Skip to content

FlowSearcher: Synthesizing Memory-Guided Agentic Workflows for Web Information Seeking

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

Why this paper was selected

Memory-guided agentic workflow synthesis for deep web information seeking

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

FlowSearcher reframes web information seeking as memory-guided agentic workflow synthesis, where queries are decomposed into subgoals and each subgoal is solved by a dynamically synthesized DAG workflow conditioned on retrieved past execution traces. Without any parameter fine-tuning or RLHF, it consistently outperforms strong ReAct-style baselines on GAIA, BrowseComp, and GPQA-Diamond.

Problem

Existing web research agents rely on ReAct-style linear tool-use chains that cannot branch, backtrack, or structurally revise search plans as evidence evolves. Additionally, agents operate in episode isolation—execution traces are discarded after each task, preventing any accumulation of reusable strategic experience across queries.

Method

FlowSearcher operates on a two-level hierarchical trajectory:

High level (orchestration): The agent iteratively decomposes query \(Q\) into subquestions \(\{\mu_i\}\) and synthesizes a typed DAG workflow \(G_i\) for each, drawn from a predefined library of building blocks (search, browse, summarize, ensemble, parallel, etc.). Workflow construction is conditioned on prior trajectory \(\Gamma_{<i}\) and memory \(M_{i-1}\):

\[P(\Gamma \mid Q, M_0) = \prod_{i=1}^{K} P(\mu_i \mid Q, \Gamma_{<i}, M_{i-1}, \theta_\mu)\, P(G_i \mid Q, \Gamma_{<i}, M_{i-1}, \theta_G, \mu_i)\]

Low level (execution): Each node \(v \in V(G_i)\) executes an action sequence against the web environment, factorized as \(P(\alpha, o \mid \mu_i, M_{i-1}) = \prod_{t=1}^{K_v} P(\alpha_t, o_t \mid \alpha_{<t}, o_{<t}, \mu_i, M_{i-1})\).

Structured Compositional Memory: A three-level hierarchy stores: - Node-level: node type/variant, action sequences \(\alpha^{(v)}\), and outputs \(o^{(v)}\) - Graph-level: DAG topology \(G_i\), success indicator \(\gamma_i \in \{0,1\}\), per-tool usage vector \(n_i \in \mathbb{N}^{|T|}\) - Task-level: full query \(Q_j\), task-success flag \(\xi_Q\), and all graph entries

Retrieval: A weighted cosine similarity over text embeddings balances original-query and subquery relevance with mixing factor \(\delta\):

\[R(Q^*, \mu^*; \zeta) = \arg\text{top-k}_{M} \left[ \delta \frac{E(Q^*)\cdot E(Q)}{|E(Q^*)||E(Q)|} + (1-\delta)\frac{E(\mu^*)\cdot E(\mu)}{|E(\mu^*)||E(\mu)|} \right]\]

Retrieved graph- and node-level traces are distilled into insights \(\xi_G\) and \(\xi_v\) by instructor modules and injected into orchestration and node-execution prompts, respectively.

Key Contributions

  • Two-level trajectory formulation that decouples high-level workflow synthesis from low-level node execution, enabling query-adaptive DAG planning distinct from flat ReAct chains.
  • A compact building-block library enabling flexible composition of multi-path search strategies (parallel, ensemble, swift-hit, break-down) without any supervised training.
  • A three-level Structured Compositional Memory (node / graph / task) with a unified weighted-embedding retrieval operator that injects past experience into both workflow orchestration and execution—achieving "co-optimization" without parameter updates.

Results

  • GAIA (103 text-only, Qwen-2.5-32B): FlowSearcher 48.5% avg vs. WebThinker-Base 40.7%, WebDancer 44.2% (+11.5% over WebThinker-Base reported in abstract aggregate).
  • GAIA (QwQ-32B): FlowSearcher 56.3% avg vs. WebThinker-RL 48.5%, WebDancer 50.0%.
  • BrowseComp avg (Qwen-2.5-32B): FlowSearcher 60.1% vs. WebThinker-Base 53.0% (+7.1pp); with QwQ-32B: 71.2% vs. WebDancer 70.7%.
  • GPQA-Diamond avg (Qwen-2.5-32B): FlowSearcher 5.6% vs. Vanilla ReAct 0.0%; with QwQ-32B: 8.0%; with GPT-4o-mini: 11.8%.
  • FlowSearcher (QwQ-32B) on GAIA Level 1: 69.2% vs. WebThinker-RL 50.0%; Level 2: 53.8% vs. WebThinker-RL 16.7%.
  • All FlowSearcher variants achieve BrowseComp scores \(\geq 1.9\%\) while all ReAct baselines (except WebDancer-32B) remain at 0–2.4%.
  • No RLHF or task-specific fine-tuning used; gains are purely from workflow synthesis and memory reuse.

Limitations

  • Memory retrieval quality degrades on truly novel query types with no prior similar task entries—cold-start scenario unaddressed.
  • Retrieval uses textual embeddings; structural similarity of workflow graphs is not directly encoded, relying on indirect proxy through subquery text.
  • BrowseComp and GPQA scores remain low in absolute terms (5–12% on GPQA), indicating fundamental headroom limited by backbone reasoning, not framework design.
  • SerpAPI and Jina Reader API are external paid services; reproducibility depends on access to these.
  • Evaluation restricted to text-only GAIA questions; multimodal web content excluded.

Relevance to Agentic AI / LLM Agents

FlowSearcher directly addresses one of the core open problems in agentic AI: how to escape the ReAct/linear-chain bottleneck and enable non-linear, revisable, experience-informed planning without retraining. The three-level memory hierarchy is a significant step toward agents that genuinely accumulate and reuse strategic knowledge across episodes, a capability that is largely absent from current tool-use agent frameworks. The decoupling of workflow synthesis (structure) from execution (behavior) establishes a compositional paradigm that complements RL-based approaches and is compatible with any LLM backbone. This work is directly relevant to researchers building long-horizon research agents, agentic workflow compilers (cf. AFLOW, AutoFlow), and memory-augmented agent architectures.