acceptodds
Under review as a conference paper at ICLR 2027

daVinci-Searcher: Eliciting Deep Research Capabilities Data Efficiently

14 views

Abstract

Deep research agents are increasingly important for solving complex, information-intensive tasks that require sustained search and evidence integration, and training on synthetic trajectories has enabled open-source agents to achieve performance competitive with proprietary systems. We ask: *Can a small set of high-quality trajectories elicit strong deep research capabilities ***data efficiently*** through supervised fine-tuning alone?* We introduce **daVinci-Searcher**, which combines scalable task synthesis with trajectory selection. Our Wikidata-based pipeline constructs challenging, verifiable multi-entity questions, using obfuscated entities and cross-entity constraints to control difficulty and encourage sustained search. We further identify *loop-prone reasoning* in teacher trajectories, where excessively long assistant turns revisit the same hypotheses without progress. Supported by external judge evaluations, our *loop-aware trajectory filtering* uses maximum per-turn reasoning length as a proxy for looping and improves performance without requiring correct final answers. We find a practical sweet spot at 2,000 training trajectories across 4B and 35B students, with no consistent gains from further scaling. Our 35B agent achieves 66.4% on BrowseComp, 60.8% on BrowseComp-zh, and a 62.8% macro average across six research benchmarks, outperforming the same backbone fine-tuned on the substantially larger S1-DR and Red-Searcher datasets.

Primary area:
generative models
Keywords:
Deep Research, Synthetic Data, Agent, Post Training

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_RKYTOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a data-efficient approach to train deep research agents. The authors generate challenging multi-hop questions and filter training trajectories based on maximum per-turn reasoning length as a proxy for looping reasoning. Experiments show that with only 2k trajectories and SFT alone, the resulting 35B model achieves strong performance across 6 benchmarks.

Critical strengths
  1. The overall performance improvements are quite substantial, especially considering that the model is trained on only 2k trajectories.

  2. The paper provides some interesting and useful insights into training data quality, for example, incorrect trajectories do not necessarily hurt performance, while trajectories with looping reasoning are harmful. The "less is more" finding is also interesting.

  3. The authors conducted a fairly extensive set of experiments, including comparisons across multiple benchmarks, different model sizes, and several ablation studies.

Critical weaknesses
  1. My main concern/question is that the contribution of discard-all context management is quite substantial. I'm not entirely sure how this relates to the main claim of data efficiency, since discard-all is an inference-time strategy adopted from prior work. While the experiments demonstrate the effectiveness of the training data, a large part of the reported performance also seems to depend on this additional inference-time strategy.

  2. The maximum per-turn reasoning length seems to be a coarse proxy for looping reasoning, as in that it might discard some genuinely difficult trajectories that require long reasoning without being repetitive. I wonder how many such trajectories are filtered out. The paper also doesn't clearly specify how the filtering threshold is chosen.

  3. This might be minor, but although the final training set contains only 2k trajectories, generating these trajectories could still be quite expensive, especially given the large number of tool calls involved. It would be helpful to report the total cost of generating these 2k trajectories, including the discarded ones, to better understand the end-to-end data efficiency.

Questions for authors
  1. Have the authors looked into the cases that triggered discard-all context management? I'm curious whether these cases also contain looping reasoning. If so, simply excluding looping trajectories from SFT doesn't explicitly teach the model how to recognize or recover from such situations. It would be interesting to understand whether the loop-aware filtering actually reduces looping during inference, or whether discard-all is still needed to handle this failure mode.

  2. How is the threshold for loop-aware filtering selected?

Reviewer_Y89cOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents daVinci-Searcher, a data pipeline for training deep research agents through supervised fine-tuning. It builds challenging Wikidata-based questions by increasing the number of hidden entities, then filters teacher trajectories using maximum per-turn reasoning length as a signal for repetitive reasoning. With 2,000 training trajectories, the 35B model achieves a 62.8% average across six benchmarks, compared with 46.9% for the base model.

Critical strengths
  1. The method achieves strong results with a small SFT dataset. The evaluation across six benchmarks and two model sizes provides useful evidence that carefully selected trajectories can improve deep research agents. The comparison with larger training datasets also supports the practical value of the approach.

