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

Fathom-DeepResearch: Unlocking Long Horizon Information Retrieval and Synthesis for SLMs

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

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.

Builds Fathom-DeepResearch, two Qwen3-4B-based models: Fathom-Search-4B, trained with RL (RAPO, a GRPO extension) on a self-play-generated DuetQA dataset and a steerable step-level tool-call reward, and Fathom-Synthesizer-4B, SFT-trained to turn search traces into cited reports.

Problem

Open-source deep research agents lag proprietary ones; multi-turn RL with tools is unstable under GRPO, sparse correctness rewards cause redundant tool calling, training data lacks search-dependent multi-hop questions, and open-ended report synthesis is under-addressed.

Approach

RAPO adds dataset pruning of saturated prompts, inverse-frequency advantage scaling, and a per-prompt replay buffer to GRPO at zero extra rollout cost; a steerable step-level reward classifies each search/query call (unique, redundant, exploration, verification) to control tool-use breadth, depth and horizon; DuetQA is built via multi-agent self-play with live-web dependence verification.

Key results

• Fathom-Search-4B (Stage-2) averages 52.1% on the five DeepSearch benchmarks versus 47.3% for the best open-source baseline II-Search-4B, reaching 90.0 SimpleQA, 64.8 FRAMES, 50.0 WebWalker.
• On DeepResearch-Bench, Fathom-DeepResearch scores 45.47 overall RACE, outperforming listed open-source agents and several closed systems such as Claude-3.7-Sonnet w/Search (40.67) and Perplexity-DeepResearch (42.25).
• In the Stage-1 ablation, RAPO improves over GRPO on Seal0 (19.8 vs 14.4) and WebWalker (39.0 vs 33.8) while reducing average tokens from 9,000 to 5,000.

Surprising findings

• The steerable step-level reward raises token usage from 5,500 to 14,500 on average, indicating longer trajectories accompany the accuracy gains.
• Fathom-Search-4B Stage-2 is compared against o3 with search, which still leads on several DeepSearch benchmarks such as SimpleQA (96.0) and FRAMES (86.8).

Limitations

• RAPO shows limited test-time scaling; with vanilla reward in Stage 2 accuracy saturates before about 6,000 tokens.
• Reliance on trajectory replacement in the replay buffer anchors learning to low-entropy traces, which may hinder adaptation to longer reasoning horizons.
• Synchronous training pipelines are inefficient and brittle at scale.

Experimental setup

Datasets: DuetQA S1 math data MuSiQue training split DeepResearch-SFT SimpleQA FRAMES WebWalker Seal0 MuSiQue HLE AIME-25 GPQA-Diamond

Benchmarks: SimpleQA FRAMES WebWalker Seal0 MuSiQue HLE AIME-25 GPQA-Diamond MedQA DeepResearch-Bench

Metrics: Pass@1 accuracy RACE score FACT citation accuracy effective citation count average tokens

Baselines: o3 GPT-4o Jan-Nano II-Search-4B Qwen3-4B ZeroSearch Search-o1 R1-Searcher WebSailor GRPO

Builds on: GRPO ReCall DAPO ReAct

Closest papers by meaning

• HierSearch: A Hierarchical Enterprise Deep Search Framework Integrating Local and Web Searches (AAAI 2026 · similarity 0.90)
• Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization (NeurIPS 2026 · similarity 0.90)
• Empowering Efficiency and Efficacy in WebAgent via Enabling Info-Rich Seeking (ICLR 2026 · similarity 0.89)
• Quest: Training Frontier Deep Research Agents with Fully Synthetic Tasks (NeurIPS 2026 · similarity 0.89)
• WideSearch: Benchmarking Agentic Broad Info-Seeking (ICLR 2026 · similarity 0.89)
• DeepSearch: Overcome the Bottleneck of Reinforcement Learning with Verifiable Rewards via Monte Carlo Tree Search (ICLR 2026 · similarity 0.88)