acceptodds
Under review as a conference paper at ICLR 2027

HintMR: Eliciting Stronger Mathematical Reasoning in Small Language Models

Abstract

Small language models (SLMs) often struggle with complex mathematical reasoning, where errors made early in a solution can compound over long reasoning trajectories. We introduce HintMR, a hint-assisted reasoning framework that elicits stronger reasoning in SLMs by providing localized guidance throughout multi-step problem solving. At each reasoning step, a separate language model generates a context-aware hint conditioned on the problem and the solver's accumulated reasoning history. Rather than revealing the full solution, these hints provide targeted intermediate guidance that helps the SLM stay on a productive reasoning trajectory while continuing to generate the solution itself. We further introduce Adaptive HintMR (AdaHintMR), an efficient variant that selectively invokes the hint model only when additional guidance is likely to be useful. AdaHintMR estimates the solver's step-level uncertainty using token-level Shannon entropy and requests a new hint only when uncertainty exceeds a predefined threshold, allowing the SLM to reason independently on confident steps. Experiments across six mathematical reasoning benchmarks show that HintMR consistently improves reasoning accuracy over standard prompting and independent SLM reasoning. When paired with GPT-5.4 as the frontier hinter, HintMR substantially improves the reasoning performance of SLM solvers across multiple benchmarks, outperforming GPT-5.4 used directly as a standalone solver by up to 6.67 percentage points. These gains show that using a frontier model for targeted intermediate guidance can be more effective than relying on it to solve the entire problem directly. AdaHintMR further reduces reliance on the frontier model by selectively requesting hints only when the solver exhibits high uncertainty. Across the evaluated solvers and benchmarks, this reduces API cost by up to 72.4% relative to HintMR while maintaining strong reasoning performance. Overall, our results show that targeted intermediate hints can effectively elicit stronger mathematical reasoning from SLMs, while uncertainty aware hinting preserves much of this benefit with substantially lower frontier model inference cost.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Small language models, Mathematical reasoning, Chain-of-thought reasoning, Adaptive inference, Hint-based learning, Multi-step reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HhuAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces HintMR, a framework that improves mathematical reasoning in small language models through context-aware hints generated by a separate language model at each reasoning step. The authors further propose AdaHintMR, which selectively requests hints based on the solver's token-level Shannon Entropy to reduce inference costs. Experiments across six mathematical reasoning benchmarks and three small language models demonstrate improvements over independent reasoning and several hint-based baselines. AdaHintMR further reduces frontier-model API costs while generally maintaining competitive reasoning accuracy.

Critical strengths
  1. Effective intermediate guidance. HintMR provides context-aware hints based on the solver's reasoning history, improving mathematical reasoning accuracy across multiple benchmarks and small language models.

  2. Improved inference efficiency. AdaHintMR selectively invokes the hinter, reducing frontier-model API costs while generally maintaining competitive accuracy.

  3. Broad empirical evaluation. The experiments cover six mathematical benchmarks, three SLM solvers, and both frontier and local hinters, demonstrating the applicability of the framework across different model configurations.

Critical weaknesses
  1. Limited novelty. The central idea of improving reasoning through intermediate hints has been explored in prior work, including HSP, TMATH, ConciseHint and Step Guided Reasoning. HintMR mainly combines context-aware hint generation with stepwise reasoning, while AdaHintMR adopts entropy-based triggering, a commonly used uncertainty estimation technique in existing methods. The methodological novelty over these approaches remains limited.

  2. Limited evaluation on smaller models. Although the paper focuses on SLMs, the evaluated solvers are relatively capable models ranging from 3.8B to 9B parameters. Experiments on smaller models (e.g., 0.5B–3B) are needed to assess whether the proposed framework remains effective for more resource-constrained SLMs.

  3. Incomplete computational efficiency evaluation. The efficiency analysis mainly considers frontier-model API costs, without accounting for the computational overhead and latency of repeated solver and hinter interactions. This makes it difficult to assess the overall efficiency of the proposed framework.

  4. Lack of hint quality analysis. The framework relies on model-generated hints, but the paper provides limited analysis of their correctness and usefulness. Incorrect or misleading hints may propagate errors through subsequent reasoning steps, particularly when using weaker local hinters.

Questions for authors
  1. Smaller models. Could the authors extend the evaluation to smaller models (e.g., 0.5B–3B) to demonstrate the effectiveness of HintMR in more resource-constrained settings?

  2. Computational efficiency. Could the authors provide a more comprehensive efficiency analysis, including total inference cost and latency for both the solver and hinter?

  3. Hint quality. How does the quality of generated hints affect reasoning performance? Could the authors analyze the frequency and impact of incorrect or misleading hints, particularly when using weaker local hinters?

