acceptodds
Under review as a conference paper at ICLR 2027

DualSpec: Accelerating Deep Research Agents via Dual-Process Action Speculation

Abstract

Large language model-based deep research agents have been increasingly popular for addressing long-horizon information-seeking tasks, but they often incur high end-to-end latency due to extensive reasoning and frequent tool use. Speculation frameworks aim to reduce latency by overlapping action execution with reasoning; however, existing approaches typically rely on uniform speculation strategies and strict action matching, which limits inference speedups and robustness. In this work, we revisit the speculate-verify paradigm for deep research agents through the lens of action heterogeneity. We show that Search and Visit actions exhibit fundamentally different reasoning and model capacity requirements: entropy-based analysis reveals that Search decisions have higher uncertainty and benefit significantly from explicit reasoning, whereas Visit decisions have lower entropy and depend primarily on model capacity. Motivated by this dual-process characteristic, we propose DualSpec, a heterogeneous speculation framework equipped with a lightweight, confidence-based semantic verifier. Experiments across multiple models and benchmarks demonstrate that DualSpec achieves up to 3.28 end-to-end speedup while maintaining accuracy comparable to fully reasoning agents.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Deep Research Agents, Speculative Execution, Inference Acceleration

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_AQmcOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DualSpec, a framework for reducing the inference latency of deep research agents. The authors observe that different tool actions have different reasoning requirements: Search actions benefit more from explicit reasoning, while Visit actions can often be generated directly by a larger model without reasoning. Based on this observation, DualSpec generates two candidate actions in parallel, using a small reasoning model and a large non-reasoning model. It then selects a candidate and uses the large model as a semantic verifier to decide whether to accept the action or fall back to full reasoning. The method is evaluated using MiroThinker and Qwen3 model pairs on three deep research benchmarks, reporting latency reductions while maintaining accuracy comparable to the base model.

Critical strengths

1. A useful observation about different agent actions. I think distinguishing the reasoning requirements of Search and Visit actions is a reasonable and practically motivated idea. The analysis suggests that applying the same inference strategy to every tool call may not be the most efficient choice.

2. Considering semantic verification instead of exact action matching. I appreciate the attempt to move beyond strict action matching, since different search queries can be semantically similar without being identical. Using a lightweight verification step to avoid unnecessary full reasoning is also a potentially useful direction for improving agent efficiency.

Critical weaknesses

1. Limited methodological novelty and questionable efficiency of the proposed design.

My main concern is that DualSpec does not seem to introduce a substantially new speculation mechanism. The framework mainly combines two existing ideas: generating candidate actions with different models and using a larger model to verify them. Similar model cascading and verification strategies have already been studied in previous work, such as SpecReason and other speculative agent frameworks.

More importantly, I find the actual efficiency of the design somewhat questionable. As described in Section 5.2, the system selects between the two drafts based on the action type and reasoning length produced by the small model. This means that even when the non-reasoning large-model draft is selected, the system still needs to wait for the small model to finish its reasoning before making that decision. I am not convinced that this is an efficient way to implement the proposed System 1 and System 2 distinction.

Also, unlike previous speculative agent approaches that overlap tool execution with model reasoning, DualSpec appears to execute tools only after the draft has been verified. In this sense, the framework looks more like a model routing and verification pipeline than speculative execution. I think the authors should better clarify the distinction and explain where the actual latency savings come from.

The theoretical analysis is also relatively weak. It largely relies on the standard observation that conditioning reduces uncertainty. Moreover, the lower token uncertainty of Visit actions may partly result from selecting URLs already present in the context, rather than reflecting a fundamentally different reasoning process. I do not find the theoretical contribution particularly convincing.

2. Insufficiently convincing experimental comparisons and unclear practical benefits.

Although the reported speedups are promising, I have several concerns about the fairness and interpretation of the results.

First, DualSpec uses two models running on separate GPUs, while the base model is evaluated on a single GPU. Since the main contribution is inference acceleration, I think a hardware-matched comparison is important. For example, the authors should compare against a base model deployed with the same GPU resources before concluding that the framework provides a substantial efficiency advantage.

