acceptodds
Under review as a conference paper at ICLR 2027

SABER: Stage-Adaptive Bridging and Early Routing for Large Reasoning Models

Abstract

Large reasoning models (LRMs) achieve strong performance through extended chains of thought, but their long reasoning trajectories incur substantial inference overhead. Existing early-exit methods reduce this cost by terminating reasoning once a correct answer becomes extractable, implicitly equating answerability with response readiness. In this work, we find that LRMs frequently become answerable well before they are ready to produce a high-quality response, revealing a Solution-Response Gap between answerability and response readiness, during which further computation consolidates the reasoning process and prepares the final response. To better exploit the information encoded within this gap, we introduce SABER, a stage-adaptive framework designed to harness the intermediate reasoning signals that bridge answerability and response readiness. SABER realizes this objective through two complementary modules: SCOUT identifies when the reasoning state becomes answerable, while BRIDGE efficiently exploits the remaining consolidation through compact recurrent latent computation. Across multiple reasoning benchmarks and different LRM backbones, SABER reduces average token usage by 22.1–49.6% while consistently outperforming competing compression methods in accuracy and quality. Beyond these performance gains, further analyses reveal heterogeneous consolidation patterns associated with distinct reasoning behaviors and varying contributions to response quality. These findings provide the research community with a functional perspective on reasoning consolidation, shifting the focus beyond when to stop toward what computation remains useful in achieving response readiness.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Reasoning Models, Efficient Reasoning, Early Exit, Latent Reasoning

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_qD6rOpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper distinguishes the point at which a reasoning prefix permits correct answer generation from a later point yielding a higher-quality response. SCOUT predicts a handoff within this interval, and BRIDGE replaces subsequent explicit reasoning with 32 recurrent latent steps while keeping the backbone frozen. Experiments across three backbones report 22.1–49.6% average token reductions, competitive accuracy, and improved quality-adjusted accuracy relative to evaluated compression methods.

Critical strengths
  • A useful evaluation perspective. Assessing final-response quality alongside answer correctness exposes failures that accuracy-only evaluations miss. The distinction is practically relevant even if its interpretation as separate computational stages requires further evidence.
  • Promising practical efficiency gains. The accounting includes latent steps and final-response generation, and Appendix G measures end-to-end latency including SCOUT and BRIDGE overhead. The frozen-backbone design and evaluation across multiple model sizes strengthen the practical contribution.
Critical weaknesses
  • The central construct of response readiness is insufficiently validated. Although the paper provides an operational definition, human validation of judge rankings, and a qualitative example, readiness is defined retrospectively by the highest observed rubric score rather than by an independently specified standard of response adequacy. A later response scoring better does not establish that the earlier response was not already correct, complete, and readable. Moreover, the human study evaluates selected tAt_A–tRt_R pairs, rather than directly establishing SABER’s benefit over immediate exit on representative matched prefixes. A blinded human comparison should assess whether SABER makes otherwise inadequate responses acceptable, while controlling for response length and distinguishing genuine consolidation from continued reasoning or stylistic refinement.
  • The distinctive contribution of gap-guided routing and supervision is insufficiently isolated. Adaptive switching between explicit and latent reasoning already appears in SwiReasoning[1], which is neither discussed nor evaluated. Hidden-state alignment for latent CoT distillation is also established in CODI[2]. SABER’s potentially distinctive contribution is therefore its answerability/readiness supervision, rather than hybrid computation or latent distillation alone. However, Direct Exit, Explicit-32, and Random-32 do not control for the benefit of training on selected high-quality teacher responses. A comparison against SwiReasoning at comparable budgets, plus the same BRIDGE architecture trained using ordinary completed-trajectory teachers instead of tRt_R-selected teachers, would directly test the claimed advance.

References

[1] SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs

[2] CODI: Compressing Chain-of-Thought into Continuous Space via Self-Distillation

Questions for authors
  1. At the selected tAt_A, how often is the correct answer already explicitly established in the retained reasoning prefix, and how much additional computation occurs inside the sampled final response?
  2. How much performance is lost when BRIDGE uses an ordinary full-trajectory teacher rather than the quality-selected tRt_R teacher, holding the student prefixes, architecture, and training budget fixed?
6 comments

W1. Response Readiness and Human Validation of Quality Gains

Thank you for your constructive feedback on response readiness and human validation. In response to your concerns, we clarify our definition and provide additional blinded human evaluation to substantiate the quality gains achieved by BRIDGE.

1. Operationalizing Response Readiness

Regarding your concern that a higher score later in a trajectory does not establish that the earlier response was inadequate, we emphasize that our quality criteria are specified independently of boundary selection, rather than defined through relative response rankings.  Drawing on FLASK, Prometheus, and ReasonEval, we defined four dimensions before selecting tRt_R: answer convergence, reasoning directness, non-redundancy, and local coherence. Each deduction is tied to observable defects (Appendix I), such as reopening settled conclusions, repeating derivations, or presenting inconsistent explanations.

Across MATH500, AIME 2024, and AIME 2025, Qwen3.5-Flash assigned correct responses at tAt_A mean scores of 3.18/4 (1.5B) and 2.65/4 (7B). On questions with both tAt_A and tRt_R, 64.4% (4B), 72.5% (1.5B), and 85.4% (7B) exhibited at least one concrete quality defect at tAt_A.

Thus, while locating tRt_R along a trajectory is an offline search step for training supervision, the transition from tAt_A to tRt_R directly remedies genuine deficits in completeness and clarity rather than chasing relative score rankings.

2. Blinded Matched-Prefix Human Evaluation

Building on this defect-level analysis, we further conducted a blinded human comparison to determine how often BRIDGE turns an inadequate Direct Exit response into one that readers judge acceptable.

We restricted the evaluation to problems with an observed rubric-score gain in both benchmark groups. We sampled 100 qualifying MATH500 problems and included all 44 qualifying problems across AIME 2024 and AIME 2025.