5 comments

We sincerely thank the reviewer for recognizing the effectiveness of context-aware intermediate guidance in improving mathematical reasoning accuracy, the inference efficiency of AdaHintMR in reducing frontier-model API costs while maintaining competitive accuracy, and the broad empirical evaluation across six mathematical benchmarks, three SLM solvers, and both frontier and local hinters. We also appreciate the reviewer’s concerns regarding novelty, smaller model backbones, end-to-end efficiency, and the potential impact of incorrect hints. We address each concern below.

Q1: How does HintMR differ from prior hint-based and step-guided reasoning methods (HSP, TMATH, ConciseHint, Step Guided Reasoning), and what is new beyond combining hint generation with stepwise reasoning and entropy-based triggering?

A1: We acknowledge that hint generation, stepwise reasoning, and uncertainty estimation are established techniques. Our contribution is context-aware solver–hinter collaboration with adaptive hint invocation, rather than the individual components.

(a) Comparison with prior methods. The key differences are summarized below.

Prior MethodWho Provides GuidanceWhenDifference from HintMR / AdaHintMR
HSP [1]Solver itselfOnce, before solvingA separate, stronger hinter provides hints throughout reasoning
TMATH [2]Fine-tuned hint-generation modelLearned during training; generated for math problemsInference-time collaboration, with hints conditioned on the solver's evolving reasoning
ConciseHint [3]Textual hint-generation mechanismAdaptively during reasoningImproves mathematical accuracy through external guidance, rather than primarily reducing reasoning length
Step Guided Reasoning [4]Solver itselfDuring stepwise reasoningSeparates solver and hinter, using an external model for context-aware guidance

Unlike one-time or self-generated guidance, HintMR uses a separate hinter conditioned on the solver's accumulated reasoning, supporting both frontier and local model collaboration.

(b) Novelty of AdaHintMR. AdaHintMR uses step-level token entropy to selectively invoke the hinter, rather than requesting hints at every step. This reduces unnecessary external-model calls and API expenditure without additional training. (c) Empirical evidence. Our original submission already compares against HSP and TMATH (Table 3) and evaluates adaptive variants (Table 1). Table 1 summarizes these results alongside other baselines.

Table 1: Accuracy (%) Comparison with Existing Reasoning Baselines

MethodAIME-24AIME-25AIME-26AMC23MATH-500Minerva-Math
No Hint (Qwen3.5-9B)26.6713.3326.6755.0074.4053.68
Self-Consistency33.3326.6730.0080.0082.0065.07
HSP (One-Time Hint)36.6720.0020.0077.5083.8044.49
Single GPT-5.4 Hint43.3346.6750.0087.5076.2058.82
TMATH26.6730.0026.6757.5075.8058.50
Qwen3.8-27B (Standalone)26.6733.3336.6780.0080.4064.39
HintMR (Qwen3.5-9B + GPT-5.4)76.6750.0066.6792.5088.8066.18
AdaHintMR (Qwen3.5-9B + GPT-5.4)76.6763.3366.6792.5086.4066.18

Results are from Tables 1 and 3 of our original submission. Single GPT-5.4 Hint is an additional baseline.

Overall, our contribution is improving SLM mathematical reasoning through context-aware guidance from a separate hinter, while reducing external API costs through adaptive hint invocation.

[1] Jinlan Fu, Shenzhen Huangfu, Hang Yan, See-Kiong Ng, and Xipeng Qiu. Hint-before solving prompting: Guiding llms to effectively utilize encoded knowledge. arXiv preprint arXiv:2402.14310, 2024.

[2] Changyong Qi, Yuang Wei, Haoxin Xu, Longwei Zheng, Peiji Chen, and Xiaoqing Gu. TMATH: A dataset for evaluating large language models in generating educational hints for math word problems. In Proceedings of the 31st International Conference on Computational Linguistics, pp. 5082–5093, 2025.

[3] Siao Tang, Xinyin Ma, Gongfan Fang, and Xinchao Wang. Concisehint: Boosting efficient reasoning via continuous concise hints during generation. arXiv preprint arXiv:2506.18810, 2025.

[4] Lang Cao, Yingtian Zou, Chao Peng, Renhong Chen, Wu Ning, and Yitong Li. Step guided reasoning: Improving mathematical reasoning using guidance generation and step reasoning. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp. 21112–21129, 2025.

Q2: Does HintMR also improve smaller, more resource-constrained solvers in the 0.5B–3B range?

A2: To evaluate the effectiveness of HintMR on smaller solvers, we conducted additional experiments using Qwen3.5-2B with GPT-5.4 as the hinter, following the same reasoning protocol as our main experiments. Table 2 presents the results across five mathematical reasoning benchmarks.

