acceptodds
Under review as a conference paper at ICLR 2027

RUBATO: Decoupled Clocks Allow Full-Duplex Models to Think as Fast as They Can

Abstract

Recent Speech Full-Duplex (FDX) models are capable of engaging in naturalistic conversation by listening and speaking at the same time. Building on this foundation, recent architectures introduce internal thinking streams to enhance their intelligence. To guarantee these streams process successfully in real time, existing methods enforce a fixed decoding rate, which assumes a hardware profile capable of a specific latency. We argue this static approach is severely suboptimal. Because internal thinking is intrinsically asynchronous, an ideal FDX model should sustain real-time conversational flow while allowing each stream to decode at its own hardware-adaptive rate. Thinking streams must process as fast as the hardware allows to deliver accurate answers faster. To solve this, we propose RUBATO, a novel multistream positional encoding that assigns independent clocks to entirely decouple the streams. This structural isolation empowers the model to zero-shot interpolate across any inference decoding speed. It drastically simplifies dataset construction by requiring training on only two extreme decoding speeds while smoothly adapting to any intermediate rate during deployment. RUBATO demonstrates consistent robustness across five spoken reasoning tasks. On Spoken GSM8K, a baseline model collapses to 1.7% accuracy at interpolated speeds. Conversely, RUBATO-Full sustains 48.7% to 51.1% accuracy across all rates, leveraging faster hardware to achieve a 6.5× shorter time-to-answer with a constant time-to-first-response. This advantage holds on BIG-Bench Audio and amplifies significantly during complex full-duplex interruptions. Specifically, when a user interjects while the model is actively speaking and reasoning, RUBATO-Full achieves 40.0% to 45.1% accuracy, whereas the baseline fails completely.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Speech, Reasoning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_piFuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

RUBATO addresses positional inconsistencies that arise when variable numbers of reasoning tokens are interleaved with real-time audio and text streams. It assigns independent positional clocks to the real-time and thinking streams and compares Split, Gated, and Full implementations. RUBATO-Full computes each query–key interaction using relative positions on the key token’s clock across the full rotary dimension of each attention head. The main demonstrated result is robustness to unseen fixed and randomly varying thinking budgets, measured in reasoning tokens per audio frame, after training on two endpoint budgets.

Critical strengths
  1. The paper clearly explains how a single positional clock introduces changes in positional distances when the thinking rate varies. The independent-clock construction directly addresses this issue, and the diagrams help explain the distinction.
  2. The experiments cover fixed and randomly varying budgets, including single-turn reasoning and synthetic overlapping conversations. RUBATO-Full largely preserves accuracy at unseen rates while reducing time-to-answer as the per-frame thinking budget increases.
  3. The Split, Gated, and Full variants make the architectural choices easier to assess. The spoken-filler quality evaluation also adds evidence beyond reasoning accuracy and latency.
  4. The method can use existing fused attention kernels, and the optimized multi-KV implementation shows small decoding overhead in the reported batch-size-one H200 benchmark (Appendix C).
Critical weaknesses
  1. End-to-end evidence for hardware-adaptive operation is limited. The experiments establish robustness to prescribed thinking budgets. Appendix C provides a decode microbenchmark, and Appendix E estimates feasible budgets from hardware characteristics. An end-to-end streaming measurement under changing compute availability would provide stronger support for the claim that the system adapts its thinking rate while meeting audio deadlines.
  2. The practical alternative of intermediate-budget training is not evaluated. The endpoint-trained Single-Clock baseline is appropriate for testing interpolation under the same training layouts. However, a comparison with Single-Clock trained on a small set of intermediate budgets would clarify the practical advantage of the proposed training recipe. Section 4.6 varies the proportions of the two endpoint layouts and does not answer this question.
  3. Training variability is not reported. Figure 3 includes Wilson confidence intervals, but these do not capture variation across independently trained models. Repeated runs would strengthen the comparisons among RUBATO variants and the conclusions from the training-layout mixture ablation. In addition, RUBATO-Full already has a substantial advantage at the trained budgets on BIG-Bench Audio, so its absolute accuracy gains should be distinguished from its resistance to degradation at unseen budgets.
Questions for authors
  1. Building on the measurements in Appendix C and the estimates in Appendix E, could you provide an end-to-end streaming experiment under changing compute availability, reporting achieved thinking throughput and the fraction of missed 80 ms audio deadlines?
  2. How does endpoint-trained RUBATO-Full compare with Single-Clock trained on the endpoints plus a small set of intermediate budgets under a matched training budget? Reporting the additional data-construction effort would help assess the practical benefit of the two-endpoint recipe.
  3. Could you report variability across independent training runs for the main comparisons and representative training-layout mixtures, including whether the relative performance of Full and Gated remains consistent?
4 comments

