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

AI summary

Overview

Research area: Long-horizon agentic search (deep research systems) and context management for LLM agents.

Technical level: Intermediate — assumes familiarity with LLM agents, tool use, RAG, and benchmarking, but explains the design clearly.

Scope: The paper diagnoses why existing agentic search frameworks fail on long trajectories and proposes a minimal three-component framework (Slim) that keeps context concise while scaling to many tool calls.

What This Paper Is About

Long-horizon agentic search systems — like deep research tools — must iteratively search the web across many turns and synthesize information from many sources. The paper shows that popular open-source frameworks (ReAct, Search-o1, HuggingFace OpenDeepResearch, GPT-Researcher) fail on more than half of long-trajectory tasks, mainly because they fill their context window with noisy scraped content, exhaust their tool budgets, or stop prematurely. The authors introduce Slim (Simple Lightweight Information Management), a bare-bones framework built on three components — search, browse, and periodic summarization — that outperforms far more complex systems at a fraction of the cost and tool calls.

Key Contributions

1. A diagnosis of failure modes in existing agentic search frameworks. The authors run ReAct, Search-o1, HF-ODR, and GPT-R on BrowseComp with fixed turn budgets, categorize every trajectory outcome (correct, exceeded context, exceeded budget, early stopping, no tool used, misc. error), and show that context limits and tool budgets are the dominant bottlenecks — not model capability.

2. The Slim framework. A single-agent design with three deliberately simple components: a search tool returning only titles, URLs, and snippets; a browse tool that scrapes one chosen URL and returns only the section most relevant to a query; and a summarization module that periodically compresses the entire conversation every n turns.

3. A systematic trajectory-level error taxonomy and automated analysis pipeline. The authors manually examine trajectories to build a shared taxonomy of failure modes (confirmation bias, unfocused search, inefficient search, abstention, answer ignored, hallucination), then build a pipeline combining rule-based heuristics and LLM-as-a-judge to annotate any framework's trajectories.

4. Broad empirical validation across models and benchmarks. Slim is evaluated with proprietary API models (o3, o4-mini, Claude-4-Sonnet), open-weight models (GPT-OSS-120B, GLM-4.7 Flash), and an RL-trained deep research model (Tongyi-DeepResearch-30B) on five datasets.

Main Findings

• Slim dominates on performance-per-cost with o3. With o3 as the base model, Slim achieves 56% on BrowseComp and 33% on HLE, beating the best open-source framework by 8 and 6 absolute points respectively, while using 4–6x fewer tool calls and lower overall cost. It also leads by 15 points on DeepSearchQA and 26 points on GAIA.

• Gains generalize to smaller and open-weight models. With GLM-4.7 Flash, Slim beats Search-o1 by 10 points on BrowseComp at one-third the cost. With GPT-OSS-120B, Slim reaches 70.5 cost-units versus Search-o1's 45.7 while scoring 20.0 versus 6.0 on BrowseComp. With o4-mini, gains hold across all datasets.

• ReAct fails through context overflow; Search-o1 fails through budget exhaustion. ReAct's single retrieval tool returns full content for all top-10 results, blowing the context window early. Search-o1's failures are almost entirely tool-budget hits, because it scrapes every result even when only a fraction are relevant.

• HF-ODR frequently stops too early or skips tools entirely. In 10% of BrowseComp trajectories HF-ODR used no tools at all, suggesting its complex multi-agent prompt workflow suppresses the base model's tool-calling behavior.

• Slim sharply reduces hallucination. In trajectory-level analysis at matched cost budgets, Slim shows a 19.0% hallucination rate on incorrect, non-abstaining samples versus 46.8% for Search-o1, 56.7% for ReAct, and 96.2% for HF-ODR. The authors attribute this to Slim's ability to choose which URLs to browse, reducing noisy context.

• New failure modes surface at longer horizons. Slim and Search-o1 both show elevated "answer ignored" rates (30.7% and 26.0%), meaning they encounter the correct answer during long trajectories but fail to use it. Slim also shows a high abstention rate (27.7%), an explicitly acknowledged limitation.

Methodology in Plain English

The authors first formalize agentic search as a loop where an LLM calls tools (each call is a turn, bounded by a tool budget T) against a corpus of web documents. They note that standard frameworks treat retrieval as one bundled tool: search plus scrape everything.

They then run four baseline frameworks on BrowseComp with controlled turn budgets and hand-categorize every trajectory's outcome. This reveals that context overflow and budget exhaustion account for most failures.

