acceptodds
Under review as a conference paper at ICLR 2027

SpecSplice: Dynamic Composition of Parametric and Retrieval-based Drafts for Agentic Speculative Decoding

Abstract

Agentic workloads interleave novel text with repeated spans, making parametric and retrieval drafters complementary for speculative decoding. Within a single round, parametric draft tokens can expose retrieval matches unavailable from the committed context alone. Existing hybrids limit these opportunities through round-level source selection or prescribed attachment rules. We introduce SpecSplice, which jointly selects candidates from both sources under a shared node budget, treating the root and every node of a supplied parametric draft as potential splice points. A unified contribution score estimates each candidate's survival probability, accounting for the parametric prefix leading to a retrieved node. Score bounds enable a best-first algorithm to interleave retrieval with node selection, skipping queries that cannot affect the selected tree. We prove that the resulting tree maximizes the total contribution score over this candidate space under the node budget, exactly as if retrieval were queried at every splice point. Across four agentic workloads and three target models, SpecSplice achieves up to a 6.91× speedup over autoregressive decoding and 13–37% higher throughput than the strongest baseline on every agentic model–workload pair.

Primary area:
foundation or frontier models, including LLMs
Keywords:
speculative decoding, inference acceleration, LLM Agent, retrieval-augmented decoding, efficient LLM inference

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

SpecSplice jointly selects parametric and retrieved draft candidates under a shared node budget. Prefix-aware scores enable best-first construction with on-demand retrieval, with optimality guaranteed for the defined surrogate objective. Experiments show 12.8–37.1% higher decoding throughput than the strongest baseline across 12 agentic model–workload pairs.

Critical strengths
  1. The method addresses a concrete limitation of existing hybrids: parametric prefixes can expose useful retrieval matches within a decoding round. Unified scoring and dynamic splicing provide a coherent solution with a sound surrogate-optimality argument.

  2. Throughput gains are consistent across models and agentic workloads. Component ablations, alternative drafting backends, official harnesses, and batching experiments strengthen the empirical evidence.

Critical weaknesses
  1. Moderate novelty. The method builds on established ideas in hybrid drafting, path-based scoring, and best-first tree construction. Its main advance is a unified, prefix-aware selection framework with dynamic retrieval at parametric nodes, making the methodological contribution useful but incremental.

  2. Parameters and node budgets are optimized per model–workload pair, but the manuscript does not clarify whether tuning uses held-out data. This limits confidence in performance on unseen tasks.

  3. Important evaluation details are missing, including sample counts, selected budgets, measurement variability, retrieval-index initialization, and precise timing boundaries. These omissions hinder reproducibility.

Questions for authors
  1. Were hyperparameters selected independently of the reported evaluation tasks? How well do fixed settings transfer to unseen tasks?

  2. Does throughput measurement include retrieval, tree construction, and index updates? Please clarify sample counts, repeated measurements, and index-reset policies.

3 comments
AuthorsOpenReview ↗

Response (1/2): Methodological contribution

We sincerely thank the reviewer for the careful and balanced review. Your comments on tuning and evaluation details helped us identify where clearer reporting and additional evidence were needed.

W1. Methodological contribution

We agree that hybrid drafting, path scoring, and best-first construction are established. Our contribution is to formulate hybrid draft construction as a joint optimization over individual parametric and retrieved nodes, under a shared verification budget and an explicit accepted-length surrogate.

The hybrid schemes compared in the paper combine drafts through source selection or prescribed composition rules. Such rules can leave useful combinations of individual candidates unexplored. In our formulation, candidate inclusion is the optimization decision: the selected nodes jointly determine the source allocation and splice structure. The candidate space includes continuations retrievable at the root and at every eligible parametric node (Section 2.3).

This joint optimization requires a comparable contribution estimate for the different evidence of the two sources. Parametric probabilities provide predictive confidence, and retrieval frequencies provide historical support. A retrieved node contributes only if its preceding parametric prefix is accepted, so its retrieval support must be evaluated together with prefix survival. Eq. (4) makes this dependency explicit on a common contribution scale, and Eq. (5) maximizes the sum of these contributions over the combined candidate space (Sections 3.1–3.2).

