Search More, Think Less: Rethinking Long-Horizon Agentic Search for Efficiency and Generalization

TL;DR

SMTL (“Search More, Think Less”) is a 30B-A3B (Qwen3-30B-A3B-Instruct-2507) deep-research agent that replaces sequential ReAct reasoning with a parallel agentic workflow: at every step the agent maintains a task plan, dispatches multiple subtasks concurrently to a small toolset (web search + page crawl), and periodically refreshes the plan, dropping pre-plan context when the 128K window overflows. The framing thesis is the opposite of MiroThinker’s “interaction depth is a third scaling axis” claim — SMTL argues the bottleneck is information density per step, not number of steps, and shows the agent matches or beats MiroThinker-v1.0 on BrowseComp/GAIA/Xbench/WebWalker while reducing the average number of reasoning steps on BrowseComp by 70.7%. A unified data-synthesis pipeline (LightRAG knowledge graph → controlled BFS subgraphs → hierarchical question construction with cross-validated constraints) produces both deterministic deep-search QA and open-ended deep-research questions, enabling a single agent that hits 48.6 / 75.7 / 82.0 / 76.5 / 85.1 / 51.4 on BrowseComp / GAIA / Xbench-DS / WebWalker / FRAMES / SEAL-0 and 45.9 overall on Deep Research Bench RACE.

Key claims

The “scaling bottleneck” of current deep-research agents is reliance on linear, sequential reasoning, not the underlying base model — parallel task decomposition + concurrent tool execution + plan-driven context management is proposed as a different scaling regime [§1, §3].
Parallel agentic workflow: at each timestep, the agent picks ready-to-execute subtasks from a pending set and executes them concurrently; observations are aggregated into a unified reasoning state $s_{t+1} = \mathrm{Aggregate}(\{(a_i, o_i)\}_{i=1..N})$ [§3, Eq. 1].
Dynamic plan refinement runs every N=5 interaction steps by default, removing completed subtasks and adding new ones; on 128K-context overflow without a confirmed answer, the agent does a forced plan refinement and drops all pre-plan context, continuing from the refreshed plan [§6, Context Management during Inference].
Two-track data synthesis from one pipeline: deterministic Deep Search QA from hierarchical question construction over BFS subgraphs (2–5 hops) of a LightRAG-built knowledge graph, plus open-ended Deep Research questions over the same subgraphs with report-style answers [§4.1, §4.2].
Subgraph design enforces “rich topology” — two k-hop nodes that share a common (k-1)-hop parent, inducing cyclic dependencies that require cross-validation rather than linear chains [§4.1].
Training is SFT (3.5 epochs, batch 128, lr 5e-6, cosine, 64K seq) → RL with a modified REINFORCE Leave-One-Out (RLOO): token-level loss (from DAPO), sequence-level importance sampling for inference/training mismatch correction, and explicit filtering of environment-failure and length-overflow trajectories from the advantage estimator [§5.1, §5.2].
SFT distillation sources differ by task type: Deep Search trajectories distilled from DeepSeek-V3.2, Deep Research trajectories from GPT-5; SFT keeps trajectories ≤64K tokens, requires ≥3 tool calls per step on average, and for tasks with multiple correct trajectories keeps only the shortest [§5.1].
SMTL-100 (30B-A3B, 100-step budget) reaches BrowseComp 43.6 / GAIA 74.8 / Xbench-DS 80.0 / WebWalker 74.9 / FRAMES 84.3 / SEAL-0 50.5 — beating Tongyi-DeepResearch-30B (43.4 / 70.9 / 75.0 / 72.2 / 90.6) on most axes and MiroThinker-v1.0-30B (41.2 / 73.5 / 70.6 / 61.0 / 85.4 / 46.8) across the board [Table 1].
SMTL-300 (300-step budget) lifts BrowseComp from 43.6 → 48.6 (+5.0) and Xbench-DS from 80.0 → 82.0 [Table 1, §6.2].
Compared to MiroThinker-v1.0, SMTL with a 100-step max reduces the average number of reasoning steps on BrowseComp by 70.7% while improving accuracy [Abstract, §6.2]. Authors elsewhere claim 78% step reduction and up to 2.6× inference-latency reduction [§1].
On Deep Research Bench RACE, SMTL-100 scores 45.9 overall (Comp 42.1 / Depth 45.6 / Inst 49.6 / Read 45.5), competitive with Kimi-Researcher (44.6) and Tongyi-DeepResearch-30B (45.7) [Table 1].

