Laser: Governing Long-Horizon Agentic Search via Structured Protocol and Context Register
Paper: arXiv:2512.20458 · Published: 2025-12-23 · Category: Agentic Search
Overview
This post summarizes the Laser framework, proposed for stabilizing and scaling long-horizon agentic search in LLM-based systems.
Background
Recent advances in Large Language Models (LLMs) and Large Reasoning Models (LRMs) have enabled agentic search systems that interleave multi-step reasoning with external tool use. However, existing frameworks largely rely on unstructured natural-language reasoning and accumulate raw intermediate traces in the context, which often leads to:
- Unstable reasoning trajectories
- Context overflow
- Degraded performance on complex multi-hop queries
- Architecture: Agentic paradigms are making retrieval count and strategy learnable objects, beyond fixed cascade pipelines.
- Evaluation: Process-level metrics (task success, citation accuracy, multi-hop chain completeness) complement static nDCG.
- Engineering: Latency, cost, interpretability, and safety remain hard constraints for production deployment; readers should verify quantitative results against the original PDF tables.
- Agentic Information Retrieval (arXiv:2410.09713)
- Synergizing RAG and Reasoning: A Systematic Review
- AceSearcher: Bootstrapping Reasoning and Search for LLMs via Reinforcement Learning
Key Ideas
Laser consists of two coordinated components:
1. Symbolic action protocol — Organizes agent behaviors into three spaces: planning, task-solving, and retrospection. Each action is specified with explicit semantics and a deterministic execution format, enabling structured reasoning and reliable action parsing. Intermediate decisions become interpretable and traceable, supporting explicit retrospection and fine-grained control over reasoning trajectories.
2. Compact context register — Stores only the essential states of the reasoning process, allowing the agent to reason over long horizons without uncontrolled context expansion.
Results
Experiments on Qwen2.5/3-series models across challenging multi-hop QA datasets show that Laser consistently outperforms existing agentic search baselines under both prompting-only and fine-tuning settings.
Original Abstract
> Recent advances in Large Language Models (LLMs) and Large Reasoning Models (LRMs) have enabled agentic search systems that interleave multi-step reasoning with external tool use. However, existing frameworks largely rely on unstructured natural-language reasoning and accumulate raw intermediate traces in the context, which often leads to unstable reasoning trajectories, context overflow, and degraded performance on complex multi-hop queries. In this study, we introduce Laser, a general framework for stabilizing and scaling agentic search. Laser defines a symbolic action protocol that organizes agent behaviors into three spaces: planning, task-solving, and retrospection. Each action is specified with explicit semantics and a deterministic execution format, enabling structured and logical reasoning processes and reliable action parsing. This design makes intermediate decisions interpretable and traceable, enhancing explicit retrospection and fine-grained control over reasoning trajectories. In coordination with parsable actions, Laser further maintains a compact context register that stores only essential states of the reasoning process, allowing the agent to reason over long horizons without uncontrolled context expansion. Experiments on Qwen2.5/3-series models across challenging multi-hop QA datasets show that Laser consistently outperforms existing agentic search baselines under both prompting-only and fine-tuning settings, demonstrating that Laser provides a principled and effective foundation for robust, scalable agentic search.