Thank you for highlighting our clear exposition of the single-clock problem, the robustness demonstrated in our fixed and randomly varying budget evaluations, the comparative insights from the Split/Gated/Full ablations, and the low decoding overhead achieved on existing kernels. We address your aligned weaknesses and questions below.

W1/Q1. End-to-end streaming under changing compute

To clarify the deployment paradigm. In a live setting, the per-frame thinking budget BfB_f is not determined in advance. Instead, a scheduler manages decoding directly. It allows the model to process thinking tokens until the next 80 ms audio deadline approaches; it then pauses the thinking stream, ingests the incoming user audio, decodes the required real-time tokens for that frame, and resumes thinking in the subsequent frame. Because of this, internal reasoning never delays the real-time audio streams. The achieved thinking throughput simply emerges as the resulting sequence of BfB_f. RUBATO's decoupled clocks allow the model to seamlessly handle arbitrary, fluctuating per-frame budgets because the real-time streams maintain their own independent positional clock.

From the model's perspective, deployment is mathematically equivalent to processing a randomized sequence of per-frame budgets BfB_f. Therefore, our frame-varying evaluations are a direct test of the changing compute scenario. When BfB_f is drawn uniformly from [1,16][1,16] at every frame, RUBATO-Full achieves 52.7% on Spoken GSM8K and 88.9% on BIG-Bench Audio, compared to a near-collapse of 15.5% and 47.0% for Single-Clock. We agree that building a live, end-to-end systems benchmark (handling streaming audio I/O and measuring actual performances across various GPU architectures) would provide excellent supplementary evidence. However, it requires substantial systems engineering beyond the architectural scope of this paper, and we plan to explore it in future work. We will add a detailed explanation of the scheduler and deployment mechanics to Section 4.2 and expand our Limitations section accordingly.

W2/Q2. Single-Clock trained on intermediate budgets

Thank you for this suggestion. We ran this exact experiment matching both data-construction and training compute. Please see the Common Response for the detailed results.

In brief, adding intermediate budgets helps Single-Clock (raising mean accuracy from 15.1% to 48.9%), but the returns diminish quickly and fail to close the gap with RUBATO-Full. At its explicitly trained budgets, Single-Clock remains 2–8 points below RUBATO-Full, and at unseen budgets, the gap grows to 6–13 points.

W3/Q3. Variability across training runs

To robustly measure variability, we retrained the models using two additional random seeds (three seeds in total).

Main comparison. Accuracy (%), mean ± s.d. over three seeds.