Second, some of the simpler baselines already achieve competitive results. In Table 1, generating Visit actions with the large model without reasoning achieves 64.08% accuracy on GAIA, slightly higher than the original base model and the reported DualSpec result. In Table 2, LLM-direct is actually faster than DualSpec on Miro/GAIA, although with lower accuracy. On Miro/XBench, LLM-direct achieves 65% accuracy in 492 seconds, compared with DualSpec's 66% in 480 seconds. This is a rather small difference considering that DualSpec requires an additional model and verification procedure.

Finally, the cost results are not uniformly favorable. For example, Table 3 shows that DualSpec is approximately 11% more expensive than SPAgent for the MiroThinker 72B + 30B-A3B configuration under the authors' pricing assumptions. I think this limitation deserves more discussion.

Additional concerns about presentation.

The presentation needs improvement. Several figures are difficult to understand, particularly Figure 6, where the verification workflow is not clearly illustrated. There are also noticeable citation formatting and spacing issues throughout the paper.

I also found some important implementation details unclear. For example, the exact thresholds used for draft selection and verification are not reported, and the procedure for constructing the development split is insufficiently explained.

In addition, Section 5.3 describes a verifier that evaluates reasoning coherence and action usefulness, while the actual prompt in Appendix A.1 mainly checks whether the trajectory is progressing and includes a criterion about final-answer grounding. I find this inconsistency confusing, especially since the verifier operates on intermediate tool actions.

Questions for authors
  1. Method design and actual speculation: Since the routing decision depends on the small model's generated action type and reasoning length, does DualSpec need to wait for the small model to complete its reasoning even when selecting the non-reasoning large-model draft? Also, how does the proposed framework differ from standard model cascading when tool execution happens only after verification?

  2. Fairness of latency comparisons: Since DualSpec uses two GPUs while the base model uses one, could the authors provide comparisons under the same hardware budget, such as running the base model with two GPUs? How much of the speedup comes from the proposed method rather than additional hardware resources?

  3. Effectiveness compared with simpler alternatives: Table 1 shows that using the large model without reasoning for Visit already achieves strong results, while Table 2 shows that LLM-direct performs competitively in both accuracy and latency. What additional benefit does the full DualSpec framework provide over these simpler configurations under matched resource budgets?

4 comments

We thank the reviewer for their careful review and comments on the drafting-verification workflow, hardware-matched comparisons, simpler baselines, cost trade-offs, and presentation clarity. We have revised the paper in light of this feedback.

W1 / Q1

Waiting for the small model's routing signal introduces zero or negligible idle waiting on Visit steps. Average single-step generation times on GAIA-Text-103 are reported below.

Model pairSLM w/ reasoning (s/step)LLM w/o reasoning (s/step)
MiroThinker 72B + 8B2.377.95
Qwen3 32B + 4B6.974.22

On MiroThinker, SLM w/ reasoning finishes faster than LLM w/o reasoning: Search steps proceed immediately at 2.37s without waiting for the 72B model, while Visit steps incur zero extra waiting since the concurrent 72B draft is already running. On Qwen3, the per-step gap is under 2.8s while keeping long syntheses above τthink=400\tau_{\text{think}} = 400 tokens immediately available.

Stage-wise runtime breakdowns further show where the end-to-end savings come from, reported in mean seconds per query and shares of total latency.

StageMiro 72B BaseMiro 72B+8B DualSpecQwen 32B BaseQwen 32B+4B DualSpec
Base / fallback reasoning884 (84.9%)160 (26.5%)71.0 (89%)31.0 (68%)
Parallel drafting—244 (40.4%)—9.2 (20%)
Semantic verification—54 (8.9%)—0.9 (2%)
Tool execution157 (15.1%)146 (24.2%)9.0 (11%)4.6 (10%)
Total1041 (100%)605 (100%)80 (100%)46 (100%)

Executing tools after verification is a deliberate choice: computing the Yes/No log-odds requires a single forward prefill taking only 2%–8.9% of runtime, so speculating web calls before this check saves negligible time while wasting live API quota on the roughly 20% rejected steps.

