FORT-Searcher: Shortcut-Resistant Search Agent

• FORT-Searcher is a deep search agent that uses shortcut-resistant data synthesis to ensure answers remain inaccessible until sufficient evidence is gathered.
• It employs the FORT pipeline, which constructs complex evidence graphs and formulates verifiable questions that prevent easy shortcut exploitation.
• Empirical results show that its methodology delays answer exposure and improves performance compared to other open-source search agents.

Searching arXiv for FORT-Searcher and closely related search-agent papers to ground the article. Searching arXiv for the specific paper and related deep search agent work. FORT-Searcher is a deep search agent and, more fundamentally, the name associated with a shortcut-aware data synthesis program for training such agents. It is introduced together with FORT, a “Framework of Shortcut-Resistant Training-Data Synthesis,” to address a specific failure mode in search-agent training: structurally complex questions can still be easy in practice if a model can reach the answer through a shortcut route. The central claim is that realized search difficulty is governed not only by graph size or hop count, but by whether the answer remains unavailable until sufficient evidence has been acquired through search. On this basis, FORT defines shortcut-aware difficulty, operationalizes four shortcut risks, synthesizes training data that suppresses them, and uses the resulting trajectories to train FORT-Searcher with supervised fine-tuning only (Deng et al., 10 Jun 2026).

1. Definition and problem setting

FORT-Searcher belongs to the class of deep search agents: LLM-based, tool-empowered systems that interact with external environments, gather evidence over multiple turns, and synthesize scattered information before answering. The work defines the training problem in terms of verifiable questions whose answers remain unavailable until sufficient evidence has been acquired through search, so that the resulting trajectories reward multi-turn query formulation, evidence discovery, evaluation, and integration rather than shallow retrieval (Deng et al., 10 Jun 2026).

A central distinction is drawn between intended or structural difficulty and realized difficulty. The paper argues that increasing structural complexity—such as adding more hops or widening an evidence graph—does not guarantee that a search process will remain hard in practice. A question may still collapse through what the work calls a “cheaper identifying route,” or through solver prior knowledge. This motivates a shortcut-aware formulation of difficulty in which the effective challenge is determined by the cheapest route that identifies the unique gold answer under the retrieval interface, rather than by surface graph properties alone (Deng et al., 10 Jun 2026).

The task is formalized as a multi-constraint agentic retrieval problem

$q=(\mathcal{X},\mathcal{C}_q,\Sigma),$

where $\mathcal{X}$ is the answer space, $\mathcal{C}_q$ is the set of constraint clues expressed by the question, and $\Sigma$ is the retrieval interface. For any subset $\mathcal{P}\subseteq\mathcal{C}_q$, the remaining candidate pool is

$$\mathrm{Ans}(\mathcal{P}) = \left\{ x\in\mathcal{X}: \bigwedge_{c_i\in\mathcal{P}} c_i(x)=1 \right\}.$$

A well-posed question satisfies $\mathrm{Ans}(\mathcal{C}_q)=\{y^\star\}$, where $y^\star$ is the gold answer (Deng et al., 10 Jun 2026).

This framing positions FORT-Searcher within a broader shift in search-agent research from maximizing apparent task complexity to controlling realized search cost. A plausible implication is that data synthesis quality, rather than model size alone, becomes a first-order determinant of search-agent behavior.

2. Shortcut-aware difficulty framework

The shortcut-aware framework centers on the difference between the cost a solver must pay if it has no prior knowledge and cannot guess unsupported answers, and the lower cost a concrete solver may realize in practice. A search trajectory is written as $\tau=(\theta_1,\dots,\theta_{|\tau|})$, where $|\tau|$ is the number of retrieval actions. The pure-posterior cost is defined as

$\mathcal{X}$0

For an identifying subset $\mathcal{X}$1 with $\mathcal{X}$2, the shortest valid evidence-acquisition route cost is

$\mathcal{X}$3

The structural lower bound is then

$\mathcal{X}$4

The realized successful-trajectory cost of a concrete solver $\mathcal{X}$5 is

$\mathcal{X}$6

with solver-side cost reduction

$\mathcal{X}$7

so that

$\mathcal{X}$8

This separates route-level shortcut structure from solver-level exploitation of prior knowledge (Deng et al., 10 Jun 2026).