BenchmarkVariantB=1[1,2]2[2,3][3,4]481632B=∞\infty
BIG-Bench AudioSingle-Clock66.4 ± 14.043.1 ± 12.138.8 ± 11.336.2 ± 10.036.3 ± 11.135.6 ± 10.150.6 ± 7.861.3 ± 2.868.0 ± 2.278.2 ± 8.1
Gated74.0 ± 8.174.6 ± 9.475.4 ± 10.473.0 ± 12.272.5 ± 12.672.2 ± 12.370.2 ± 13.369.7 ± 16.068.7 ± 15.776.4 ± 10.6
Full84.3 ± 2.585.4 ± 4.384.9 ± 4.784.0 ± 5.784.1 ± 4.384.1 ± 4.284.2 ± 5.284.8 ± 4.883.4 ± 3.586.3 ± 2.0
Spoken GSM8KSingle-Clock39.5 ± 11.210.5 ± 6.87.7 ± 4.24.5 ± 2.94.7 ± 1.75.6 ± 1.312.8 ± 0.720.7 ± 4.629.5 ± 5.044.2 ± 8.3
Gated45.3 ± 5.746.1 ± 4.846.9 ± 5.147.2 ± 5.046.1 ± 4.445.7 ± 4.046.0 ± 4.345.9 ± 3.743.1 ± 3.346.5 ± 4.8
Full47.1 ± 2.447.2 ± 3.747.1 ± 2.245.9 ± 5.647.2 ± 4.446.5 ± 4.045.7 ± 5.847.8 ± 3.147.0 ± 3.349.2 ± 2.1

Training-layout mixture ablation. Spoken GSM8K accuracy (%), mean ± s.d. over three seeds.

MixtureVariantB=1[1,2]2[2,3][3,4]481632B=∞\infty
21.9%Gated49.6 ± 3.949.0 ± 4.847.6 ± 6.046.9 ± 8.345.7 ± 6.645.5 ± 6.045.9 ± 5.446.8 ± 6.743.3 ± 8.351.9 ± 2.5
Full53.0 ± 2.454.4 ± 1.254.0 ± 2.455.1 ± 3.553.5 ± 3.454.6 ± 2.552.8 ± 3.851.8 ± 4.149.7 ± 6.657.2 ± 3.0
4.4%Gated40.4 ± 17.343.1 ± 14.745.0 ± 12.845.1 ± 10.146.7 ± 7.347.6 ± 7.448.3 ± 3.347.4 ± 0.341.4 ± 0.455.0 ± 2.6
Full53.7 ± 0.054.3 ± 0.854.5 ± 0.454.5 ± 0.654.9 ± 0.255.2 ± 1.453.0 ± 2.650.3 ± 5.048.4 ± 4.657.4 ± 2.5

Across both experiments, RUBATO-Full demonstrates significantly higher robustness across training seeds compared to RUBATO-Gated, and the main trends hold. Thank you for raising the specific point regarding BIG-Bench Audio. We found that RUBATO-Full's absolute lead over Single-Clock at the explicitly trained budgets is indeed a result of training noise. Specifically, one of the Single-Clock seeds manages to match Full there (achieving 82.4% vs. 82.2% at B=1B=1). However, Single-Clock's severe performance degradation at unseen budgets is consistent across every seed. Its worst unseen-budget accuracy is 34.2 ± 9.3%, compared to 82.8 ± 4.2% for Full. This structural robustness to interpolation, rather than an absolute accuracy gain at the trained endpoints, is the advantage of RUBATO, and we will update our claims in the text to reflect this clearly. We will include these multi-seed results in the revised appendix.

AuthorsOpenReview ↗

Follow-up on Rebuttal

Dear Reviewer piFu,

Thank you again for your time and constructive feedback.

As the end of the discussion is approaching, we wanted to gently check if our rebuttal has resolved your concerns. Please let us know if any further clarification is needed. We are fully available to answer any remaining questions before the deadline.

Best regards,

The Authors

Thank the authors for the detailed response; it has resolved most of my concerns.

Reviewer_jo4rOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies a mismatch in full-duplex speech models: audio advances at a fixed real-time rate, whereas internal reasoning should use whatever decoding capacity is available. RUBATO gives real-time and thinking tokens separate positional clocks and compares three ways of incorporating them into RoPE. RUBATO-Full, selects the key token's clock for each attention pair and retains the full rotary frequency basis. Models are trained at two budget extremes and evaluated at unseen budgets. Experiments on spoken reasoning and overlapping questions show substantially more stable accuracy than a shared-clock baseline, with shorter time to answer as the assigned thinking budget increases.

