Deep-research agents repeatedly decide what to search, read, and verify, often using large cloud-hosted language models for action proposal. However, privacy and local control motivate running agents on users' devices, where limited memory and compute favor smaller models and complicate on-device training. Yet inference-time guidance faces three obstacles: exploring alternatives through real execution adds tool calls, prompted self-evaluation is bounded by the agent's judgment, and model-specific guidance risks repeated adaptation costs. We introduce Seer, an inference-time framework providing local guidance that separates the agent's action proposal from learned consequence prediction and value estimation. From a shared corpus of real research trajectories, we train two models: a tool-response simulator on observed call--response pairs and a value estimator on histories labeled with discounted terminal correctness. The simulator and estimator are trained offline and reused across agents without further training. Moreover, Seer executes only the selected action and adds only real tool responses to the committed history, avoiding tool calls on discarded candidates at additional inference cost. Across multiple deep-research benchmarks, Seer substantially improves agent performance for different language models. For instance, GAIA accuracy rises from 13.6% to 41.7% for Qwen3.5-4B and from 22.3% to 50.5% for Qwen3.5-9B. Furthermore, our analyses suggest that broader candidate coverage can compensate for imperfect simulation, while prediction and scoring errors limit the benefits of wider imagined search. These findings highlight the potential of reusable models for prediction and evaluation to strengthen small research agents through extra local computation without the real execution of the discarded candidates.
Figures & tables
Figure 1: How can we provide reusable guidance that strengthens small agents without model-specific retraining or the real execution of every discarded candidate?
Figure 2: Seer overview. Monte Carlo branching collects real trajectories to train the simulator on tool responses and the estimator on discounted terminal correctness. At inference, these shared models guide search over imagined histories, while only selected actions are executed to obtain real evidence.
Model
Base
RAP@4
Seer
Qwen3.5-4B
8.3
8.0
18.8
Qwen3.5-9B
14.2
13.9
25.2
Table 1: Preliminary comparison of RAP@4 and Seer ( k=4 ) using average accuracy across benchmarks (%). RAP@4 is omitted from later tables as it barely improves on the Base Agent.
Qwen3.5-4B
Qwen3.5-9B
Qwen3.5-35B-A3B
Benchmark
Base
BoN
AV
S4
S32
Base
BoN
AV
S4
S32
Base
BoN
AV
S4
S32
GAIA
13.6
21.4
26.2
29.1
41.7
22.3
35.9
35.0
35.9
50.5
57.3
53.4
47.6
62.1
57.3
BC-200
1.0
4.5
3.5
5.0
7.0
2.5
5.5
6.5
5.5
9.5
22.0
24.0
24.0
24.5
22.0
SEAL-0
4.5
9.9
8.1
10.8
16.2
10.8
23.4
18.9
22.5
26.1
27.0
28.8
33.3
31.5
34.2
XBench-DS
20.0
36.0
35.0
41.0
43.0
28.0
38.0
38.0
47.0
60.0
64.0
67.0
64.0
66.0
69.0
BC-ZH
2.4
7.6
5.2
8.0
9.7
7.6
11.4
13.5
15.2
16.3
36.3
37.7
34.9
36.0
43.3
Table 2: Comparisons of final-answer accuracy (%) across Qwen and Gemma agents using Base Agent, BoN@4, AV@4, and Seer ( Sk , candidate width k ). Seer improves over Base Agent in 26 of 30 settings at k=32 , reusing the simulator and estimator across agents.
Qwen3.5-122B-A10B
Qwen3.5-397B-A17B
Benchmark
Base
BoN
AV
S4
S32
Base
BoN
AV
S4
S32
GAIA
67.0
64.1
61.2
68.0
71.8
68.9
74.8
69.9
76.7
72.8
BC-200
31.5
32.5
35.0
35.0
39.0
34.5
37.5
37.5
37.0
42.5
SEAL-0
32.4
28.8
31.5
34.2
36.9
31.5
36.9
36.0
37.8
40.5
XBench-DS
66.0
70.0
61.0
71.0
72.0
75.0
73.0
75.0
79.0
80.0
BC-ZH
39.4
38.1
36.3
40.5
43.6
50.2
49.8
49.8
55.7
55.4
Table 3: Comparisons of final-answer accuracy (%) on larger agents. Seer transfers to 122B and 397B agents without retraining, with S32 outperforming Base Agent in all cases.
Qwen3.5
Qwen3.6
Gemma-4
Variant
4B
9B
35B-A3B
122B-A10B
397B-A17B
35B-A3B
e2b-it
e4b-it
Base + search ( k=4 )
16.5
44.7
62.1
71.8
73.8
74.8
26.2
19.4
+ estimator
38.8
45.6
66.0
72.8
71.8
75.7
31.1
29.1
+ simulator (Seer)
29.1
35.9
62.1
68.0
76.7
78.6
30.1
24.3
Direct scoring
23.3
31.1
61.2
68.0
72.8
76.7
28.2
21.4
Table 4: Comparisons of cumulative component additions on GAIA across agents (accuracy %, k=4 ). Adding the estimator improves over prompted scoring, while subsequently adding the simulator avoids real calls for discarded candidates at an accuracy cost.
Figure 3: Comparisons of final-answer accuracy and real tool-call counts on BrowseComp-200 across agents. The vertical axis (Performance) reports accuracy (%). Seer achieves the highest accuracy while keeping tool-call counts close to those of Base Agent.
