Assess-then-Search Workflow

Updated 5 July 2026

Assess-then-Search Workflow is a control paradigm that evaluates task sufficiency using explicit state representations before triggering search actions.
It is applied across domains such as legal question answering, multimodal search, and literature retrieval to iteratively refine search strategies based on evidence.
Empirical results indicate enhanced search quality and budget efficiency, albeit with increased orchestration complexity and added latency.

Assess-then-Search Workflow is a decision-first search paradigm in which a system evaluates sufficiency, uncertainty, validity, risk, or task state before committing budget to retrieval, exploration, browsing, scanning, or other search actions. In recent arXiv work, it appears as a recurring control principle across deep web search, legal question answering, multimodal visual search, scientific literature search, analytical search, interactive search guidance, and long-horizon scientific automation. Despite substantial variation in tools and domains, the shared structure is stable: assess the current state, choose or refine a search strategy, execute targeted search, update state, and repeat until a finish or sufficiency condition is met (Shi et al., 10 Jun 2026, Wang et al., 31 Aug 2025, Li et al., 22 May 2026, Li et al., 1 Jul 2026, Tu et al., 12 Feb 2026).

1. Conceptual definition and contrast classes

Assess-then-Search is explicitly defined in several systems as the principle of evaluating what is missing, uncertain, invalid, or still unresolved before launching additional search. In legal question answering, it is “the principle of first evaluating what is missing, uncertain, or potentially invalid in the current evidence for a legal query, and only then triggering targeted retrieval to close those gaps” (Wang et al., 31 Aug 2025). In deep web search, the same pattern is implemented by alternating an assessment phase over branch-local search trees with a search phase that commits budget to exploit, explore, or prune actions (Shi et al., 10 Jun 2026). In high-resolution image perception, the scheduler first asks whether the current visual context is sufficient, then attempts expert-assisted search, and only on expert failure triggers semantic-aware scanning (Li et al., 22 May 2026).

This pattern is usually positioned against three alternatives. The first is greedy continuation, in which an agent simply extends the currently best-looking trajectory; TreeSeeker argues that this can repeatedly extend weak continuations and therefore introduces explicit branch-and-return control (Shi et al., 10 Jun 2026). The second is undisciplined exploration, where alternatives are tried without a disciplined controller; TreeSeeker contrasts its risk-aware balancing of exploit, explore, and prune with budget dispersion on disconnected trials (Shi et al., 10 Jun 2026). The third is search-then-assess or single-pass RAG, in which broad retrieval occurs before sufficiency and validity are checked; L-MARS contrasts its iterative Judge-gated loop with both one-shot retrieve-then-summarize pipelines and workflows that verify only after synthesis (Wang et al., 31 Aug 2025).

Analytical Search generalizes the same idea at the system level by reframing search as “an evidence-governed, process-oriented analytical workflow that explicitly models analytical intent, retrieves evidence for fusion, and produces verifiable conclusions through structured, multi-step inference” (Tu et al., 12 Feb 2026). This suggests that Assess-then-Search is not merely a prompt pattern but a broader control architecture for search systems with high accountability, large search spaces, or delayed verification.

2. Control-loop structure and state representations

A defining property of Assess-then-Search systems is that assessment is performed over an explicit, persistent representation of task state rather than over a single transient prompt. TreeSeeker decomposes a root query into sub-goals $G=\{g_1,\dots,g_K\}$ , a goal dependency DAG $D_G$ , and candidate paths $P_i$ , then maintains one sub-goal tree $T_i$ per goal inside TreeMem. Each tree contains a root goal state, first-level path branches, and deeper trace nodes, so evidence, conflicts, uncertainty, progress, and failure cues remain attached to the branch that produced them (Shi et al., 10 Jun 2026).

PaperPilot formalizes literature search as workflow induction over an executable DAG $W=(V,E)$ , where each node is an instantiated operator $v=(o,\theta)$ and each edge represents typed data flow. The multi-turn loop induces a workflow $G_t$ , executes it, presents assessed results, incorporates feedback, and refines the workflow itself via $G_{t+1}=\mathrm{Refine}(G_t,f_t,H_{t+1})$ (Li et al., 1 Jul 2026). L-MARS uses a centralized WorkflowState with deterministic, type-checked transitions, holding the evolving query, accumulated SearchResult objects, iteration history, and agent outputs (Wang et al., 31 Aug 2025). Search-o1 uses a distinct but related state split: the main reasoning chain, search-query markers, retrieved documents, and a separate Reason-in-Documents stream that returns a concise “Final Information” snippet to be reinserted into the main reasoning process (Li et al., 9 Jan 2025).