Unlike sequential model cascading, which runs a small model first and invokes a larger model upon failure along one axis, DualSpec concurrently runs two complementary pathways and reuses the base model for single-token prefill verification. We have clarified this in Section 5.1 and revised Section 4 to present the entropy proxy results as empirical observations, noting the impact of URL copying while relying on discrete URL hit rates in Figure 4b and interventions in Table 1.

W2 / Q2

Our main setup placed the two models on separate GPUs to model cloud API deployments such as DSP, where independent endpoints serve multi-batch traffic at stable throughput and are evaluated via latency and metered token cost in Table 3; DSP and SPAgent use the same two-GPU setup in Figure 7. For local single-GPU deployment, where interactive agents typically serve single-request (batch = 1) workloads, we co-located both models on one 80GB A100 GPU under batch = 1 on GAIA-Text-103 (added in Appendix D).

Model pairDeploymentGPUsMean latency (s/query)Speedup vs. 1-GPU Base
MiroThinker 72BBase110411.00×
MiroThinker 72B + 8BDualSpec co-located16231.67×
MiroThinker 72B + 8BDualSpec separate endpoints26051.72×
Qwen3 32BBase1801.00×
Qwen3 32B + 4BDualSpec co-located1481.67×
Qwen3 32B + 4BDualSpec separate endpoints2461.74×

Single-GPU co-location retains a 1.67× speedup on both model pairs, closely matching separate endpoints. Together with the super-linear 2.6× average speedup on 72B + 30B-A3B and the 49.5%–66.6% token cost reduction in Table 3, this confirms that the gains come from removing roughly 80% of base-model reasoning tokens rather than extra hardware.


W2 / Q3

Table 1 is an offline diagnostic study with Oracle action-type knowledge (clarified in Section 3.1): full base-model reasoning first determines each step's action type before selectively replacing Visit steps. At actual inference time, the next action type is unknown prior to reasoning, so DualSpec runs the small reasoning model to predict the action type online while simultaneously drafting Search.

Moreover, LLM-direct in Table 2 is not a verifier-free baseline—all variants in Table 2 already include our semantic verifier and full-reasoning fallback, a clarification restored in Section 6.3. Even with identical verification and fallback, LLM-direct loses 3.9 percentage points on Miro/GAIA and is both less accurate and slower than DualSpec across all other three settings:

Pair / DatasetLLM-direct (+ Verifier): Pass@1 / LatencyDualSpec: Pass@1 / Latency
Miro 72B+8B / GAIA59.2% / 575s63.1% / 605s
Miro 72B+8B / XBench65.0% / 492s66.0% / 480s
Qwen 32B+4B / GAIA27.1% / 67s30.1% / 46s
Qwen 32B+4B / XBench25.0% / 46s27.0% / 41s

In Table 3, DualSpec costs 16.9% and 17.6% less than SPAgent on 72B + 8B and 32B + 4B. On 72B + 30B-A3B, SPAgent lowers cost by 11.0% only by skipping verification entirely in early turns, which degrades accuracy and increases latency in Figure 7b, whereas DualSpec still cuts base-model cost by 49.5% while preserving accuracy at a 2.6× speedup (discussed in Section 6.2 and Appendix B).


Presentation and Implementation Details

  • Verifier prompt alignment: Because each ReAct step emits either a tool call or a Final Answer and drafts may conclude prematurely or hesitate, the prompt in Appendix A.1 audits all steps via stagnation and drift rules for tool calls (Criteria 1 and 3) and grounding for final answers (Criterion 2). Section 5.3 now states all three criteria explicitly.
  • Thresholds and calibration split: Using 20 GAIA-Text-103 questions without inspecting ground-truth answers, we calibrated τ\tau solely to match an unlabeled roughly 20% step-fallback rate (τ=1.5\tau = 1.5 for both MiroThinker-72B pairs, τ=0.4\tau = 0.4 for Qwen3-32B, and shared τthink=400\tau_{\text{think}} = 400 tokens; Section 6.1), and reused these exact values zero-shot on XBench-DeepSearch and Seal-0.
  • Figure 6 and citations: We expanded the Figure 6 caption to clarify parallel drafting, routing, verification, and fallback step by step, and fixed all citation and spacing issues.

