Memory and context management over long temporal horizons

Appending all past observations and actions to the context soon exceeds the context window, adds irrelevant or redundant information, and raises inference cost.

• Trajectory memorization and management: memory search systems that select relevant information from long, accumulated contexts.
• Context filtration and compression: summarizing long contexts or filtering noise while keeping the relevant information.
• Consistent personalization: keeping user personas consistent over time and at scale.

Optimization over long temporal horizons

Step-wise supervision is scarce over long horizons, and reinforcement learning with outcome-only rewards leads to unstable training and poorly optimized intermediate steps.

• Optimization with search calls: issuing search queries and using retrieval pipelines effectively and efficiently over long horizons.
• Intermediate supervision and synthetic data generation: training data or annotations for intermediate steps.
• Learning with sparse rewards: handling delayed and sparse rewards in long-horizon processes.

Evaluation over long temporal horizons

Most work evaluates only final answers, but errors accumulate over long horizons, so intermediate steps such as search quality also need evaluation.

• Evaluation data curation with intermediate steps: datasets for assessing intermediate search results, decisions, and actions.
• Multi-dimensional evaluation: efficiency, latency, robustness, computational cost, and error propagation beyond final-task effectiveness.
• Transparency and explainability: measured at intermediate steps and in the final answer.

Multilingual and multimodal scenarios

Evidence spread across languages and modalities must stay aligned and consistent in memory over many steps.

• Multilingual: cross-lingual alignment when evidence is in multiple languages.
• Multimodal: reasoning over text, images, and other modalities with consistent representations and memory.

Long horizons in domain-specific scenarios

We welcome work on law, finance, medicine, science, education, and other domains, including domain knowledge integration and domain-specific models and applications.

Human-centered long-horizon information tasks

How do users inspect and correct what a system remembers, and how do trust and reliance evolve over weeks of interaction? We welcome user studies, interaction design, and qualitative work.

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 systems for long-horizon information tasks?