Table 2: Accuracy (%) of Qwen3.5-2B with GPT-5.4 as the Hinter

MethodAIME-24AIME-25AIME-26AMC23MATH-500
No Hint20.0016.6713.3352.5075.00
HintMR30.0020.0020.0057.5082.00
AdaHintMR26.6720.0020.0055.0082.00

As shown in Table 2, both HintMR and AdaHintMR improve Qwen3.5-2B accuracy across all five benchmarks. For example, HintMR improves AIME-2024 accuracy from 20.00% to 30.00% and MATH-500 accuracy from 75.00% to 82.00%. AdaHintMR also improves upon No Hint, although its accuracy is slightly lower than HintMR on AIME-2024 and AMC23.

Overall, these additional results demonstrate that the benefits of intermediate guidance extend to smaller 2B-parameter solvers. We also observe that larger solvers generally achieve stronger performance with hint-based guidance. For example, on AIME-2024, HintMR achieves 30.00% with Qwen3.5-2B, 80.00% with Phi-4-mini (3.8B), 76.67% with Gemma-2-9B, and 76.67% with Qwen3.5-9B. Similarly, on MATH-500, accuracy reaches 82.00%, 87.80%, 87.60%, and 88.80%, respectively. While performance does not increase strictly with parameter count, these results suggest that larger, more capable solvers can generally utilize intermediate guidance more effectively.

Q3: How efficient is the full framework once solver and hinter computation, repeated calls, and wall-clock latency are included?

A3: We would like to clarify that Table 2 of our original submission reports external GPT-5.4 API costs only, rather than total end-to-end computational costs. Our local experiments were conducted on institutional shared GPUs, which incurred no direct monetary charges. To provide a more comprehensive efficiency analysis, Table 3 reports average solver-generated tokens and wall-clock inference latency for No Hint, HintMR, and AdaHintMR, using Qwen3.5-9B as the solver and GPT-5.4 as the hinter, alongside direct GPT-5.4 API inference.

Table 3: Solver Token Usage and Inference Latency Across Mathematical Reasoning Benchmarks

DatasetNo Hint TokensHintMR TokensAdaHintMR TokensNo Hint Latency (s)HintMR Latency (s)AdaHintMR Latency (s)GPT-5.4 Direct Latency (s)
AIME-20249,2038,6426,3437575193398.8
AIME-20258,16811,9688,17862269747511.2
AIME-20266,89911,1768,75557864958111.6
AMC-20234,6223,9832,7834242421465.5
MATH-5002,6481,9562,016130.7107.4106.52.5
Minerva-Math2,6791,3071,40197.273.577.12.8

Token counts represent solver-generated output tokens and exclude hinter input/output tokens. All local experiments used a single NVIDIA A100-SXM4-80GB.

Computational overhead. AdaHintMR generally reduces solver token usage and inference latency compared with standard HintMR. For example, on AIME-2024, solver tokens decrease from 8,642 to 6,343, while latency decreases from 519 to 339 seconds. On AMC-2023, latency decreases from 242 to 146 seconds. Compared with No Hint, AdaHintMR also reduces latency on five of the six benchmarks. However, the improvements are not uniform: on AIME-2026, AdaHintMR uses more solver tokens than No Hint and has slightly higher latency.

Hinter overhead and API efficiency. Repeated hinter calls introduce additional token consumption and latency. AdaHintMR reduces this overhead by selectively requesting hints based on step-level entropy. As reported in Table 2 of our original submission, GPT-5.4 API cost decreases from 1.21¢ to 1.01¢ on AIME-2024, 1.44¢ to 1.04¢ on AIME-2025, and 0.83¢ to 0.32¢ on AMC-2023 compared with HintMR.

Overall, AdaHintMR improves the efficiency of solver–hinter collaboration by reducing external API costs and, on most benchmarks, wall-clock latency compared with HintMR. However, direct GPT-5.4 inference remains substantially faster, as it is accessed through a hosted API, whereas our solver runs locally on a single GPU.

Q4: How do incorrect or misleading hints, especially from weaker local hinters, affect the solver's reasoning accuracy?

A4: To evaluate hint quality, we conducted an additional analysis of hints generated by GPT-5.4 and Qwen3.8-27B, using GPT-6.1-sol as the annotator. Each hint was classified into four categories: (i) correct and useful, (ii) correct but vague, (iii) incorrect or misleading, and (iv) answer leakage. We clarify that the hinter instructions in Appendix B.6 explicitly prohibit revealing final answers or complete solutions, ensuring that the solver remains responsible for deriving the solution. Table 4 summarizes the results.

Table 4: Hint Quality Analysis Across Different Hinter Models (%)

HinterMethodCorrect/UsefulCorrect but VagueIncorrect/MisleadingAnswer Leakage
GPT-5.4HintMR72.018.59.50.0
GPT-5.4AdaHintMR68.124.57.40.0
Qwen3.8-27BHintMR33.060.96.10.0
Qwen3.8-27BAdaHintMR37.152.510.40.0

Percentages are based on GPT-6.1-sol annotations of sampled hints.