Dear Reviewer AQmc,

Thank you again for your review. Our responses and the revised manuscript (with changes highlighted in blue) are posted above. W1/Q1 provides single-step drafting latencies and stage-wise runtime breakdowns showing why waiting for the small model adds zero or negligible overhead, why tools execute after single-token prefill verification, and how DualSpec differs from sequential cascading, while reframing Section 4 as empirical observations. W2/Q2 adds hardware-normalized single-GPU co-location experiments demonstrating a 1.67× speedup on both model pairs on one A100 GPU (Appendix D). W2/Q3 clarifies that Table 1 is an offline Oracle action-type diagnostic (Section 3.1), confirms that LLM-direct in Table 2 already uses our verifier and fallback, and discusses the Table 3 cost–accuracy trade-off (Section 6.2 and Appendix B). We also aligned Section 5.3 with the Appendix A.1 prompt, documented all thresholds, expanded the Figure 6 caption, and fixed citation formatting.

Could you let us know whether these additions address your concerns, or which points remain open? We remain available to answer any further questions during the discussion. If your concerns are resolved, we would be grateful if you could reflect this in your assessment.

Best regards,

The authors

I thank the authors for their thorough response and want to confirm that I have read it. I will keep my score.

Reviewer_y7wMOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DualSpec, a method for accelerating deep research agents through action-specific speculation. The main motivation is that search benefits more from explicit reasoning, while visit benefits more from model capacity. DualSpec combines a small reasoning model with a large model that generates actions without reasoning, selecting between their drafts according to the action type and reasoning length. A lightweight large-model critic accepts drafts using a confidence score, with full reasoning used as a fallback. Experiments on public benchmarks cover three model pairs and report 1.33–3.28× end-to-end speedups with generally comparable accuracy. Action-level replacement studies and an intervention-rate analysis provide additional validation.

Critical strengths
  1. The action-specific design is supported by direct interventions. Table 1 selectively changes the generation strategy for search or visit and measures end-to-end accuracy. Small-model reasoning preserves search performance for both model families, while large-model direct generation approximately preserves Visit performance. This provides stronger support for the routing policy than action similarity or reasoning-length measurements alone.

  2. The reported speedups are substantial and evaluated across several settings. Figure 7 covers three model pairs, three benchmarks, and comparisons with DSP and SPAgent. Table 2 reports concrete latency reductions while retaining 63.1% accuracy. The intervention-rate analysis further shows how fallback recovers accuracy.

Critical weaknesses
  1. The theoretical interpretation in Section 4 is stronger than the analysis supports. Eq. (5) concerns conditioning within a fixed joint distribution; it does not directly compare the model’s reasoning and no-reasoning modes. The token-level proxy in Eq. (2) can also be affected by output formatting and URL copying. The authors should clarify the assumptions connecting these quantities or present the entropy results as empirical observations. The action-level experiments in Table 1 remain useful evidence for the design.

  2. Table 2 does not clarify whether LLM-direct and SLM-CoT use the same semantic verifier and fallback policy as DualSpec. This matters because lightweight verification with thresholded fallback is already used in SpecReason, making heterogeneous drafting an important distinction of this work. If verification is shared across the variants, the paper should state this explicitly and report their fallback rates. Otherwise, a comparison with verification held fixed is needed to quantify the benefit of action-specific drafting.

Questions for authors
  1. Do the drafting variants in Table 2 share the same verifier, threshold-selection procedure, and fallback policy? What are their respective fallback rates?

  2. What assumptions connect the no-reasoning policy used in the experiments to the marginal policy in Eq. (4)? How should the conclusion following Eq. (5) be qualified if they differ?

  3. Which GAIA examples are used for threshold tuning, and are they separate from GAIA-Text-103? What verifier and reasoning-length thresholds are used for each model pair?

3 comments

The authors did not submit a response during the rebuttal period. My questions regarding the theoretical framing in Section 4 and whether the single-drafter baselines in Table 2 share the same semantic verifier and fallback policy remain unaddressed. I will therefore keep my initial score.

We thank the reviewer for their careful review and comments on the theoretical interpretation in Section 4, the verification setup in Table 2, and the threshold calibration protocol. We have revised the paper in light of this feedback.