Critical strengths

1.The paper identifies a useful and concrete problem. The distinction between elapsed audio time and progress in internal reasoning is well motivated. With a shared position counter, inserting more thinking tokens changes the positional distance between audio tokens, even though their physical timing is unchanged. This makes a change in decoding speed a distribution shift for the model. 2.The design choices are well supported. The progression from Split to Gated to Full helps explain why separate counters alone are insufficient. Split still introduces rotations from an unrelated clock, while Gated and Full remove this dependence. The differences become clearer in overlapping conversations and filler-audio quality, beyond single-turn answer accuracy. 3.The endpoint-training result is promising. On the main Spoken GSM8K sweep, RUBATO-Full maintains 48.7%-51.1% accuracy while the shared-clock baseline can fall to 1.7%. The fixed and frame-varying budget evaluations directly test the intended interpolation behavior, and the benefit persists when a second question overlaps the first. The multi-KV-cache implementation also makes a reasonable effort to preserve decoding efficiency, with a small attention overhead in the reported benchmark.

Critical weaknesses

1.The separation between training and test questions is unclear. Section 4.1 states that test leakage is avoided, but both the Spoken BIG-Bench Hard training data and the BIG-Bench Audio evaluation are derived from BBH questions. I could not find a clear explanation of how the original questions were separated between training and evaluation. Generating different spoken versions of the same question would not make the test independent. 2.The comparison leaves a simple alternative unexplored. All methods are trained at two extreme thinking budgets and then tested at intermediate budgets. This provides a clear comparison under the chosen training setup. However, it remains unclear whether adding a few intermediate budgets to Single-Clock's training would substantially reduce the performance gap. Since avoiding the preparation of training data for different decoding speeds is an important motivation, this comparison would help establish the practical advantage of the proposed method. 3.The conversation setting is narrower than the motivation. The overlapping benchmark is constructed by combining two independent mathematical questions. This is a useful test of handling a second question while processing the first, but everyday interruptions can also correct an earlier detail or change the original request. The current experiments do not establish whether the model can update its ongoing answer in these situations. The distinction between handling independent overlapping questions and responding to changes in a conversation should be clearer when discussing the results.

Questions for authors

1.Could the authors explain how the original BBH questions were separated between training and evaluation? In particular, can any evaluation question appear in the training data in a different spoken or paraphrased form? 2.How would Single-Clock perform if its training data included a small number of intermediate thinking budgets, under comparable training resources? 3.Are the conclusions about interruptions intended to cover a user correcting the original question, or only independent overlapping questions? If the former, a simple example in which the second utterance changes the required answer would help support that claim.

2 comments

We greatly appreciate your assessment that the problem we identify is concrete and useful, that the progression from Split to Gated to Full supports our design choices, and that the endpoint-training results are promising. We address your aligned weaknesses and questions below.

W1/Q1. Separation of BIG-Bench training and test questions

We sincerely apologize for the confusion, which stems from typos in our paper. Our BIG-Bench training questions are derived from the original BIG-bench dataset, not from BIG-Bench Hard. Specifically, we took questions from the same four tasks as BIG-Bench Audio from the broader BIG-bench dataset, and rigorously filtered out every single question that appears in the BIG-Bench Audio evaluation set. To balance the dataset, we then generated additional questions for each task. We have verified every BIG-Bench training question against all 1,000 BIG-Bench Audio questions to ensure zero test leakage. We will correct this typo and clarify the separation in the final text.

W2/Q2. Single-Clock trained on intermediate budgets

Thank you for raising this point. We conducted this experiment at a matched data-construction and training cost. Please refer to the Common Response above for the setup and empirical results.

In short, Single-Clock falls short of RUBATO-Full even when augmented with intermediate budgets. Mean accuracy plateaus at 48.9%, and adding B=4B=4 to the 1,2,3,∞\\{1,2,3,\infty\\} mix yields no further gains. At explicitly trained budgets, Single-Clock remains 2–8 points behind RUBATO-Full, and at unseen intermediate budgets, the gap widens to 6–13 points.