  2. The paper studies a concrete problem in synthetic training data: repetitive reasoning without progress. External loop scores and the filtering ablation support the proposed length-based signal. The analyses of dataset size and answer correctness provide further insight into which trajectories are useful for training.

Critical weaknesses
  1. The analysis supports reasoning length as a practical signal for looping, but the frequency of the problem in the collected data remains unclear. Figure 7 shows a strong correlation between length and loop scores, yet does not show how well the chosen cutoff separates repetitive reasoning from useful long reasoning. A brief analysis of the retained and removed examples would help explain what the filter removes and when it is useful.

  2. The benefit of entity scaling is not directly evaluated through downstream training. Appendix J shows that high-hop tasks require more tool calls, and Appendix E supports the value of the proposed task source under the same teacher and filtering pipeline. These results are useful, but do not isolate whether increasing the number of hidden entities leads to better training data. A direct comparison would strengthen the evidence for this design choice.

Questions for authors
  1. What fraction of trajectories does the length filter remove? Could the existing loop-score analysis help describe repetition in the retained and removed groups? A few examples near the cutoff would also help show whether useful long reasoning is sometimes removed.

  2. Have you compared downstream SFT performance using low-hop and high-hop tasks with the same teacher, filtering procedure, and number of training examples? Such a comparison would help show whether the additional task difficulty leads to better learning.

Reviewer_wUA8OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes to use only a small set of training trajectories to improve deep research agents using supervised fine-tuning alone. The key idea is to filter trajectories generated by the teacher model using the maximum reasoning length with a turn. In this way, redundant/repetitive reasoning can be mitigated. The authors construct challenging questions from Wikidata and synthesize 2,000 training trajectories. Two models (4B and 35B) trained on these trajectories achieve good performance across 6 benchmarks.

Critical strengths
  • The performance improvement is impressive. The proposed daVinci-Searcher consistently outperforms recent baselines (e.g., OpenResearcher, OpenSeeker, and Red-Searcher) with similar backbone size.
  • Besides performance comparison, the authors conduct some informative analyses, such as the filtering strategy, answer correctness, data size across student models, and context management.
Critical weaknesses
  • The connection between entity count and research difficulty is plausible sounding but not fully verified. Specifically, checking answer uniqueness in Wikidata does not show that all entity hidden entities must be resolved. A subset of hints may already indicate the answer.
  • The filtering experiment does not fully isolate the harm of redundant/repetitive reasoning. This is because filtering longer turns will change several factors (e.g., training data size and tool call distribution) simultaneously. However, it is not reported whether these factors roughly match before and after filtering.
Questions for authors
  • For Weakness 1 above, I would suggest the authors remove individual constraints and test whether the questions remain solvable. This helps validate the proposed difficulty control.
  • Do training data size and tool call distribution roughly match before and after filtering? Is there a way to make the experiment more controlled than Table 2?
4 comments
AuthorsOpenReview ↗

Response to Reviewer wUA8 (1/2): Entity Scaling and Task Difficulty

Dear Reviewer wUA8,

Thank you for your thoughtful questions about task difficulty and constraint redundancy. Below, we provide a controlled empirical comparison and a model of how entity scaling increases evidence-integration demands.

w1. A unique Wikidata answer may be identifiable from a subset of clues, so adding hidden entities may not increase difficulty.

q1. Do questions remain solvable when individual constraints are removed one at a time?

Empirical evidence. Appendix J compares low-hop tasks with 3–5 obfuscated entities along the root path and high-hop tasks with 7–9, using the same teacher and trajectory generator. We summarize the tool-call statistics and corresponding teacher accuracy below.

TrajectoryAvg. tool callsP50 tool callsP90 tool callsCorrect ratio (%)
Low-hop11.6763068.9
High-hop (ours)44.60378051.0
S1-DR*12.7872865.3

S1-DR* denotes S1-DR questions with trajectories regenerated by the same teacher.

Increasing the number of hidden entities raises average tool calls by 3.82× and lowers teacher accuracy by 17.9 percentage points. Both measurements demonstrate greater retrieval and reasoning demands.


