GRIT 2026

Overview

Generative information retrieval (GenIR) reformulates retrieval as sequence generation, directly generating identifiers of relevant targets. Beyond document retrieval, this paradigm has expanded to generative recommendation, multimodal retrieval, code retrieval, and entity retrieval. GRIT focuses on the real-world readiness of GenIR beyond benchmark accuracy. The workshop brings together research on foundations and probing, reliability and robustness, continual adaptation, scalability and efficiency, evaluation beyond relevance, and deployment experience, with the goal of advancing GenIR systems that are understandable, dependable, adaptable, and practical.

Scope

In scope: systems where generation is the retrieval mechanism itself, including document and passage retrieval, generative recommendation, multimodal and cross-modal retrieval, entity retrieval, product search, and code retrieval. Generated outputs may include titles, URLs, entity names, semantic codes, or other identifiers that map back to retrieval targets.

Out of scope: systems that use generative models only as auxiliary components, such as query expansion, summarization, or answer generation. Standard RAG pipelines are outside the core scope unless retrieval itself is performed through generation.

Topics of Interest

We welcome academic, industrial, and position papers on topics including, but not limited to:

Foundations and Probing

What information about the target collection is encoded in model parameters? How can we probe, interpret, and diagnose retrieval-by-generation systems, and how do generation likelihood, relevance, preference, and utility relate to each other?
Reliability and Robustness

How reliable are generated identifiers under noisy, ambiguous, adversarial, or out-of-distribution inputs? How should systems handle invalid, duplicated, hallucinated, or poorly calibrated generations?
Resilience and Continual Adaptation

How can GenIR systems support dynamic collections, insertions, deletions, updates, cold-start targets, temporal shifts, and machine unlearning without costly retraining or catastrophic forgetting?
Scalability and Efficiency

What are the training, indexing, updating, serving, latency, memory, throughput, and energy costs of GenIR at realistic scales, compared with sparse, dense, and hybrid retrieval infrastructures?
Evaluation Beyond Relevance

What benchmarks, metrics, and diagnostic protocols are needed beyond accuracy? How should we evaluate updateability, coverage, diversity, novelty, long-tail accessibility, fairness, privacy, robustness, and user- or session-level utility?
Deployment and Industrial Experience

What can be learned from real deployments, system analyses, monitoring practices, hybrid architectures, success stories, failure cases, and negative results involving retrieval-by-generation systems?

Call for Papers

GRIT solicits short papers of up to 4 pages, including references, in the ACM two-column format. Authors may use the ACM proceedings template or the Overleaf template. Please follow the official SIGIR-AP 2026 paper submission instructions. We welcome case studies, experience reports, position and vision papers, open problems, preliminary ideas, research findings, system analyses, benchmarks, and negative results related to the real-world readiness of retrieval-by-generation systems.

Submit via EasyChair →

Important Dates

All deadlines are 11:59 PM AoE (Anywhere on Earth) unless noted otherwise.

Paper submission due: September 25, 2026
Notification: October 15, 2026

Review Process

Each submission will receive at least two reviews. Decisions will consider relevance, clarity, technical soundness where applicable, discussion potential, and balance across workshop topics. Standard conflict-of-interest procedures will be followed. Accepted papers will be presented as lightning talks, posters, demonstrations, or breakout discussion inputs.