Three structural factors shape the cheapest identifying route. Subset selectivity is

$\mathcal{X}$9

so small $\mathcal{C}_q$0 for a small subset can identify the answer early. Evidence dispersion is

$\mathcal{C}_q$1

so co-coverage by a single page or snippet lowers the effective route cost. Dependency depth is

$\mathcal{C}_q$2

which captures the longest serial dependency chain and is reduced by exposed intermediate constants (Deng et al., 10 Jun 2026).

The paper states the proposition

$\mathcal{C}_q$3

hence

$\mathcal{C}_q$4

It also gives a collapse corollary: if some identifying subset can be verified by a single initially executable query, then $\mathcal{C}_q$5 (Deng et al., 10 Jun 2026).

The framework operationalizes four shortcut risks.

Shortcut risk	Mechanism	Effect on difficulty
Single-clue selectivity	One clue or small clue subset is overly identifying	Shrinks exploration
Evidence co-coverage	One item verifies multiple intended constraints	Lowers evidence-acquisition cost
Exposed constants	Intermediate names, dates, or numbers appear on the question surface	Shortens dependency chains
Prior-knowledge binding	Model names the answer before retrieval anchors it	Reduces realized cost below structural cost

To diagnose realized effects at scale, the paper introduces trajectory signatures. Realized solving cost is

$\mathcal{C}_q$6

answer hit time uses

$\mathcal{C}_q$7

and dataset average

$\mathcal{C}_q$8

The observable proxy for prior binding is the prior-shortcut rate

$\mathcal{C}_q$9

Later $\Sigma$0 indicates longer pre-answer search, while large gaps between $\Sigma$1 and $\Sigma$2 indicate post-hit verification or detours (Deng et al., 10 Jun 2026).

3. FORT synthesis pipeline

FORT constructs shortcut-resistant training data in four stages: graph initialization, evidence graph construction, question formulation, and adversarial refinement. The synthesis workspace is an internal evidence graph whose nodes are entities and whose edges are verified facts with sources and dependencies. The stated objective is to preserve route-level difficulty while reducing solver-level shortcuts (Deng et al., 10 Jun 2026).

In graph initialization, the root entity $\Sigma$3 is selected from Wikidata, with preference for long-tail entities, including entities without English Wikipedia pages, to reduce prior-knowledge binding. The system also uses cycle-based seeds: it pre-mines a cycle library from Wikidata, builds a pair index, enumerates small cycles, removes duplicates and unsuitable hub-based or redundant cycles, and stores an inverted index $\Sigma$4 such that $\Sigma$5 lists retained cycles containing entity $\Sigma$6. If $\Sigma$7, one cycle is used as the initial subgraph $\Sigma$8; otherwise the seed is a single-node graph (Deng et al., 10 Jun 2026).

Evidence graph construction targets evidence co-coverage and single-clue selectivity. Starting from $\Sigma$9, expansion proceeds under a depth limit $\mathcal{P}\subseteq\mathcal{C}_q$0 and a node-addition budget $\mathcal{P}\subseteq\mathcal{C}_q$1, expanding deepest nodes first to preserve serial dependencies. Facts are collected from heterogeneous sources, including Wikidata, web pages, structured databases, and Google Scholar or Maps. The system performs multi-source enrichment, prefers distributing root-node facts across sources, and avoids verifying multiple selected facts on the same page (Deng et al., 10 Jun 2026).

The paper’s evidence graph construction algorithm is:

$\mathcal{X}$32

Derived facts are used to reduce exact-match shortcuts and include coincidence bridging, count aggregation, numerical relation, and meta-fact extraction. Fact verification filters contradictions and entity-consistency failures such as similar names, ambiguous abbreviations, temporal or geographic drift, and series–edition mismatch. Fact selection then prefers generic facts that are reliable but non-representative in isolation, so that only combinations of clues identify the answer (Deng et al., 10 Jun 2026).