W3/Q3. Independent overlapping questions vs. corrections

Our interruption experiments specifically test independent overlapping questions, where a second question arrives while the model is actively reasoning and speaking about the first. We do not claim the model can dynamically revise an ongoing answer if the user issues a correction to the original prompt. We focused on independent overlapping questions because it serves as the most direct stress test of core full-duplex mechanical ability, meaning the model must simultaneously listen, think, and speak.

We completely agree that conversational corrections are a critical real-world scenario. However, handling them requires specialized training and evaluation data where the second utterance fundamentally alters the required answer, and this data, to our knowledge, is not publicly available yet. Because RUBATO is an architectural modification governing positional encoding, it is entirely agnostic to the conversational scenario. A model trained on correction data could leverage RUBATO without modification. We will make this scope boundary explicitly clear in our discussion.

AuthorsOpenReview ↗

Follow-up on Rebuttal

Dear Reviewer jo4r,

Thank you again for your time and constructive feedback.

As the end of the discussion is approaching, we wanted to gently check if our rebuttal has resolved your concerns. Please let us know if any further clarification is needed. We are fully available to answer any remaining questions before the deadline.

Best regards,

The Authors

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

This paper proposes RUBATO, a novel positional encoding method that enables full-duplex speech LLMs to adapt to fluctuations in decoding speed at inference time, mitigating the resulting mismatch between training and inference. The paper compares the three proposed variants on single-turn and overlapping QA tasks. Despite being trained only on extreme budget settings, RUBATO-Full maintains strong robustness across unseen intermediate budgets.

Critical strengths
  • The proposed method is novel and well-motivated. Independent clocks directly address the identified positional mismatch. Selecting rotations according to the key’s clock removes irrelevant positional terms, while Full preserves the original RoPE frequency basis. Compatibility with existing attention kernels improves practical feasibility.
  • Under the same two-endpoint training setup, RUBATO-Full maintains 48.7%-51.1% accuracy on Spoken GSM8K, whereas Single-Clock falls to 1.7% at some intermediate budgets. Overlapping-question and speech-quality evaluations further substantiate the benefit beyond ordinary single-turn accuracy.
Critical weaknesses
  • The evaluation lacks comparisons with relevant baselines. Results from recent half-duplex and full-duplex speech models on the two benchmarks would help contextualize the proposed model’s performance. Given differences in training, I do not expect RUBATO to achieve the best results, but measuring these performance gaps would provide a clearer picture of its practical value.
  • The comparisons focus on models trained at the two extreme budgets. Baselines trained on intermediate budgets are missing. Without this comparison and matched cost measurements, the advantage over broader speed augmentation and the claimed data-preparation savings remains uncertain. The authors could reuse the filler sentences generated for B=1 while randomly sampling B during training. If Single-Clock also achieves strong robustness under this setup, the benefits of the proposed method would require further justification.
Questions for authors
  • My understanding is that, during training, samples with B=1B=1 require the model to generate a filler sentence before answering, whereas samples with B=∞B=\infty allow it to produce the final answer directly. How does the model learn to generate fillers of appropriate lengths for intermediate budgets at test time? Given this training setup, I would expect its fillers to follow the style and length seen in the B=1 samples.
3 comments

Thank you for recognizing that RUBATO is novel, well-motivated, and practical on existing attention kernels, and for highlighting how RUBATO-Full demonstrates strong robustness across unseen budgets, extending its benefits to overlapping questions and speech quality. We address the weakness and questions in the following.

W1. Missing reference results from recent half- and full-duplex speech models

We agree that contextualizing our results against existing models is valuable. We summarize these comparisons below.

