acceptodds
Under review as a conference paper at ICLR 2027

Sequential Anytime Valid Tests for Efficient Inference-Time Alignment

13 views

Abstract

Inference-time alignment methods, such as Best-of-N (BoN), improve language model outputs by generating many candidates per prompt and selecting among them using a reward model. The cost is largely wasted, since a fixed budget is applied uniformly across prompts, even though the computation required by different prompts varies widely. Adaptive methods halt generation once the sampled answers appear settled, but their thresholds are tuned heuristics with no control over the probability of stopping at the wrong answer. Instead, we ask for a rule that takes that probability – a permitted error level – as input, and returns the number of tokens spent to meet it. To build such a rule, we aggregate candidates by their final answers using a reward-weighted self-consistency (RWSC) score and treat the decision of when to stop as a sequential test over the resulting clusters. Wald's sequential probability ratio test (SPRT) is the natural tool for such a test, and when applied to the RWSC clusters, it yields SPRT-RWSC, which performs well empirically. Yet its guarantee does not hold here, since the leading answer shifts as sampling proceeds, and the run is continuously monitored. Our main contribution replaces Wald's test with an anytime-valid construction that resolves this issue: it maintains a betting supermartingale (an e-process) for every answer the policy produces, bounding the probability of ever certifying a non-dominant answer by , uniformly over time and across all answers, however many the policy produces. We call the resulting rule e-RWSC: the RWSC scores drive the bets, which affect only how quickly a dominant answer is certified, not the validity of the guarantee; and certification takes samples. Experiments across several language models, datasets, and reward models confirm the predicted scaling of per-response cost, and show that accuracy improves over baselines at substantially lower token cost.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Test-Time Alignment, Best-of-N Selection, e-processes, Wald's SPRT

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_ZwjkOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the computational inefficiency of fixed sampling budgets in inference-time alignment methods such as Best-of-N. It proposes RWSC, which combines reward model scores with answer-level self-consistency, and develops two sequential stopping methods: SPRT-RWSC and e-RWSC. In particular, e-RWSC uses per-answer e-processes and reward-informed Kelly betting to provide anytime-valid statistical guarantees while improving sampling efficiency. Experiments across multiple language models, reward models, and reasoning benchmarks demonstrate promising accuracy-efficiency trade-offs.

Critical strengths

1. A principled adaptive stopping mechanism with statistical guarantees.

The main strength is the application of anytime-valid e-processes to a dynamically expanding set of candidate answers. The proposed approach maintains error control even when the leading answer changes during sampling. Compared with heuristic stopping rules, this provides a stronger theoretical justification and is a key reason I view the paper positively.

2. A thoughtful integration of reward models with practical efficiency gains.

The reward-informed betting mechanism uses reward model predictions to accelerate evidence accumulation without directly affecting the validity of certification. This is an appealing design choice. The empirical results are also encouraging. For example, on MATH500 under a large sampling budget, e-RWSC improves accuracy by 8.2 percentage points while reducing average candidate generation by approximately 61.7% compared with Vanilla BoN.

Critical weaknesses

1. Proposition 1 appears to omit important conditions for its sampling complexity guarantee.

Proposition 1 assumes λ⋆<1/p†\lambda^\star < 1/p^\dagger, whereas the actual algorithm restricts bets to λn≤c/p†\lambda_n \leq c/p^\dagger, with c<1c<1. This weaker assumption does not ensure that the implemented strategy achieves the unconstrained Kelly-optimal growth rate. For example, when p†=0.5p^\dagger=0.5, c=0.8c=0.8, and θ=0.95\theta=0.95, the optimal bet is 1.8, exceeding the implemented cap of 1.6.

Moreover, the claimed asymptotic sampling complexity is difficult to reconcile with a fixed finite Nmax⁡N_{\max}, since N/log⁡(1/α)N/\log(1/\alpha) must approach zero as α→0\alpha\to0. These issues suggest that Proposition 1 needs stronger assumptions and a clearer distinction between unconstrained certification time and budget-limited stopping.

2. The reliability evaluation and empirical superiority claims need further qualification.

Theorem 1 controls the probability of incorrectly certifying a non-dominant answer, rather than the correctness of the final response. Furthermore, when the sampling budget is exhausted, the algorithm returns an uncertified RWSC answer. Comparing early stopping decisions against a 512-sample consensus does not directly validate the theoretical certification event.

