Skip to content

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

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Efficient deep research for small LMs; long-horizon synthesis without large model dependency

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Fathom-DeepResearch is an end-to-end agentic deep-research system composed of two 4B-parameter models (Fathom-Search-4B and Fathom-Synthesizer-4B) built on Qwen3-4B that enables reliable long-horizon web-search-augmented reasoning beyond 20 tool calls. The system introduces RAPO (a stabilized multi-turn RL algorithm), DUETQA (a search-dependent synthetic QA dataset), and a steerable step-level reward to control exploration/verification behavior. It achieves state-of-the-art performance among open-weights DeepResearch systems on multiple benchmarks.

Problem

Open-source DeepResearch agents suffer from four compounding failures: (1) GRPO training instability when tool responses induce distribution shift, saturating group-relative advantages and causing malformed generations; (2) reward hacking under correctness-only sparse signals, collapsing into repetitive identical tool calls; (3) training datasets (TriviaQA, HotpotQA, 2WikiMH) that are solvable from parametric knowledge without live search; (4) overemphasis on closed-form QA at the expense of open-ended synthesis requiring evidence grounding and multi-perspective integration.

Method

Fathom-Search-4B is trained via a two-stage RLVR pipeline on top of the RECALL framework:

  • DUETQA dataset: 5K multi-hop QA pairs generated by two frontier search-enabled LRMs (O3, O4-mini) acting as generators, with a non-search model (GPT-4o) for obfuscation and baseline verification. Questions enforce live-web-search dependency (\(P(a|q,M_{\text{no-search}}) \ll P(a|q,M_{\text{search}})\)) and are grounded in \(k \in [5,7]\) themes from a 200+ taxonomy. Two independent search-enabled verifiers must agree while the non-search model fails.

  • RAPO (Reward-Aware Policy Optimization): A zero-overhead extension of GRPO addressing "bad groups" (zero reward variance \(\sigma_R=0\)). Three modifications: (i) dataset pruning — drop prompts with SolveRate \(\geq 0.9\); (ii) advantage scaling — rescale token-level advantages for good groups by \(\frac{G}{G_{\text{good}}}\); (iii) replay buffer — if all rollouts fail, inject a cached successful trajectory \(o^*\) with \(R(q,o^*)>0.5\) to restore variance.

  • Steerable Step-Level Reward: An LLM-as-judge (GPT-4.1) classifies each tool call as UNIQUE SEARCH, REDUNDANT SEARCH, EXPLORATION, VERIFICATION, or REDUNDANT QUERY. The reward for correct rollouts penalizes redundancy via \(\rho = (n_{\text{redS}}+n_{\text{redQ}})/T\); for incorrect rollouts it credits diverse exploration via \(\Delta_S\) and \(\Delta_Q\): $\(r_i = \begin{cases} 0.1 R_i^{\text{format}} + \max\!\bigl((1-\rho),\,0.5\bigr) & \text{if } a_{\text{pred}}^{(i)} = a_{\text{gt}} \\ 0.1 R_i^{\text{format}} + c_1 \min\!\bigl(1,\tfrac{\Delta_S}{C_S}\bigr) + c_2 \min\!\bigl(1,\tfrac{\Delta_Q}{C_Q}\bigr) & \text{otherwise} \end{cases}\)$ with \(c_1=c_2=0.2\), \(C_S=8\), \(C_Q=16\), \(B_v=1\) cross-check per claim.

  • Training: Stage 1 — 10 epochs on DUETQA with vanilla RAPO reward. Stage 2 — 2 more epochs on a mixed dataset (DUETQA + math + MuSiQue + MedQA, adversarially filtered to 5,471 instances) using the steerable reward to push tool-use horizon beyond 20 calls.

Fathom-Synthesizer-4B is SFT-trained on DEEPRESEARCH-SFT (2,500 samples distilled from GPT-5) using a Plan-then-Write protocol: the model first produces a private <think> plan \(\pi = (\pi^{\text{decomp}}, \pi^{\text{map}}, \pi^{\text{insight}})\), then generates a citation-dense public report. YaRN RoPE scaling (factor 2.0) extends the context to 65,536 tokens.

Key Contributions

  • DUETQA: 5K open-source search-dependent multi-hop QA dataset with multi-agent self-play generation and cross-model verification
  • RAPO: Zero-overhead GRPO modification with dataset pruning, advantage scaling, and replay buffers that stabilizes multi-turn RL
  • Steerable Step-Level Reward: Classifies tool calls by cognitive behavior and marginal utility, enabling explicit control over search trajectory breadth, depth, and horizon
  • RL-Zero two-stage training framework: Staged curriculum that first instills format adherence, then extends the tool-use horizon steerably
  • DEEPRESEARCH-SFT: Synthetic SFT corpus for converting search traces into structured citation-dense reports via explicit plan-then-write
  • Fathom-Search-4B and Fathom-Synthesizer-4B: Open-weights 4B models achieving SOTA in the open-source DeepResearch category

Results

DeepSearch benchmarks (Fathom-Search-4B Stage-2 vs. best open-source competitor II-Search-4B): - SimpleQA: 90.0 vs. 88.2 - FRAMES: 64.8 vs. 58.7 - WebWalker: 50.0 vs. 40.8 - Seal0: 22.5 vs. 17.1 - MuSiQue: 48.9 vs. 31.8 - Average DeepSearch: 55.2 vs. 47.3

General reasoning (Fathom-Search-4B Stage-2): - HLE: 9.5, AIME-25: 70.0, GPQA-Diamond: 75.4, MedQA: 53.8 (avg 52.1)

DeepResearch-Bench (Fathom-DeepResearch overall score): - 42.98 overall — matches Kimi-Researcher (44.96) and Doubao-DeepResearch (44.84) in the open-source category - Surpasses GPT-4o-Search-Preview (35.10) and GPT-4.1 w/Search (33.46) - Trails closed-source DeepResearch agents: Gemini-2.5-Pro DeepResearch (48.88), OpenAI-DeepResearch (46.98)

Efficiency: Fathom-Search-4B achieves higher accuracy at fewer average tool calls than contemporaries (per Fig. 1).

Limitations

  • Fathom-Synthesizer-4B lags best closed-source systems on citation accuracy (E. Cit. metric: 38.3 vs. 111.21 for LangChain Open-DeepResearch) and Relevance/Completeness dimensions
  • DUETQA generation relies on frontier models (O3, O4-mini, GPT-4o) making the pipeline costly despite claimed scalability improvements
  • The steerable step-level reward requires a GPT-4.1 LLM-as-judge for classifying every tool call during training, adding inference cost
  • Evaluation on DeepResearch-Bench shows the system ties with or is slightly below Kimi-Researcher and Doubao-DeepResearch in most sub-dimensions
  • Context window extended via YaRN (to 65,536) rather than natively trained; potential degradation on very long traces not fully characterized
  • All results use Serper API for live web search, creating a runtime dependency not available offline

Relevance to Agentic AI / LLM Agents

This work directly advances the core open problem in agentic LLMs: how to train a small model to reliably execute long-horizon, multi-step tool-use without reward hacking or training instability. RAPO and the steerable step-level reward are general-purpose RL stabilization techniques applicable beyond search to any multi-turn tool-augmented agent. The DUETQA pipeline demonstrates a scalable approach to constructing verifiable, search-forcing training data — a critical bottleneck for agentic RLVR. The Plan-then-Write synthesizer pattern (Fathom-Synthesizer-4B) is directly relevant to the broader shift from retrieval-augmented generation toward deep-research agents capable of structured, citation-grounded synthesis, bridging DeepSearch and report-generation into a unified agentic system.