For each of the 144 problems, Direct Exit and BRIDGE generated their responses from the same SCOUT-selected prefix. Ten evaluators independently rated both responses for correctness, completeness, and readability with method identity concealed and presentation order randomized, using an explicit adequacy scale (1 = unusable, 2 = requires a substantive correction or addition, 3 = acceptable with minor defects, and 4 = fully satisfactory). We counted a change when at least 6 of 10 evaluators recorded that change in their paired scores, after which evaluators inspected the shared prefix to identify the primary reason for each improvement.

3. Genuine Consolidation vs. Continued Reasoning

BRIDGE frequently converts inadequate Direct Exit responses into acceptable ones, rather than merely refining already acceptable answers. As detailed in the table below, the vast majority of evaluated pairs achieved a human-rated score increase, and most of these gains directly crossed the explicit adequacy threshold (from 1–2 to 3–4).

The observed quality gains primarily reflect genuine consolidation of existing reasoning, rather than continued inference, stylistic refinement, or longer responses. Prefix inspection confirms that improvements are overwhelmingly driven by synthesizing and organizing existing derivations, with only minor fractions attributable to new inference steps or superficial stylistic edits. Furthermore, most improved BRIDGE responses were no longer than their Direct Exit counterparts, ruling out response length as a confounding factor.

\hline \text{Measure} & \text{Outcome} & \text{Pairs (\\%)} \\\\ \hline \text{Human ratings} & \text{Adequacy gain }(1\text{–}2 \to 3\text{–}4) & \mathbf{107/144\\ (74.3\\%)} \\\\ & \text{Any score increase} & \mathbf{122/144\\ (84.7\\%)} \\\\ \hline \text{Reason for gain} & \text{Consolidation} & 92/122\\ (75.4\\%) \\\\ & \text{Necessary new inference} & 15/122\\ (12.3\\%) \\\\ & \text{Stylistic refinement only} & 15/122\\ (12.3\\%) \\\\ \hline \text{Response length} & \text{BRIDGE no longer than Direct Exit} & 104/122\\ (85.2\\%) \\\\ \hline \end{array}$$

W2 & Q2. Isolating the Contribution of Gap-Guided Routing and Quality-Selected Teacher Supervision

SABER's distinctive contribution is not hybrid computation or hidden-state alignment alone, but answerability–readiness supervision that jointly targets reasoning efficiency, accuracy, and final-response quality. Unlike accuracy-oriented switching in SwiReasoning [1] or latent distillation in CODI [2], SABER trains SCOUT to route within [tA,tR][t_A,t_R] and BRIDGE to consolidate toward quality-selected correct responses at tRt_R. To substantiate this distinction, we conducted both suggested experiments: a budget-matched comparison with SwiReasoning and a controlled ablation of teacher supervision.

1. Comparison with SwiReasoning at Comparable Token Budgets

We varied SwiReasoning's switch-count limit across all three backbones and benchmarks. Figure 11, newly added to Appendix E, presents the resulting accuracy–token curves. For the table below, we selected the SwiReasoning configuration closest to SABER's mean token usage in each setting.

At comparable generation budgets, SABER achieves higher accuracy in seven of nine combinations and higher conditional response quality in eight. These results demonstrate that SABER generally achieves a more favorable accuracy–quality trade-off than the representative hybrid baseline at similar reasoning costs.

\hline & & \text{MATH500} & & \text{AIME24} & & \text{AIME25} & \\\\ \text{Backbone} & \text{Method} & \text{Acc.}\uparrow & \text{Qual.}\uparrow & \text{Acc.}\uparrow & \text{Qual.}\uparrow & \text{Acc.}\uparrow & \text{Qual.}\uparrow \\\\ \hline \text{R1-Distill-Qwen-1.5B} & \text{SwiReasoning} & 79.4 & 3.822 & 23.3 & 3.464 & 23.3 & 3.500 \\\\ & \text{SABER} & \mathbf{84.4} & \mathbf{3.823} & \mathbf{31.7} & \mathbf{3.600} & \mathbf{27.1} & \mathbf{3.875} \\\\ \hline \text{R1-Distill-Qwen-7B} & \text{SwiReasoning} & 85.6 & 3.576 & \mathbf{63.3} & \mathbf{3.539} & \mathbf{46.7} & 3.268 \\\\ & \text{SABER} & \mathbf{91.9} & \mathbf{3.677} & 57.1 & 3.359 & \mathbf{46.7} & \mathbf{3.432} \\\\ \hline \text{Qwen3-4B-Thinking} & \text{SwiReasoning} & 88.4 & 3.457 & 80.0 & 3.312 & 60.0 & 3.333 \\\\ & \text{SABER} & \mathbf{97.8} & \mathbf{3.965} & \mathbf{84.2} & \mathbf{3.781} & \mathbf{80.4} & \mathbf{3.750} \\\\ \hline \end{array}$$ **2. Isolating the Benefit of Readiness-Guided Teacher Supervision** To distinguish the contribution of teacher selection from that of the BRIDGE architecture, we fixed the student handoff prefixes, BRIDGE architecture, and training budget, varying only the source of teacher supervision. The full-trajectory teacher provides a correct response from an ordinarily completed reasoning trajectory, whereas the readiness-guided teacher provides the highest-scoring correct response sampled at $t_R$. The results show that ordinary completed-trajectory supervision can improve accuracy on AIME24 and AIME25 over Direct Exit, but does not consistently improve response quality or QAA. In contrast, **readiness-guided supervision achieves the highest accuracy, quality, and QAA across all three benchmarks**, outperforming the full-trajectory teacher under the same BRIDGE architecture and training budget. This demonstrates that the gains cannot be attributed solely to latent distillation or access to correct teacher responses; the readiness-guided selection of supervision targets provides additional benefits. $$\begin{array}{lccccccccc} \hline & \text{MATH500} & & & \text{AIME24} & & & \text{AIME25} & & \\\\ \text{Method} & \text{Acc.}\uparrow & \text{Qual.}\uparrow & \text{QAA}\uparrow & \text{Acc.}\uparrow & \text{Qual.}\uparrow & \text{QAA}\uparrow & \text{Acc.}\uparrow & \text{Qual.}\uparrow & \text{QAA}\uparrow \\\\ \hline \text{Direct Exit} & 96.8 & 3.870 & 3.746 & 77.9 & 3.710 & 2.891 & 72.5 & 3.730 & 2.704 \\\\ \text{BRIDGE (full-trajectory teacher)} & 96.5 & 3.874 & 3.738 & 82.5 & 3.745 & 3.090 & 79.6 & 3.697 & 2.943 \\\\ \text{BRIDGE (quality-selected at } t_R\text{)} & \mathbf{97.8} & \mathbf{3.965} & \mathbf{3.878} & \mathbf{84.2} & \mathbf{3.781} & \mathbf{3.184} & \mathbf{80.4} & \mathbf{3.750} & \mathbf{3.015} \\\\ \hline \end{array}$$ **References** [1] SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs [2] CODI: Compressing Chain-of-Thought into Continuous Space via Self-Distillation

