The First Workshop on Long-Horizon Information Seeking
at the 20th ACM International Conference on Web Search and Data Mining (WSDM 2027)
19 February 2027, Hong Kong

About the workshop

Long horizons have emerged as a key characteristic of information seeking systems in the agentic era. We define them at two levels.

  • At the level of a single complex query, a system may need to carry out a large number of search and reasoning rounds across multiple resources. For example, a complex user question in deep research can require over one hundred search calls.
  • Beyond individual queries, users issue sequences of related queries over multiple days or months, where later queries build on information accumulated from earlier interactions. For example, a user managing a health concern may first search for symptoms, later compare treatments, and return over months to interpret new test results.

We are excited to announce the first Workshop on Long-Horizon Information Seeking (LongSeek), which will be held in person at the 20th ACM International Conference on Web Search and Data Mining (WSDM 2027) in Hong Kong on 19 February 2027. LongSeek brings together the information retrieval (IR), natural language processing (NLP), and machine learning (ML) communities for long-horizon information seeking. The workshop covers memory, optimization, evaluation, and beyond.

Call for Papers

Themes and Topics

Topics of interest include, but are not limited to, the following. We are also open to other topics. Negative results and analyses of failure cases are also welcome.

Memory management over long temporal horizons

Appending all past observations (e.g., retrieved documents) and actions to the context can cause context-window overflow, introduce irrelevant or redundant information, and increase inference costs.

  • Context filtering and compression: How can long, accumulated contexts be filtered or compressed while preserving information needed for subsequent search, reasoning, and decision-making?
  • Long-term memory organization and retrieval: How can agents organize and update long-term memory, including file-based memory? How can an agent manage information flow between its long-term memory and its active context? How can agents search long-term memory to retrieve and select relevant information from long, accumulated interaction histories?
  • Consistent personalization: How can agents maintain consistent personalization across interactions while adapting to changes in users' preferences and needs?

Optimization over long temporal horizons

Limited step-wise supervision, sparse or delayed rewards, and high computational costs make long-horizon agents challenging to train and optimize.

  • Synthetic training data generation: How can we generate high-quality long-horizon training trajectories that include a user question, intermediate search and reasoning steps, and a final answer?
  • Reinforcement learning with sparse and delayed rewards: How can agents learn stable and effective policies from sparse or delayed rewards? How can rewards be designed and credit assigned to intermediate search, reasoning, and other tool-use decisions?
  • Optimizing agents for efficiency: How can we train agents to allocate search and reasoning budgets, avoid redundant search tool calls, and decide when to stop while balancing task effectiveness, latency, and computational cost?
  • Continual learning and self-evolution: How can agents improve from feedback and accumulated experience across tasks and over time while retaining previously acquired capabilities?

Evaluation over long temporal horizons

Final-answer accuracy alone does not capture intermediate search and reasoning quality, cumulative costs, or safety and privacy risks over long horizons.

  • Evaluation datasets with intermediate annotations: How can we build evaluation datasets with relevance judgments for retrieved documents and annotations for intermediate search queries and reasoning steps? How can we construct such datasets at scale?
  • Efficiency and robustness: How can we evaluate the trade-offs between task effectiveness, latency, and computational cost? How can we assess robustness to noisy evidence and tool failures as search trajectories and interaction histories grow?
  • Transparency and explainability: How can we trace the evidence behind agents' intermediate decisions and final answers over long horizons? How can we assess whether explanations faithfully reflect the search and reasoning process? How can we evaluate whether these explanations are understandable to humans?
  • Safety and privacy: How can we evaluate whether agents continue to follow safety constraints over long search and reasoning trajectories, including under prompt injection? How can we detect privacy leakage through reasoning traces, tool calls, and final answers across steps and sessions?

Long horizons in multilingual and multimodal scenarios

Long-horizon search and reasoning become more challenging when user queries and retrieved evidence span multiple languages and modalities.

  • Multilingual: How can evidence retrieved early in one language guide later queries and reasoning in other languages? How can agents integrate evidence from different languages over long search trajectories without changing the meaning of the original information or losing track of its sources?
  • Multimodal: How can agents use visual clues from earlier images, tables, or videos to guide later search and reasoning? How can they decide when to retain, compress, or revisit visual evidence while preserving details needed in later steps?