Question formulation then chooses an answer node $\mathcal{P}\subseteq\mathcal{C}_q$2 and a subgraph $\mathcal{P}\subseteq\mathcal{C}_q$3. Intermediate names are withheld and rendered as descriptions such as “the artist,” “the institution,” or “the work,” so downstream queries are not executable initially. Exact values are fuzzed into truthful but less directly searchable constraints through category generalization, range relaxation, meta-attribute description, arithmetic encoding, and contrastive exclusion. The formulation is explicitly required to remain verifiable and resolvable through evidence, avoiding ambiguity and over-fuzzing (Deng et al., 10 Jun 2026).

4. Adversarial refinement and shortcut suppression

Adversarial refinement calibrates realized difficulty in actual search. A strong search-agent adversary is run on each draft question, trajectory signatures are measured, and the question is repaired if it is shortcut-prone, ambiguous, or over-fuzzed. The acceptance criteria require that the adversary answer correctly, use at least a minimum desired solving cost threshold, exhibit sufficiently late answer hit time, and show no prior binding, meaning the model must not name the answer before evidence (Deng et al., 10 Jun 2026).

Repair actions depend on failure mode. If a draft is solved too quickly or exposes the answer early, the system replaces co-covered evidence, removes overly selective facts, and withholds or fuzzes exposed constants. If it is prior-bound, the root entity is replaced with a long-tail alternative or the evidence path is strengthened. If it is unsolved within budget, interpreted as over-hard or underspecified, the system narrows fuzzed clues, removes ambiguity, or restores necessary constraints (Deng et al., 10 Jun 2026).

The paper gives two case studies. In one shortcut repair, broadening “shot down by the Soviet Union” to “shot down by a certain country during the Cold War,” and relaxing a runtime digit-sum constraint from “sum to 10” to “sum to a multiple of 5,” removed a cross-domain triangulation shortcut. In one narrowing repair for over-fuzzed clues, “a certain European capital” was replaced with “the capital of a federal republic in Central Europe,” and “a certain year” with “mid-20th century,” to restore solvability while maintaining obfuscation (Deng et al., 10 Jun 2026).

This stage is central to the paper’s claim that shortcut resistance cannot be guaranteed by static graph design alone. The work treats actual search trajectories, not only graph templates, as the criterion for difficulty. This suggests that the FORT pipeline is best understood not as a pure synthesis engine, but as an overview-and-calibration loop.

5. FORT-Searcher agent and training procedure

FORT-Searcher is trained on FORT-generated trajectories with supervised fine-tuning only. The backbone is Qwen3-30B-A3B-Thinking-2507, with approximately 3B active parameters at inference and a 256K context window. Within a rollout, all tool-call results are retained so that accumulated evidence can be reused. If the model reaches a predefined turn limit without producing a final answer, the interaction history is cleared and the agent restarts from the original question. The rollout limits are 300 for BrowseComp and BrowseComp-ZH, and 200 for xbench-DeepSearch-2505, xbench-DeepSearch-2510, and Seal-0 (Deng et al., 10 Jun 2026).

The training recipe is standard supervised token-level learning rather than a custom reinforcement objective. The reported optimization setup uses bf16 precision; Adam with $\mathcal{P}\subseteq\mathcal{C}_q$4, $\mathcal{P}\subseteq\mathcal{C}_q$5, $\mathcal{P}\subseteq\mathcal{C}_q$6, weight decay $\mathcal{P}\subseteq\mathcal{C}_q$7, and gradient clipping $\mathcal{P}\subseteq\mathcal{C}_q$8; a cosine schedule with peak learning rate $\mathcal{P}\subseteq\mathcal{C}_q$9, minimum learning rate $$\mathrm{Ans}(\mathcal{P}) = \left\{ x\in\mathcal{X}: \bigwedge_{c_i\in\mathcal{P}} c_i(x)=1 \right\}.$$0, and $$\mathrm{Ans}(\mathcal{P}) = \left\{ x\in\mathcal{X}: \bigwedge_{c_i\in\mathcal{P}} c_i(x)=1 \right\}.$$1 warmup iterations; tensor parallelism $$\mathrm{Ans}(\mathcal{P}) = \left\{ x\in\mathcal{X}: \bigwedge_{c_i\in\mathcal{P}} c_i(x)=1 \right\}.$$2; pipeline parallelism $$\mathrm{Ans}(\mathcal{P}) = \left\{ x\in\mathcal{X}: \bigwedge_{c_i\in\mathcal{P}} c_i(x)=1 \right\}.$$3; context parallelism $$\mathrm{Ans}(\mathcal{P}) = \left\{ x\in\mathcal{X}: \bigwedge_{c_i\in\mathcal{P}} c_i(x)=1 \right\}.$$4; expert parallelism $$\mathrm{Ans}(\mathcal{P}) = \left\{ x\in\mathcal{X}: \bigwedge_{c_i\in\mathcal{P}} c_i(x)=1 \right\}.$$5 for the MoE architecture; sequence parallelism; activation recomputation; and a distributed optimizer. Hyperparameters are $$\mathrm{Ans}(\mathcal{P}) = \left\{ x\in\mathcal{X}: \bigwedge_{c_i\in\mathcal{P}} c_i(x)=1 \right\}.$$6 epochs, global batch size $$\mathrm{Ans}(\mathcal{P}) = \left\{ x\in\mathcal{X}: \bigwedge_{c_i\in\mathcal{P}} c_i(x)=1 \right\}.$$7, and maximum sequence length $$\mathrm{Ans}(\mathcal{P}) = \left\{ x\in\mathcal{X}: \bigwedge_{c_i\in\mathcal{P}} c_i(x)=1 \right\}.$$8 tokens (Deng et al., 10 Jun 2026).