W1 / Q2

Connecting Eq. (5) to the no-reasoning mode assumes that direct generation πdirect(a∣s)\pi_{\text{direct}}(a \mid s) approximates the marginal policy πmarg(a∣s)=∑zπ(z∣s)π(a∣s,z)\pi_{\text{marg}}(a \mid s) = \sum_z \pi(z \mid s)\pi(a \mid s, z) without distribution shift. Because the two policies can differ in practice, Eq. (5) serves as an information-theoretic motivation rather than a strict inequality across modes: omitting zz shifts the policy and raises uncertainty on non-local Search queries, whereas πdirect\pi_{\text{direct}} stays concentrated on context-grounded Visit selections.

We also agree that Eq. (2) is influenced by output formatting and URL copying, although comparing each action type with and without reasoning cancels fixed schema tokens. Following your suggestion, Section 4 now presents the entropy proxy results as empirical observations under these qualifications, relying primarily on Figure 4 and Table 1 to justify our routing design.


W2 / Q1

All variants in Table 2 share the exact same semantic verifier, threshold-selection procedure, and fallback policy—a clarification trimmed for space and restored in Section 6.3, the Table 2 caption, and Appendix C.2. Results on GAIA-Text-103 with step-level fallback rates are shown below.

PairDatasetVariantPass@1 (%)Latency (s)Fallback rate (%)
Miro 72B+8BGAIABase63.11041—
Miro 72B+8BGAIALLM-direct59.257514.4%
Miro 72B+8BGAIASLM-CoT56.365117.6%
Miro 72B+8BGAIADualSpec63.160518.1%
Qwen 32B+4BGAIABase29.180—
Qwen 32B+4BGAIALLM-direct27.16711.2%
Qwen 32B+4BGAIASLM-CoT25.23217.9%
Qwen 32B+4BGAIADualSpec30.14619.2%

At comparable roughly 18%–19% fallback rates, DualSpec outperforms SLM-CoT by 6.8 and 4.9 percentage points while running faster on MiroThinker by skipping small-model reasoning on Visit. LLM-direct triggers slightly fewer fallbacks because omitting reasoning allows weak Search queries to pass verification, reducing accuracy and adding extra search turns on Qwen3 (67s vs. 46s).


Q3

Because GAIA-Text-103 comprises all 103 text-only validation questions in GAIA, we sampled 20 questions to calibrate τ\tau solely to match an unlabeled roughly 20% step-fallback rate, without inspecting ground-truth answers or task accuracy (detailed in Section 6.1). The exact thresholds used across all experiments are listed below.

Model pairCritic modelVerifier threshold τ\tauReasoning threshold τthink\tau_{\text{think}} (tokens)
MiroThinker 72B + 8BMiroThinker-72B1.5400
MiroThinker 72B + 30B-A3BMiroThinker-72B1.5400
Qwen3 32B + 4BQwen3-32B0.4400

Here τ\tau depends only on the critic model's log-odds scale, while τthink=400\tau_{\text{think}} = 400 tokens is fixed across all pairs. Reusing these exact thresholds zero-shot on XBench-DeepSearch and Seal-0, alongside the flat accuracy plateau across 15%–30% fallback rates in Appendix C, confirms that DualSpec is insensitive to threshold tuning.

Dear Reviewer y7wM,

Thank you for following up! As the AoE discussion window is still open, we have just posted our detailed response above and uploaded the revised manuscript (with changes highlighted in blue), which directly addresses all of your questions:

  1. Section 4 framing (W1 / Q2): Following your suggestion, we have reframed Section 4 as empirical observations rather than a formal guarantee, explicitly qualifying both the approximation gap between direct no-reasoning generation and the marginal policy in Eq. (4)–(5) and the formatting/URL-copying effects in Eq. (2).
  2. Shared verifier in Table 2 (W2 / Q1): Yes—all single-drafter baselines in Table 2 (LLM-direct and SLM-CoT) share the exact same semantic verifier, threshold-selection procedure, and full-reasoning fallback policy as DualSpec. We have restored this explanation in Section 6.3 and the Table 2 caption, and reported the step-level fallback rates and exact threshold values (τ\tau and τthink\tau_{\text{think}}, Q3) in our response above and Appendix C.2.