These systems differ in granularity but share a common design choice: assessment requires inspectable state. In TreeSeeker the relevant unit is a branch-local subtree; in PaperPilot it is a typed operator graph; in L-MARS it is a workflow state with evidence and jurisdictional metadata; in Search-o1 it is a reasoning prefix plus retrieved-document analysis (Shi et al., 10 Jun 2026, Li et al., 1 Jul 2026, Wang et al., 31 Aug 2025, Li et al., 9 Jan 2025). This suggests that explicit state is the mechanism by which Assess-then-Search becomes controllable, auditable, and revisable.

3. Assessment signals, scoring functions, and sufficiency gates

Assessment is implemented through explicit signals, checklists, or scoring rules. TreeSeeker uses operation-level textual UCB over candidate operations $a\in\{\mathrm{Exploit},\mathrm{Explore},\mathrm{Prune}\}$ , with

$\phi(a,\beta_a,s_a)=(V_a,U_a,R_a), \qquad V_a,U_a,R_a\in\{\mathrm{Low},\mathrm{Medium},\mathrm{High}\},$

and the parameter-free score

$D_G$ 0

where the discrete mapping is $D_G$ 1. Assessment therefore combines expected progress, expected information gain, and risk of wasted budget, and it also enforces a finish gate if evidence already suffices for the root task (Shi et al., 10 Jun 2026).

L-MARS operationalizes assessment through a Judge Agent that checks source authority, date recency, jurisdiction match, and contradictions, and then returns a binary sufficiency decision. Its explicit uncertainty metric is

$D_G$ 2

where $D_G$ 3 denote hedging, temporal vagueness, citation sufficiency, jurisdictional specificity, and decisiveness; lower is better (Wang et al., 31 Aug 2025).

CVSearch defines sufficiency at the view level as

$D_G$ 4

with a default threshold $D_G$ 5. If $D_G$ 6, it answers directly; otherwise it triggers search. During dynamic bottom-up search it uses a decaying threshold

$D_G$ 7

and prioritizes nodes by

$D_G$ 8

where $D_G$ 9 is Visual Complexity, $P_i$ 0 is existence confidence, and $P_i$ 1 is max child priority (Li et al., 22 May 2026).

SRR-Judge moves assessment to the step level of search-integrated reasoning. At step $P_i$ 2, the judge returns $P_i$ 3, where $P_i$ 4, and the policy proceeds with the original step only if the best candidate satisfies $P_i$ 5 with $P_i$ 6 (Zhang et al., 8 Feb 2026). The judged dimensions are clarity and conciseness, logical structure, query appropriateness or answer faithfulness, and coverage and improvement potential (Zhang et al., 8 Feb 2026).

Across these implementations, assessment is not limited to confidence estimation. It can mean sufficiency gating, risk scoring, structural validation, legal authority checking, image-existence probing, or step-level quality rating (Shi et al., 10 Jun 2026, Wang et al., 31 Aug 2025, Li et al., 22 May 2026, Zhang et al., 8 Feb 2026).

4. Search actions after assessment

Once assessment identifies what remains unresolved, search is executed in a targeted rather than indiscriminate form. TreeSeeker commits budget to three operations: Exploit, which deepens a promising branch; Explore, which tests uncertain sibling paths; and Prune, which compresses a failed continuation into a failure cue and returns to an earlier branch point. Guardrails enforce backtracking when a path has been re-selected many times without useful progress: consider backtrack at $P_i$ 7 prior selections and must backtrack at $P_i$ 8 with unresolved looping (Shi et al., 10 Jun 2026).