Q1. What Has the Model Established by tAt_A, and How Much Computation Occurs in the Final Response?

1. What Has Been Established by tAt_A?

To directly examine what the model has established by tAt_A, we conducted additional prefix-level statistics and direct-answer probing. Using Qwen3-4B-Thinking, we examined the 509 usable tAt_A prefixes across MATH500, AIME24, and AIME25. We found that 270/509 (53.0%) of these prefixes already explicitly contained the correct final answer. For the remaining 239 prefixes where the answer had not yet been articulated, we further tested whether it could be elicited without additional reasoning by appending </think>The Final Answer is. This direct-answer probe successfully recovered the correct answer for 153/239 (64.0%) of those remaining prefixes——a phenomenon consistent with recent hidden-state probing findings showing that reasoning models internally encode correct solutions before explicitly verbalizing them (Zhang et al. [1]).

Combining these complementary signals, 423/509 (83.1%) of tAt_A prefixes either explicitly contained the correct answer or yielded it directly without intermediate reasoning. These analyses show that by tAt_A, the model has either explicitly written the final answer in the prefix or largely established the solution internally.

2. How Much Computation Remains in the Final Response?

Our analysis indicates that responses at tAt_A do not offload substantial unfinished reasoning into the final output, leaving at most minor tail-end steps. To quantify how much additional computation occurs inside the sampled final response, we conducted additional response-length statistics and found that, compared with final responses at tRt_R on the same questions, mean response length at tAt_A differed by only 107 tokens.

Crucially, even when such minor residual calculation remains at tAt_A, it falls squarely within our formulation rather than contradicting it. In SABER, consolidation is not restricted to passive restatement; as analyzed in Section 4.5, once the core solution path is settled, BRIDGE is explicitly capable of absorbing routine, deterministic tail steps in latent space.

References

[1] Reasoning Models Know When They're Right: Probing Hidden States for Self-Verification. (Zhang et al.)

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

This paper introduces the Solution–Response Gap, where the point at which a model can produce a correct answer t_A may precede the point at which it can produce a high-quality final response t_R. Based on this observation, the authors propose SABER: SCOUT predicts when to exit explicit reasoning using intermediate hidden states, and BRIDGE performs 32 recurrent latent steps before generating the final response. Experiments cover three reasoning models, three mathematical benchmarks, and GPQA-Diamond. SABER reduces reasoning tokens by 22.1%–49.6% and improves accuracy in some settings.

Critical strengths
  1. Valuable problem formulation. The paper extends reasoning compression beyond deciding when to stop, distinguishing answerability from response readiness.

  2. Promising empirical results. Across multiple models and benchmarks, SABER reduces explicit reasoning tokens while maintaining or improving performance in several settings.

  3. Broad evaluation. The experiments include cross-domain evaluation, ablations, sensitivity analyses, human validation, latency measurements, and case studies.

Critical weaknesses
  1. Complex supervision pipeline and unclear cost–benefit trade-off. Constructing SCOUT and BRIDGE supervision requires sampling responses at intermediate reasoning boundaries, evaluating response quality, and identifying tA and tR. The additional annotation and training costs are not sufficiently quantified, making it difficult to assess how the inference-time savings offset the upfront investment.

  2. Insufficient component-level ablations. Although the paper includes layer-aggregation, threshold, latent-step, and post-handoff analyses, it does not fully isolate the contributions of key design choices. For SCOUT, the individual contributions of state-transition features and temporal modeling remain unclear. For BRIDGE, the necessity of teacher-confidence weighting and hidden-state alignment is not established by dedicated ablations.

  3. Missing closely related work. The paper omits DTSR, SwiReasoning, and Selective Latent Thinking. The latter two already adaptively combine explicit and latent reasoning, making Section 2’s characterization of prior methods overly broad. While SABER’s response-readiness-based consolidation offers a meaningful distinction, clearer positioning and comparison with a representative hybrid baseline are needed to substantiate its quality–efficiency advantage.

[1] Xiang, Y., et al. When Is Thinking Enough? Early Exit via Sufficiency Assessment for Efficient Reasoning. ACL, 2026.

[2] Shi, D., et al. SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs. ICLR, 2026.

[3] Xie, H., et al. Selective Latent Thinking: Adaptive Compression of LLM Reasoning Chains. arXiv:2605.25745.

Questions for authors
  1. Could the authors isolate the contributions of state-transition features and the causal Transformer, and compare interval-based supervision with simpler stopping objectives?

  2. Could the authors ablate the individual loss terms, particularly teacher-confidence weighting and hidden-state alignment, to clarify their contributions?

  3. Could the authors quantify the additional data-construction and training costs, and discuss how these compare with the inference-time savings?

5 comments

W1 & Q3: Offline Construction and Training Costs versus Inference-Time Savings

SABER incurs offline supervision and training costs to learn reusable inference modules, while its inference savings accumulate over subsequent requests.

