acceptodds
Under review as a conference paper at ICLR 2027

When to Think Deeply: Response-Conditioned Inhibitory Routing for LLM Reasoning

Abstract

Selective reasoning in language models must balance opportunities to correct an available answer against additional computation and harmful replacements. Inspired by inhibitory control in the human brain, we introduce Inhibitory Deliberation for Problem Reasoning (IDPR), which separates answer formation from answer release. A fast policy generates a candidate, and a response-conditioned gate either releases it or inhibits its release to invoke slow reasoning. Decomposed supervision uses paired fast–slow outcomes to learn slow-over-fast quality gain, generation cost, and corrective potential from the question, fast answer, and generation evidence. Calibration selects the most accurate of three scoring rules at fixed 8% coverage. On a fixed-policy benchmark of 5,000 mathematical problems, IDPR raises accuracy from 47.90% to 48.92% at an 8.20% slow-call rate, with 519.42 average generated tokens. The gain over Always-Fast is significant under a Holm-adjusted paired McNemar test (), with 111 corrected errors and 60 harmful replacements. With response-and-evidence inputs and encoder family held fixed, IDPR-Matched's joint supervision-and-scoring scheme achieves 1.40 percentage points higher accuracy than RouteLLM-Q+R+E while reducing slow calls from 9.46% to 7.14% at their respective validation-calibrated thresholds.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM reasoning, Inhibitory control, Test-time compute allocation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_sCPdOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Inhibitory Deliberation for Problem Reasoning (IDPR), a response-conditioned routing method that determines whether to return a fast-generated answer or invoke additional slow reasoning. The method employs a controller trained on paired fast–slow outcomes to predict answer quality improvement, generation cost, and corrective potential, followed by validation-based calibration for routing decisions. Experiments on mathematical reasoning problems from GSM8K and Mixture-of-Thoughts show that IDPR improves accuracy over the always-fast policy and several routing baselines at a low slow-call rate.

Critical strengths
  1. The paper addresses a practical and well-motivated problem of deciding when additional reasoning is worthwhile, considering both potential corrections and harmful replacements.

  2. The proposed framework is straightforward and interpretable, leveraging paired fast–slow outcomes to guide routing decisions.

  3. The paper provides statistical significance tests and analyses of corrective and harmful replacements, offering useful insights into the behavior of different routing strategies.

Critical weaknesses
  1. Limited methodological novelty. The core ideas of response-conditioned routing and learning from fast–slow outcome differences have been explored in prior work. The proposed method largely combines existing techniques, while its specific methodological contributions remain insufficiently distinguished from related approaches.

  2. Insufficient ablations and baseline comparisons. The paper lacks key ablations to validate the necessity and individual contributions of its proposed components, particularly the quality, cost, and correction supervision, given that the final routing decision relies only on predicted quality gain. Moreover, direct comparisons with closely related learned deferral and cascading methods are missing.

  3. Limited experimental scope and generalizability. The main experiments rely on a single fast–slow model pair and mathematical problems from GSM8K and Mixture-of-Thoughts, without evaluation across diverse models or task domains. Most additional analyses revisit the same experimental results rather than providing independent validation, leaving the generalizability and robustness of the proposed method unclear.

Questions for authors
  1. How does IDPR meaningfully differ from existing learned deferral and cascading methods, given that response-conditioned routing, relative correctness prediction, corrective supervision, and cost-aware decisions have all been explored in prior work? What is the specific methodological contribution beyond combining these existing components, and what evidence establishes its necessity?

  2. Given that the final routing decision relies only on predicted quality gain, what are the actual contributions of the auxiliary cost and correction supervision? In particular, can the authors compare quality-only training against the full multi-task design under controlled settings? If these additional components provide negligible or even negative gains, what justifies their necessity in the proposed framework? More broadly, the method introduces multiple components and design choices, yet lacks sufficient ablations and sensitivity analyses to justify key settings, such as loss weights and other hyperparameters.

  3. Why does the evaluation primarily focus on heuristic baselines and RouteLLM adaptations, while closely related learned deferral and cascading methods are not directly compared? How can the claimed effectiveness of IDPR be established without comparisons to these more directly relevant approaches under matched experimental settings?

  4. Can the authors evaluate IDPR on additional fast–slow model pairs and independent benchmarks beyond mathematical reasoning? Given that the main evaluation relies on a single model pair and a custom in-domain benchmark, the current results seem insufficient to establish whether the observed improvements generalize beyond this specific experimental setting.

Reviewer_Ld3bOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a two-stage cascade inference-time compute method for mathematical reasoning, termed IDPR. A fast policy generates a candidate answer, and then a probe is used to decide whether to output that answer or to call a slower, more capable model. The controller predicts, from the input query, quick-model output and numerical features, the change in correctness between models, and resamples around 8% of answers using a calibrated threshold. On a composite dataset they show IDPR increases accuracy while maintaining around 8% of inference with the slow model.