The empirical advantages are also not universal. For example, on MATH500 with Qwen2.5-3B and ArmoRM at Nmax⁡=24N_{\max}=24, ConSol achieves 75.4% accuracy with 8.02 average candidates, compared with 74.4% and 15.59 for e-RWSC. While e-RWSC offers stronger statistical guarantees, its empirical accuracy-efficiency advantage should be stated more carefully.

Questions for authors
  1. Could the authors revise Proposition 1 to account for the actual betting cap c/p†c/p^\dagger and distinguish unconstrained certification time from budget-limited stopping? Should the sampling complexity instead depend on the truncated betting strategy's achievable growth rate?

  2. Could the authors report certification rates, budget exhaustion rates, and the accuracy of certified versus uncertified outputs across different α\alpha and Nmax⁡N_{\max}? Could independent sampling estimates of answer probabilities also be used to assess the event controlled by Theorem 1?

  3. How do the authors characterize e-RWSC's advantages in settings where ConSol or SPRT-RWSC achieves both higher accuracy and lower sampling costs? Confidence intervals, generated token counts, and end-to-end inference latency would help clarify the practical cost of obtaining statistical guarantees.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Zwjk

Thank you for your feedback. We address the concerns below.

Proposition 1 statement: Thanks for catching this discrepancy in the Proposition 1 statement. We will revise the statement as: "λ⋆(θa⋆)≤c/p†\lambda^\star(\theta_{a^\star}) \le c/p^\dagger, where c∈[0,1)c \in [0,1)..." Note that Step (i) of the proof already uses this condition, and the result will remain unchanged. Accordingly, in Lemma 1, we will state Lemma 1 for λ∈[0,c/p†]\lambda \in [0, c/p^\dagger] and modify the admissibility condition to λ⋆(θ)≤c/p†\lambda^\star(\theta) \le c/p^\dagger. The proof remains unchanged.

On Nmax⁡N_{\max}, the reviewer is right that the α→0\alpha\to 0 limit concerns the unbudgeted certification time. We will state Proposition 1 for Nmax⁡=∞N_{\max}=\infty, and state that the deployed rule stops at min⁡{N,Nmax⁡}\min\lbrace N,N_{\max}\rbrace. Accordingly, we will state Theorem 1 for any budget Nmax⁡∈N∪{∞}N_{\max} \in \mathbb{N} \cup \lbrace\infty\rbrace, so truncation affects only whether certification occurs, never its validity. The proofs remain unchanged.

E-RWSC certification rates, budget exhaustion rates, and the accuracy of certified versus uncertified outputs: We report the results on MATH500 (Qwen2.5-3B base model + Armo reward model).

Scaling Across Budget NN (α=0.10\alpha = 0.10)

Metric / Budget (NN)243264128256512
Certified Accuracy (%)99.2%98.9%97.6%96.1%95.3%94.5%
Exhausted / Fallback Acc (%)46.6%48.2%47.1%48.2%47.3%47.1%
Certification Rate (%)52.8%54.8%59.2%61.8%63.6%65.2%
Budget Exhaustion Rate (%)47.2%45.2%40.8%38.2%36.4%34.8%

Sensitivity Across Error Boundary α\alpha (Nmax⁡=512N_{\max} = 512)

Metric / Nominal α\alpha0.010.050.100.200.300.400.50
Certified Accuracy (%)94.7%94.4%94.5%94.2%94.2%93.9%93.7%
Exhausted / Fallback Acc (%)48.1%47.8%47.1%47.4%47.7%48.0%47.9%
Certification Rate (%)63.8%64.4%65.2%65.4%65.6%65.8%66.2%
Budget Exhaustion Rate (%)36.2%35.6%34.8%34.6%34.4%34.2%33.8%

Empirical Assessment of Theorem 1: As per your suggestion, we evaluated Theorem 1 using independent empirical estimates θ^a⋆=1N∑i=1N1(yi∈a⋆)\hat{\theta}_ {a^\star} = \frac{1}{N}\sum_ {i=1}^N \mathbf{1}(y_ i \in a^\star) computed over the full N=512N=512 independent Monte Carlo sample pool. Across all α∈[0.01,0.40]\alpha \in [0.01, 0.40], we observe that the empirical false certification rate (θ^a⋆≤p†\hat{\theta}_{a^\star} \le p^\dagger) is zero, whereas for α=0.45,0.50\alpha=0.45, 0.50, it is 0.4%0.4\%. This verifies finite-sample error control under Theorem 1. We will replace the Figure in the main text with this new result.