Hint quality and reasoning accuracy. GPT-5.4 produces substantially more correct and useful hints (68.1–72.0%) than Qwen3.8-27B (33.0–37.1%). In contrast, the weaker local hinter generates a much larger proportion of correct but vague and incorrect/misleading hints (62.9–67.0% combined), which may provide insufficient guidance or redirect the solver toward incorrect reasoning. This offers a plausible explanation for the lower accuracy observed when Qwen3.5-9B is paired with Qwen3.8-27B as the hinter. Conversely, the substantially higher proportion of correct and useful hints from GPT-5.4 provides more actionable intermediate guidance, consistent with the stronger reasoning accuracy achieved by the same Qwen3.5-9B solver.

These findings align with our original experimental results. On AIME-2024, HintMR achieves 76.67% accuracy with GPT-5.4 as the hinter, compared with 40.00% using Qwen3.8-27B. Similarly, on AIME-2025, AdaHintMR achieves 63.33% with GPT-5.4 versus 30.00% with Qwen3.8-27B. Thus, the observed relationship between hint quality and reasoning accuracy suggests that the effectiveness of solver–hinter collaboration depends strongly on the hinter's ability to provide correct, actionable guidance, although the aggregate annotations do not establish a direct causal relationship. No explicit answer leakage was detected in the annotated samples.

Reviewer_3UVKOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes HintMR, in which an independent hinter provides a brief, local hint for the next step based on the problem and the solver's reasoning history, while the solver remains responsible for constructing the complete solution. AdaHintMR retains the initial hint and uses the solver's average Shannon entropy over the top-20 token probabilities, with a threshold of τ=0.1, to decide when to request another hint. The authors evaluate Qwen3.5-9B, Phi-4-mini, and Gemma-2-9B as solvers, with Qwen3.8-27B or GPT-5.4 as the hinter, on six benchmarks: AIME-2024/25/26, AMC-2023, MATH-500, and Minerva-Math. Baselines include no hints, a single hint, Self-Consistency, and TMATH. The local-model configuration outperforms No Hint on all six benchmarks. With GPT-5.4 as the hinter, the adaptive method generally reduces API token costs, by up to 72.4% relative to HintMR, although its accuracy trade-off is not always better. The method is clear and interesting, but total computation/latency is not accounted for, statistical uncertainty is not reported for the small test sets, and GPT hint leakage, entropy-threshold calibration, and baseline/contamination controls need further attention.

Critical strengths
  1. Clear solver–hinter roles and adaptive gating. The method requests local hints only when needed while retaining the solver's responsibility for carrying out the solution. It is simple and requires no model retraining.
  2. Evaluation of effectiveness and cost across models and datasets. Six benchmarks, three solvers, and local/frontier hinters provide reasonably broad directional evidence, particularly that AdaHintMR can retain much of the benefit with fewer frontier-model calls.
Critical weaknesses
  1. Experimental robustness. AIME and AMC have small sample sizes, making a single evaluation susceptible to random fluctuations. The authors should report AVG@8 or AVG@16 results on AIME and AMC, averaging accuracy over independent generations, together with confidence intervals. A larger dataset, such as Olympiad Bench, could also be included. GPT hints may contain nearly complete solutions, so it remains unclear whether the gains come from local stepwise hinting or additional information from a stronger model.

  2. Concerns about the credibility of the experimental results. For Qwen3.5-9B on AIME-2026, the public MathArena evaluation reports 92.5%, whereas the No Hint result in Table 1 is only 26.67%. In Table 3, Qwen3.8-27B achieves only 33.33% on AIME-2025, while Red Hat's public evaluation of the original BF16 model reports 95.42%. Different evaluation configurations may affect results, but such large gaps require an explanation: are they caused by thinking mode, generation budgets, the stepwise reasoning protocol, or the evaluation implementation? The 16,000-token budget in Table 6 is explicitly specified only for AdaHintMR; the no-hint baseline's budget still needs clarification. The authors should first reproduce a reasonable standalone baseline and then validate their method. Improvements over a poorly performing baseline may not transfer to a model operating at its normal capability, directly affecting the credibility of the paper's main conclusions.

  3. Fairness of experimental comparisons. The method uses two models for reasoning and introduces additional hint calls. The authors should therefore compare it with mainstream multi-agent and test-time scaling (TTS) methods under the same or similar total inference cost. Costs should include both solver and hinter inputs/outputs and end-to-end latency. A single-hint baseline under the same budget should also be included to distinguish the benefits of stepwise interaction from those of additional strong-model information.

