Search agents extend large reasoning models with multi-round access to external knowledge, yet most existing systems largely conflate retrieving evidence with consuming it. Retrieved documents are commonly incorporated according to fixed or rank-based policies, despite the fact that evidence utility is conditioned on the agent's current reasoning state. Consequently, a lower-ranked candidate may be critical for the current step, while a higher-ranked candidate may provide redundant, irrelevant, or misleading information. We formulate this overlooked problem as evidence admission, which asks, at each search step, which retrieved candidates should be allowed to enter the agent's evolving reasoning context. To enable such selective control, we introduce SEEDS (Selective Evidence Entry During Search), a preview-guided evidence admission framework for search agents. SEEDS first exposes lightweight previews of retrieved candidates and then lets the agent jointly determine which candidates and how many should be fully inspected, conditioned on its current reasoning state. Unlike fixed or adaptive prefix policies, SEEDS can select arbitrary subsets with flexible cardinality, and it can be instantiated directly with an existing reasoning agent or further optimized through learning. Across nine knowledge-intensive QA benchmarks under both local corpus and live web search, SEEDS consistently improves search effectiveness while substantially reducing full-evidence consumption. Experimental analyses show that SEEDS more effectively selects useful candidates beyond the corresponding retrieval prefix. These results establish evidence admission as a complementary dimension of agentic search, separating what information is retrieved from what information is ultimately allowed to shape reasoning.
Figures & tables
Figure 1: Limitation of retrieval-rank prefix selection in multi-round search-augmented reasoning on QA tasks. Using a ReAct-style search agent with QwQ-32B, fixed- k is compared with dyn- k , where the search agent adaptively selects the prefix size at each search round.
Figure 2: Overview of Seeds compared with conventional RAG and prefix rank-based agentic search. Conventional RAG injects top-ranked documents before reasoning, while agentic search retrieves iteratively but still injects a rank prefix at each round. Seeds presents lightweight previews, allowing the reasoning model to choose which candidates , and how many , to expand based on its current reasoning state. Only expanded documents enter subsequent reasoning rollouts.
Single-hop QA
Multi-hop QA
Avg.
Method
NQ
TriviaQA
HotpotQA
2Wiki
MuSiQue
Bamboogle
(all)
EM
F1
EM
F1
EM
F1
EM
F1
EM
F1
EM
F1
EM ↑
F1 ↑
#Doc ↓
Without Retrieval
Direct Reasoning
28.0
39.7
66.4
72.0
27.0
36.6
32.6
38.5
10.2
19.2
48.0
63.0
35.4
44.8
0
Retrieval-augmented Reasoning
RAG k=3
39.8
49.7
60.8
66.6
31.0
39.7
22.0
28.9
8.4
17.4
32.8
45.0
32.5
41.2
3.0
Table 1: Main results (EM%, F1%, #Doc per question) of Seeds compared with baselines on open-domain QA benchmarks using the local Wikipedia dump corpus, with the reasoning backbone QwQ-32B model.
BrowseComp
BrowseComp-zh
HLE
Avg.
Method
Acc
F1
Acc
F1
Acc
F1
Acc ↑
F1 ↑
#Pages ↓
Search-o1
3.32
6.09
14.19
21.30
10.34
10.60
9.28
12.66
27.81
Seeds -o1 (ours)
5.29
8.42
16.26
23.85
13.79
6.66
11.78
12.98
8.30
Table 2: Results (Acc%, F1%, #Pages per question) on web browsing and expert-level question benchmarks with live web search, QwQ-32B.
Figure 3: Evidence admission probability (%) of documents with different retrieval ranks.
Figure 4: Evidence admission quality. (a) Hit rate is the probability that an answer-bearing document is admitted when it is available in the preview candidate pool. (b) Density is the proportion of answer-bearing documents among all documents admitted into the reasoning trajectory.
Table 7
Method
Clean
ρ=30%
ρ=50%
ρ=80%
ReAct dyn- k
36.9
37.0 (+0.1)
36.4 (-0.5)
33.4 (-3.5)
ReAct k=3
37.5
38.2 (+0.7)
36.0 (-1.5)
36.8 (-0.7)
ReAct k=10
35.3
36.0 (+0.7)
35.7 (+0.4)
36.9 (+1.6)
Seeds -ReAct
41.6
42.1 (+0.5)
42.3 (+0.7)
42.9 (+1.3)
Injected noise admitted
6.4%
7.5%
7.3%
Table 5: Robustness comparison (EM %) under preview candidate-pool perturbations.