In comparative context, this training philosophy differs from recent multimodal or web search agents that rely on SFT-then-RL pipelines, such as VSearcher, which adopts an SFT-then-RL procedure with GRPO in real web environments (Zhang et al., 3 Mar 2026), and from openly released SFT-only search systems such as OpenSeeker, which emphasizes fact-grounded QA synthesis and denoised trajectory synthesis over $$\mathrm{Ans}(\mathcal{P}) = \left\{ x\in\mathcal{X}: \bigwedge_{c_i\in\mathcal{P}} c_i(x)=1 \right\}.$$9k synthesized samples (Du et al., 16 Mar 2026). FORT-Searcher’s distinguishing claim is narrower and more specific: the decisive factor is shortcut-resistant synthesis rather than broader scale or reinforcement learning (Deng et al., 10 Jun 2026).

6. Empirical performance, diagnostics, and position in the literature

FORT-Searcher is evaluated on BrowseComp, BrowseComp-ZH, xbench-DeepSearch-2505, xbench-DeepSearch-2510, and Seal-0, with Hugging Face domains blocked during testing. Its reported scores are BrowseComp $\mathrm{Ans}(\mathcal{C}_q)=\{y^\star\}$0, BrowseComp-ZH $\mathrm{Ans}(\mathcal{C}_q)=\{y^\star\}$1, xbench-05 $\mathrm{Ans}(\mathcal{C}_q)=\{y^\star\}$2, xbench-10 $\mathrm{Ans}(\mathcal{C}_q)=\{y^\star\}$3, Seal-0 $\mathrm{Ans}(\mathcal{C}_q)=\{y^\star\}$4, and overall $\mathrm{Ans}(\mathcal{C}_q)=\{y^\star\}$5, which the paper describes as the best overall performance among comparable-size open-source search agents (Deng et al., 10 Jun 2026).

Agent	BrowseComp	BrowseComp-ZH	xbench-05	xbench-10	Seal-0	Overall
FORT-Searcher	72.2	75.0	80.8	57.2	46.0	66.2
MiroThinker-1.7-mini	67.9	72.3	77.2	57.2	48.2	64.6
Qwen3.5-35B-A3B	61.0	69.5	77.4	50.3	41.4	59.9
OpenSeekerV2	46.0	58.1	78.0	43.4	41.4	53.4
Tongyi DeepResearch	43.4	46.7	75.0	47.5	45.8	51.7

The context-reset mechanism materially affects performance. With versus without context reset, the paper reports BrowseComp $\mathrm{Ans}(\mathcal{C}_q)=\{y^\star\}$6 vs. $\mathrm{Ans}(\mathcal{C}_q)=\{y^\star\}$7, BrowseComp-ZH $\mathrm{Ans}(\mathcal{C}_q)=\{y^\star\}$8 vs. $\mathrm{Ans}(\mathcal{C}_q)=\{y^\star\}$9, xbench-05 $y^\star$0 vs. $y^\star$1, xbench-10 $y^\star$2 vs. $y^\star$3, and Seal-0 $y^\star$4 vs. $y^\star$5 (Deng et al., 10 Jun 2026). This suggests that restart-based exploration of alternative routes is important when long trajectories accumulate unproductive context.