Boundary probing, response-quality scoring, and teacher-target selection are performed strictly once offline to train the reusable SCOUT and BRIDGE modules. Decoupling offline supervision curation from online inference is a common practice in recent reasoning-efficiency literature, for instance:

  • CODI [1] uses GPT-4-generated GSM8k-Aug data and constructs 8.1k correctness-filtered CommonsenseQA reasoning examples using GPT-4o-mini.
  • s1 [2] uses the Google Gemini to generate reasoning traces and answers, then curates 1,000 examples for supervised fine-tuning. Paper.
  • ConPress [3] generates and filters concise, correct reasoning trajectories offline for supervised fine-tuning.

We transparently quantify all offline workloads—including API annotation, data generation, and lightweight module training—and demonstrate that they are rapidly offset by online inference savings. As detailed in the table below, scoring responses with Qwen3.5-Flash incurs a total API cost of \$63.60 across backbones. Keeping backbone parameters frozen, training our lightweight SCOUT and BRIDGE modules requires 9.22, 15.16, and 10.21 A100 GPU-hours for the 1.5B, 4B, and 7B backbones, respectively.

\hline \text{Backbone} & \text{Response evaluations} & \text{Scoring cost (USD)} & \text{Total training (GPU-hours)} \\\\ \hline \text{R1-Distill-Qwen-1.5B} & 60{,}114 & 19.89 & 9.22 \\\\ \text{Qwen3-4B-Thinking} & 70{,}848 & 21.43 & 15.16 \\\\ \text{R1-Distill-Qwen-7B} & 66{,}576 & 22.29 & 10.21 \\\\ \hline \text{Total} & 197{,}538 & 63.60 & 34.58 \\\\ \hline \end{array}$$ **In contrast to these modest upfront investments, online deployment yields substantial, enduring latency reductions that quickly amortize the training cost.** As reported in Appendix G, SABER achieves end-to-end latency reductions of 90.43 s (MATH500), 61.57 s (AIME24), and 68.83 s (AIME25) per query, incorporating SCOUT monitoring, BRIDGE computation, and final-response generation. Assuming single-GPU inference and uniform pricing, the 15.15 GPU-hours required for Qwen3-4B BRIDGE training are fully amortized after just **742 requests**. While full pipeline amortization additionally encompasses offline generation, API scoring, feature extraction, and SCOUT training, these upfront investments are rapidly recovered within a fraction of standard deployment volume. ### References [1] Shen et al. CODI: Compressing Chain-of-Thought into Continuous Space via Self-Distillation. EMNLP 2025. [2] Muennighoff et al. s1: Simple test-time scaling. EMNLP 2025. [3] Deng et al. ConPress: Learning Efficient Reasoning from Multi-Question Contextual Pressure. ICML 2026.

W2 & Q1: Isolating SCOUT’s State-Transition Features, Temporal Modeling, and Interval-Based Supervision

The additional ablation study of SCOUT shows that interval-based supervision keeps mean stopping positions after the answerability boundary on all three benchmarks, while state-transition features and the causal Transformer reduce the average post-answerability delay.

Following your suggestion, we separately examined the architecture and supervision:

  • No state-transition features: Removes state-difference features while retaining interval first-stop supervision.
  • No causal Transformer: Retains the local MLP but removes causal temporal modeling, using the same interval first-stop supervision.
  • Pointwise stopping BCE: Retains the full SCOUT architecture but replaces interval first-stop supervision with a pointwise binary classification objective.

We report the mean signed stopping offset. Positive values indicate stopping after the answerability boundary on average; negative values indicate stopping before it. The average column gives the equally weighted mean across the three benchmarks.

\hline \text{Variant} & \text{Supervision} & \text{MATH500} & \text{AIME24} & \text{AIME25} & \text{Average} \\\\ \hline \text{Full SCOUT} & \text{Interval first-stop} & 479 & 581 & 858 & \textbf{639} \\\\ \text{No state-transition features} & \text{Interval first-stop} & 521 & 1550 & 707 & 926 \\\\ \text{No causal Transformer} & \text{Interval first-stop} & 488 & 569 & 1610 & 889 \\\\ \text{Full SCOUT} & \text{Pointwise stopping BCE} & 436 & -2587 & -4259 & -2137 \\\\ \hline \end{array}$$ **Interval first-stop supervision prevents catastrophic premature stopping.** Because inference handoff is irreversibly triggered by the *first* stopping signal, pointwise BCE suffers from a severe train-test mismatch: evaluating steps independently accumulates false-positive hazards over long reasoning horizons. While tolerable on shorter MATH500 traces, this cumulative hazard causes catastrophic early triggering on long-horizon AIME24 and AIME25 problems, handing off well before the problem becomes answerable. By directly penalizing the cumulative probability of premature exit prior to $t_A$, our interval first-stop objective eliminates this failure mode and guarantees safe post-answerability handoffs across all benchmarks. **State-transition features and causal temporal modeling act complementarily to bound post-answerability delay and maintain stopping stability.** Removing state-transition features deprives SCOUT of local velocity signals (i.e., how rapidly the hidden state is converging), whereas stripping the causal Transformer leaves the MLP blind to long-range trajectory context and vulnerable to local fluctuations. In both cases, ablating either component destabilizes boundary recognition on reasoning traces and substantially inflates the overall average post-answerability delay. Combining both mechanisms couples local transition dynamics with global trajectory history, achieving the lowest and most consistent overall delay.

W2 & Q2: Isolating the Contributions of Teacher-Confidence Weighting and Hidden-State Alignment

The ablation study of BRIDGE’s training objective confirms that both teacher-confidence weighting and hidden-state alignment are essential and complementary, with the full objective achieving the highest accuracy and quality-adjusted accuracy (QAA) across all three benchmarks.