The controlled ablations show the value of treating hybrid composition as joint node selection. At a common B = 64 with retrieval fixed at the root, unified scoring raises MAT by 17.9% over Selection and 26% over Graft. Extending retrieval to all eligible parametric nodes, with scoring unchanged, adds a further 10.9% (Section 4.3, Figure 5(a–b)). These comparisons isolate the benefit of joint budget allocation and of the expanded candidate space.

Appendix J makes the resulting composition concrete: SpecSplice combines both sources within a verification step, which reduces the verification steps for the identical output from 12 to 9 (Figure 10(b)).

The same contribution score also supplies the path monotonicity and query bounds for efficient construction: SpecSplice skips queries that cannot affect selection while retaining the defined surrogate optimum over the supplied candidate space (Section 3.2, Propositions 1–2; Appendix A.2).

AuthorsOpenReview ↗

Response (2/2): Tuning protocol, transfer, and evaluation details

W2 / Q1. Tuning and transfer

Yes, the original hyperparameters were selected on the evaluation tasks. For each model–workload pair, every method's node budget and any method-specific parameter were selected to maximize throughput (Appendix C, Table 3). We should have been more explicit about this protocol, and we agree that it alone does not establish transfer to unseen tasks.

Following your suggestion, we added a transfer study (Section 4.5; Appendix I, Table 10). Tuning separately on Spider2-DBT and BFCL-v4 with Qwen3.5-9B selects the same value, λ = 0.1, which we reuse unchanged on the remaining agentic pairs. Speedup and MAT decrease by at most 2.0% and 1.0%, and SpecSplice remains 12–37% faster than the strongest baseline in Table 1 on all 12 agentic pairs. The original budgets are unchanged, so this study isolates the transfer of λ.

Budgets are examined separately: on Qwen3.5-9B/Spider2-DBT at batch size 1, SpecSplice has higher throughput than Selection at every tested common budget from 16 to 256, on both GPUs (Appendix F, Figure 9(a)).

The transfer holds for the agentic workloads. On Spec-Bench, λ = 0.1 lowers the two Qwen models' speedups by 7.7–9.1% relative to their tuned weights, so a smaller weight works better there (Table 10).

W3 / Q2. Timing and reproducibility details

We agree that these details should have been clearer, and they are now stated in the paper.

The metric is decode throughput. Measured decoding time includes drafting, retrieval, suffix-tree updates, tree construction, and verification (Section 4.1, "Metrics"). Table 2 separates drafting, meaning the DFlash forward pass plus construction including retrieval, from the remaining work, which is mainly target verification (Section 4.4).

The global suffix tree starts empty for each model–workload–method run and grows as each request's output is generated (Appendix C, "Retrieval index"). A separate per-request tree covers the current prompt and output (Appendix B.2).

Appendix C ("Workloads") now reports the evaluated sizes: 300 SWE-bench Lite instances, 68 Spider2-DBT tasks, 665 BFCL-v4 memory and web-search entries, 480 Spec-Bench questions, and 160 Gaia2 mini scenarios. The runtime breakdown separately averages over 100 Gaia2 tasks (Tables 2 and 7).

For variability, Appendix D.1 (Table 5) reports five sampling runs each for SpecSplice and Selection on Qwen3.5-9B/Gaia2. Mean speedup is 4.91 ± 0.09× for SpecSplice and 3.53 ± 0.07× for Selection, and the geometric-mean gain is 39.4% (95% CI 35.5–43.3%).

Reviewer_RjizOpenReview ↗

Recommendation After the Author Response

Thank you for your response. It addressed my questions, and I appreciate the clarifications. My original assessment remains unchanged, and I continue to lean toward acceptance.

Reviewer_Fn75OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

SpecSplice construct draft trees using parametric (neural nets) and retrieval drafters. In contrast with previous work that either select one drafter each round or use prescribed composition rules, SpecSplice is maximizing a surrogate objective used by OPT-Tree and DDTree. For the retrieval drafters, the surrogate probability is based on the fraction of matches. Experiments show substantial gains in terms of accepted length and overall speedup, especially in agentic benchmarks, with no degradation in non-agentic benchmarks.

