AI summaryExtends FAIR principles to federated learning by adding an observability layer that produces signed trust evidence from each participant per round. Demonstrates a three-level trust model and a Python prototype for evidence generation and trust assessment, enabling future participant selection and aggregation.
Generative Recommenders Forget the Long Tail: Restoring Memorization in Retrieval and Ranking
Generative recommenders replace the retrieve-then-rank cascade with a single model: beam search over semantic item identifiers performs retrieval, and the same likelihood performs ranking. We show that this shared score makes them forget long-tail items twice. In retrieval, beam…
Directional Posterior Feedback Fusion for Query Reformulation in Dense Retrieval
AI summaryDirectional Posterior Feedback Fusion (DIFF) models query feedback as directional evidence in the query’s tangent space, jointly inferring a controlled update via posterior fusion without retraining the retriever. It outperforms prior methods on dense retrieval benchmarks, showing superior query reformulation.
Beyond the Vote: When Do LLMs Add Value to Retrieval-Based Tabular Classification?
Retrieval-based in-context learning (ICL) enables large language models (LLMs) to predict tabular labels from similar examples, but it remains unclear whether LLM inference improves on direct voting over those same examples. We address this question by comparing four…
X-Embedding: Two-Stage Training for Multilingual Patent Prior-Art Retrieval
Multilingual patent prior-art retrieval requires identifying X-category references that are particularly relevant to the novelty or inventive step of a claimed invention, beyond topical similarity across languages and document views. We present X-Embedding, a two-stage framework…
pplx-embed-v2-late Technical Report
In this report, we introduce pplx-embed-v2-late, a pair of 0.6B and 9B multilingual and multimodal late-interaction models. Both models achieve state-of-the-art results within their respective size classes across a range of text, image, and visual document retrieval benchmarks.…
SOAR: Social and Opinion-Aware Recommendations
One of the enduring challenges in online platforms is recommendation. Both communities and advertisers face difficulties in analyzing complex user behavior, identifying habits and evolving preferences, and providing timely recommendations that are relevant to user interests. In…
Real PMIDs, Wrong Papers: Harness Design for Biomedical Claim-Verification Agents
Literature-searching language agents are becoming a standard component of scientific assistants for the life sciences, yet there is little controlled evidence on which parts of their harness matter. We run a controlled ablation on biomedical claim verification (SciFact) with the…
ERCLIP: Small-Data Adaptation of CLIP for Free-Length Image-Text Retrieval
CLIP's 77-token text encoder truncates detailed image descriptions. We present ERCLIP, a small-data adaptation of CLIP for long-caption image--text retrieval. During training, a long caption is split into sentences that the CLIP text encoder encodes independently; the sentence…
What the Author Wrote, What the Agent Reads: Silent Cue Loss in Enterprise Document Ingestion
Enterprise agents act on documents. Before an agent can follow a policy, answer a compliance question or file a ticket, some ingestion layer must turn a Word file or PDF into tokens - and that layer is where an agent deployed in the wild silently loses information. We show that…
Think Harder Where It Matters: Item-Level Adaptive Recursion for Sequential Recommendation
Existing sequential recommenders usually allocate a fixed computational budget to every item in a user's interaction history, regardless of the item's contribution to the final prediction. We challenge this de facto design with two diagnostic findings. (1) Non-uniform…