  4. Effectiveness of entropy gating. Table 4 groups steps only by whether the entire trajectory ultimately answers correctly, which does not establish that high-entropy steps need hints more. The authors should include random and fixed-interval triggering controls with the same number of hints or cost, along with sensitivity analysis of the threshold τ, to verify that the gains come from entropy-based timing rather than simply reducing hint calls.

Questions for authors
  1. Please report AVG@8 or AVG@16 results and confidence intervals for AIME and AMC, and check whether hints contain the final answer or nearly complete solutions.
  2. How do you explain the large gaps between the Qwen baselines and public evaluations? Please specify thinking mode, each method's generation budget, and the stepwise invocation settings, and revalidate the method against standalone baselines operating at their normal capability.
  3. Under the same total inference budget, does the method still outperform multi-agent, TTS, and single-hint approaches? Please report solver/hinter inputs and outputs, total cost, and end-to-end latency.
  4. Under the same hint budget, does entropy gating outperform random or fixed-interval triggering? Please provide threshold sensitivity analysis to validate the effectiveness of hint timing.
7 comments

We sincerely thank the reviewer for the detailed and constructive feedback. We especially appreciate the recognition of HintMR's clear separation between the solver and hinter, the training-free adaptive gating mechanism, and the evaluation across six benchmarks, three SLM solvers, and local/frontier hinters. We address the concerns regarding statistical robustness, baseline comparability, budget fairness, and entropy gating below.

Q1: How statistically reliable are the gains on AIME and AMC, and do frontier-model hints reveal final answers or nearly complete solutions?

A1: To evaluate the statistical reliability of our results, we conducted additional experiments using AVG@8, averaging accuracy across eight independent runs on AIME-2024/2025/2026 and AMC-2023 with the same model checkpoints and decoding settings. We computed 95% confidence intervals using problem-level bootstrap resampling, where problems were sampled with replacement and accuracy was averaged across the eight runs for each bootstrap sample. Table 1 reports the mean accuracy and bootstrap confidence intervals for HintMR and AdaHintMR.

Table 1: AVG@8 Accuracy (%) and 95% Bootstrap Confidence Intervals

DatasetHintMR (AVG@8)HintMR 95% CIAdaHintMR (AVG@8)AdaHintMR 95% CI
AIME-202472.5[59.8, 85.2]73.8[60.4, 85.8]
AIME-202547.5[35.6, 59.4]58.3[46.2, 70.0]
AIME-202667.5[54.1, 80.9]70.8[57.1, 83.8]
AMC-202388.8[82.3, 94.8]87.2[80.3, 92.8]

Statistical reliability. AdaHintMR achieves higher average accuracy on all three AIME benchmarks, with the largest improvement on AIME-2025 (58.3% vs. 47.5%). HintMR performs slightly better on AMC-2023 (88.8% vs. 87.2%). Although the confidence intervals overlap, these results provide additional evidence that HintMR and AdaHintMR maintain competitive reasoning performance across independent runs, while accounting for variability in the sampled problems.

Hint leakage. We would like to clarify that the hinter instructions in Appendix B.6 explicitly prohibit revealing final answers or complete solutions, requiring the hinter to provide concise, context-aware guidance while leaving the solver responsible for deriving the answer. To further investigate potential answer leakage, we conducted an additional hint-quality analysis using GPT-6.1-sol as the annotator. The analysis detected no answer leakage in the sampled hints generated by GPT-5.4 and Qwen3.8-27B under either HintMR or AdaHintMR. Table 2 presents the hint-quality distributions across both methods and hinter models.

Table 2: Hint Quality Analysis Across Different Hinter Models (%)

HinterMethodCorrect/UsefulCorrect but VagueIncorrect/MisleadingAnswer Leakage
GPT-5.4HintMR72.018.59.50.0
GPT-5.4AdaHintMR68.124.57.40.0
Qwen3.8-27BHintMR33.060.96.10.0
Qwen3.8-27BAdaHintMR37.152.510.40.0

Percentages are based on GPT-6.1-sol annotations of sampled hints.

Overall, the additional AVG@8 results provide stronger evidence of the consistency of HintMR and AdaHintMR across independent runs. Furthermore, the explicit hinter instructions and the absence of detected answer leakage across hints support our approach of providing targeted intermediate guidance rather than directly revealing solutions.

Q2: How can the credibility of the experimental results be established given the large performance gaps between the Qwen baselines and public evaluations, and how do differences in inference settings and generation budgets explain these discrepancies??

A2: We would like to clarify that our experiments and the public evaluations use different inference settings and evaluation protocols, making direct comparisons difficult. Table 3 summarizes the differences between these settings.