E-RWSC Advantages vs. ConSol & SPRT-RWSC: ConSol performed better than our method on only one dataset (MATH500) for the Armo reward model, but E-RWSC maintains a consistently higher accuracy across datasets and models. Even on this benchmark, switching to a better reward model like Skywork-V2, E-RWSC achieves 78.6% accuracy compared to ConSol's 75.4%.

As you suggested, to further characterize e-RWSC's advantages, we compare Certified Accuracy, certification rate, and generated token counts on the MATH500 dataset (Qwen2.5-3B base model + Armo reward model).

MethodOverall Acc (%)Certification Rate (%)Certified Acc (%)Certified StepsCertified Tokens / Query
ConSol74.8%100.0%74.8%12.39.1k
SPRT-RWSC76.8%100.0%76.8%21.813.6k
CITE77.4%85.8%83.7%60.839.1k
E-RWSC78.0%65.2%94.5%27.513.1k

We observe that when E-RWSC certifies, its accuracy is 94.5% spending only 13.1k13.1\text{k} tokens per query, compared to ConSol's 74.8%74.8\% (9.1k9.1\text{k} tokens) and SPRT-RWSC's 76.8%76.8\% (13.6k13.6\text{k} tokens). This shows that E-RWSC gives more reliable certified decisions.

While ConSol and SPRT-RWSC stop aggressively on small budgets, they suffer from premature convergence under compute scaling: ConSol halts at step ∼12\sim 12 even with Nmax⁡=512N_{\max}=512, plateauing at 74.8%74.8\% accuracy. In contrast, E-RWSC adaptively allocates compute to difficult queries, thereby scaling overall accuracy to 78.0%78.0\% (+3.2%+3.2\%).

Moreover, the CPU overhead of E-RWSC's e-process update is merely 0.014 ms0.014\text{ ms} per step, which is <0.0002%<0.0002\% of a new candidate generation time on the GPU (7,144.3 ms7,144.3\text{ ms}), and thus negligible.

We thank the reviewer for the constructive feedback. We have revised the manuscript with new results and clarifications, and we hope these, together with our responses, address your concerns. If any point remains unclear, please let us know; we would be glad to clarify further during the discussion period. If you feel your concerns have been resolved, we would be grateful if you would consider updating your assessment.

Reviewer_ZwjkOpenReview ↗

Comments on Rebuttal

Thank you for the detailed rebuttal and the substantial revisions. The authors have addressed most of my major concerns.

In particular, I appreciate the correction to Proposition 1, which now properly accounts for the constrained Kelly betting strategy and distinguishes the unbounded certification-time result from budget-limited stopping.

The new empirical evaluation is also a significant improvement. The authors now distinguish certified from uncertified outputs, report certification and budget exhaustion rates, and provide a more direct empirical assessment of the false-certification event controlled by Theorem 1. The comparison with MMC, together with the additional 24B model evaluation and computational overhead analysis, further strengthens the empirical study.

I particularly appreciate the analysis of certified accuracy, which better illustrates the reliability-efficiency trade-off offered by E-RWSC.

Some limitations remain. In particular, certification guarantees dominance under the generation distribution rather than factual correctness, and the empirical advantages over existing adaptive methods are not uniform across all benchmarks and budget settings. The applicability to genuinely open-ended tasks also remains to be empirically validated.

Nevertheless, I believe the authors have adequately addressed the main technical and experimental concerns raised in my review. The revised paper is substantially stronger, and I am updating my recommendation to Weak Accept (3).

Reviewer_W37KOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies adaptive stopping for inference-time sampling. It first introduces reward-weighted self-consistency (RWSC), which aggregates generations by their final answers, and then formulates stopping as a sequential testing problem. The main method, E-RWSC, maintains an e-process for each observed answer and controls the probability of certifying an answer whose generation probability is below a chosen threshold. Experiments compare the method with fixed-budget and early-stopping baselines across several models, reward models, and reasoning benchmarks.

