REDSearcher: A Scalable and Cost-Efficient Framework for Long-Horizon Search Agents

TL;DR

REDSearcher is a 30B-A3B (Qwen3-MoE) long-horizon search agent trained with a three-stage recipe — complex task synthesis → capability mid-training → SFT + Agentic RL post-training — that targets both text-only and multimodal deep-research benchmarks. Task synthesis builds dual-constrained difficulty over knowledge-graph DAGs (topological complexity × evidence dispersion) and runs a cascaded verifier pipeline to produce 20K+ trajectories at >85% accuracy, of which 10K text + 5K multimodal SFT samples are released. The text agent reports 42.1 / 57.4 (with Context Management) on BrowseComp, 49.8 / 58.2 on BrowseComp-zh, 80.1 on GAIA, and 33.3 on HLE; the multimodal variant tops MM-BrowseComp 23.5, BC-VL 57.2, and BrowseComp-VL 31.2, beating other 30B-class multimodal agents. The release covers SFT data (text + MM), an RL dataset demo, models (forthcoming), and the DeepTraceHub evaluation/synthesis framework.

Key claims

Three-stage pipeline: (1) Complex Task Synthesis via DAGs built from knowledge-graph entities and web-walk hyperlinks, with explicit difficulty control by topological complexity and Query Fuzzing (entity/attribute anonymization); (2) Capability Mid-Training to strengthen knowledge, planning, function calling, and long-horizon interaction; (3) SFT + Agentic RL post-training, with a local simulated environment for rapid algorithmic iteration [Project page, README §Overview].
The synthesis pipeline produces 20K+ verified trajectories at >85% accuracy via a cascaded verifier — LLM difficulty check, QA-graph alignment, Google retrieval verification, hallucination detection, agent rollout confirmation, and answer uniqueness validation [Project page Fig. 2].
Two open SFT datasets released: 10K text trajectories (Zchu/REDSearcher_SFT_10K) and 5K multimodal trajectories (honglyhly/REDSearcher_MM_SFT_5K), plus an RL dataset demo (Zchu/REDSearcher_RL_1K) [README §Timeline, §Data Processing].
SFT uses ms-swift 3.12.3 + Megatron-LM with qwen3_moe model type, TP=2 / EP=8 / PP=2, context-parallel=4, max length 131,072, lr 3.08e-5 cosine with 5% warmup, packing on, sequence parallel on, flash-attn on [README §Train].
Text-only deep-research results (30B-A3B, with Context Management figures marked ): BrowseComp 42.1 / 57.4, BrowseComp-zh 49.8 / 58.2*, GAIA 80.1, HLE 33.3, Overall 51.3 [Project page Table 1].
REDSearcher (30B-A3B) outperforms other 30B-A3B open deep-research agents on most axes: WebResearcher-30B (37.3 / 45.2 BrowseComp/-zh), WebSailorV2-30B (35.3 / 44.1, GAIA 74.1), Tongyi DeepResearch-30B (43.4 / 46.7, GAIA 70.9, HLE 32.9) [Project page Table 1].
Reported as competitive with larger open deep-research systems: GLM-4.7 (52.0 / 66.6* on BrowseComp), DeepSeek-V3.2 (51.4 / 67.6*), LongCat-Flash-Thinking (56.6 / 73.1*) at the price of substantially smaller active parameters [Project page Table 1].
Multimodal results: REDSearcher-MM-RL (30B-A3B class) reaches MM-BrowseComp 23.5, BC-VL 57.2, MMSearch+ 26.6, MMSearch 72.9, LiveVQA 79.3, HLE-Text 25.3, HLE-VL 25.6, BrowseComp 31.2, BrowseComp-ZH 44.5 — surpassing MMSearch-R1, WebWatcher, DeepEyesV2, Vision-DeepResearch baselines [Project page Table 2].
RL stage gains over MM-SFT-only baseline: BC-VL 55.3 → 57.2, MMSearch+ 20.2 → 26.6, MMSearch 70.3 → 72.9, BrowseComp 30.1 → 31.2 (small but consistent across 9 multimodal benchmarks) [Project page Table 2].
Trajectory synthesis and evaluation are factored into a separate open framework, DeepTraceHub, which deploys agent / summarizer / LLM-as-judge models via vLLM or SGLang and supports ReACT and DeepSeek Thinking-with-Tools reasoning loops [README §Trajectory Synthesis & Evaluation].