Slim replaces the bundled retrieval with two separate tools. Search returns only the top-k titles, URLs, and short snippets, so the model must decide what is worth reading. Browse takes a URL and a query and returns only the passage of that page with the highest similarity to the query, so a single page costs a fraction of the tokens that full scraping would. Finally, a summarization module re-writes the entire conversation history into a summary every n turns — not just the search results, as prior work does — keeping the trajectory short enough to allow many more turns within the same context window.

For evaluation, they sample 300 instances per dataset from BrowseComp, GAIA (text-only subset of 103), HLE, DeepSearchQA, and HealthBench, and measure accuracy, tool calls, and dollar cost (LLM tokens plus tool fees).

For the error analysis, they manually inspect trajectories to derive a taxonomy, then automate annotation with a mix of hand-written rules and an LLM judge that inspects search queries, results, browsed content, and final answers.

Why This Matters

Impact on research: The paper argues that the bottleneck in deep research agents is context and tool-interface design, not raw model capability. It provides the community with both a simple, strong baseline and — arguably more valuable — a reusable error taxonomy and automated trajectory-analysis pipeline for comparing agentic systems beyond headline accuracy numbers. This shifts evaluation from "did it get the answer" toward "how did it fail."

Real-world applications:

• Deep research assistants for analysts, journalists, and lawyers who need multi-source synthesis over long investigation trails.
• Enterprise knowledge agents that must search internal corpora plus the open web under cost and latency constraints.
• Medical and scientific literature review (HealthBench and HLE results directly target these domains), where hallucination reduction is a safety requirement.
• Cost-sensitive deployments where a 4–6x reduction in tool calls and token spend makes agentic search economically viable at scale.

Industry relevance: The finding that complex multi-agent orchestration (HF-ODR, GPT-R) can underperform a two-tool single agent is directly actionable for product teams building deep research features. The result that summarization of the whole trajectory beats per-search summarization gives a concrete, cheap design pattern.

Future Directions

• Making agents use answers they already found. "Answer ignored" was a leading failure mode for both Slim and Search-o1, meaning correct information was in the context but not used — an attention/aggregation problem the paper leaves open.

• Reducing abstention. Slim's 27.7% abstention rate was its most prominent remaining weakness; the authors explicitly defer this to future work.

• Optimizing the summarization schedule. The paper uses a simple fixed-interval heuristic and notes that memory-selective mechanisms may not be necessary, but the optimal compression policy is unexplored.

• Applying the error taxonomy more broadly. The pipeline is framework-agnostic; extending it to compare RL-trained agents and larger model families, or using it as a training signal, is a natural next step.

Target Audience

Researchers and engineers building LLM agents for search, retrieval, and deep research; practitioners choosing or designing tool interfaces for agentic pipelines; and evaluation researchers who need interpretable, trajectory-level diagnostics rather than single accuracy scores. Readers interested in prompt-level agent design will find the tool-interface ablations particularly useful, while those focused on RL training for search agents will benefit from the error taxonomy as a diagnostic and possibly a reward-shaping tool.

Authors’ abstract

Long-horizon agentic search requires iteratively exploring the web over long trajectories and synthesizing information across many sources, enabling powerful applications like deep research systems. In this work, we show that popular agentic search frameworks struggle to scale to long trajectories primarily due to context limitations--they accumulate long, noisy content, hit context window and tool budgets, or stop early. We therefore introduce SLIM (Simple Lightweight Information Management), a simple framework that separates retrieval into distinct search and browse tools, and periodically summarizes the trajectory, keeping context concise while enabling longer, more focused searches. Across a wide range of long-horizon tasks, SLIM achieves comparable performance at substantially lower cost and far fewer tool calls than strong open-source frameworks with both proprietary and open-weight models, including RL-trained models for deep research. Specifically, with o3 as the base model, SLIM achieves 56% on BrowseComp and 33% on HLE, outperforming all open-source frameworks by 8 and 6 absolute points, respectively, while incurring 4-6x fewer tool calls. With GLM-4.7 Flash, SLIM achieves 10 points improvement over the next best open-source framework, Search-o1, on BrowseComp using a third of the cost. To systematically understand failure modes in long-horizon agentic search, we develop an automated fine-grained trajectory analysis pipeline and error taxonomy, and find that SLIM exhibits significantly fewer hallucinations than prior systems. We hope our analysis framework and simple tool design inform future long-horizon agents.