Critical strengths

The motivation makes sense because different prompts may require different numbers of samples, and using the same budget for all of them can therefore be wasteful. The empirical study also covers several models, reward models, datasets, and reasonably relevant baselines.

Critical weaknesses
  • The central objective is not presented consistently. The abstract and introduction motivate the method as avoiding a "wrong answer," while the theorem only certifies that an answer has generation probability above a threshold, explicitly not that it is correct. The paper also moves between correctness, dominance, consensus, and alignment without clearly separating them. This makes the scope of the guarantee difficult to understand and, at times, misleading.
  • The empirical validation does not measure the error event controlled by Theorem 1. Figure 3 defines an error as disagreement with a finite-budget consensus at N=512N=512, whereas the theorem concerns certification of an answer whose true generation probability is at most p†p^\dagger. These events are not equivalent. Moreover, 174 of 500 MATH500 problems exhaust the budget and receive an uncertified fallback answer, but certified and fallback cases are mixed in the reported performance. The paper also repeatedly claims token savings while reporting only the number of sampled responses.
  • The novelty relative to the closest anytime-valid self-consistency methods is not fully convincing. In particular, MMC uses e-processes to handle a changing empirical leader and an unknown answer set, yet it is not included in the empirical comparison. The remaining distinction appears to be mainly reward-informed betting and certification of an absolute majority, which is a more restrictive target than mode certification.
Questions for authors

See Weakness

3 comments
AuthorsOpenReview ↗

Response to Reviewer W37K

Thank you for your feedback. We address the concerns below.

Wording. We agree. Theorem 1 certifies dominance under the policy, θa>p†\theta_a>p^\dagger, not correctness. Phrases such as "stopping at the wrong answer" in the abstract and introduction blur this distinction. Remark 1 states it. In the revision, we will define the controlled error once in the introduction as certifying an answer that the policy does not generate with probability above p†p^\dagger, and use that wording throughout. We will centralize our theme on strict dominance; phrases like "stopping at wrong answer" will be changed to "non-dominant answer" to ensure correctness and clarity.

Empirical Assessment of Theorem 1: We re-evaluated Theorem 1 using independent empirical estimates θ^a⋆=1N∑i=1N1(yi∈a⋆)\hat{\theta} _ {a^\star} = \frac{1}{N}\sum _ {i=1}^N \mathbf{1}(y _ i \in a^\star) computed over the full N=512N=512 independent Monte Carlo sample pool. Across all α∈[0.01,0.40]\alpha \in [0.01, 0.40], we observe that the empirical false certification rate (θ^a⋆≤p†\hat{\theta} _ {a^\star} \le p^\dagger) is zero, whereas for α=0.45,0.50\alpha=0.45, 0.50, it is 0.4%0.4\%. This verifies finite-sample error control under Theorem 1. We will replace the Figure in the main text with this new result.

Separately reporting certified accuracy on MATH500: As per your suggestion, we report certification Rate (the percentage of prompts certified early), Certified Accuracy (accuracy on this certified subset) and Certified Tokens per Query (average token cost per certified prompt).

On confident prompts where early stopping triggers (65.2% of the benchmark), E-RWSC achieves peak certified accuracy of 94.5% while spending only 27.5 steps (13.1k tokens) on average, substantially higher reliability than MMC (88.9%, 17.8k tokens) and CITE (83.7%, 39.1k tokens), as well as other baselines:

MethodOverall Acc (%)certification Rate (%)Certified Acc (%)Certified Steps (τˉ\bar{\tau})Certified Tokens per Query
AdaBoN69.4%63.6%76.4%305.2144.8k
BEACON70.0%20.6%71.8%48.421.4k
Pandora’s Box68.6%100.0%68.6%4.02.0k
ConSol74.8%100.0%74.8%12.39.1k
CITE77.4%85.8%83.7%60.839.1k
MMC77.4%61.4%88.9%32.917.8k
SPRT-RWSC (Ours)76.8%100.0%76.8%21.813.6k
E-RWSC (Ours)78.0%65.2%94.5%27.513.1k

We will put this result as a Figure in Section 4.

Novelty relative to MMC. Conceptually, MMC certifies the mode of plain majority voting and has no channel for an external reward signal. Our contribution is a construction in which the reward model drives the bets. Because Theorem 1 holds for any predictable bet, a miscalibrated or hacked reward model can slow certification, but it cannot change what is certified or the error level. RWSC also lets the reward determine which response is returned, which MMC does not address.

