Research map / Multi-agent systems and games / LLM agents

Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes WebLeaper, which synthesizes entity-intensive information-seeking tasks from Wikipedia tables (Basic, Union, Reverse-Union), filters trajectories by ISR/ISE, and trains with SFT plus GRPO using a hybrid F-score reward.

Problem

Information-seeking LLM agents search inefficiently, partly because synthetic training tasks contain too few target entities, limiting learning of efficient search behavior and biasing efficiency measurement.

Approach

Model IS as a tree-structured reasoning problem populated from Wikipedia tables so that many target entities fit in a bounded context; build Union and Reverse-Union variants via maximal biclique enumeration; keep only trajectories with high information-seeking rate and efficiency; and use a soft-F-score reward in GRPO.

Key results

• WebLeaper-Reverse-Union C reaches 38.8 on BrowseComp, 73.2 on GAIA, 72.0 on xbench-DS, and 48.6 on Seal-0 in the base/comprehensive setting, versus 15.7 for WebExplorer-8B on BrowseComp.
• On WideSearch, WebLeaper-Reverse-Union C achieves 4.0 SR, 31.0 Row F1, and 48.8 Item F1, the best among open-source agents in the table.
• Union-5k and Reverse-Union-10k improve average score over WebSailor-V2-5k (44.05) by +3.26 and +4.34 points respectively, while Basic-5k drops by -7.64.

Surprising findings

• The Basic synthetic variant substantially underperforms the WebSailor-V2-5k baseline, with GAIA dropping by 26.91 points, suggesting shortcut patterns in simple tasks.
• On WideSearch, the ISR-only, ISE-only, and ISR+ISE trajectory filters give comparable results within variance.

Limitations

• Relies on a single open-source model to generate trajectories and an LLM to synthesize questions, with generator identities not fully specified.
• Reported gains on WideSearch SR remain very low in absolute terms (4.0).

Experimental setup

Datasets: Wikipedia tables WebSailor-V2

Benchmarks: BrowseComp GAIA xbench-DeepSearch Seal-0 WideSearch

Metrics: pass@1 success rate row F1 item F1

Baselines: Claude-4-Sonnet OpenAI-o3 OpenAI DeepResearch ASearcher-Web-32B DeepDive-32B DeepDiver-V2-38B MiroThinker-32B-DPO-v0.2 Kimi-K2-Instruct-1T WebExplorer-8B WebDancer-QwQ-32B WebSailor-32B WebShaper-QwQ-32B

Builds on: ReAct GRPO WebSailor PPO

Closest papers by meaning

• Demystifying and Enhancing the Efficiency of Large Language Model Based Search Agents (ICLR 2026 · similarity 0.91)
• InfoMosaic-Bench: Evaluating Multi-Source Information Seeking in Tool-Augmented Agents (ICLR 2026 · similarity 0.90)
• InteractComp: Evaluating Search Agents With Ambiguous Queries (ICML 2026 · similarity 0.90)
• UIS-Digger: Towards Comprehensive Research Agent Systems for Real-world Unindexed Information Seeking (ICLR 2026 · similarity 0.90)
• Mind2Web 2: Evaluating Agentic Search with Agent-as-a-Judge (NeurIPS 2025 · similarity 0.89)
• Fathom-DeepResearch: Unlocking Long Horizon Information Retrieval and Synthesis for SLMs (ICLR 2026 · similarity 0.89)