L-MARS triggers retrieval when the Judge flags insufficiency or a query classifier predicts that external evidence will reduce uncertainty. Search is then source-specific and adaptive: Serper web search for breadth and deduplication, enhanced content extraction for depth, CourtListener for case law, and a local BM25 index for curated documents (Wang et al., 31 Aug 2025). Search-o1 similarly lets the model emit search queries when it encounters an “uncertain knowledge point,” then performs web retrieval and a separate document-reasoning pass before reintegrating only a concise “Final Information” result into the reasoning chain (Li et al., 9 Jan 2025).

CVSearch uses a staged search policy. If global context is insufficient, it first performs expert-assisted search using SAM 3 proposals. If the proposals are empty or fail category-wise count verification, semantic-aware scanning is triggered. That scanning phase combines Semantic Guided Adaptive Patching with Dynamic Bottom-Up Search, prunes low-complexity regions using $P_i$ 9 with default $T_i$ 0, and iterates from fine to coarse until sufficiency is re-established (Li et al., 22 May 2026).

PaperPilot’s search operators include keyword_search, citation_expand, and web_resolve, while assessment operators include filter, score, top_k, above, llm_rerank, nli_filter, fine_read, and extract_evidence. User feedback edits the workflow itself, such as switching citation direction to forward, adding filter(year ≥ 2022), or changing reranking criteria (Li et al., 1 Jul 2026). Analytical Search describes a similar but more general retrieval module: recall-oriented multi-path retrieval across structured and unstructured sources, with explicit counterevidence and control retrieval for downstream verification (Tu et al., 12 Feb 2026).

5. Domain-specific instantiations

The workflow recurs across domains, but the assessment object and the search action differ substantially.

System	Assessment stage	Search stage
TreeSeeker	Textual UCB over branch-local states	Exploit, Explore, Prune/return
L-MARS	Judge checks sufficiency, jurisdiction, temporal validity	Targeted Serper, CourtListener, local BM25 retrieval
CVSearch	Global sufficiency and expert-failure checks	Expert proposals, then semantic-aware scanning
PaperPilot	Workflow and candidate-set assessment under evolving intent	Keyword search, citation expansion, evidence extraction
Search-o1	Self-triggered uncertainty in chain-of-thought	Web retrieval plus Reason-in-Documents
TSAgent	Multimodal analysis of convergence diagnostics and geometry	Replanned GO/NEB/VFA search actions

TreeSeeker’s domain is deep web search over sub-goal DAGs (Shi et al., 10 Jun 2026). L-MARS applies the pattern to legal QA, where assessment emphasizes authority, jurisdiction, and temporal validity (Wang et al., 31 Aug 2025). CVSearch applies it to high-resolution image perception, where assessment decides whether the global view, expert proposals, or semantic scanning should consume the next compute budget (Li et al., 22 May 2026). PaperPilot applies it to multi-turn scientific literature search by turning feedback into edits of an executable workflow DAG (Li et al., 1 Jul 2026). Search-o1 uses it inside long-chain reasoning, where search is a conditional response to knowledge insufficiency rather than a fixed retrieval prelude (Li et al., 9 Jan 2025). TSAgent transfers the same logic to DFT-level transition-state search, where the assessment side is a multimodal analysis of scalar convergence diagnostics and atomic geometries, and the search side consists of replanned GO, NEB, and VFA steps (Madhavan et al., 13 May 2026).

The breadth of these applications supports the view that Assess-then-Search is a general control pattern for problems in which search is costly, delayed, or failure-prone.

6. Empirical performance and observed effects

Reported results consistently associate explicit assessment with improved search quality or budget efficiency. TreeSeeker on XBench-DeepSearch, BrowseComp, and BrowseComp-ZH achieves 56.3, 47.0, and 43.0, compared with Flash-Searcher at 50.7, 43.0, and 40.3, and its ablations show drops to 52.0 without Textual UCB, 48.0 without Explore & Prune, and 51.3 without TreeMem leaf trace (Shi et al., 10 Jun 2026). Its operation frequencies on XBench-DS are Exploit 51.39%, Explore 43.45%, and Prune 5.17%, whereas removing Textual UCB shifts the distribution to Exploit 36.93%, Explore 61.08%, and Prune 1.98% (Shi et al., 10 Jun 2026).