The more distinctive evidence concerns dataset difficulty. On 200 sampled questions per dataset using the same adversary and budget, FORT yields $y^\star$6, $y^\star$7, and $y^\star$8. By contrast, OpenSeeker yields $y^\star$9, $\tau=(\theta_1,\dots,\theta_{|\tau|})$0, and $\tau=(\theta_1,\dots,\theta_{|\tau|})$1; REDSearcher yields $\tau=(\theta_1,\dots,\theta_{|\tau|})$2, $\tau=(\theta_1,\dots,\theta_{|\tau|})$3, and $\tau=(\theta_1,\dots,\theta_{|\tau|})$4; DeepResearch-9K yields $\tau=(\theta_1,\dots,\theta_{|\tau|})$5, $\tau=(\theta_1,\dots,\theta_{|\tau|})$6, and $\tau=(\theta_1,\dots,\theta_{|\tau|})$7; DeepDive yields $\tau=(\theta_1,\dots,\theta_{|\tau|})$8, $\tau=(\theta_1,\dots,\theta_{|\tau|})$9, and $|\tau|$0; InfoSeek yields $|\tau|$1, $|\tau|$2, and $|\tau|$3 (Deng et al., 10 Jun 2026). The paper interprets this as evidence that FORT delays answer exposure and induces longer pre-answer search rather than merely longer post-hit detours.

A controlled training-data analysis strengthens that interpretation. Four 12K-example training settings are compared. Open-source data at $|\tau|$4 gives $|\tau|$5, $|\tau|$6, BrowseComp $|\tau|$7, and BrowseComp-ZH $|\tau|$8, whereas FORT at the same approximate solving cost gives $|\tau|$9, $\mathcal{X}$00, BrowseComp $\mathcal{X}$01, and BrowseComp-ZH $\mathcal{X}$02 (Deng et al., 10 Jun 2026). This is the paper’s main empirical support for the claim that longer trajectories alone are insufficient; what matters is delaying answer exposure and reducing prior binding.

The cumulative shortcut-resistance ablation on 2K synthesized questions shows that removing any component makes questions easier. The full system yields accuracy $\mathcal{X}$03, $\mathcal{X}$04, $\mathcal{X}$05, and $\mathcal{X}$06. Removing fuzzing changes these to $\mathcal{X}$07, $\mathcal{X}$08, $\mathcal{X}$09, and $\mathcal{X}$10, representing the largest difficulty drop among the ablations. Removing generic-fact selection yields $\mathcal{X}$11, $\mathcal{X}$12, $\mathcal{X}$13, and $\mathcal{X}$14; removing source diversity yields $\mathcal{X}$15, $\mathcal{X}$16, $\mathcal{X}$17, and $\mathcal{X}$18; removing derived-fact construction yields $\mathcal{X}$19, $\mathcal{X}$20, $\mathcal{X}$21, and $\mathcal{X}$22; removing long-tail entity selection yields $\mathcal{X}$23, $\mathcal{X}$24, $\mathcal{X}$25, and $\mathcal{X}$26; and removing cycle construction yields $\mathcal{X}$27, $\mathcal{X}$28, $\mathcal{X}$29, and $\mathcal{X}$30 (Deng et al., 10 Jun 2026).

Within the broader literature, FORT-Searcher is positioned against methods that scale question generation by graph expansion, topology control, or web-graph reverse engineering, such as OpenSeeker (Du et al., 16 Mar 2026), and against agentic systems emphasizing multimodal RL and tool-use policies, such as VSearcher (Zhang et al., 3 Mar 2026) and M$\mathcal{X}$31Searcher (Yu et al., 14 Jan 2026). Its specific contribution is not a new planner or reward function, but a theory-and-synthesis account of shortcut-resistant search difficulty. A plausible implication is that FORT-Searcher’s main legacy may be methodological: it provides a vocabulary and a set of measurable proxies for asking whether a search benchmark is genuinely search-heavy, rather than merely structurally elaborate.