Figure 9
NQ
HotpotQA
2Wiki
HLE
Configuration
↑ EM
↓ #Doc
↑ EM
↓ #Doc
↑ EM
↓ #Doc
↑ Acc.
↓ #Pages
Seeds -o1 (QwQ-32B)
40.2
9.8
39.8
13.2
33.6
16.6
13.8
8.3
Seeds -o1 (DeepSeek)
43.7
9.2
45.2
12.2
58.4
14.6
45.1
12.9
Search-o1 (DeepSeek)
45.1
19.2
44.2
26.2
57.2
37.6
45.1
22.9
Table 6: Generalization across different reasoning backbones and search scenarios.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Preview
NQ
TriviaQA
HotpotQA
2Wiki
Title only
38.0
67.2
33.8
37.8
Prefix, 30 tokens
37.4
67.2
35.2
39.2
Prefix, 60 tokens
38.6
67.0
37.6
32.4
Prefix, 90 tokens
41.0
68.0
35.6
28.6
Query window (adopted)
39.6
69.6
35.6
31.2
Chunk summary
38.2
68.6
35.8
35.2
Appendix
Table 7: Preview ablation: strict EM (%, QwQ-32B, n=500 per dataset).
Method
NQ
TriviaQA
HotpotQA
2Wiki
MuSiQue
Bamboogle
Avg.
ReAct k=3
1.70
1.32
2.75
4.47
3.02
2.06
2.55
ReAct k=10
0.96
0.89
1.63
2.66
1.70
1.10
1.49
ReAct dyn- k
1.77
1.53
2.71
5.33
2.74
1.92
2.67
Seeds -ReAct
1.16
1.23
2.19
3.26
2.41
1.78
2.00
Search-o1
2.04
1.84
3.37
4.73
3.60
2.55
3.02
Seeds -o1
2.07
1.80
3.38
4.52
3.72
2.70
3.03
Appendix
Table 8: Average number of search calls per question on QwQ-32B. Avg. is the macro-average across the six benchmarks.
Method
NQ
TriviaQA
HotpotQA
2Wiki
MuSiQue
Bamboogle
Avg.
RAG k=3
3.0
3.0
3.0
3.0
3.0
3.0
3.0
RAG k=10
10.0
10.0
10.0
10.0
10.0
10.0
10.0
ReAct k=3
5.1
4.0
8.3
13.4
9.1
6.2
7.7
ReAct k=10
9.6
8.9
16.3
26.6
17.0
11.0
14.9
ReAct dyn- k
5.7
4.7
7.9
14.3
8.0
5.3
7.7
Seeds -ReAct
3.3
3.3
2.7
2.9
3.0
3.2
3.1
Appendix
Table 9: Average number of full documents inspected per question. A document is counted when its full content is passed to a reader. Avg. is the macro-average across the six benchmarks.
Method
NQ
TriviaQA
HotpotQA
2Wiki
MuSiQue
Bamboogle
Avg.
RAG k=3
452
454
475
502
467
460
468
RAG k=10
1,512
1,518
1,584
1,671
1,563
1,526
1,562
ReAct k=3
806
622
1,370
2,270
1,482
972
1,254
ReAct k=10
1,518
1,405
2,704
4,438
2,767
1,744
2,429
ReAct dyn- k
902
747
1,336
2,448
1,308
861
1,267
Seeds -ReAct
512
513
461
527
511
643
528
Appendix
Table 10: Average number of full-document tokens consumed per question, measured with the model tokenizer. The metric excludes reasoning, preview, query, and generated-output tokens. Avg. is the macro-average across six benchmarks.
Subject
n
Search-o1
Seeds -o1
Mathematics
187
12.83
13.90
Computer science
52
9.62
13.46
Humanities
45
13.33
17.78
Biology / medicine
42
11.90
19.05
Other
42
7.14
11.90
Physics
40
0.00
5.00
Appendix
Table 11: HLE text-only accuracy (%) by subject (QwQ-32B). n is the number of questions.