Critical strengths
  1. The objective in (5) is elegant. It is an extension of OPT-Tree/DDTree surrogate objective for both parametric and retrieval drafters.
  2. Principled tree construction, optimizing (5).
  3. Experiments show substantial gains in terms of accepted length and overall speedup, especially in agentic benchmarks, with no degradation in non-agentic benchmarks.
  4. The method is easy to implement, given existing parametric and retrieval drafters.
Critical weaknesses
  1. Lack of analysis of the retrieval latency. If I understand correctly, retrieval drafters runs on cpu, so there might be a significant latency when transferring data from GPU to CPU and vice versa.
  2. There's no code or release statement. Low level details could be very important here, and reference code will be very valuable.
Questions for authors

Would you consider releasing the source code for the experiments?

1 comment
AuthorsOpenReview ↗

Response (1/2): Retrieval latency

We sincerely thank the reviewer for the careful review and for recognizing the objective and tree construction. Your comments helped us identify an omission in the retrieval-cost analysis, which we address below.

W1. Retrieval latency and decoding-step overhead

Retrieval and tree construction take 0.8 ms per step, or 2.9% of the step. The full step is 0.4 ms (1.5%) longer than Selection's, and all reported throughput already includes this cost. We agree that the paper needed an explicit cost analysis.

Throughput uses the complete decoding-step time, including drafting, verification, and suffix-tree updates (Section 4.1, "Metrics"). Drafting comprises the DFlash forward pass and construction until the verification tree is final, including retrieval. The remaining time is mainly target verification. SpecSplice's additional work is retrieval at parametric prefixes and joint tree construction (Section 4.4, "Drafting overhead").

We measured 100 Gaia2 tasks with Llama-3.1-8B-Instruct under greedy decoding, at B = 64 and batch size 1, on an RTX PRO 6000 with an Intel Xeon Gold 6534 CPU (Appendix G, "Setup"). Construction takes 0.8 ms, or 2.9% of the step (Table 7, "Cold start"). Total drafting takes 4.4 ms for SpecSplice and 4.1 ms for Selection, and the full step takes 27.9 ms and 27.5 ms (Section 4.4, Table 2).

Deferred querying becomes more useful as retrieval becomes costlier. Under cold start, the index accumulates responses from one workload, and exhaustive querying would add only 0.6 ms (2.2%) to SpecSplice's step (Table 2). To emulate an index shared across workloads, we prewarmed it with outputs from other agentic workloads. With 32K requests prewarmed, construction takes 1.5 ms (5.3%) for SpecSplice and 4.0 ms (13.2%) for exhaustive querying. Exhaustive querying returns the same tree but lowers throughput by 6.6% (Table 7, "32K"; "Index size").

In the budget and batch-size sweeps, SpecSplice's construction share stays below 5%, while exhaustive querying reaches 7.8% at B = 256 and 8.8% at batch size 4 (Table 7, "Budget and batch size"; summarized in Section 4.4).

Reviewer_9soYOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces SpecSplice, a hybrid speculative decoding technique for agentic workloads. The key idea is to use a model-based parametric speculator together with a retrieval-based model-free speculator. Compared to existing hybrids (e.g. SuffixDecoding's own hybrid speculator), SpecSplice builds a single tree mixing parametric and retrieved nodes, competing for a shared node budget, rather than picking one source per round. SpecSplice is evaluated using three different models and 4 agentic workloads on a RTX PRO 6000, and reports up to 6.91x speedup compared to vanilla decoding, and a 13–37% improvement over the best available speculation baselines.

Critical strengths
  • Well-motivated, simple mechanism (Splicing at any parametric node, ranked jointly; lazy queries keep cost bounded)
  • Broad, mostly convincing eval (12 agentic cells, robustness evaluation)
Critical weaknesses
  • λ is tuned on the evaluation traces themselves.
  • overheads and generality are underexplored
  • limited novelty with respect to existing hybrid speculator mechanisms (SuffixDecoding, Suffix Automaton, etc)
Questions for authors
  • how sensitive is the perforamnce to the choice of λ? Can you provide the speedup with fixed λ across all workloads. or with λ tuned on a held-out workload not part of the evaluation?
  • for the sampling experiments, each method follows a different trajectory. how do you account for those differences and how do you separate any performance gains from differences in output length or repetitiveness of the output?
  • how does SpecSlice perform in a concurrent setting with many requests?
  • how is the global suffix tree populated?
3 comments
AuthorsOpenReview ↗

Response (1/2): λ sensitivity and transfer, runtime overhead and generality

We sincerely thank the reviewer for the positive assessment and the thoughtful questions. They helped us identify where the manuscript needed clearer explanations and additional evidence. We address each point below.