Following your suggestion, we evaluated three controlled loss ablations under identical training schedules and target responses:

  • No-confidence: Removes teacher-confidence weighting while retaining unweighted response cross-entropy (CE) and hidden-state alignment.
  • No-alignment: Removes hidden-state alignment while retaining confidence-weighted response CE.
  • CE-only: Removes both components, retaining only unweighted response CE.
\hline & & \text{MATH500} & & \quad & & \text{AIME24} & & \quad & & \text{AIME25} & \\\\ \smash{\raise 0.7em{\text{Method}}} & \text{Acc.}\uparrow & \text{Qual.}\uparrow & \text{QAA}\uparrow & & \text{Acc.}\uparrow & \text{Qual.}\uparrow & \text{QAA}\uparrow & & \text{Acc.}\uparrow & \text{Qual.}\uparrow & \text{QAA}\uparrow \\\\ \hline \text{No-confidence} & 96.7 & 3.891 & 3.763 & & 82.1 & \textbf{3.785} & 3.107 & & 78.3 & 3.685 & 2.885 \\\\ \text{No-alignment} & 96.8 & 3.900 & 3.776 & & 81.7 & 3.740 & 3.056 & & 80.0 & 3.723 & 2.978 \\\\ \text{CE-only} & 96.5 & 3.886 & 3.750 & & 80.8 & 3.770 & 3.046 & & 77.5 & 3.693 & 2.862 \\\\ \textbf{Ours (BRIDGE)} & \textbf{97.8} & \textbf{3.965} & \textbf{3.878} & & \textbf{84.2} & 3.781 & \textbf{3.184} & & \textbf{80.4} & \textbf{3.750} & \textbf{3.015} \\\\ \hline \end{array}$$ **Removing either component reduces accuracy and QAA on every benchmark, and removing both produces the largest drop in both metrics.** The two components provide different training signals: - **Hidden-State Alignment:** Response cross-entropy supervises BRIDGE through final-response prediction. Hidden-state alignment additionally matches the student’s pre-response hidden state to the teacher’s after the end-of-thinking marker, directly supervising the state from which response generation begins. - **Teacher-Confidence Weighting:** Standard response cross-entropy gives each target token equal weight. The additional teacher-confidence-weighted term emphasizes tokens assigned higher probability by the teacher, while retaining the standard cross-entropy term for all target tokens. Together, these terms guide both the state prepared for response generation and the relative emphasis placed on response-token targets during training.
Reviewer_JS4QOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the distinction between a reasoning model being able to produce a correct answer and being ready to produce a settled, concise, and coherent final response. It defines the intervening reasoning span as the Solution–Response Gap. The authors identify answerability and response-readiness boundaries by truncating chains of thought, repeatedly sampling final responses, and scoring response quality along four dimensions. SABER then uses SCOUT to predict when to hand off explicit reasoning from intermediate hidden states, followed by BRIDGE, which performs 32 recurrent latent computation steps through a frozen backbone to replace subsequent explicit reasoning. Experiments cover three reasoning models, three mathematical benchmarks, and transfer to GPQA-Diamond. Table 1 reports a 22.1%–49.6% reduction in average generation steps while generally maintaining strong accuracy and response quality. The paper also explores associations between latent representations and subsequent explicit reasoning behavior.

Critical strengths
  1. The problem characterization and measurement validation are reasonably thorough. The paper goes beyond treating correctness after truncation as evidence that all subsequent computation is unnecessary, examining repetition, repeated revision, and incoherence in final responses. Appendix B compares three quality judges and includes human comparisons of 150 response pairs per judge, with 10 annotators per pair, as well as independent regeneration at the originally selected boundaries. In Table 5, 86.9% and 84.1% of the originally positive cases retain strictly positive quality gains under the two judges, respectively. These checks provide important support for the observed phenomenon.

  2. The method achieves practically meaningful compression and accounts for actual inference overhead. In Table 1, Qwen3-4B-Thinking reduces average generation steps by 34.1% while improving average accuracy by approximately 1.7 percentage points; the other two models also show useful quality–efficiency trade-offs. Appendix E.1 explicitly counts recurrent latent steps as computation steps, and Appendix G reports end-to-end latency including online monitoring and latent computation. This makes the efficiency claims more credible, although the latency results currently cover only single-request execution with batch size 1.

Critical weaknesses
  1. The ablations do not fully separate the effects of stage-specific latent consolidation from those of additional training on high-quality responses. BRIDGE is supervised using the highest-scoring correct response at the response-readiness boundary, with response-level cross-entropy and hidden-state alignment losses (page 5, Equations 9–10). However, Direct Exit, Explicit-32, and Random-32 in Figure 4 and Appendix E.6 do not receive the same response supervision. These results show that trained BRIDGE outperforms these untrained controls, but do not establish whether the gains primarily arise from recurrent latent computation, stage-specific handoff selection, or distillation of better final responses. For example, it remains unclear whether an input-independent learned soft prompt or a small non-recurrent adapter trained on the same targets would achieve similar results. This affects my assessment of the necessity of the individual design choices and the proposed mechanistic interpretation.

  2. The evidence for three latent modes does not yet support a strong functional interpretation. Appendix F.1 first projects 428 representations into a two-dimensional t-SNE space and then fits a GMM with three components specified in advance. It does not report clustering stability in the original representation space, component-count selection, or reproducibility across random seeds. Moreover, all 100 Rapid Closure examples come from MATH500, while 52 of the 56 AIME examples fall into Extended Consolidation. Dataset composition and problem difficulty may therefore explain part of the separation. The mean quality gains across modes in Table 11 are 0.023, 0.017, and 0.064, but uncertainty estimates for these paired differences are not provided. Classifying explicit continuations offers useful correlational evidence, but does not establish three stable, distinct consolidation functions in latent space. The claims about functional organization on page 9 should be more cautious.

Questions for authors
  1. With the SCOUT handoff prefixes, final-response training targets, and 32-step computation budget held fixed, how would an input-independent learned soft prompt or a non-recurrent adapter perform? Can controls trained with the same supervision establish an additional benefit from recurrent BRIDGE and stage-specific consolidation?

  2. Do the three latent modes reproduce in the original representation space or PCA space, across random seeds and different numbers of clustering components? After controlling for dataset and handoff position or trajectory length, do differences in behavioral composition and paired quality gains remain, and what are their confidence intervals?