We agree that MMC should serve as an empirical baseline, and we compared our methods against it. E-RWSC outperforms MMC in both token usage and accuracy across all benchmarks. In MATH500, E-RWSC delivers 94.5% certified accuracy at scale (N=512N=512) vs. 88.9% for MMC, while stopping in fewer steps (13.1K vs. 17.8K for MMC). See the above table. In the revised draft, we will update all empirical comparisons with MMC performance.

We thank the reviewer for the constructive feedback. We have revised the manuscript with new results and clarifications, and we hope these, together with our responses, address your concerns. If any point remains unclear, please let us know; we would be glad to clarify further during the discussion period. If you feel your concerns have been resolved, we would be grateful if you would consider updating your assessment.

AuthorsOpenReview ↗

Request to review the rebuttal

Dear Reviewer,

Thank you again for your reviews. As the discussion period is nearing its end, we wanted to briefly check in. We have addressed each of your concerns and uploaded a revised draft incorporating these changes. We would be grateful to know whether our responses and the revision address your concerns, and whether they change your assessment of the paper. We are happy to clarify any remaining points.

The authors

Reviewer_ETJzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper develops a new approach to dynamically decide the number of samples in inference-time alignment methods, such as Best-of-N and self-consistency.

Critical strengths
  • The proposed method is theoretically grounded and has clear error control.

  • It is robust to reward hacking that appears in Best-of-N

Critical weaknesses
  • The empirical improvement in terms of budget does not seem significant compared to existing adaptive methods. It will be helpful to clarify the main advantages of the proposed method compared to existing baselines.

  • The cost of the proposed method is not discussed. It is unclear whether the sequential hypothesis testing and clustering will add additional overhead.

  • The tested models are relatively small (e.g., 3B-8B). It will be better to show comparison on larger and newer models.

  • It is unclear whether the proposed method can be applied to open-ended tasks, as it performs clustering of responses based on exact final answer match.

  • The paper is difficult to follow in general. Many terminologies and design choices are not well explained.

Questions for authors

See weaknesses

3 comments
AuthorsOpenReview ↗

Response to Reviewer ETJz

Thank you for your feedback. We address the concerns below.

Improvement over adaptive methods. Most existing adaptive rules use thresholds tuned per policy, reward model, and benchmark, with no control over the probability of stopping on a non-dominant answer. E-RWSC takes an error level α\alpha as input and guarantees that the probability of ever certifying such an answer is at most α\alpha for any reward model (Theorem 1), while compute scales with prompt difficulty (Proposition 1). We will state this up front. Empirical improvements are summarized below:

  • Against the reward-based adaptive baselines, the savings are substantial: At Nmax⁡=24N_{\max}=24 on MATH500, AdaBoN uses 18.8 samples and BEACON 22.6, both at ≤71.4% accuracy. E-RWSC uses 15.6 samples at 74.4%. The gap widens with the budget. At Nmax⁡=512N_{\max}=512, AdaBoN uses 380 samples (69.4%) and BEACON 417 (70.0%), while E-RWSC uses 196 (78.0%).

  • Against reward-agnostic baselines, ConSol is cheaper than E-RWSC at small budgets but plateaus at large ones (74.8% at Nmax⁡=512N_{\max}=512), whereas E-RWSC keeps spending compute on hard prompts and reaches 78.0%. Compared with CITE, the closest anytime-valid baseline, E-RWSC uses fewer samples across all four benchmarks at Nmax⁡=24N_{\max}=24 and is more accurate on three (e.g., MMLU: 76.7% with 14.2 samples vs. 72.2% with 21.6 samples). On MATH500, CITE is 1.0 point more accurate but uses 39% more samples.

  • To further characterize e-RWSC's advantages, we perform Certified Accuracy, certification rate, and generated token count comparisons on the MATH500 dataset (Qwen2.5-3B base model + Armo reward model). When E-RWSC stops early (65.2% of queries), its certified answers are correct 94.5% of the time and, on average, use 13.1k tokens. This compares with 74.8% for ConSol (9.1k tokens) and 83.7% for CITE (39.1k tokens). Please see the table below. We will add this comparison in the revised draft.