We agree with Q1 that relations in the question graph can be redundant, here we also present a theoretical model shows that increasing the number of hidden entities still makes synthesized questions harder by reducing the probability of a correct answer.

Let nn count hidden entities, CiC_i the clues for entity ii, CC all clues, and A(S)A(S) candidate answers from a Wikidata query using clues SS. Given ∣A(C)∣=1|A(C)|=1, define k=∑i=1n1 ⁣[∣A(C∖Ci)∣>1].k=\sum_{i=1}^{n}\mathbf{1}\!\left[|A(C\setminus C_i)|>1\right].

kk counts indispensable clue groups. For example, n=8n=8 and k=3k=3 indicate three indispensable groups among eight hidden entities. Our construction preserves a linear backbone, excluding diamonds or other complex topologies that introduce shortcuts. On this backbone, we model each added entity as independently adding an indispensable group with fixed probability α>0\alpha>0, giving E[k∣n]=αn\mathbb{E}[k\mid n]=\alpha n.

With one correct candidate, dd distractors, and score advantage Δ\Delta, softmax selection gives p(d,Δ)=eΔeΔ+d=11+de−Δ.p(d,\Delta)=\frac{e^{\Delta}}{e^{\Delta}+d}=\frac{1}{1+d e^{-\Delta}}.

Let LL be context length in thousands of tokens and ρ\rho the confusable-candidate density per thousand tokens, so d(L)≈ρLd(L)\approx\rho L. The solution path integrates all kk indispensable groups; applying the chain rule with a common conditional selection probability yields

Pground(k,L)≈(1+ρLe−Δ)−k.P_{\mathrm{ground}}(k,L)\approx\left(1+\rho L e^{-\Delta}\right)^{-k}.

Each additional indispensable group multiplies success by a factor below one. Independent backbone extensions therefore lower expected answer accuracy as the number of hidden entities grows. At 0.9 success per group, success falls from 72.9% for three groups to 43.0% for eight. This explains the observed increase in task difficulty.

AuthorsOpenReview ↗

Response to Reviewer wUA8 (2/2): Separating the Two Filtering Components

Dear Reviewer wUA8,

Thank you for suggesting a more controlled analysis of trajectory filtering. Below, we separate the two filtering components while holding the training set size fixed.

w2. Filtering long turns may also change training set size and tool-call distribution, making the effect of repetitive reasoning hard to isolate.

q2. How do training set size and tool-call distribution compare before and after filtering? Can Table 2 control for both?

Training set size is already fixed at 2,000 trajectories in the Table 2 condition. This applies to Table 2 and to the expanded ablation below. We vary two components independently: token-based filtering uses maximum per-turn reasoning length, and turn-based filtering selects trajectories by their interaction count.

Token-basedTurn-basedBrowseComp-zhBrowseComp
✗✗39.132.6
✓✗57.361.4
✗✓44.238.9
✓✓60.866.4

The token-based component produces the larger gain. With turn-based filtering disabled, it improves BrowseComp-zh by 18.2 percentage points and BrowseComp by 28.8 points. Turn-based filtering alone improves the same benchmarks by 5.1 and 6.3 points. Combining both components achieves the strongest results, adding 3.5 and 5.0 points over token-based filtering alone.

The turn-based component directly selects the interaction-count distribution. The token-only comparison establishes the benefit of the reasoning-length proxy at a fixed data budget with turn-based filtering switched off. Together, the four conditions show the contribution of each component and the additional benefit of combining them.

Best,
The authors of Paper 25444

I thank the authors for their detailed response, which addresses most of my concerns. (There is a small typo in the first formula in the theoretical model, which I assume "1!" should be "1", or the indicator function.) After reading the other two reviews, I decide to maintain my positive rating of 3.

Thank you for taking the time to review our rebuttal. We sincerely appreciate your recognition of our work and your suggestions regarding task difficulty and the inclusion of a more detailed ablation study. We will revise our manuscript accordingly.

Thank you again for your time and thoughtful feedback!

Comments on the paper

1 comment

I thank the authors for the detailed responses. I decided to maintain my positive rating of 3.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.