  3. Please provide the training and threshold-selection configurations for the main baselines, their validation budgets, and question-level paired uncertainty estimates for the key differences in Table 1. At operating points matched for generation steps, does SABER retain the same accuracy and quality advantages over thoroughly tuned early-stopping methods? Each AIME test set contains only 30 questions, and repeated generations across four random seeds do not increase the number of independent test questions. This information could therefore affect my assessment of the smaller accuracy differences.

4 comments

W1 & Q1: Isolating the Gains of Recurrent BRIDGE Beyond Response Supervision

The additional ablation study of BRIDGE shows that recurrent BRIDGE outperforms both controls under the same supervision across all three benchmarks.

Following your suggestion, we added two trained controls to distinguish the contribution of recurrent latent computation from that of response supervision:

  • Learned soft prompt: 32 trainable embeddings shared across all inputs. Their values are input-independent and optimized using the same supervision as BRIDGE.
  • Non-recurrent adapter: Generates 32 latent embeddings conditioned on the initial handoff state, without receiving updated backbone states during latent computation.
  • Recurrent BRIDGE (Ours): Generates each latent embedding from the backbone state updated by the preceding latent step, allowing subsequent embeddings to adapt to the evolving computation.

Across the three trained variants, we used the same SCOUT-selected prefixes, training data, teacher supervision, and optimization budget. Each variant inserts 32 latent embeddings before the same end-of-thinking marker, followed by final-response generation. Direct Exit is included as a reference.

\hline & & \text{MATH500} & & \quad & & \text{AIME24} & & \quad & & \text{AIME25} & \\\\ \smash{\raise 0.7em{\text{Method}}} & \text{Acc.}\uparrow & \text{Qual.}\uparrow & \text{QAA}\uparrow & & \text{Acc.}\uparrow & \text{Qual.}\uparrow & \text{QAA}\uparrow & & \text{Acc.}\uparrow & \text{Qual.}\uparrow & \text{QAA}\uparrow \\\\ \hline \text{Direct Exit} & 96.80 & 3.870 & 3.746 & & 77.92 & 3.710 & 2.891 & & 72.50 & 3.730 & 2.704 \\\\ \text{Learned Soft Prompt} & 97.08 & 3.887 & 3.774 & & 82.08 & 3.760 & 3.086 & & 77.92 & 3.746 & 2.919 \\\\ \text{Non-recurrent Adapter} & 97.50 & 3.925 & 3.827 & & 79.17 & 3.737 & 2.958 & & 74.58 & 3.739 & 2.789 \\\\ \text{Ours (BRIDGE)} & \textbf{97.80} & \textbf{3.965} & \textbf{3.878} & & \textbf{84.17} & \textbf{3.781} & \textbf{3.182} & & \textbf{80.42} & \textbf{3.750} & \textbf{3.016} \\\\ \hline \end{array}$$ **As shown in the table, while both controls improve over Direct Exit, recurrent BRIDGE consistently delivers the strongest performance across all benchmarks and metrics**, with particularly pronounced advantages on the challenging AIME24 and AIME25 tasks. Under identical teacher supervision, this superiority highlights the limitations of both baselines: the Learned Soft Prompt relies on a static sequence incapable of adapting to instance-specific handoff states, while the Non-recurrent Adapter’s single-pass projection remains vulnerable to representation drift over long horizons. By contrast, BRIDGE couples instance-dependent conditioning with recurrent backbone feedback to progressively calibrate each latent embedding against updated hidden states, effectively mitigating drift and confirming the critical role of recurrent state consolidation.

W2 & Q2. Clustering Stability, Behavioral Differentiation, and Mode-Specific Quality Gains

Following your suggestions, we evaluated clustering reproducibility across representation spaces, random seeds, and component counts, and tested whether differences in behavioral composition and paired quality gains (with 95% CIs) remain robust after controlling for dataset identity, handoff position, and trajectory length.

The three latent modes are stably reproduced across random seeds and validated by BIC component selection. We fitted Gaussian mixture models across different component counts and random seeds in the original representation space, in PCA space, and in the two-dimensional t-SNE space. We used the Bayesian information criterion (BIC) to select the component count and the adjusted Rand index (ARI) to measure across-seed stability and agreement with the original grouping. As shown, both PCA and t-SNE projections select three components, achieving strong across-seed stability and high agreement with the original grouping.