W1 / Q1. Source weight selection, sensitivity, and transfer

The reviewer is right that λ was selected on the evaluation tasks, jointly with the node budget (Appendix C, Table 3). We should have discussed its sensitivity and transfer. Following the reviewer's suggestion, we added two analyses. They show that a value selected on Spider2-DBT and BFCL-v4 transfers to the other evaluated agentic workloads and models with little performance loss.

Transfer (Appendix I, Table 10). Tuning Qwen3.5-9B separately on Spider2-DBT and on BFCL-v4 selects the same value, λ = 0.1. We reused this value unchanged on the ten remaining agentic model–workload pairs, keeping each pair's original node budget. Compared with per-pair tuning, speedup decreases by at most 2.0% and MAT by at most 1.0%. SpecSplice remains 12–37% faster than the strongest baseline in Table 1 on all 12 agentic pairs.

Sensitivity (Section 4.5, Figure 7). For Qwen3.5-9B at B = 48, every tested λ from 0.08 to 0.3 keeps MAT within 2% of its maximum on each of the four agentic workloads. The transferred value lies within this region.

Scope. On Spec-Bench, the smaller tuned weights perform better than λ = 0.1 (Appendix I, Table 10). The transfer conclusion therefore concerns the evaluated agentic workloads.

Section 4.5 summarizes these analyses under "Weight sensitivity" and "Single weight for all pairs". Appendix C specifies the tuning protocol and the selected budgets.

W2. Runtime overhead and generality

The overhead is small. A SpecSplice step is 0.4 ms (1.5%) longer than a Selection step, and the gains persist under λ reuse, across node budgets, and across batch sizes. We agree that the paper needed clearer evidence on both points.

Overhead. All reported throughput uses the complete decoding-step time, including index updates (Section 4.1, "Metrics"). Drafting comprises the DFlash forward pass and tree construction, including retrieval. The remaining time is mainly target verification. SpecSplice's additional work is retrieval at parametric prefixes and joint tree construction. On Llama-3.1-8B/Gaia2 at B = 64 and batch size 1, the full step takes 27.9 ms for SpecSplice and 27.5 ms for Selection (Section 4.4, "Drafting overhead"; Table 2).

Generality. The overhead above is measured in one setting (Llama-3.1-8B/Gaia2, B = 64, batch size 1). The following results show that the gains and the low overhead are not specific to that setting: the gains hold with a single λ across models and agentic workloads and across node budgets and batch sizes, and the overhead stays low with a larger shared index, budget, and batch size.

Across models and workloads with a fixed λ. As described in W1, a single λ = 0.1 reused across the agentic pairs reduces speedup by at most 2.0% relative to per-pair tuning (Section 4.5; Table 10).

Across index sizes, budgets, and batch sizes: cost. When we prewarm the global index with outputs from 32K requests on other agentic workloads, SpecSplice spends 5.3% of the step on construction, compared with 13.2% for exhaustive querying. In the cold-start sweeps, this share stays below 5% as B grows to 256 or the batch size grows to 4. Deferred querying therefore reduces construction cost in these settings while preserving the selected tree (Appendix G, Table 7).

Across budgets and batch sizes: throughput. On Qwen3.5-9B/Spider2-DBT at batch size 1, SpecSplice has higher throughput than Selection at every tested budget from 16 to 256, on both the RTX PRO 6000 and the B200 (Appendix F, Figure 9(a)). With each method at its own throughput-optimal budget, SpecSplice is 15–31% faster at batch sizes 1–4 on the RTX PRO 6000 and 1–16 on the B200 (Figure 9(b)).

AuthorsOpenReview ↗

Response (2/2): Novelty, sampling variation, concurrency, and the global suffix tree

W3. Novelty relative to existing hybrids

We agree that SpecSplice builds on established drafting components. Our contribution is to choose where to splice by comparing candidates' estimated contributions to accepted length, using evidence from both sources.