Long horizons in domain-specific scenarios

We welcome domain-specific models and applications in law, finance, medicine, science, education, enterprise, and other domains.

  • Domain-specific information tasks: How can agents support legal case analysis, financial research, scientific literature review, and other tasks that require evidence to be gathered, verified, and reused over many search and reasoning steps?
  • Domain knowledge and tools: How can agents combine domain knowledge, specialized information sources, and professional tools over long horizons?

Human-centered long-horizon information tasks

Long-horizon information tasks require users to guide and review agents' work as goals and evidence evolve. We welcome user studies, interaction design, and qualitative work.

  • User control and collaboration: How can users revise search goals, inspect and correct what agents remember, and redirect ongoing work across steps and sessions? When should agents ask for clarification or return control to users?
  • Oversight and interaction design: How can interfaces help users review long search and reasoning trajectories, find errors, and intervene without inspecting every intermediate step?
  • Trust and reliance over time: How do trust, reliance, and users' willingness to verify agents' work evolve over days or weeks of interaction? How can agents support appropriate reliance as their capabilities and users' needs change?

Deployment and industrial experience

We welcome reports from real deployments of agents for long-horizon information tasks, including system architectures, workflow integration, monitoring and failure recovery, human oversight, and operational costs.

Paper type

We welcome paper submissions in various types, including original research, resource, reproducibility, industry, position, survey, and demonstration papers.

Papers can be up to 5 pages long, including references. Figures, tables, proofs, appendices, acknowledgments, and references all count toward the page limit.

Paper submission & selection

Paper template: papers must use the ACM formatting, sigconf style, and be submitted as PDF; use \documentclass[sigconf,anonymous,review]{acmart}.

Each paper will be peer-reviewed by at least two program committee (PC) members. The review is double-blind, so authors must anonymize their submissions.

Accepted papers will be published in the WSDM 2027 Companion Proceedings.

Authors of accepted papers will be invited to present at the workshop, so that all authors have a chance to speak. All presentations will be held in person in Hong Kong, in accordance with WSDM 2027 policy. There will be no provision for remote presentations.

Note that accepted papers that cannot be presented in person will be excluded from the Companion Proceedings. A non-archival option is available on request.

Best paper award

A best paper award will be selected based on the PC's evaluations and announced at the end of the workshop. The awards will be presented at the closing session of the workshop.

Important Dates

  • Submission deadline: 30 November 2026 (11:59 PM, AoE)
  • Acceptance notification: 18 December 2026 (11:59 PM, AoE)
  • Camera ready: To be announced
  • Workshop date: 19 February 2027

All deadlines are at 11:59 PM AoE (Anywhere on Earth).

Keynote Speakers

To be announced.

Workshop Schedule

Date: Friday, 19 February 2027 (half day)
Venue: Cordis, Hong Kong (555 Shanghai St, Mong Kok, Hong Kong), co-located with WSDM 2027
Time slot: To be announced
Room: To be announced
Program: To be announced

Workshop Organizers

Chuan Meng

Chuan Meng

The University of Edinburgh

United Kingdom

chuan.meng@ed.ac.uk

Fengran Mo

Fengran Mo

Rochester Institute of Technology

United States

fengran.mo@rit.edu

Jinghan Zhang

Jinghan Zhang

Clemson University

United States

jinghaz@clemson.edu

Zhenrui Yue

Zhenrui Yue

Google DeepMind

United States

yuezrhb@google.com

Yan Wang

Yan Wang

Yale University

United States

yan.wang.yw937@yale.edu

Jimin Huang

Jimin Huang

The University of Manchester
The Fin AI

United Kingdom / United States

jimin.huang@thefin.ai

Chen Luo

Chen Luo

Amazon

United States

cheluo@amazon.com

Mohammad Aliannejadi

Mohammad Aliannejadi

University of Amsterdam

The Netherlands

m.aliannejadi@uva.nl

Michael Bendersky

Michael Bendersky

Databricks

United States

michael.bendersky@databricks.com

Jeff Dalton

Jeff Dalton

The University of Edinburgh

United Kingdom

jeff.dalton@ed.ac.uk

Program Committee

To be announced.
Copyright © LongSeek: The First Workshop on Long-Horizon Information Seeking