Qwen3.5-4B
Qwen3.5-9B
Qwen3.5-35B-A3B
Qwen3.5-122B-A10B
Benchmark
k=4
k=8
k=16
k=32
k=4
k=8
k=16
k=32
k=4
k=8
k=16
k=32
k=4
k=8
k=16
k=32
GAIA
29.1
35.9
39.8
41.7
35.9
38.8
42.7
50.5
62.1
64.1
59.2
57.3
68.0
68.9
72.8
71.8
BC-200
5.0
5.5
6.0
7.0
5.5
7.0
9.0
9.5
24.5
27.5
27.0
22.0
35.0
34.0
36.5
39.0
SEAL-0
10.8
10.8
15.3
16.2
22.5
24.3
23.4
26.1
31.5
29.7
31.5
34.2
34.2
35.1
36.0
36.9
XBench-DS
41.0
39.0
43.0
43.0
47.0
50.0
51.0
60.0
66.0
69.0
71.0
69.0
71.0
68.0
71.0
72.0
BC-ZH
8.0
9.3
9.0
9.7
15.2
15.9
15.9
16.3
36.0
41.5
41.0
43.3
40.5
39.8
41.9
43.6
Table 5: Comparisons of Seer accuracy (%) across candidate widths k on all five benchmarks, with the best and second-best results within each model and benchmark shown in bold and underlined, respectively, including ties. Larger k does not consistently improve performance, highlighting the need to select a width for each agent and task.
Model
Dataset
k=4
k=8
k=16
k=32
397B
GAIA
25.0
21.1
11.8
10.0
122B
BC-200
40.9
21.4
17.6
22.5
397B
XBench-DS
30.8
7.9
0.0
0.0
Table 6: Estimator–oracle agreement (%): how often both select the same candidate action. Agents are Qwen3.5 variants.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Symbol
Meaning
Research interactions and resource constraints
q,D
Research request and task distribution, with q∼D .
G,πG
Research agent and its action distribution, including search guidance.
π
Action-proposal distribution of the frozen agent model.
t,st
Real-round index and committed history, initially containing q and agent instructions.
at,a^
Selected action (tool use or final submission) and submitted answer, respectively.
Appendix
Table 7: Notation used in the paper.
Case
Base
BoN
AV
S32
1. OpenCV contributor
11 (I)
39 (I)
50 (I)
3 (C)
2. Hiking location
54 (I)
178 (I)
124 (I)
17 (C)
3. UN parliamentary seats
40 (I)
135 (I)
104 (I)
16 (C)
4. NBA final-game margin
14 (I)
62 (I)
47 (I)
8 (C)
5. Encoder-layer comparison
11 (C)
—
—
2 (C)
6. Elevator-accident article
—
21 (C)
—
2 (C)
Appendix
Table 8: Case outcomes and real tool-call counts. C denotes a correct answer; I denotes an incorrect or unanswered result; — indicates a comparison not included in the case report. Base, BoN, AV, and S32 follow the notation in § 6 .
Autonomous research agents use language models and tools to search literature, design and run experiments, analyze results, and support scientific discovery.
Deep research agents face vast, interdependent, and pervasively uncertain information. Existing systems explore what evolving intermediate representations should look like, but leave their evolution to the LLM's implicit reasoning. Without explicit regulation, the intermediate layer is easily contaminated by mixed-quality information, and errors propagate along its dependencies, so model scale often ends up substituting for absent regulation. We argue that an agent's mental model should instead evolve through explicit feedback that continuously aligns task understanding with reality, and identify three regulatory loops: interpretive update, deviation feedback, and schema revision. We realise this in VeriTrace, a cognitive-graph framework that explicitly implements the three loops. Using matched Qwen3.5-27B backbones, VeriTrace improves over the strongest matched baseline by an average of 4.82 pp on DeepResearch Bench (DRB) Insight (1.83 pp Overall) and by 5.9 pp Overall win rate on DeepConsult. With Config-DeepSeek, it achieves the strongest reproducible open-source result on DRB.
Haolang Zhao, Yunbo Long, Lukas Beckenbauer +1
Department of Engineering, University of Cambridge · TUM School of Management, Technical University of Munich · The Alan Turing Institute
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
The DeepResearch Agent System is a large language model system engineered for deep information retrieval, multi-step reasoning, and autonomous research tasks. Built upon a sparse activation architecture with 30 billion total parameters of which only 3 billion are activated per token, the system achieves state-of-the-art performance on multiple agent search benchmarks while delivering 3.2 times faster inference compared to dense counterparts of equivalent scale. The system supports a 128K-token context window with hierarchical attention mechanisms that yield 18.7% accuracy and 23.4% recall improvements over standard long-context approaches. A dual-mode reasoning engine provides both a ReAct paradigm for basic multi-step problem solving and an IterResearch mode for high-performance iterative research with up to 20 reasoning steps, collectively delivering a 31.2% accuracy improvement over single-pass baselines. Multi-tool coordination integrates retrieval, computation, web search, and file parsing modules to achieve 92.1% tool-use accuracy. A reinforcement learning optimization framework based on the GRPO algorithm provides token-level policy gradients that improve training stability by 35% and accelerate convergence by 42%. An automated data synthesis pipeline with seed-based expansion achieves a 92.5% usability rate. Benchmark results include 87.3% on Humanity's Last Exam, 85.3% on BrowserComp Chinese, and 91.2% on WebWalkerQA. The system is fully open-sourced, including data synthesis, training, and inference code, and supports applications in academic research, business analysis, R&D support, and education.
Yong Huang, Yulu Huang, for the team Collaboration