REDSearcher: Scalable Deep-Search Framework

• REDSearcher is a unified framework that co-designs task synthesis, midtraining, and posttraining to overcome the scarcity of high-quality search trajectories.
• It employs dual-constrained task synthesis and tool-augmented queries to generate complex, high-difficulty tasks while reducing reliance on costly live interactions.
• By integrating a simulated environment for low-cost reinforcement learning experiments, REDSearcher achieves state-of-the-art performance on deep-search benchmarks.

Searching arXiv for REDSearcher and closely related search-agent work to ground the article with current papers. REDSearcher is a unified framework for training long-horizon search agents. It is designed to address two bottlenecks in deep-search optimization: the scarcity of high-quality, challenging trajectories and the high cost of collecting such trajectories through live, interaction-heavy tool use. The framework co-designs complex task synthesis, midtraining, and posttraining, and is presented as a scalable and cost-efficient approach for optimizing LLMs for deep search tasks (Chu et al., 15 Feb 2026).

1. Definition and research setting

REDSearcher is situated in the transition of LLMs from general-purpose knowledge engines to real-world problem solvers. In that setting, deep search is treated not as passive recall from parametric memory but as a task regime in which optimization depends on high-quality search trajectories, reward signals, and repeated interaction with external tools (Chu et al., 15 Feb 2026).

The framework’s stated scope is long-horizon search-agent optimization. In the paper’s formulation, the central difficulty is not merely final-answer quality; it is the construction of scalable long-horizon tasks and the acquisition of sufficiently informative trajectories for downstream training. REDSearcher is therefore defined less as a single model than as an end-to-end training framework spanning task generation, capability preparation, and reinforcement-learning-oriented experimentation (Chu et al., 15 Feb 2026).

A useful way to characterize REDSearcher is as a pipeline for making deep-search supervision more abundant and less expensive. This suggests that the framework treats search competence as a data-generation and systems problem as much as an objective-design problem.

2. Core bottlenecks in deep-search optimization

The paper identifies an extreme sparsity problem: high-quality search trajectories and reward signals are difficult to obtain at scale. The source of that sparsity is twofold. First, long-horizon task construction is itself difficult to scale. Second, collecting trajectories is expensive because rollouts require interaction-heavy external tool calls (Chu et al., 15 Feb 2026).

These bottlenecks have direct consequences for training. If high-difficulty tasks are rare, the supervision signal for long-horizon behavior becomes thin. If each useful rollout depends on costly live interaction, posttraining iteration becomes slow and expensive. REDSearcher is explicitly designed against this backdrop, with each major component targeting one part of that dual constraint (Chu et al., 15 Feb 2026).

The paper’s problem framing is important because it shifts emphasis away from isolated prompting or retrieval heuristics and toward the economics of supervision. A plausible implication is that long-horizon search performance depends on the joint availability of difficult tasks, usable tool trajectories, and affordable experimental loops.

3. Framework architecture

REDSearcher introduces four named improvements. First, it frames task synthesis as a dual-constrained optimization. In that formulation, task difficulty is precisely governed by graph topology and evidence dispersion, with the goal of enabling scalable generation of complex, high-quality tasks (Chu et al., 15 Feb 2026).

Second, the framework introduces tool-augmented queries. Their stated purpose is to encourage proactive tool use rather than passive recall. This is a significant design choice: the search agent is not merely exposed to tools as optional peripherals, but is conditioned toward active interaction with them during search (Chu et al., 15 Feb 2026).

Third, during midtraining REDSearcher strengthens three atomic capabilities: knowledge, planning, and function calling. The paper states that this substantially reduces the cost of collecting high-quality trajectories for downstream training. In other words, midtraining is used as a capability-conditioning stage that makes later search rollouts cheaper to acquire and more likely to be useful (Chu et al., 15 Feb 2026).

