FORT-Searcher: Synthesizing Shortcut-Resistant Search Tasks for Training Deep Search Agents
Summary
FORT-Searcher introduces a framework for synthesizing shortcut-resistant training data for deep search agents by identifying and mitigating four shortcut risks. The resulting agent, trained via supervised fine-tuning, achieves state-of-the-art performance among comparable open-source search agents.
View Cached Full Text
Cached at: 06/12/26, 02:52 AM
Paper page - FORT-Searcher: Synthesizing Shortcut-Resistant Search Tasks for Training Deep Search Agents
Source: https://huggingface.co/papers/2606.12087 Published on Jun 10
#2 Paper of the day Authors:
,
,
,
,
,
,
,
,
,
,
Abstract
A framework for creating shortcut-resistant training data for deep search agents by identifying and mitigating four shortcut risks in data synthesis processes.
Trainingdeep search agentsrequires verifiable questions whose answers remain unavailable until sufficient evidence has been acquired through search. Existing synthesis methods often increase apparent difficulty by enriching graph structures, but structural complexity alone does not guarantee realized search difficulty: the intended search process can collapse through a cheaper identifying route. We formalize this gap with ashortcut-aware difficulty frameworkand identify four actionableshortcut risks: evidence co-coverage, single-clue selectivity, exposed constants, and prior-knowledge binding. To diagnose their realized effects, we usetrajectory signaturesincluding solving cost, answer hit time, and prior-shortcut rate. Guided by this framework, we introduceFORT, a Framework of Shortcut-Resistant Training-Data Synthesis.FORTconstructs shortcut-resistant training data by controllingshortcut risksacross entity selection, evidence graph construction, question formulation, and adversarial refinement. Experiments show thatFORTinduces longer pre-answer search and fewer shortcut patterns than existing open-source deep search datasets. Using the resulting trajectories, we trainFORT-Searcher withsupervised fine-tuning(SFT) only, and it achieves the best overall performance among comparable-size open-source search agents on challengingdeep search benchmarks. Relevant resources will be made available at https://github.com/RUCAIBox/FORT-Searcher.
View arXiv pageView PDFGitHub4Add to collection
Get this paper in your agent:
hf papers read 2606\.12087
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2606.12087 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2606.12087 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2606.12087 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
SlimSearcher: Training Efficiency-Aware Web Agents via Adaptive Reward Gating
SlimSearcher is a framework that improves efficiency in deep research agents by combining Pareto-efficient trajectory filtering and adaptive reward shaping, reducing tool-call rounds by 17-58% while maintaining accuracy on benchmarks like GAIA, BrowseComp, and XBenchDeepSearch.
QUEST: Training Frontier Deep Research Agents with Fully Synthetic Tasks
QUEST is an open family of deep research agents trained with synthetic data and reinforcement learning, achieving strong performance across diverse long-horizon search tasks, approaching frontier closed-source agents.
Auto-FL-Research: Agentic Search for Federated Learning Algorithms
Auto-FL-Research introduces a constrained coding-agent workflow for automatically searching and evaluating federated learning algorithmic recipes, showing performance gains on multiple healthcare and LEAF tasks while also exposing seed-sensitive and search-selected failure cases.
S1-DeepResearch: Beyond Search, Toward Real-World Long-Horizon Research Agents
This paper introduces S1-DeepResearch-32B, an open-source model and 15K trajectory dataset for deep research agents, achieving state-of-the-art performance across 20 benchmarks by jointly modeling information acquisition, knowledge synthesis, and planning.
FastContext: Training Efficient Repository Explorer for Coding Agents
FastContext introduces specialized exploration models that separate repository exploration from code solving in LLM agents, reducing token consumption by up to 60% while improving resolution rates on software engineering benchmarks.