Table 3: Comparison of Public Qwen Evaluations and Our Experimental Settings

Model / BenchmarkPublic Accuracy (%)Our Accuracy (%)Public Evaluation SettingOur Evaluation Setting
Qwen3.5-9B / AIME-202692.5026.67MathArena; standalone reasoning; 192K maximum output tokens; 31,184 average output tokens; temperature 1.0, top-p 0.95; 4 runs per problemStepwise reasoning without hints; 16K token cap; 6.8K average output tokens; single-run evaluation
Qwen3.8-27B / AIME-202595.4233.33Red Hat AI BF16 reference; standalone reasoning through vLLM; 69,632 maximum sequence length; 8 evaluation seedsStepwise reasoning without hints; 16K token cap; 4.3K average output tokens; single-run evaluation

Public evaluation settings are taken from MathArena and Red Hat AI's published results.

Evaluation setting differences. As shown in Table 3, the public evaluations use different inference protocols, generation budgets, and evaluation runs. For example, MathArena allows up to 192K output tokens, averaging 31,184 tokens on AIME-2026, whereas our evaluation uses a 16K-token cap, averaging approximately 6.8K tokens. Similarly, Red Hat AI reports results averaged across eight evaluation seeds, while our original submission reports single-run accuracy. We respectfully agree that using the public evaluation configurations could substantially improve standalone model performance, although reproducing these settings would be necessary to determine the extent of the differences.

However, we emphasize that our primary objective is to evaluate whether targeted intermediate hints improve SLM mathematical reasoning under a consistent inference protocol. Therefore, the No Hint baseline follows the same generation framework without external guidance, allowing us to isolate the benefits of intermediate hinting. For example, on AIME-2026, Qwen3.5-9B accuracy improves from 26.67% (No Hint) to 40.00% with AdaHintMR using Qwen3.8-27B as the local hinter, and further to 66.67% with GPT-5.4 as the frontier hinter. These results demonstrate the effectiveness of intermediate guidance within our experimental setting.

Q3: Under the same total inference budget, does HintMR outperform multi-agent, test-time scaling (TTS), and single-hint approaches? Please report solver/hinter inputs and outputs, total cost, and end-to-end latency.

A3: We would like to clarify that our original submission already includes comparisons with Self-Consistency, HSP (Hint-before-Solving Prompting), TMATH, and standalone Qwen3.8-27B in Table 3 (Section 5.3), along with the No Hint baseline in Table 1. All methods were evaluated under the same configured inference budget, although their actual computational costs differ. To further assess the benefits of step-level guidance, we include an additional Single GPT-5.4 Hint baseline. While HintMR involves two-model collaboration, our focus is specifically on targeted intermediate guidance between a solver and a hinter, rather than general-purpose multi-agent coordination. Therefore, broader multi-agent approaches fall outside the primary scope of our baseline comparisons.

(a) Comparison with existing baselines. Table 4 summarizes the accuracy of the methods already evaluated in our original submission, alongside the additional Single GPT-5.4 Hint baseline.

Table 4: Accuracy (%) Comparison with Existing Reasoning Baselines

MethodAIME-24AIME-25AIME-26AMC23MATH-500Minerva-Math
No Hint (Qwen3.5-9B)26.6713.3326.6755.0074.4053.68
Self-Consistency33.3326.6730.0080.0082.0065.07
HSP (One-Time Hint)36.6720.0020.0077.5083.8044.49
Single GPT-5.4 Hint43.3346.6750.0087.5076.2058.82
TMATH26.6730.0026.6757.5075.8058.50
Qwen3.8-27B (Standalone)26.6733.3336.6780.0080.4064.39
HintMR (Qwen3.5-9B + GPT-5.4)76.6750.0066.6792.5088.8066.18
AdaHintMR (Qwen3.5-9B + GPT-5.4)76.6763.3366.6792.5086.4066.18

Results are from Tables 1 and 3 of our original submission. The Single GPT-5.4 Hint baseline is an additional experiment.

As shown in Table 4, HintMR and AdaHintMR outperform Self-Consistency, HSP, TMATH, and standalone Qwen3.8-27B across all six benchmarks when GPT-5.4 is used as the hinter. For example, on AIME-2025, AdaHintMR achieves 63.33% accuracy, compared with 26.67% for Self-Consistency, 20.00% for HSP, and 30.00% for TMATH. The additional Single GPT-5.4 Hint baseline provides one initial frontier-model hint without subsequent guidance, allowing us to further isolate the benefits of repeated step-level hinting.

(b) Computational cost and inference latency. We would like to clarify that the cost comparisons in Table 2 of our original submission reflect external GPT-5.4 API expenditure only, rather than total end-to-end computational costs. Our local experiments were conducted on institutional shared GPUs, which incurred no direct monetary charges. To provide a more comprehensive efficiency analysis, Table 5 reports average solver-generated tokens and wall-clock inference latency for No Hint, HintMR, and AdaHintMR, using Qwen3.5-9B as the solver and GPT-5.4 as the hinter, alongside direct GPT-5.4 solving through API calls.