Fourth, the framework builds a local simulated environment that enables rapid, low-cost algorithmic iteration for reinforcement learning experiments. This environment is part of the framework’s posttraining logic: instead of depending exclusively on expensive live deployment conditions, REDSearcher supports lower-cost experimentation and iteration inside a simulated setup (Chu et al., 15 Feb 2026).

Taken together, these components define REDSearcher’s central design principle: task synthesis, capability preparation, and reinforcement-learning experimentation are co-designed rather than optimized independently.

4. Data assets and research infrastructure

To facilitate future research on long-horizon search agents, the paper states that REDSearcher will release four classes of artifacts: 10K high-quality complex text search trajectories, 5K multimodal trajectories, 1K text RL query set, and code and model checkpoints (Chu et al., 15 Feb 2026).

The release plan is notable because it spans both text-only and multimodal settings. It also separates trajectory data from a text RL query set, indicating support not only for supervised imitation-style training but also for reinforcement-learning experimentation. This suggests an intent to establish REDSearcher as a reusable research substrate rather than only a one-off benchmark submission.

The emphasis on released trajectories is consistent with the paper’s diagnosis of the field. If high-quality long-horizon data are the main bottleneck, then public trajectory corpora become a primary research contribution in their own right.

5. Reported performance and benchmark position

Across both text-only and multimodal search-agent benchmarks, REDSearcher is reported to achieve state-of-the-art performance (Chu et al., 15 Feb 2026). The abstract does not enumerate benchmark-specific numbers, but it presents the result as evidence that the combined strategy of task synthesis, midtraining, and low-cost RL iteration is competitive at the frontier of long-horizon search-agent optimization.

A later comparison in "OpenSeeker-v2: Pushing the Limits of Search Agents with Informative and High-Difficulty Trajectories" reports RedSearcher-30B at 42.1% on BrowseComp, 49.8% on BrowseComp-ZH, and 34.3% on Humanity’s Last Exam, and also notes that RedSearcher trajectories average 36.01 steps (Du et al., 5 May 2026). In that comparison, OpenSeeker-v2 exceeds those scores with a simpler SFT-only recipe driven by high-difficulty trajectories, which places REDSearcher in the role of a strong baseline or reference system for subsequent work (Du et al., 5 May 2026).

This comparative position is informative. It indicates that REDSearcher helped define a performance frontier for long-horizon search agents, while later papers used it to test competing claims about training complexity, data difficulty, and trajectory length.

6. Relation to adjacent search-agent frameworks and naming ambiguity

REDSearcher belongs to a broader family of systems concerned with tool-using search agents, but neighboring frameworks emphasize different failure modes and remedies. "RE-Searcher: Robust Agentic Search with Goal-oriented Planning and Self-reflection" focuses on robustness in complex search environments, introducing explicit search goals and self-reflection to resist spurious cues and unstable trajectories (Fu et al., 30 Sep 2025). By contrast, REDSearcher’s emphasis is on scalable task synthesis, midtraining of atomic capabilities, and cost-efficient reinforcement-learning iteration (Chu et al., 15 Feb 2026).

OpenSeeker-v2 provides a different contrast. It argues that, when training trajectories are sufficiently informative and high-difficulty, a straightforward SFT objective is enough, and it explicitly compares against RedSearcher-30B as another 30B-scale search agent (Du et al., 5 May 2026). That comparison does not negate REDSearcher’s contribution; rather, it sharpens an active research question about whether the principal lever is training-pipeline complexity or trajectory quality.

The name itself also admits possible confusion. A separate line of work uses REDSearcher to denote a framework for redescription set construction in redescription mining rather than long-horizon search agents (Mihelčić et al., 2016). A plausible implication is that the term should be interpreted contextually: in current agent literature, REDSearcher refers to the 2026 framework for scalable and cost-efficient long-horizon search-agent training (Chu et al., 15 Feb 2026), whereas similarly named systems in other subfields address distinct problems and methodologies.