L-MARS on LegalSearchQA reaches accuracy 0.98 and U-Score 0.39 in multi-turn mode, compared with GPT-4o at 0.89 and 0.55 and with its own simple mode at 0.96 and 0.42; the gain comes with higher latency, 55.67s versus 13.62s in simple mode and 1.69–3.84s for baseline LLMs (Wang et al., 31 Aug 2025). CVSearch shows the same trade-off in multimodal search: on Qwen2.5-VL-7B it reports 90.1 accuracy and 1.02 throughput on V* Bench, 76.6 and 3.77 on HR-Bench 4K, and 75.6 and 1.92 on HR-Bench 8K, outperforming both expert-only and rigid scan-based alternatives in accuracy while remaining substantially faster than ZoomEye and RAP scanning (Li et al., 22 May 2026).

PaperPilot-9B improves over the base Qwen3.5-9B toolset agent under multi-turn interaction, increasing Hit@5 from 58.0 to 77.0, MRR from 47.5 to 59.4, nDCG@10 from 26.8 to 32.5, and reducing workflow execution errors from 9.5% to 0% (Li et al., 1 Jul 2026). SRR-Judge reports inference-time gains for QwQ-32B from 5.4 to 9.1 on BrowseComp, 23.4 to 29.2 on BrowseComp-ZH, and 46.0 to 53.3 on Xbench, while iterative SRR-guided fine-tuning reaches 16.2, 38.3, and 61.3 on those benchmarks (Zhang et al., 8 Feb 2026). TSAgent, in a scientific automation setting, achieves 83% accuracy on a 100-example subset of OC20NEB and 70% success on a 10-example held-out comparison against a human-expert average of $T_i$ 1 (Madhavan et al., 13 May 2026).

These results do not imply that assessment is free. Several papers explicitly report increased orchestration or inference overhead. The empirical pattern is narrower: when search spaces are deep, multi-step, and failure-prone, explicit assessment tends to convert compute or tool budget into higher-quality search trajectories (Shi et al., 10 Jun 2026, Wang et al., 31 Aug 2025, Li et al., 22 May 2026, Li et al., 1 Jul 2026, Zhang et al., 8 Feb 2026, Madhavan et al., 13 May 2026).

7. Limitations, controversies, and research directions

Assess-then-Search does not eliminate the core uncertainties of search. TreeSeeker notes signal miscalibration and path dependence: if $T_i$ 2, $T_i$ 3, or $T_i$ 4 are misjudged, the controller may over-exploit or over-explore, and behavior can be sensitive to prompt phrasing and backtrack thresholds (Shi et al., 10 Jun 2026). L-MARS reports orchestration complexity, latency, and Judge over-conservatism, including cases where the Judge over-rejects partially relevant sources and causes unnecessary iterations (Wang et al., 31 Aug 2025). CVSearch remains slower than a single-pass MLLM or pure expert-assisted pipeline and depends on the quality of the expert and clustering stages; highly cluttered or domain-shifted scenes can degrade performance (Li et al., 22 May 2026). SRR-Judge identifies a weak-to-strong supervision gap: the judge cannot reliably supervise much stronger agents, and applying it to DeepSeek-V3.1 degraded performance (Zhang et al., 8 Feb 2026).

A second limitation is that assessment often depends on handcrafted or prompt-induced criteria. TreeSeeker uses prompt-elicited textual UCB signals rather than learned calibration (Shi et al., 10 Jun 2026). PaperPilot’s robustness depends on the coverage and quality of the typed operator library (Li et al., 1 Jul 2026). Analytical Search remains primarily a conceptual framework and therefore leaves concrete optimization and evaluation protocols as research directions (Tu et al., 12 Feb 2026). TSAgent still requires reasonable initial and final geometries and incurs substantial DFT compute cost (Madhavan et al., 13 May 2026).

The most common future directions are already visible across the literature: multimodal integration, learned calibration of assessment signals, adaptive summary intervals, dynamic path generation, richer verification, and stronger interaction between explicit workflows and user or tool feedback (Shi et al., 10 Jun 2026, Li et al., 22 May 2026, Li et al., 1 Jul 2026, Tu et al., 12 Feb 2026). A plausible implication is that future search systems will differentiate less by whether they search and more by how explicitly they assess before searching, while searching, and before deciding to stop.