Method

The agent stack has three explicit subsystems. The Complex Task Synthesis stage extracts directed acyclic graphs (DAGs) from a knowledge graph (entities + relations) and from web-walk hyperlink trajectories. Each node is enriched with multi-source evidence; reasoning paths sampled from the DAG become candidate questions. Difficulty is jointly controlled by graph-topological complexity (number of hops, branching) and evidence dispersion (how spread out the supporting evidence is across the corpus). Query Fuzzing anonymizes entities/attributes to force the agent to actually retrieve rather than recall from parametric memory. A six-step cascaded verifier — LLM difficulty check, QA-graph alignment, Google retrieval verification, hallucination detection, agent rollout confirmation, answer uniqueness — gates outputs to >85% accuracy. Mid-training then continues pretraining on data designed to strengthen the atomic agentic capabilities (knowledge, planning, function calling, long-horizon interaction) before any task-specific SFT. Post-training runs SFT on the synthesized trajectories (Megatron-LM via ms-swift, qwen3_moe, 131K context, lr 3.08e-5) followed by Agentic RL on Slime, with a local simulated environment to drive iteration speed.

DeepTraceHub is the inference + evaluation harness. It deploys three model roles via vLLM or SGLang — the Agent (REDSearcher or a stronger synthesizer like GPT-OSS / DeepSeek-V3.2 for synthesis), a lightweight Summarizer (e.g. Qwen3-30B-A3B-Instruct) for web page compression, and an LLM-as-Judge (e.g. GPT-OSS) for automated scoring against ground truth. It supports ReACT and DeepSeek Thinking-with-Tools reasoning loops, multi-process execution, and emits trajectory JSONs formatted for both SFT and RL training together with aggregated statistics.

Results

On the four text-only deep-research axes, REDSearcher-30B-A3B reaches Overall 51.3 — between OpenAI-o3 (49.6) and GPT-5-Thinking-high (59.1), and competitive with Gemini-3-Pro (52.5) — while remaining one of the smallest active-parameter open agents in the comparison [Project page Table 1]. With Context Management it lifts BrowseComp 42.1 → 57.4 and BrowseComp-zh 49.8 → 58.2, narrowing the gap to closed-source deep-research systems. Against the direct 30B-A3B peers it beats Tongyi DeepResearch-30B on GAIA (80.1 vs 70.9) and HLE (33.3 vs 32.9) and is roughly comparable on BrowseComp (42.1 vs 43.4) and BrowseComp-zh (49.8 vs 46.7) without CM — and clearly ahead with CM. The multimodal variant produces consistent SFT→RL deltas of +1 to +6 absolute points across MM-BC, BC-VL, MMS+, MMS, LiveVQA, HLE-T, HLE-VL, BC, BC-ZH, with the largest gain on MMSearch+ (20.2 → 26.6, +6.4), showing the RL stage is not redundant with SFT even on synthesized multimodal trajectories [Project page Table 2].

Why it’s interesting

REDSearcher is the third filed 30B-A3B parameter-space deep-research agent on the wiki and the first one to commit to a multimodal variant trained with the same recipe — making it directly comparable to Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization (SMTL, Qwen3-30B-A3B with parallel agentic workflow) and OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis (Nemotron-3-Nano-30B-A3B with offline-retriever-only SFT). All three agree on the base-model size and the “synthesize trajectories + GRPO-family RL” template but pull on different design knobs: SMTL on per-step information density (parallel decomposition), OpenResearcher on environment cost (offline corpus eliminates search API spend), REDSearcher on task-synthesis difficulty control (DAG topology + evidence dispersion + Query Fuzzing) and a six-step cascaded verifier. The DAG-with-cyclic-constraints + Query Fuzzing recipe is structurally close to SMTL’s “rich-topology BFS subgraphs with cross-validation conditions” but explicitly adds entity/attribute anonymization as an anti-recall mechanism, which neither sibling reports. For Luma, the more transferable contribution is probably the cascaded verifier (six independent gates, with an agent rollout confirmation and answer uniqueness check), which addresses the “synthesized trajectories overfit to their synthesizer” open question raised on Tool-Use Agents — measuring quality via downstream agent solvability rather than just LLM-judge correctness. The multimodal variant also extends VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents‘s point that vision tool-use is its own design surface, but here with a search rather than perception/manipulation toolset.