Table 5: Solver Token Usage and Inference Latency Across Mathematical Reasoning Benchmarks

DatasetNo HintHintMRAdaHintMRNo Hint Latency (s)HintMR Latency (s)AdaHintMR Latency (s)GPT-5.4 Direct Latency (s)
AIME-20249,2038,6426,3437575193398.8
AIME-20258,16811,9688,17862269747511.2
AIME-20266,89911,1768,75557864958111.6
AMC-20234,6223,9832,7834242421465.5
MATH-5002,6481,9562,016130.7107.4106.52.5
Minerva-Math2,6791,3071,40197.273.577.12.8

Computational overhead. AdaHintMR reduces solver token usage and inference latency relative to both HintMR and No Hint on several benchmarks. On AIME-2024, solver tokens decrease from 9,203 (No Hint) and 8,642 (HintMR) to 6,343 (AdaHintMR), while latency decreases from 757 and 519 to 339 seconds, respectively. Similar improvements are observed on AMC-2023, MATH-500, and Minerva-Math relative to No Hint. All local experiments used a single NVIDIA A100-SXM4-80GB. However, both hint-based methods incur substantially higher latency than direct GPT-5.4 API inference.

Overall, our original submission demonstrates that HintMR and AdaHintMR achieve substantially higher accuracy than several established reasoning baselines under the same configured inference budget. The additional computational analysis further shows that AdaHintMR can reduce solver-side token usage and inference latency relative to HintMR and No Hint on several benchmarks, while reducing external API expenditure relative to standard HintMR.

Reviewer_yrvDOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2(+1)
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces HintMR, a framework to enhance small language models (SLMs) on their math reasoning ability. It uses a proprietary model as a hinter to provide hint after reading an SLM’s reasoning at each step, after which the SLM continues solving the problem with the provided hint. Results show boosted performance on several math benchmarks using SLMs like qwen3.5-9b, gemma2-9b and phi-4-mini, and GPT-5.4 as the hinter.

Critical strengths
  1. The proposed framework seems useful to enhance math reasoning ability on small models.
Critical weaknesses
  1. Tab2 shows API cost but doesn’t include GPU cost. The cost comparison only makes sense if assuming that local GPUs are free, otherwise it could be misleading.

  2. How does it perform when GPT-5.4 is used as both the solver and hint provider? This is essentially a step-wise self guidance system.

  3. It would be meaningful to test how it performs when both hinter and solver are small models (eg gemma-2-9b). It would be interesting to have a ranking for different settings (increased performance) like:

  • 1 model, end2end solver
  • 1 model, self-reflection solver
  • 2 models, 1 as hinter and 1 as solver, same model
  • 2 models, 1 as hinter and 1 as solver, hinter is a stronger/larger model
  1. There lacks systematic analysis on why the proposed/baseline methods help/fail. Eg when the method fails, is it due to incorrect hints, or the solver not using the hints correctly?

  2. How’s the quality of hints generated by different models (eg gpt5.4, qwen38-27b, TMATH)?

Questions for authors

See weakness above.

4 comments

We sincerely thank the reviewer for recognizing the usefulness of HintMR in improving mathematical reasoning in small language models. We also appreciate the questions regarding computational cost, different solver–hinter configurations, and the quality of generated hints. We address these concerns below.

Q1: How does HintMR account for local GPU costs and inference latency, rather than reporting only external API expenditure?

A1: We would like to clarify that the cost comparisons in Table 2 of our original submission reflect external GPT-5.4 API costs only, rather than total end-to-end computational costs. Our local experiments were conducted on institutional shared GPUs, which incurred no direct monetary charges. To provide a more comprehensive efficiency analysis, Table 1 below reports average solver-generated tokens and wall-clock inference latency for No Hint, HintMR, and AdaHintMR, using Qwen3.5-9B as the solver and GPT-5.4 as the hinter, alongside direct GPT-5.4 solving through API calls.

Table 1: Solver Token Usage and Inference Latency Across Mathematical Reasoning Benchmarks

DatasetSolver Tokens: HintMRSolver Tokens: AdaHintMRSolver Tokens: No HintNo Hint Latency (s)HintMR Latency (s)AdaHintMR Latency (s)GPT-5.4 Direct Latency (s)
AIME-20248,6426,3439,2037575193398.8
AIME-202511,9688,1788,16862269747511.2
AIME-202611,1768,7556,89957864958111.6
AMC-20233,9832,7834,6224242421465.5
MATH-5001,9562,0162,648130.7107.4106.52.5
Minerva-Math1,3071,4012,67997.273.577.12.8

AdaHintMR generally reduces solver token usage and inference latency compared with both HintMR and No Hint. On AIME-2024, solver tokens decrease from 9,203 (No Hint) and 8,642 (HintMR) to 6,343 (AdaHintMR), while latency drops from 757 and 519 to 339 seconds, respectively. Similar improvements are observed on AMC-2023, MATH-500, and Minerva-Math, although AdaHintMR uses slightly more tokens than HintMR on the latter two benchmarks. All local experiments used a single NVIDIA A100-SXM4-80GB.