SystemTypeSpoken GSM8K acc. (%)BIG-Bench Audio acc. (%)
RUBATO-Full (ours)FD48.7–51.186.1–90.0
Single-Clock (ours)FD1.7–50.024.8–71.4
MiniCPM-o 4.5 (chat)HD70.762.4
MiniCPM-o 4.5 (duplex)FD62.549.5
MoshiFD4.614.2
GPT-realtime-2.1HD93.294.6

(FD = full-duplex, HD = half-duplex. For our models, the range reflects performance across the evaluated thinking budgets.)

While these numbers are not an apples-to-apples comparison, due to significant differences in model size, architecture, and proprietary training datasets, they provide helpful context. With that caveat in mind,`RUBATO-Full perform well on BIG-Bench Audio and moderate on Spoken GSM8K. We attribute part of the BIG-Bench Audio advantage to our training data covering the same four task types. We will include this table and context in the revised paper.

W2. Single-Clock trained on intermediate budgets

Thank you for suggesting this experiment. We ran it with strictly matched data-construction and training costs. Please see the Common Response for the full setup and results.

In summary, Single-Clock only approaches RUBATO-Full at the specific budgets it was trained on. Adding intermediate budgets raises its mean accuracy from 15.1% to 48.9%, but the gain saturates (adding B=4B=4 to 1,2,3,∞\\{1,2,3,\infty\\} provides no further benefit). Even at its trained budgets, Single-Clock remains 2–8 points below RUBATO-Full. At unseen budgets, the gap expands to 6–13 points for the best variants.

Q1. How does the model learn to produce fillers of appropriate length at intermediate budgets?

You raise a great point regarding the style. At intermediate budgets, the model's fillers do follow the stylistic patterns of the B=1B=1 training data. However, their duration does not remain fixed to the B=1B=1 length. Filler length varies intrinsically within the B=1B=1 data itself, as each filler lasts roughly as long as its corresponding thinking trace, and these trace lengths vary widely across examples. Consequently, the model learns a functional behavior (to "hold the floor" by producing fillers until its internal thinking trace concludes) rather than speaking for a rigid duration. This learned behavior transfers naturally to intermediate budgets, where the trace simply closes sooner, prompting the model to terminate the filler and deliver the answer.

AuthorsOpenReview ↗

Follow-up on Rebuttal

Dear Reviewer VBJB,

Thank you again for your time and constructive feedback.

As the end of the discussion is approaching, we wanted to gently check if our rebuttal has resolved your concerns. Please let us know if any further clarification is needed. We are fully available to answer any remaining questions before the deadline.

Best regards,

The Authors

Thank you for running the additional experiments. The results address some of my concerns. I will keep my score unchanged and lean toward accepting the paper. I still have some concerns about the experiment with intermediate thinking budgets. Only 4.4% of the training examples contain spoken fillers, and these examples are split across several budget settings. This may leave Single-Clock with too few examples per setting to learn to handle different budgets.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common Response to Reviewers

We sincerely thank all reviewers for recognizing the novelty and contributions of our work, and for their constructive feedback. A shared question among all three reviewers was how the Single-Clock baseline performs when trained on intermediate thinking budgets in addition to the two endpoints, under a matched computational cost. We ran this experiment and report it here, and we address the remaining concerns in the individual responses.

Single-Clock trained on intermediate thinking budgets in addition to the two endpoints, under a matched computational cost

Setup. We train Single-Clock on the two endpoints plus one, two, or three intermediate budgets, comparing it with RUBATO-Full trained strictly on the endpoints. We use the 4.4% training-layout mixture, where 4.4% of the training examples contain spoken fillers ( B≠∞B\neq\infty ). To match data-construction and training costs, every model uses the exact same number of spoken-filler examples, divided approximately equally among the bounded training budgets, with new fillers generated for each B>1B>1 example. The B=∞B=\infty examples remain identical across all models, and all models in this experiment are trained exclusively on Spoken GSM8K.

Results. Spoken GSM8K accuracy (%) with 95% Wilson confidence intervals. Training budgets are in braces; [a,b][a,b] denotes a per-frame budget drawn uniformly from [a,b][a,b], as detailed in the paper.

