Large language model search agents are often trained with synthetic questions whose difficulty is increased through larger evidence graphs, additional hops, and longer trajectories. These global properties, however, are only indirect proxies for the local retrieval capabilities required during search. To address this mismatch, we introduce latent anchor reasoning, which consists of resolving an unnamed retrieval anchor from descriptive specifications and transferring the recovered anchor into a subsequent information demand. This primitive retrieval unit decomposes deep search into chains of coupled operations and organizes question construction around anchor resolution and relation transfer, without prescribing a canonical search path. Based on this formulation, we propose PrimeSeeker, a capability-oriented framework that constructs web-grounded anchor structures and jointly derives a question and a reference evidence skeleton. The skeleton preserves supporting evidence from construction and guides expert generation through extractive highlights of current tool observations. These highlights are removed before supervised fine-tuning, while the skeleton is subsequently reused to audit reference-step coverage for reinforcement-learning rewards. We construct 9,221 expert trajectories, training a 30B search agent. Across five deep-search benchmarks, PrimeSeeker achieves strong performance, while reference-step optimization further improves the supervised policy. The resulting trajectories exhibit low retrieval redundancy, and fixed-budget evaluation shows strong solution coverage with substantially fewer tool calls than long-horizon systems.
Figures & tables
Figure 1: Illustration of latent anchor reasoning. Descriptive specifications first identify an unnamed play; the recovered anchor makes the ceremony relation queryable, leading to a second latent anchor and ultimately the name of the requested location. The decomposition specifies local retrieval capabilities without prescribing a unique search trajectory.
Figure 2: Overview of PrimeSeeker . The constructed anchor structure yields both a question q and a reference evidence skeleton Dq . During expert generation, Dq extracts matching evidence from the current observation. These annotations are removed before SFT, while Dq remains external to the policy and later provides reference-step coverage signals during reinforcement learning.
Model
BS-EN
BS-ZH
DS2505
DS2510
GAIA
P@1
P@3
P@1
P@3
P@1
P@3
P@1
P@3
P@1
P@3
Advanced tool-augmented models
OpenAI o3
49.7
–
58.1
–
67.0
–
–
–
70.5
–
DeepSeek-V3.1
30.0
–
49.2
–
71.0
–
–
–
63.1
–
GPT-5
54.9
–
63.0
–
77.8
–
75.0
–
76.4
–
DeepSeek-V3.2
67.6
–
65.0
–
78.0
–
55.7
–
75.1
–
Table 1: Reported performance on five deep-search benchmarks. Underlined values are the best among reported models in each block; bold values are the best among PrimeSeeker variants.
Figure 3: Pass@3 across tool call and context budgets. Solid lines denote PrimeSeeker- Dq , and pale dashed lines denote PrimeSeeker-SFT. A question is solved if correct within the cutoff.
Model
DS2505
DS2510
Standard P@1
Budgeted Solve Rate
Consumption Avg. Calls
Correct Avg. Calls
Standard P@1
Budgeted Solve Rate
Consumption Avg. Calls
Correct Avg. Calls
OpenSeeker-v1
74.0
83.0
34.5
16.2
–
57.0
109.5
46.4
MiroThinker
77.2
85.0
99.9
70.4
57.2
58.0
144.2
72.6
PrimeSeeker- Dq
76.0
91.0
21.8
16.3
41.0
68.0
68.6
33.5
Table 2: Performance under a shared 300-call budget on DS2505 and DS2510. Pass@1 values for OpenSeeker-v1 and MiroThinker are from their original reports; all remaining results are obtained by rerunning each system with its original inference procedure.
Dataset
Samples
Avg. Tools
Avg. Tokens
Search Dup.
Visit Dup.
DeepResearch-9K
12,974
25.0
12.3k
3.1%
13.1%
OpenSeeker
11,677
47.0
73.9k
1.6%
10.3%
REDSearcher
10,001
65.1
56.4k
0.2%
14.5%
PrimeSeeker
9,221
19.5
37.8k
0.1%
8.0%
Table 3: Aggregate statistics of expert search trajectories.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Expert
K
Unguided
Random Highlight
Dq -guided
Acc.
Calls
Acc.
Δ Acc.
Calls
Δ Calls
Acc.
Δ Acc.
Calls
Δ Calls
Qwen3
2
52.0
6.45
28.0
-24.0
70.52
+64.07
54.0
+2.0
9.09
+2.64
Qwen3
3
57.0
7.01
34.0
-23.0
72.10
+65.09
61.0
+4.0
9.87
+2.86
DeepSeek-V4
2
51.0
24.67
24.0
-27.0
50.12
+25.45
56.0
+5.0
21.99
-2.68
DeepSeek-V4
3
59.0
28.68
24.0
-35.0
57.18
+28.50
66.0
+7.0
24.59
-4.09
Appendix
Table 4: Comparison of random and Dq -guided highlights against unguided expert generation.