We hope our response and revisions resolve your concerns, and we would be glad to answer any further questions during the discussion period.

Best regards,

The Authors

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

This paper presents DualSpec, a framework for reducing the latency of deep research agents. It generates candidate actions using a small model with reasoning and a large model without reasoning, favoring the former for Search and the latter for Visit, selecting between them based on the action type and the length of the small model’s reasoning. A large-model critic checks each selected draft and falls back to full reasoning when confidence is low. Experiments on GAIA-Text-103, XBench-DeepSearch, and Seal-0 report 1.33–3.28× speedups over fully reasoning base agents with comparable accuracy.

Critical strengths
  1. The action-level interventions provide useful evidence for action heterogeneity: Search benefits more from explicit reasoning, while Visit depends more on model capacity. These results give a clear empirical motivation for using an SLM with reasoning for Search and an LLM without reasoning for Visit.
  2. DualSpec achieves meaningful end-to-end latency reductions across two model families and three benchmarks, reporting 1.33–3.28× speedup with comparable pass@1 accuracy. The framework requires no additional model training, making it practical to integrate into existing deep research agents.
Critical weaknesses
  1. The methodological novelty appears incremental. Action-level speculation, fast and slow inference, and semantic verification with fallback have precedents in DSP, SPAgent, and SpecReason. The main addition is the empirical Search/Visit distinction and the corresponding action-aware selection rules. While useful, the paper does not sufficiently establish the scope and significance of this addition beyond adapting existing mechanisms to these two action types.
  2. The ablations do not sufficiently isolate the benefits of the proposed designs. Table 2 does not clarify whether the drafting variants share the same semantic verifier and fallback procedure. The long-reasoning preservation rule is not separately evaluated, nor is semantic verification compared with action matching under a fixed drafting policy. These omissions make it difficult to attribute the reported gains to the individual contributions.
Questions for authors
  1. Do the drafting variants in Table 2 use the same semantic verifier and fallback procedure? A comparison under identical verification settings would help isolate the benefit of heterogeneous drafting.
  2. How is τ_"think" selected, and how does removing the long-reasoning preservation rule affect accuracy and latency?
  3. Are the main results averaged over multiple runs? Please clarify the evaluation protocol and report variability in accuracy and latency.
  4. Could the authors provide a runtime breakdown of drafting, verification, fallback, and tool execution to quantify where the latency savings come from?
3 comments

We thank the reviewer for their careful review and comments on the methodological scope, ablation design, evaluation protocol, and runtime breakdown. We have revised the paper in light of this feedback.

W1

DualSpec differs from prior speculative frameworks in both critical-path execution and action-aware routing:

  • DSP: Verifies small-model drafts via edit distance against the base model, keeping full base-model reasoning on the critical path and limiting speedups to overlapping tool calls.
  • SPAgent: Skips verification in fixed early turns without quality control, then reverts to strict action matching with base-model reasoning on the critical path.
  • SpecReason: Offloads steps uniformly by model capacity in single-model, tool-free reasoning.
  • DualSpec: Pairs heterogeneous drafting with single-token prefill log-odds verification, bypassing base-model autoregressive decoding on accepted steps.

Beyond Search and Visit, this design addresses a broader division in tool-using agents between high-entropy synthesis actions requiring deliberative reasoning and context-grounded selection actions relying primarily on model capacity. Tables 1 and 2 show that uniform degradation fails on one of the two regimes, whereas DualSpec accelerates both without sacrificing accuracy.


W2 / Q1

All variants in Table 2 share the exact same semantic verifier, threshold-selection procedure, and fallback policy—a clarification trimmed for space and restored in Section 6.3 and Appendix C.2. Results on GAIA-Text-103 with step-level fallback rates are reported below.