The hybrids we compare against or discuss compose the two sources through source-level or predefined rules, rather than by scoring a retrieved continuation jointly with the parametric prefix before it (Section 2.3). SuffixDecoding's hybrid (our Selection baseline) chooses between whole drafts using the retrieval score. SAM-Decoding uses match length to select a source. Graft sets the retrieval budget from draft-model confidence and fills it with a prescribed template. These rules also fix where retrieval can attach, such as the root or a predefined boundary in Graft and RASD, or the draft tail in Ouroboros. They can therefore leave additional combinations unexplored: a parametric prefix can expose a useful retrieval match elsewhere within the same draft.

Choosing among these additional splice points requires considering both sources together. A retrieved continuation contributes only if its parametric prefix is accepted; otherwise, its nodes occupy budget without adding accepted tokens. Prefix confidence alone does not show whether the continuation retrieved after it is worth that budget. The value of a retrieved candidate therefore depends jointly on prefix survival and acceptance along the retrieved continuation (Section 3.1, Eq. (2)).

SpecSplice estimates this joint contribution from parametric probabilities and retrieval frequencies on a common scale (Eq. (4)). Candidates retrieved at the root and at every eligible parametric node then compete directly with parametric candidates under one budget. This expands the composition space, and splice placement and budget allocation follow the same contribution criterion (Sections 3.1–3.2).

Figure 5 isolates the two effects at a common B = 64. With root-only retrieval, unified scoring improves MAT by 17.9% over Selection and by 26% over Graft. Allowing all eligible splice points, with scoring unchanged, adds a further 10.9% (Section 4.3, Figure 5(a–b)).

Appendix J shows how this differs from source-level selection (Figure 10(b)); colors identify the drafting source, and vertical bars mark verification steps. Selection drafts the opening words with DFlash in one step and switches to retrieval in the next. SpecSplice combines both sources within a step, so the identical output takes 9 verification steps instead of Selection's 12.

Q2. Trajectory differences under sampling

The gain is much larger than the run-to-run variation, and it extends across most of the task distribution. We should have explained how trajectory variation is accounted for.

The main results use identical greedy trajectories, which directly controls output length and repetitiveness (Section 4.1; Table 1). Under sampling, individual trajectories differ across methods and runs. However, lossless verification means that every method samples from the target model's distribution, as autoregressive decoding does, so output length and repetitiveness follow the same distribution, and their differences reflect sampling randomness.

To measure the remaining variation, we ran SpecSplice and Selection five times each on Qwen3.5-9B/Gaia2 at T = 0.6 (Appendix D.1, Table 5). The geometric-mean gain over Selection is 39.4% in speedup (95% CI 35.5–43.3%) and 37.9% in MAT (95% CI 34.0–41.9%). The speedup coefficient of variation is 1.9% for both methods, and every SpecSplice run is faster than every Selection run.

A stable aggregate gain could still be dominated by a few highly repetitive tasks. We therefore examined per-task CDFs that weight every task equally (Figure 8(a–b)). SpecSplice's MAT and speedup curves lie to the right across most of the distribution, while the five curves of each method closely overlap.

Q3. Concurrent requests

Yes, the advantage persists under batching. On Qwen3.5-9B/Spider2-DBT, SpecSplice is 15–31% faster than Selection at batch sizes 1–4 on the RTX PRO 6000 and 1–16 on the B200. Each method uses its own throughput-optimal budget at each batch size, so the advantage holds as batching changes the preferred budget. We should have made these results more visible. Section 4.2 now summarizes them, and Appendix F (Figure 9(b)) gives the setup and full results.

Q4. Global suffix tree population

The main evaluation uses a cold-start global index that grows online. The global suffix tree starts empty for every model–workload–method run, and each request's output is inserted into it as it is generated (Appendix C, "Retrieval index"). A separate per-request tree indexes the current prompt and output (Appendix B.2). The prewarmed index appears only in Appendix G, where it emulates a shared serving index to study how construction cost scales (Table 7).

Dear Reviewer 9soY,

We sincerely thank you again for your thoughtful review. As the discussion period ends soon, we would like to kindly ask whether our response addresses your concerns.

In our response, we added a transfer study of the source weight λ (Section 4.5; Appendix I, Table 10), a runtime-cost analysis (Section 4.4; Appendix G, Table 7), and five sampling runs each for SpecSplice and Selection (Appendix D.1, Table 5). We also clarified the novelty relative to existing hybrids, the results under concurrent requests, and how the global suffix tree is populated.