Hyperparameter
Value
Stage 1: SFT
Teacher / student
QwQ-32B / Qwen2.5-7B-Base
Training / dev trajectories
900 / 100
Loss
masked NLL on policy tokens only
Optimizer
AdamW, β=(0.9,0.95) , weight decay 0.01
Learning rate / warmup
1×10−5 / 10%
Batch size / max. length
16 / 8,192 tokens
Appendix
Table 12: Optimization settings for Seeds -R1 (Qwen2.5-7B). The interaction environment is specified in Appendix E.1 .
Figure 7: Training curves for Seeds -R1 (Qwen2.5-7B, K=5 ). The horizontal axis is the optimizer update; the first ten updates train only the critic. Curves show training-batch means, not held-out evaluation scores.
Deep research requires reasoning over web evidence to answer open-ended questions, and it is a core capability for AI agents. Yet many deep research agents still rely on implicit, unstructured search behavior that causes redundant exploration and brittle evidence aggregation. Motivated by Anthropic's "think" tool paradigm and insights from the information-retrieval literature, we introduce Q+, a set of query and evidence processing tools that make web search more deliberate by guiding query planning, monitoring search progress, and extracting evidence from long web snapshots. We integrate Q+ into the browser sub-agent of Eigent, an open-source, production-ready multi-agent workforce for computer use, yielding EigentSearch-Q+. Across four benchmarks (SimpleQA-Verified, FRAMES, WebWalkerQA, and XBench DeepSearch), Q+ improves Eigent's browser agent benchmark-size-weighted average accuracy by 3.0, 3.8, and 0.6 percentage points (pp) for GPT-4.1, GPT-5.1, and Minimax M2.5 model backends, respectively. Case studies further suggest that EigentSearch-Q+ produces more coherent tool-calling trajectories by making search progress and evidence handling explicit.
Boer Zhang, Mingyan Wu, Dongzhuoran Zhou +6
Meta · Northeastern University, China · University of Oslo +5
Retrieval-augmented search agents answer multi-hop questions by repeatedly issuing search queries and accumulating evidence. This creates a stopping problem: after the necessary evidence has appeared, further retrieval often adds cost, latency, and distracting context rather than useful information. We frame stopping as evidence coverage rather than generator confidence, and introduce HALT, a lightweight verification-aware policy that leaves the search agent unchanged. Given expected hop claims, HALT stops only when cumulative evidence supports each required claim. Across three multi-hop QA benchmarks, HALT reduces redundant search while largely preserving exact match. We separate a deployable setting, where hop claims are generated from the question, from a diagnostic upper bound that uses gold supporting-fact annotations: generated claims give smaller but still exact-match-preserving savings, while gold claims show the larger savings available when hop targets are clean. Baseline comparisons and ablations show that this behavior is driven by claim-evidence alignment rather than generic sufficiency, fixed stop positions, or lexical overlap. Open-corpus pilots further suggest that HALT abstains when coverage cannot be reliably verified. Overall, evidence coverage provides a practical runtime control signal for improving retrieval-augmented agents without retraining or modifying the host agent.
Search agents adapt their queries, yet fixed search interfaces leave candidate processing and evidence presentation outside the agent's direct control. Our trajectory analysis shows that supporting passages can be retrieved yet never delivered to the agent; a same-page oracle intervention shows that changing the returned evidence can reduce subsequent search. We introduce Programmatic Search Agent (PSA), which makes a local executable computation over candidates the unit of a search action. PSA unifies a persistent candidate workspace, flexible primitive composition, and selective evidence presentation. It incrementally generates program cells that reuse candidates, execute dependent operations, and select what the agent inspects next. The runtime resolves specified data dependencies within each cell, while the agent adapts its search strategy across cells as new evidence arrives. We compare PSA with the Query-based Agent and Tool-based Agent on InfoSeek-Eval and BrowseComp-Plus using five policy backbones without task-specific training. All three interfaces share the search substrate, and the Tool-based Agent also shares PSA's primitives and persistent workspace. Relative to the Query-based Agent, PSA improves macro-averaged task success by 4.00 and 7.56 percentage points on the two benchmarks, respectively; within-backbone reductions in final-step tokens average 28.3% and 33.9%. These results support extending agent control beyond query reformulation to the processing and presentation of retrieved evidence. Code will be released subject to approval.
Jiaming Qian, Huiyan Yang, Mandi Liu +7
Zhejiang University · Yuanbao Team, Tencent · Peking University