\hline \text{Representation space}&\text{BIC-selected components}&\text{Across-seed ARI}&\text{ARI vs. original grouping}\\\\ \hline \text{Original representations}&4&0.68\pm0.05&0.54\pm0.04\\\\ \text{PCA}&3&0.83\pm0.03&0.71\pm0.03\\\\ \text{Two-dimensional t-SNE}&3&0.74\pm0.05&0.82\pm0.04\\\\ \hline \end{array}$$ **Distinct behavioral patterns across modes are genuine rather than artifacts of dataset composition or problem difficulty.** To test whether the link between the three latent modes and the three explicit continuation behaviors is driven by dataset or difficulty confounding, we evaluated their association under two controlled settings: within MATH500 alone using a $3\text{ modes} \times 3\text{ behaviors}$ contingency table, and across the pooled cohort using regression adjusted for dataset identity, handoff position, and original trajectory length. As shown in the table, latent mode membership remains strongly associated with continuation behavior within MATH500 and continues to significantly predict target continuation demand even at identical dataset and difficulty levels. $$\begin{array}{lccc} \hline \text{Behavioral-composition analysis}&\text{Effect estimate}&95\\%\text{ CI}&p\\\\ \hline \text{Within MATH500 }(3\times3,n=372)&\text{Cramér's }V=0.44&[0.37,0.51]&<0.001\\\\ \text{Pooled, covariate-adjusted}&\text{Adjusted OR}=2.56&[1.58,4.13]&<0.001\\\\ \hline \end{array}$$ **Mode-specific quality gains remain statistically significant and functionally distinct after covariate adjustment.** We compared BRIDGE and Direct Exit from the same handoff prefix on instances answered correctly under both conditions, computing the paired quality gain. With the judge blind to method and group labels, we used question-level bootstrapping to estimate 95% confidence intervals for within-group mean gains and between-group differences. Mean paired quality gains (95% CI): $$\begin{array}{lccc} \hline \text{Group}&\text{Pooled}&\text{MATH500 only}&\text{Covariate-adjusted}\\\\ \hline \text{Rapid Closure}&+0.023[+0.005,+0.041]&+0.023[+0.005,+0.041]&+0.025[+0.007,+0.043]\\\\ \text{Intermediate Stabilization}&+0.017[+0.000,+0.034]&+0.014[-0.003,+0.031]&+0.016[-0.001,+0.033]\\\\ \text{Extended Consolidation}&+0.064[+0.042,+0.086]&+0.052[+0.028,+0.076]&+0.048[+0.025,+0.071]\\\\ \hline \end{array}$$ Between-group differences in mean paired quality gains (95% CI): $$\begin{array}{lccc} \hline \text{Group contrast}&\text{Pooled}&\text{MATH500 only}&\text{Covariate-adjusted}\\\\ \hline \text{Extended}-\text{Rapid}&+0.041[+0.013,+0.069]&+0.029[-0.001,+0.059]&+0.023[-0.006,+0.052]\\\\ \text{Extended}-\text{Intermediate}&+0.047[+0.019,+0.075]&+0.038[+0.009,+0.067]&+0.032[+0.003,+0.061]\\\\ \text{Rapid}-\text{Intermediate}&+0.006[-0.019,+0.031]&+0.009[-0.016,+0.034]&+0.009[-0.016,+0.034]\\\\ \hline \end{array}$$ Across the pooled, MATH500-only, and covariate-adjusted specifications, both Extended Consolidation and Rapid Closure maintain statistically significant positive quality gains. Extended Consolidation consistently achieves the largest mean gain and preserves a significant advantage over Intermediate Stabilization across all three analyses, whereas its contrast with Rapid Closure narrows moderately after accounting for dataset and trajectory covariates. Together, these results demonstrate that even after adjusting for dataset and trajectory covariates, the three modes reliably differentiate both explicit continuation demand (Appendix F.2) and the paired interventional gain of BRIDGE over Direct Exit from identical handoff states. Operationally, this dual differentiation directly supports a functional organization of the handoff space, which we will rigorously substantiate on page 9 with these new findings.

Q3. Baseline Comparisons at Similar Generation Steps and Question-Level Uncertainty

To assess whether SABER’s gains persist at comparable generation budgets and how precisely the smaller AIME accuracy differences can be estimated from 30 questions, we compared methods at the closest observed budgets and computed question-level paired confidence intervals for the original Table 1 results.

Baseline comparisons at similar generation steps.

For CUSUM and DEER, we adjusted stopping thresholds to match SABER's mean token usage. We additionally included SwiReasoning[1], a method that combines explicit and latent reasoning, and selected its closest token-budget configuration. At comparable generation lengths, SABER consistently demonstrates superior accuracy and response quality across most backbone–benchmark combinations.

\hline &&&\text{MATH500}&&&\text{AIME24}&&&\text{AIME25}&&&\text{Avg.}&&\\\\ \text{Backbone}&\text{Method} &\text{Acc.}\uparrow&\text{Tok.}\downarrow&\text{Qual.}\uparrow &\text{Acc.}\uparrow&\text{Tok.}\downarrow&\text{Qual.}\uparrow &\text{Acc.}\uparrow&\text{Tok.}\downarrow&\text{Qual.}\uparrow &\text{Acc.}\uparrow&\text{Tok.}\downarrow&\text{Qual.}\uparrow&\text{QAA}\uparrow\\\\ \hline \text{R1-Distill-Qwen-1.5B}&\text{CUSUM} &76.6&3105&2.207&\mathbf{33.3}&12051&2.825&20.0&13186&3.792&43.30&9447&2.941&1.130\\\\ &\text{DEER} &61.8&3106&\mathbf{3.873}&13.3&12334&2.438&16.7&13001&3.800&30.60&9481&3.370&1.117\\\\ &\text{SwiReasoning} &79.4&3072&3.822&23.3&11082&3.464&23.3&12966&3.500&42.00&9040&3.595&1.553\\\\ &\text{Ours} &\mathbf{84.4}&3029&3.823&31.7&11106&\mathbf{3.600}&\mathbf{27.1}&12779&\mathbf{3.875}&\mathbf{47.73}&8971&\mathbf{3.766}&\mathbf{1.805}\\\\ \hline \text{R1-Distill-Qwen-7B}&\text{CUSUM} &86.0&2809&3.208&56.7&11766&3.426&40.0&13326&\mathbf{3.604}&60.90&9300&3.413&2.047\\\\ &\text{DEER} &85.0&2900&3.643&46.7&12914&3.054&26.7&13381&2.656&52.80&9732&3.118&1.743\\\\ &\text{SwiReasoning} &85.6&2751&3.576&\mathbf{63.3}&12839&\mathbf{3.539}&\mathbf{46.7}&12753&3.268&65.20&9448&3.461&2.276\\\\ &\text{Ours} &\mathbf{91.9}&2854&\mathbf{3.677}&57.1&12876&3.359&\mathbf{46.7}&13005&3.432&\mathbf{65.23}&9578&\mathbf{3.489}&\mathbf{2.299}\\\\ \hline \text{Qwen3-4B-Thinking}&\text{CUSUM} &78.0&3868&1.419&63.3&15154&1.303&30.0&15273&0.750&57.10&11432&1.157&0.719\\\\ &\text{DEER} &95.8&6105&3.891&83.3&18849&\mathbf{3.820}&80.0&22276&3.688&86.37&15744&3.800&3.287\\\\ &\text{SwiReasoning} &88.4&3814&3.457&80.0&14879&3.312&60.0&14398&3.333&76.13&11030&3.367&2.569\\\\ &\text{Ours} &\mathbf{97.8}&3842&\mathbf{3.965}&\mathbf{84.2}&15184&3.781&\mathbf{80.4}&14548&\mathbf{3.750}&\mathbf{87.47}&11191&\mathbf{3.832}&\mathbf{3.359}\\\\ \hline \end{array}$$ **Question-level uncertainty.** We performed 20,000 paired bootstrap resamples by sampling question IDs with replacement, retaining all eight generations per question across four seeds. In each replicate, we recomputed Qual. from correct responses within each seed and QAA from seed-averaged accuracy and quality, with equal weighting across the three benchmarks. The table reports SABER's gains over baselines, with 95% paired confidence intervals. These results demonstrate that **SABER's advantage persists after accounting for question-level uncertainty**. Moreover, the smaller AIME accuracy gains are distributed across multiple independent questions rather than driven by a single outlier, supporting the robustness of the original Table 1 comparisons without treating repeated generations as independent samples. $$\begin{array}{llccccc} \hline \text{Metric / Benchmark}&\text{Baseline}&\text{Gain}\uparrow&\text{95\\% paired CI}&\text{Improved}&\text{Unchanged}&\text{Declined}\\\\ \hline \text{QAA (3-benchmark avg.)}&\text{CUSUM}&\mathbf{+0.677}&\mathbf{[0.485,0.874]}&519&10&31\\\\ &\text{BCR}&\mathbf{+0.265}&\mathbf{[0.118,0.410]}&394&74&92\\\\ \hline \text{Acc. (AIME24)}&\text{Original}&\mathbf{+2.08}\\,\text{pp}&\mathbf{[0.00,4.58]}\\,\text{pp}&3&27&0\\\\ \text{Acc. (AIME25)}&\text{Original}&\mathbf{+3.33}\\,\text{pp}&\mathbf{[0.42,6.67]}\\,\text{pp}&5&24&1\\\\ \hline \end{array}$$ **References** [1] Shi, D., et al. SwiReasoning: Switch-Thinking in Latent and Explicit for Pareto-Superior Reasoning LLMs. ICLR, 2026.

