Mechanistic Interpretability for Information Retrieval

Modern IR systems depend on large neural models: dense retrievers, cross-encoder rankers, and LLM-based generation components in retrieval-augmented pipelines. We usually evaluate them end to end, with effectiveness metrics on benchmark queries, and learn little about the internal steps that produce a ranking or an answer. That limits how well we can diagnose failures, explain why a retriever prefers certain documents, or tell why a generator ignores retrieved evidence.

Mechanistic interpretability studies the internal computations, representations, and circuits of neural models. Its methods include probing hidden activations, circuit analysis, causal tracing, activation patching, and sparse autoencoders. A small number of IR papers have started to use them, but the work is still scattered, without shared benchmarks, terminology, or evaluation practices. This workshop brings these researchers together at ECIR, following the tutorial on mechanistic interpretability given at ECIR 2026.

• Survey and consolidate existing work at the intersection of mechanistic interpretability and IR.
• Identify open problems that are specific to retrieval and ranking settings.
• Start building a community around the topic within IR. We want to learn from how the mechanistic interpretability community formed within NLP, so that the IR community grows around shared benchmarks and terminology from the start.

The panel discussion is meant to produce a shared research agenda that summarizes open problems and promising directions.

The workshop is a half-day, in-person workshop with a keynote talk, a poster session, and a panel discussion. It is planned as non-archival: accepted papers are presented as posters and are not published in formal proceedings. This may change depending on the number of submissions.

• Chen, Merullo, Eickhoff. Axiomatic Causal Interventions for Reverse Engineering Relevance Computation in Neural Retrieval Models. SIGIR 2024.
• Parry, Chen, Eickhoff, MacAvaney. MechIR: A Mechanistic Interpretability Framework for Information Retrieval. ECIR 2025.
• Reusch, Belinkov. Reverse-Engineering the Retrieval Process in GenIR Models. SIGIR 2025.
• Wang, Qi, Chen, Eustratiadis, Verberne. How Role-Play Shapes Relevance Judgment in Zero-Shot LLM Rankers. ECIR 2026.
• Chen, Heuss, Eickhoff. Tutorial on Mechanistic Interpretability. ECIR 2026.
• Saphra, Wiegreffe. Mechanistic? BlackboxNLP 2024. On how the mechanistic interpretability community formed within NLP.
• Bereska, Gavves. Mechanistic Interpretability for AI Safety: A Review. TMLR 2024.

We welcome work on the application, adaptation, and development of mechanistic interpretability methods for IR models and pipelines, including:

• Probing and analysis of internal representations in dense retrievers and rankers
• Circuit-level analysis of relevance judgments and ranking decisions
• Interpretability of retrieval-augmented generation, including how generation components use retrieved evidence
• Causal interventions on IR model internals, such as activation patching and ablation studies
• Sparse autoencoders and feature-level analysis applied to IR models
• Relationships between mechanistic interpretability and existing IR notions of explainability
• Evaluation methodologies for interpretability claims in IR settings
• Benchmarks and datasets for interpretability research in IR
• Applications of mechanistic interpretability across IR tasks and domains, such as ad hoc retrieval, ranking, recommendation, question answering, conversational search, and domain-specific search in health, legal, scientific, and e-commerce settings

We invite short papers of up to 4 pages in the Springer LNCS format used by ECIR.

Papers that describe negative or null results are welcome. Work that is already published elsewhere is welcome too, in its original published format, with no page limit.

We accept submissions from every stage of research, from perspective papers and work in progress to published work, and we welcome submissions from students.

Review is single-blind: reviewers see author names and affiliations, and authors do not see reviewer identities. Accepted papers are presented as posters at the workshop.

The workshop is planned as non-archival, so presenting a paper at the workshop does not constitute a formal publication.

The submission system and instructions will be added here.