Please let us know if we can provide any additional clarification. Thank you very much for your time and effort.

Comments on the paper

1 comment

Dear AC,

We would like to thank you for handling our submission and for your time and effort during the discussion. As the discussion period ends soon, we would like to briefly recap our contribution and the main discussion points across the reviews.

Novelty. The core novelty of SpecSplice is node-level composition of the two sources. Each draft-tree node (token), whether parametric or retrieved, competes for the shared verification budget, so where retrieval is spliced and how much budget it receives follow from this node-level selection, rather than from choosing one source per round or attaching retrieval at prescribed locations.

In agentic workloads, a few parametric tokens often expose a retrieval match that the committed context alone cannot reach (draft-exposed matches; Figure 1, Section 2.3). An agent's trajectory interleaves newly generated content with spans that recur from earlier turns, as the agent repeats tool-call formats, revises earlier outputs, and quotes tool results. Tool calls are a typical case: recurring call syntax and argument names are interleaved with task-specific values that change between invocations, so a parametric prefix that generates such a value can expose the recurring continuation. Accordingly, SpecSplice's largest gains arise in tool-call segments, where it raises MAT by 31–46% over the strongest baseline on Qwen3.5-9B (Section 4.2, Figure 4).

Exploiting such matches requires composing both sources within the same verification round. The hybrids we compare against or discuss compose the two sources through source-level or location-level rules: Selection and SAM-Decoding choose one source per round, and Graft, RASD, and Ouroboros attach retrieval at prescribed locations. These rules can leave such matches unreached or not weighed against parametric candidates for the shared budget.

SpecSplice introduces three components for this setting:

  1. Joint candidate space. We formulate hybrid draft construction as a joint optimization over individual parametric and retrieved nodes under a shared verification budget, with retrieval allowed at the root and at every eligible parametric node (Section 2.3).
  2. Unified contribution score. A retrieved node contributes only if its parametric prefix is accepted. The score therefore combines prefix survival with retrieval support on a common scale (Section 3.1, Eq. (4)) and is nonincreasing along every draft path.
  3. Interleaved construction. Using this monotonicity and the resulting query bounds, best-first construction maximizes the accepted-length surrogate over the full candidate space while skipping queries that cannot affect the selection (Section 3.2, Propositions 1–2). In the runtime analysis at B = 64, SpecSplice issues 9.4 queries per step instead of the 64 issued by exhaustive querying, and both return the same tree (Section 4.4, Table 2).

The ablations isolate the first two components. At a common B = 64, unified scoring at the root raises MAT by 17.9% over Selection and 26% over Graft, and extending retrieval to all eligible parametric nodes adds a further 10.9% (Section 4.3, Figure 5(a–b)). Overall, SpecSplice achieves 13–37% higher throughput than the strongest baseline on every agentic model–workload pair (Table 1).

Source weight selection and transfer (Reviewers 9soY and Rjiz). We reused λ = 0.1, selected on Spider2-DBT and BFCL-v4 with Qwen3.5-9B, unchanged on the ten remaining agentic pairs. Compared with per-pair tuning, speedup decreases by at most 2.0%, and SpecSplice remains 12–37% faster than the strongest baseline on all 12 agentic pairs (Section 4.5; Appendix I, Table 10).

Runtime overhead (Reviewers Fn75 and 9soY). A SpecSplice step is 0.4 ms (1.5%) longer than a Selection step, and all reported throughput includes this cost (Section 4.4, Table 2). With 32K requests prewarmed, construction takes 5.3% of the step for SpecSplice, compared with 13.2% for exhaustive querying (Appendix G, Table 7).

Run-to-run variation under sampling (Reviewers 9soY and Rjiz). Over five sampling runs each for SpecSplice and Selection, the geometric-mean gain in speedup is 39.4% (95% CI 35.5–43.3%) (Appendix D.1, Table 5).

Evaluation details and code (Reviewers Rjiz and Fn75). Section 4.1 and Appendix C now state the timing boundaries, workload sizes, selected budgets, and retrieval-index initialization. We also shared the anonymized code in a comment visible to the reviewers and the AC.

We believe that these revisions address the main concerns raised by the reviewers. Please let us know if we can provide any additional clarification. Thank you very much for your time and effort.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.