PairDatasetVariantPass@1 (%)Latency (s)Fallback rate (%)
Miro 72B+8BGAIABase63.11041—
Miro 72B+8BGAIALLM-direct59.257514.4%
Miro 72B+8BGAIASLM-CoT56.365117.6%
Miro 72B+8BGAIADualSpec63.160518.1%
Qwen 32B+4BGAIABase29.180—
Qwen 32B+4BGAIALLM-direct27.16711.2%
Qwen 32B+4BGAIASLM-CoT25.23217.9%
Qwen 32B+4BGAIADualSpec30.14619.2%

At comparable roughly 18%–19% fallback rates, DualSpec improves pass@1 over SLM-CoT by 6.8 and 4.9 percentage points while running faster on MiroThinker by avoiding small-model reasoning on Visit. LLM-direct triggers slightly fewer fallbacks because omitting reasoning lets weak Search queries pass verification, lowering accuracy and adding search turns on Qwen3.

W2 / Q2

To isolate semantic verification and long-reasoning preservation, we fix DualSpec's heterogeneous drafter on GAIA-Text-103 (added in Appendix C.2). Based on Figure 3, we set a fixed heuristic τthink=400\tau_{\text{think}} = 400 tokens across benchmarks, since routine Visit steps use short reasoning while longer traces summarize cross-step evidence.

ConfigurationQwen 32B+4B: Pass@1 (%) / Latency (s)Miro 72B+8B: Pass@1 (%) / Latency (s)
DualSpec drafter + No verifier26.2 / 4156.3 / 587
DualSpec drafter + Action matching30.1 / 7163.1 / 876
DualSpec without τthink\tau_{\text{think}}30.1 / 5262.1 / 637
DualSpec30.1 / 4663.1 / 605

Removing verification drops pass@1 by 3.9 and 6.8 points due to compounding draft errors, whereas action matching increases latency to 71s and 876s by keeping base-model reasoning on the critical path. Removing τthink\tau_{\text{think}} discards cross-step syntheses, slowing convergence to 52s and 637s and lowering MiroThinker accuracy from 63.1% to 62.1%.

Q3

Main results report single runs on single-tenant SGLang (batch = 4) with fixed Bing and Jina backends; five independent trials on GAIA-Text-103 with MiroThinker 72B+8B are summarized below as mean ± standard deviation.

SystemPass@1 (%)Mean latency (s/query)
Base63.3 ± 0.41018 ± 97
DualSpec63.1 ± 0.7632 ± 28

Run-to-run variance is small relative to the 1.61× mean speedup, with DualSpec consistently matching base-model accuracy.


Q4

Stage-wise runtime breakdowns in mean seconds per query and shares of total latency are reported below.

StageMiro 72B BaseMiro 72B+8B DualSpecQwen 32B BaseQwen 32B+4B DualSpec
Base / fallback reasoning884 (84.9%)160 (26.5%)71.0 (89%)31.0 (68%)
Parallel drafting—244 (40.4%)—9.2 (20%)
Semantic verification—54 (8.9%)—0.9 (2%)
Tool execution157 (15.1%)146 (24.2%)9.0 (11%)4.6 (10%)
Total1041 (100%)605 (100%)80 (100%)46 (100%)

Semantic verification takes only 2%–8.9% of runtime via a single forward prefill, while parallel drafting replaces full base-model reasoning on over 80% of steps, cutting large-model reasoning time from 884s to 160s on MiroThinker and from 71.0s to 31.0s on Qwen3.

Dear Reviewer CDsX,

Thank you again for your review. Our responses and the revised manuscript (with changes highlighted in blue) are posted above. W1 clarifies how DualSpec removes base-model autoregressive reasoning from the critical path and generalizes beyond Search and Visit. W2/Q1 confirms that all variants in Table 2 share the exact same semantic verifier and fallback policy and reports their step-level fallback rates; W2/Q2 adds factorized ablations isolating semantic verification, action matching, and the long-reasoning preservation threshold τthink=400\tau_{\text{think}} = 400 (Appendix C.2). Q3 reports five-run mean and standard deviation statistics, and Q4 provides the stage-wise runtime breakdown across drafting, verification, fallback reasoning, and tool execution.

Could you let us know whether these additions address your concerns, or which points remain open? We remain available to answer any further questions during the discussion. If your concerns are resolved, we would be grateful if you could reflect this in your assessment.

Best regards,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.