Figure 4: Distributional diagnostics of expert trajectories. (a) Trajectory length distribution, shown as KDE-smoothed expected sample count per 1K-token window. (b) Tool-call distribution, shown as KDE-smoothed expected sample count per one-tool-call window. (c) CCDF of duplicated search-query counts within trajectories. (d) CCDF of duplicated visited-URL counts under the same criterion. PrimeSeeker trajectories occupy moderate tool-use ranges and exhibit lighter duplicate-retrieval tails, indicating more compact and less redundant expert supervision.
Training LLM-based search agents requires high-quality search data: tasks that demand genuine multi-hop retrieval and trajectories that use search tools effectively. Existing pipelines often depend on human-written tasks, expert demonstrations, or stronger teacher models. We present SearchMaster, a self-play framework that trains a shared LLM policy from search tasks it generates, solves, and verifies in a local search environment. The key challenge is that self-generated tasks and rollouts can yield misleading signals: pseudo multi-hop questions, success-rate difficulty estimates that ignore search depth, and rollouts with excessive opening but little targeted evidence acquisition. SearchMaster addresses these failure modes with three controls. An Evidence-Chain Generator (ECG) grounds task generation in explicit cross-document evidence chains to reduce pseudo multi-hop questions. A Search-Depth Reward (SDR) scores task difficulty by the search depth of successful rollouts rather than success rate alone, keeping retained tasks search-intensive. An Over-Opening Penalty (OOP) regulates tool use by discouraging excessive document opening, avoiding long but shallow browsing. Verified Proposer and Solver rollouts are then jointly optimized with GRPO. Across six deep-search benchmarks, SearchMaster raises average accuracy from 38.19% to 51.52% with a Qwen3.5-9B backbone and from 51.25% to 62.54% with a Qwen3.5-27B backbone. On BrowseComp-Plus, the corresponding gains are 30.1 and 17.8 percentage points. These results show that grounded and regulated self-play provides effective search-agent training data across model scales, without human-labeled QA pairs or expert demonstrations. The code is available at https://github.com/WentaoTan/SearchMaster.
Deep search is becoming a core capability of modern agent systems, yet it is typically evaluated solely based on end-to-end answer accuracy. This coupled evaluation paradigm entangles retrieval quality, long-context comprehension, evidence verification, and tool-use decisions, making it difficult to determine whether a model truly knows when and how to delegate information seeking to search. To this end: (1) We formalize this meta-capability as Delegation Intelligence in deep search and decompose it into complementary dimensions-Search Decision-Making (recognizing information insufficiency and deciding whether, when, and how to search) and Information Synthesis and Verification (aggregating evidence from multiple sources, judging source reliability, and synthesizing information under noisy, potentially adversarial conditions). (2) To enable disentangled and reproducible measurement, we develop a controllable synthesis pipeline built on document-grounded reverse engineering. This yields a general recipe for constructing controlled deep-search evaluations rather than a single fixed dataset. (3) As a concrete instantiation, we construct DelegSearchBench, together with a disentangled evaluation protocol that isolates each capability dimension by varying document composition and tool access. (4) Across representative models, we demonstrate that deep-search competence cannot be adequately characterized by final-answer accuracy alone...
Deep-research agents answer complex questions by interacting with search and browsing tools, yet they often search along a single evolving trajectory. Our trajectory-level analysis reveals a common failure mode in which the agent may encounter an early search state with several plausible directions, but follow one direction before collecting enough comparative evidence. Once this happens, subsequent tool calls tend to reinforce the same path, increasing the chance of failure when the initial direction is misleading. We further find that successful trajectories reduce this risk through two behaviors: grounding vague exploration in concrete candidates and shifting directions when the current path is weak or incomplete. Based on these findings, we propose HypoSearch, which generates lightweight hypotheses as soft search hints, explores them through bounded independent branches, and compares branch-level evidence before commitment. Across four deep-research benchmarks and three backbone models, HypoSearch consistently outperforms single-trajectory search and standard parallel baselines, improving Qwen3.5-122B from 46.7 to 60.0 on BC-small while using fewer tool calls than five independent trajectories. A pilot supervised fine-tuning study further shows that these behavioral signals can curate compact training trajectories and reduce degradation from unfiltered data.
Ruochen Zhou, Zhengyu Chen, Luan Zhang +3
City University of Hong Kong · Meituan · Independent +1