Lost in the Maze: Overcoming Context Limitations in Long-Horizon Agentic Search

2025-04-16 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Expert, extended

Summary

Slim (Simple Lightweight Information Management), a new framework for long-horizon agentic search, addresses critical context limitations in existing systems. Current frameworks like ReAct and Search-o1 struggle with long trajectories due to accumulating noisy content, hitting context window limits, or exceeding tool budgets. Slim overcomes this by separating retrieval into distinct search and browse tools and periodically summarizing the conversation history. This design keeps context concise, enabling longer, more focused searches. Benchmarking with o3 as the base model, Slim achieved 56% on BrowseComp and 31% on HLE, surpassing all open-source baselines by 8 and 4 absolute points, respectively. Crucially, it incurred 4–6x fewer tool calls and lower overall costs. The framework also includes an automated trajectory analysis pipeline and error taxonomy, showing Slim reduces hallucinations.

Key takeaway

For AI Engineers designing or optimizing long-horizon agentic search systems, you should prioritize efficient context management to overcome limitations. Implement a framework like Slim, which separates search and browse tools and periodically summarizes conversation history, to significantly reduce tool calls and costs while improving performance. This approach helps mitigate issues like context window limits and hallucinations, enabling more robust and cost-effective deep research applications.

Key insights

Slim improves long-horizon agentic search by managing context efficiently through distinct search/browse tools and periodic summarization.

Principles

• Context management is crucial for long-horizon agents.
• Separate search and browse tools for efficiency.
• Periodically summarize trajectories to reduce noise.

Method

Slim uses a search tool for snippets, a browse tool for relevant content sections from URLs, and a summarization module that compresses the entire conversation history every 'n' turns.

In practice

• Implement distinct search and browse tools.
• Periodically summarize conversation history.

Topics

• Long-Horizon Agentic Search
• Context Management
• Large Language Models
• Web Search Agents
• Information Retrieval
• Performance Optimization
• Error Taxonomy

Code references

• howard-yen/SLIM
• unclecode/crawl4ai
• assafelovic/gpt-researcher
• RUC-NLPIR/Search-o1
• huggingface/smolagents

Best for: Research Scientist, AI Architect, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Context Window Management for Long-Running Agents: Strategies and Tradeoffs — Managing context windows in long-running AI agents requires specific strategies to overcome bottlenecks and tradeoffs.
• A Late Chunking Approach for Visual Documents, Does Agentic Search Make GraphRAG Obsolete? and More! — New research enhances LLM retrieval, recommendation systems, and RAG through novel architectures and training methods.
• Implement Anthropic's Context Engineering Framework with open source models — Dynamic context engineering mitigates "context rot" in LLM agents by managing information flow and computational load.
• LLM with 12M Context Window — SubQ introduces subquadratic sparse attention, enabling massive context windows with significantly reduced compute and cost.
• OCR-Memory: Optical Context Retrieval for Long-Horizon Agent Memory — Visual encoding of agent trajectories enables long-horizon memory with reduced token costs and hallucination.
• Giving Your AI Agent Less Context Made It Better at Its Job — Less context, specifically curated, enhances AI agent performance and cost-efficiency on long-horizon tasks.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.