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

Howard Yen, Ashwin Paranjape, Mengzhou Xia, Thejas Venkatesh, Jack Hessel, Danqi Chen, Yuhao Zhang

Abstract

As AI agents transition into the era of "Deep Research," they are increasingly tasked with scouring the web for extended periods to answer complex, multi-faceted questions. However, current frameworks often fail due to poor context management, leading to information overload and prohibitive costs. We introduce SLIM (Simple Lightweight Information Management), a framework designed to shift agents from passive data consumers to active information managers. By employing surgical browsing and periodic trajectory summarization, SLIM achieves state-of-the-art results on long-horizon benchmarks while using significantly fewer resources.

The Context Management Bottleneck

Existing agentic frameworks like ReAct or Search-o1 typically follow a "search-and-dump" pattern. They scrape entire web pages and insert them directly into the model's context window. This approach leads to three critical failure modes:

  1. Context Explosion: Token limits are quickly reached, causing system crashes or tool-use cessation.
  2. Information Overload: Relevant signals are buried under "noise" (ads, navigation bars, SEO filler), triggering hallucinations.
  3. Budget Exhaustion: Redundant tool calls make the research process economically unviable without nearing the goal.

The SLIM Framework

SLIM addresses these issues through three core design principles that prioritize information density over volume:

  • Split Search and Browse: The search tool returns only titles and short snippets. The agent must explicitly use a separate browse tool to "dig deeper" into specific URLs, forcing it to reason about information value before consumption.
  • Surgical Browsing: Instead of full-page scraping, SLIM uses heuristics (e.g., ROUGE-L or BM25) to extract only the most relevant text chunks based on the agent's specific query.
  • Periodic Trajectory Summarization: Every $n$ steps (e.g., every 50 turns), SLIM condenses the entire conversation history into a concise "state of the research" summary. This clears the mental workspace while retaining essential findings.

Evaluation Datasets

The framework was evaluated on two challenging benchmarks designed for long-form reasoning:

  • BrowseComp: Requires navigating complex web structures for queries that typically take humans over 10 minutes of active searching to solve.
  • Humanity’s Last Exam (HLE): A high-level reasoning benchmark featuring expert-level questions in STEM fields that cannot be answered by internal model knowledge alone.
Examples of complex queries from the BrowseComp and Humanity’s Last Exam (HLE) datasets.

Key Results and Efficiency

Using OpenAI’s o3 as the base model, SLIM demonstrated superior accuracy and extreme efficiency. It achieved 56% on BrowseComp and 31% on HLE, outperforming existing baselines like Search-o1 and HF-Open Deep Research by 4–8 percentage points.

Performance comparison of SLIM against baselines on BrowseComp and HLE benchmarks, showing accuracy vs. cost and tool call count.

Crucially, SLIM reached these benchmarks with 4–6x fewer tool calls than Search-o1. For instance, on BrowseComp, SLIM achieved higher scores with 75 tool calls than Search-o1 did with over 300. This aggressive token management makes "Deep Research" significantly more cost-effective for developers.

Insights and Future Directions

The study reveals that many "hallucinations" in AI agents are actually context management failures—SLIM reduced hallucination rates from 96% to 19% on difficult tasks. This suggests that hallucinations often stem from models getting confused by noisy text rather than "lying."

However, challenges remain. SLIM exhibits high abstention rates on extremely difficult queries and occasionally fails to synthesize answers even when the correct information is present in its logs (the "Answer Ignored" problem). Future work will focus on improving answer-synthesis prompts and training models to better navigate the condensed research summaries that SLIM produces.

Summary updated

BibTeX
@inproceedings{yen2026lost,
  author = {Howard Yen and Ashwin Paranjape and Mengzhou Xia and Thejas Venkatesh and Jack Hessel and Danqi Chen and Yuhao Zhang},
  title = {{Lost in the Maze: Overcoming Context Limitations in Long-Horizon Agentic Search}},
  booktitle = {COLM},
  year = {2026},
  url = {https://openreview.net/forum?id=k4FBtSjHa4},
}
Figure