Method

The agent’s loop is plan → parallel-execute → refine. Given a query, the model first generates an initial task plan that decomposes the problem into a set of interrelated yet partially independent subtasks; the plan is produced before any tool call and is explicitly designed to expose parallelizable execution paths early. At each timestep, ready-to-execute subtasks (those whose dependencies are satisfied) are pulled from the pending set and run concurrently against a small reusable toolset (web search + page crawling — Appendix A). The aggregated tool observations update a unified reasoning state, and every N=5 steps the agent emits a refined plan that removes completed subtasks, re-checks unresolved dependencies, and may add new subtasks. When accumulated history reaches the 128K context budget without a confirmed final answer, the agent triggers a forced plan refinement and drops everything before the new plan — preserving the latest execution state and subtask structure while resetting the context budget. This plan-centric reset is what lets SMTL run 100–300 step trajectories under a fixed 128K window.

Data is built from a single subgraph-extraction pipeline. Raw text is collected by following multi-hop URL trajectories from the TaskCraft corpus and extracted with Jina; LightRAG instantiates a knowledge graph over the chunks; for each task, a target entity is sampled as ground-truth answer and BFS up to k hops produces a subgraph whose multi-hop nodes serve as question conditions. Hierarchical question construction starts from the k-hop frontier and aggregates outward; when multiple k-hop nodes are semantically related, those interdependencies are encoded as verifiable conditions that force cross-validation across parallel evidence paths. An LLM verifier checks for information leakage after each iteration. The same subgraphs feed Deep Search (hierarchical QA over hop-depth 2–5) and Deep Research (open-ended report-style questions). RL uses modified RLOO with token-level loss, sequence-level importance sampling for training/inference log-prob mismatch correction, and explicit removal of environment-failure and length-overflow trajectories from the advantage estimator; the reward is binary correctness as judged by an LLM, with format-violating tool calls auto-zeroed.

Results

SMTL-100 (30B-A3B) on Deep Search: BrowseComp 43.6 / GAIA 74.8 / Xbench-DS 80.0 / WebWalker 74.9 / FRAMES 84.3 / SEAL-0 50.5 — state of the art among 30B-scale open-source agentic models on BrowseComp (vs Tongyi-DeepResearch-30B 43.4, MiroThinker-v1.0-30B 41.2), Xbench-DS (vs 75.0 / 70.6), and WebWalker (vs 72.2 / 61.0). Pushing the budget to 300 steps lifts BrowseComp to 48.6 (+5.0) and Xbench-DS to 82.0 [Table 1]. On Deep Research Bench RACE, SMTL-100 scores 45.9 overall, ahead of Kimi-Researcher (44.6) and Minimax-M2 (46.1) on several sub-axes [Table 1]. The headline efficiency claim: SMTL with a 100-step budget cuts the average number of reasoning steps on BrowseComp by 70.7% vs MiroThinker-v1.0 while improving accuracy; the intro reports a 78% step reduction and up to 2.6× inference-latency reduction [Abstract, §1]. SMTL also competes with several closed-source deep-research systems (Gemini DeepResearch 59.2 on BrowseComp, OpenAI DeepResearch 51.5) on most metrics despite the much smaller backbone.

Why it’s interesting

SMTL is the cleanest published counter-thesis to MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling‘s “interaction depth is a third scaling axis” framing. Both papers train mid-size (30B / 72B) agents on synthesized multi-hop QA + GRPO-family RL and head-line BrowseComp; MiroThinker argues that lifting the per-task tool-call ceiling from <100 to ~600 is the dominant gain, while SMTL argues that steps per task is the wrong axis — what matters is information density per step (≥3 tool calls per step on average in the SFT filter, parallel decomposition at runtime), and reducing the average step count by 70-78% at iso-accuracy is the actual win. The two recipes coexist (SMTL’s plan-refinement-on-overflow is itself a kind of context-management trick MiroThinker would call “interaction-depth headroom”), but the framing disagreement is real and belongs as a **Conflict:** on Tool-Use Agents. SMTL also complements OpenResearcher: A Fully Open Pipeline for Long-Horizon Deep Research Trajectory Synthesis (the offline-corpus + SFT-only counterpart at the same 30B-A3B scale, on a different benchmark) and Close the Loop: Synthesizing Infinite Tool-Use Data via Multi-Agent Role-Playing (self-synthesized trajectories + GRPO for general tool-use, not deep research). The data-synthesis recipe — LightRAG knowledge graph + BFS subgraphs with cyclic-dependency design + hierarchical question construction with verifier — is a cleaner reusable blueprint than MiroThinker’s MultiDocQA pipeline for any Luma task where ground truth is checkable but multi-hop. The base model is the same Qwen3-30B-A3B-Instruct-2507 that several other deep-research recipes are converging on, which makes the comparison apples-to-apples in a way the field rarely is.