Critical strengths
  1. IDPR uses estimated improvement as the quantity for deciding on regeneration, which is more straightforward than uncertainty-based methods for cascades.
  2. Reporting of results is comprehensive, with improved/harmful replacement rates reported and comprehensive statistical analysis.
  3. Reporting of implementation is comprehensive.
Critical weaknesses
  1. The paper’s own cost-quality metric, UU, for every intervention, including IDPR, is lower than the always-fast baseline. This implies that UU is a flawed measure.
  2. Qwen2.5-Math-Instruct outperforms IDPR on accuracy as well as average token count.
  3. No reporting of always-slow results. This is a much-needed baseline. The paper’s own training statistics suggest it may outperform IDPR. Appendix A.1 reports 277 slow wins and 226 fast wins among the 2,000 training pairs, meaning the slow policy is roughly (277 − 226)/2000 ≈ 2.55 percentage points more accurate than the fast policy. If the test distribution is similar, always-slow would reach roughly 50.4%, above IDPR’s 48.92%. The same statistics imply an oracle router could gain up to about 13.9 points over Always-Fast, so IDPR’s 1.02-point gain captures a small fraction of the available headroom. Both always-slow and oracle-routing results should be reported.
  4. The effect of the auxiliary losses is unablated, even though it’s considered a central contribution. 5.Section 4 states that the test set was used during method development.
  5. No comparison to Gupta et al. (2024), which should be a main baseline.
  6. IDPR’s improvement over the simplest confidence baseline is not statistically significant. Table 3 shows that the 0.70-point gain over Low Avg LogProb at the same 410 slow calls has a Holm-adjusted McNemar p-value of 0.1064. The paper therefore does not establish that the learned controller outperforms routing on the fast model’s average token log-probability, a baseline that requires no training. The difference between main IDPR and IDPR-Matched is likewise non-significant (p = 0.1083). These results come from single selected runs, and the reported intervals exclude variance from controller retraining and generation resampling, so the true uncertainty is larger than shown.
Questions for authors

See weaknesses

Reviewer_2j82OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

To address the trade-off between reasoning accuracy and inference cost, this paper proposes Inhibitory Deliberation for Problem Reasoning (IDPR). A fast policy first generates a candidate answer, and a response-conditioned controller decides whether to release this answer or invoke a slow reasoning policy. The controller jointly predicts signed quality gain, generation cost, and corrective potential from paired fast–slow outcomes. A separate calibration set selects the routing score and threshold. On a custom benchmark of 5,000 mathematical problems, IDPR improves accuracy from 47.90% to 48.92% with an 8.20% slow-call rate. The paper also compares response-conditioned routing schemes and analyzes corrections, harmful replacements, and threshold transfer.

Critical strengths
  • S1. This paper studies a relevant problem in selective reasoning: additional computation can damage an already correct answer, so the routing decision should consider the value of replacement rather than fast-answer uncertainty alone. Using paired outcomes to distinguish corrections from harmful replacements is reasonable and intuitive. The formulation also makes clear that the slow policy generates a new answer from the original question.

  • S2. The paper provides a useful analysis of the observed routing gains. IDPR corrects 111 errors while introducing 60 harmful replacements, yielding 51 additional correct answers. The paired tests, source-wise results, and explicit accounting for fallback and repair tokens make the results easier to interpret. The matched-input comparison with RouteLLM-Q+R+E is also informative: IDPR-Matched improves accuracy by 1.40 percentage points with fewer slow calls at the respective calibrated thresholds.

Critical weaknesses
  • W1. The evaluation does not provide sufficiently independent evidence for the proposed controller's generalization. Section 4 acknowledges use of the 5,000-example benchmark during earlier system development; Appendix A.1 describes a custom source-balanced set, rather than the standard GSM8K test split, and unknown overlap with the slow model's Mixture-of-Thoughts training data. Disjoint controller-training and calibration partitions do not resolve these issues. The intervals also condition on fixed generations and omit retraining and generation variability. This is not evidence of demonstrated leakage, but it makes a fresh, documented test set important when the headline gain is 51 additional correct answers out of 5,000.

  • W2. The methodological advance over existing deferral objectives is not established. [When Does Confidence-Based Cascade Deferral Suffice? (NeurIPS 2023)] explicitly studies correctness-difference targets, and [A Unified Approach to Routing and Cascading for LLMs (ICML 2025)] allows quality and cost estimates to depend on observed model outputs. IDPR's response-conditioned, format-aware targets and auxiliary supervision are plausible refinements, but the main deployed score uses only predicted quality gain. The matched RouteLLM comparison changes both supervision and ranking: a preference classifier could rank by P(slow-win) minus P(fast-win), rather than slow-win probability alone. Without that control and fixed-score auxiliary-loss ablations, the value of the cost and corrective heads remains unclear. The nonsignificant difference from Low Avg LogProb after Holm correction further limits the empirical case for the added complexity.

Questions for authors
  • Q1. Which design choices were informed by the current benchmark during earlier development? Can the authors provide results on a fresh held-out set and document the source splits, deduplication procedure, and overlap with the slow model’s training data?

  • Q2. Can the authors compare quality-only training, quality-plus-cost training, and the full objective under the same encoder, inputs, checkpoint-selection procedure, and routing budget? For the preference classifier, how does ranking by P(slow-win) − P(fast-win) compare with the current slow-win-only ranking?

  • Q3. Can the authors report accuracy–cost curves and variability across training seeds on the fresh evaluation? This would help determine whether the benefit persists beyond the selected 8% calibration point. It would also help explain the practical value relative to the direct Qwen2.5-Math-7B-Instruct reference, which has higher reported raw accuracy and lower generation cost, while having lower utility under the paper’s format penalty.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.