BBRUBATO-Full 1,∞\\{1,\infty \\}Single-Clock $ \{1,\infty\} $Single-Clock 1,2,∞\\{1,2,\infty\\}Single-Clock 1,2,3,∞\\{1,2,3,\infty\\}Single-Clock 1,2,3,4,∞\\{1,2,3,4,\infty\\}
153.7 [51.0, 56.4]30.2 [27.8, 32.8]45.5 [42.8, 48.2]47.9 [45.2, 50.6]45.9 [43.2, 48.6]
[1,2]53.6 [50.9, 56.3]3.6 [2.8, 4.8]46.8 [44.1, 49.5]50.4 [47.7, 53.1]45.2 [42.5, 47.9]
254.9 [52.2, 57.6]3.3 [2.5, 4.5]49.3 [46.6, 52.0]51.7 [49.0, 54.4]46.8 [44.1, 49.5]
[2,3]55.1 [52.4, 57.8]3.6 [2.8, 4.8]43.2 [40.6, 45.9]53.1 [50.4, 55.8]50.1 [47.4, 52.8]
354.8 [52.1, 57.5]4.8 [3.8, 6.1]34.7 [32.1, 37.3]52.9 [50.2, 55.6]49.1 [46.4, 51.8]
[3,4]54.7 [52.0, 57.4]3.8 [2.9, 5.0]29.7 [27.3, 32.2]51.9 [49.2, 54.6]50.5 [47.8, 53.2]
456.4 [53.7, 59.0]4.6 [3.6, 5.9]24.2 [21.9, 26.6]50.2 [47.5, 52.8]50.1 [47.4, 52.8]
654.4 [51.7, 57.1]7.9 [6.6, 9.5]23.6 [21.3, 25.9]46.5 [43.8, 49.2]45.3 [42.6, 48.0]
855.2 [52.5, 57.8]10.3 [8.8, 12.1]24.2 [22.0, 26.6]43.8 [41.2, 46.5]43.8 [41.1, 46.5]
1255.7 [53.0, 58.4]16.6 [14.7, 18.7]25.7 [23.4, 28.1]44.2 [41.6, 46.9]42.8 [40.1, 45.5]
1653.6 [50.9, 56.3]22.0 [19.9, 24.4]28.3 [25.9, 30.8]45.5 [42.8, 48.2]46.8 [44.1, 49.5]
3253.3 [50.6, 55.9]32.8 [30.3, 35.3]35.9 [33.3, 38.5]45.7 [43.1, 48.4]47.0 [44.4, 49.7]
∞\infty58.1 [55.4, 60.8]54.3 [51.6, 57.0]54.9 [52.2, 57.6]55.2 [52.5, 57.9]55.2 [52.5, 57.9]
[1,16]54.2 [51.5, 56.9]13.4 [11.6, 15.3]23.5 [21.3, 25.8]45.0 [42.3, 47.7]45.8 [43.1, 48.5]
Mean (14 settings)54.815.135.048.947.5

Findings.

  1. Diminishing Returns: Including more training budgets improves Single-Clock, but the gains saturate quickly. Its mean accuracy rises from 15.1% to 35.0%, peaks at 48.9%, and then falls slightly to 47.5% when B=4B=4 is added.
  2. Persistent Performance Gap: Even with intermediate budgets, Single-Clock only approaches RUBATO-Full at the specific budgets it was trained on, where it still lags by 2–8 points. At unseen budgets, this gap widens to 6–13 points for the two best Single-Clock models, and up to a severe 32-point deficit for the 1,2,∞\\{1,2,\infty\\} variant.

In conclusion, while training Single-Clock on more intermediate thinking budgets yields improvements, the performance saturates and remains inferior to RUBATO-Full (which only requires training on the two extreme budgets). This experiment underscores RUBATO's core contribution: enabling the model to dynamically adapt and interpolate to arbitrary thinking budgets during inference. We will include these results in the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.