As the discussion phase is approaching its end, we kindly ask you to let us know if the above clarifications and the previously added experiments have addressed your remaining questions. We would be happy to address any additional points you may have during the remaining time of the discussion phase.

We sincerely thank you for your positive assessment of our work and for your engagement in the discussion.

Comments on the paper

1 comment

Dear AC and Reviewers,

We thank all reviewers for their thorough and constructive feedback, and appreciate that they highlighted:

  1. Our problem formulation distinguishes answerability from response readiness. Unlike prior methods focused solely on correctness, SABER also targets final-response quality for stronger overall performance (JS4Q, qD6r).
  2. The framework achieves meaningful compression and practical efficiency gains under a frozen-backbone design, accounting for inference overhead (JS4Q, rRnm, qD6r).
  3. The presentation is clear and backed by comprehensive evaluation, including cross-domain transfer, ablations, human validation, and latency analysis (JS4Q, rRnm).

We summarize our central contributions, findings, and newly added rebuttal evidence:

1. Scope and Contributions:

We shift reasoning compression from when a model can answer to when it can deliver a high-quality response, by (i) separating answerability (tAt_A) from response readiness (tRt_R) via the Solution–Response Gap; (ii) proposing SABER to keep explicit solution discovery while compressing consolidation into latent steps; and (iii) jointly evaluating accuracy, quality, and cost. Even correct answers can repeat steps or turn incoherent. Under our rubric, new rebuttal defect checks and a blinded matched-prefix human study confirm BRIDGE turns inadequate Direct Exit responses into acceptable ones via consolidation rather than added inference, style edits, or length (qD6r), distinguishing SABER from accuracy-oriented methods (rRnm, qD6r).

2. Stage-Adaptive Routing and Latent Consolidation:

SCOUT uses hidden-state transitions and causal temporal modeling to select handoff within [tA,tR][t_A,t_R]; BRIDGE then runs 32 recurrent latent steps on updated backbone states before responding, trained with frozen backbones on quality-selected tRt_R targets via response and alignment losses. New rebuttal ablations confirm each choice: under identical supervision and budgets, recurrent BRIDGE outperforms a learned soft prompt and non-recurrent adapter (JS4Q); a readiness-selected teacher beats an ordinary full-trajectory teacher on accuracy, quality, and QAA (qD6r); SCOUT's interval loss prevents premature exit on long traces while transition/temporal features bound delay; and confidence weighting and alignment complementarily supervise BRIDGE (rRnm). Latent analysis finds three modes—Rapid Closure, Intermediate Stabilization, and Extended Consolidation—linked to distinct continuations and gains. Follow-up rebuttal checks verify across-seed stability in PCA/t-SNE spaces and persistence of behavioral differences and key quality gains after covariate adjustment (JS4Q).

3. Effectiveness, Generalization, and Practical Utility:

Across three backbones on MATH500, AIME24, and AIME25, SABER reduces average generation steps by 22.1%–49.6% while preserving strong accuracy and response quality, alongside cross-domain evaluation on GPQA-Diamond. In new budget-matched comparisons, SABER generally achieves stronger accuracy–quality trade-offs than tuned early-exit baselines (CUSUM, DEER) and the newly added hybrid baseline SwiReasoning (JS4Q, rRnm, qD6r), while a new matched-prefix human study further confirms quality improvements (qD6r). Finally, our end-to-end latency analysis accounts for SCOUT monitoring, BRIDGE computation, and response generation, while newly quantified one-time offline supervision and training costs are rapidly amortized by recurring inference savings (rRnm).

We have updated the manuscript with additional baseline comparisons (Appendices E.2–E.3), BRIDGE design comparisons and component ablations (Appendices E.7 and E.9), and further analyses of latent-mode stability and quality gains (Appendices F.1 and F.3). We have also clarified related work (Appendix A.3) and improved the manuscript’s presentation throughout.

The constructive discussions during the rebuttal have substantially strengthened the rigor and quality of our paper. We are grateful to Reviewer qD6r for acknowledging our revisions and raising their score from 2 to 3, and we sincerely thank the AC and all reviewers for helping us improve this work.

Sincerely,

All authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.