    MethodOverall Acc (%)Certification Rate (%)Certified Acc (%)Certified tokens / Query
    ConSol74.8%100.0%74.8%9.1k
    CITE77.4%85.8%83.7%39.1k
    E-RWSC78.0%65.2%94.5%13.1k

Compute overhead. Clustering costs one answer extraction and one string comparison per response. The test keeps a few scalars per observed answer and updates them in O(∣An∣)O(|\mathcal A_n|) operations per response. These costs are negligible compared with the cost of generating a response. On an RTX 6000 Ada GPU, generating and scoring each candidate takes 7,144.3 ms, whereas CPU overhead per step is merely 0.20 ms for answer extraction and 0.014 ms for updating the e-process. At an algorithmic overhead of under 0.003% of generation latency, this is negligible. We will add a complete empirical latency breakdown study in Section 4 and Appendix D.

Larger model result. We expanded our comparison to include the larger 24B-parameter model Mistral-Small-24B-Instruct-2501 on MATH500 with ArmoRM (Nmax⁡=24,α=0.10N_{\max}=24, \alpha=0.10). As shown in the table below, the empirical behavior aligns with our findings on 3B–8B models. Here, Certified Accuracy denotes accuracy on the certified subset, and Average Budget is candidate usage across all queries (including budget exhaustion).

MetricPandora's BoxAdaBoNBEACONConSolCITEMMCSPRT-RWSC (Ours)E-RWSC (Ours)
Overall Accuracy (%)70.8%70.8%72.2%76.8%77.4%77.4%76.8%76.8%
Certified Accuracy (%)70.8%78.8%74.7%84.6%98.7%90.1%90.2%98.3%
Average Budget4.018.722.17.621.519.212.915.0

A clear improvement is evident in the certified accuracy. CITE also achieves a good certified accuracy, partly because the model is larger and stronger, so self-consistency toward the right answer is higher, but CITE considers many more candidates than E-RWSC to achieve similar accuracy.

Applicability to Open-ended tasks. The method needs only a fixed rule that maps each response to an equivalence class; exact string match is one such rule. Our benchmarks already go beyond numeric answers: MMLU is multiple choice, and CruxEval uses code outputs. For open-ended generation, any fixed canonicalization or semantic-equivalence function can replace exact match. Theorem 1 will remain unchanged as long as the map is applied to each response independently. Tasks without a meaningful notion of equivalent answers are outside our scope. Note that the adaptive baselines (eg., AdaBoN, CITE, MMC) also evaluate only on verifiable tasks.

Presentation. We will define all terms and design choices before use, if necessary.

We thank the reviewer for the constructive feedback. We have revised the manuscript with new results and clarifications, and we hope these, together with our responses, address your concerns. If any point remains unclear, please let us know; we would be glad to clarify further during the discussion period. If you feel your concerns have been resolved, we would be grateful if you would consider updating your assessment.

AuthorsOpenReview ↗

Request to review the rebuttal

Dear Reviewer,

Thank you again for your reviews. As the discussion period is nearing its end, we wanted to briefly check in. We have addressed each of your concerns and uploaded a revised draft incorporating these changes. We would be grateful to know whether our responses and the revision address your concerns, and whether they change your assessment of the paper. We are happy to clarify any remaining points.

The authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Revision

We thank all reviewers for their constructive feedback, and have revised the paper accordingly, using the additional page allowed for the rebuttal. The main changes are:

  • Replaced the figure 3 (now figure 4) in the main text with a new empirical assessment verifying the finite-sample error control of Theorem 1.
  • Added a complete empirical latency and compute overhead breakdown study in Section 4 and Appendix D.
  • Added new evaluation metrics reporting Certification Rate, Certified Accuracy, and Certified Tokens per Query (summarized as a new figure 2 in Section 4).
  • Expanded empirical evaluations in Table 1 to include a larger 24B-parameter model (Mistral-Small-24B-Instruct-2501).
  • Updated all empirical comparisons across benchmarks to include the MMC (Majority Marginal Certification) baseline.
  • Revised statements of Lemma 1, Proposition 1, and Theorem 1 to update the admissibility condition of Kelly bet and budget limit.
  • Clarified wording around Theorem 1 in the introduction and throughout the text, replacing phrases like "stopping at the wrong answer" with "non-dominant answer" to accurately reflect strict dominance under the policy.

We welcome any further questions during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.