Overall, these results demonstrate that AdaHintMR reduces external API expenditure relative to standard HintMR and, on several benchmarks, also reduces solver-side token usage and inference latency relative to both standard HintMR and the No Hint baseline.

Q2: How does HintMR perform when GPT-5.4 is used as both the solver and the hint provider?

A2: To evaluate whether HintMR also benefits frontier models, we conducted additional experiments using GPT-5.4 as both the solver and the hinter. In this setting, GPT-5.4 generates solutions step by step while also acting as the hinter, following the same HintMR protocol with step-level evaluation. Table 2 compares this self-hinting setting against standalone GPT-5.4 and HintMR using Qwen3.5-9B as the solver with GPT-5.4 as the hinter.

Table 2: Accuracy (%) Comparison Across Different Solver–Hinter Configurations

DatasetGPT-5.4 StandaloneGPT-5.4 Solver + GPT-5.4 HinterQwen3.5-9B Solver + GPT-5.4 Hinter
AIME-202476.6776.6776.67
AIME-202550.0053.3350.00
AIME-202660.0066.6766.67
AMC-202392.5092.5092.50
MATH-50084.4084.4088.80
Minerva-Math68.0268.0266.18

Overall, GPT-5.4 self-hinting improves accuracy over standalone GPT-5.4 on AIME-2025 (50.00% to 53.33%) and AIME-2026 (60.00% to 66.67%), while maintaining identical performance on the remaining four benchmarks. Interestingly, Qwen3.5-9B with GPT-5.4 hints matches GPT-5.4 self-hinting on three benchmarks and outperforms it on MATH-500 (88.80% vs. 84.40%). These results suggest that step-level guidance can benefit frontier models, although the improvements are modest, while smaller solvers can also achieve competitive or stronger reasoning performance through targeted hints from a frontier model.

Q3: How does HintMR perform when both solver and hinter are small models? Can you rank the settings by performance?

A3: To evaluate the effect of different solver–hinter configurations, we conducted additional experiments using Phi-4-mini (3.8B) as the solver, with Phi-4-mini, Qwen3.5-9B, and GPT-5.4 as hinters. We compare standalone reasoning, self-reflection, same-model hinting, and stronger-model hinting using both HintMR and AdaHintMR. Table 3 presents the accuracy across four mathematical reasoning benchmarks.

Table 3: Accuracy (%) Comparison Across Different Solver–Hinter Configurations

SettingSolverHinterAIME-2024AIME-2025AIME-2026AMC23
1 model, end-to-endPhi-4-miniNone10.0010.006.6742.50
1 model, self-reflectionPhi-4-miniItself10.006.676.6742.50
2 models, same model (HintMR)Phi-4-miniPhi-4-mini10.006.676.6742.50
2 models, same model (AdaHintMR)Phi-4-miniPhi-4-mini10.0010.006.6745.00
2 models, larger hinter (HintMR)Phi-4-miniQwen3.5-9B20.0016.6716.6757.50
2 models, larger hinter (AdaHintMR)Phi-4-miniQwen3.5-9B16.6716.6716.6760.00
2 models, frontier hinter (HintMR)Phi-4-miniGPT-5.480.0060.0063.3397.50
2 models, frontier hinter (AdaHintMR)Phi-4-miniGPT-5.470.0060.0073.3392.50

Overall, these results demonstrate that hint quality and hinter capability play an important role in improving SLM reasoning. Same-model hinting and self-reflection provide limited improvements over standalone reasoning, while using Qwen3.5-9B as a stronger hinter consistently improves accuracy. The largest gains are achieved with GPT-5.4 as the hinter, improving Phi-4-mini's accuracy from 10.00% to 80.00% on AIME-2024 and from 42.50% to 97.50% on AMC23. These results suggest that stronger hinters provide more effective intermediate guidance, allowing even relatively small solvers to achieve substantially better mathematical reasoning performance.

Q4: How does hint quality affect reasoning, and are failures caused by incorrect hints or by the solver failing to use valid hints?

A4: To evaluate hint quality, we conducted an additional analysis of hints generated by GPT-5.4 and Qwen3.8-27B, using GPT-6.1-sol as the annotator. We classified each hint into four categories: (i) correct and useful, (ii) correct but vague, (iii) incorrect or misleading, and (iv) answer leakage. As specified in Appendix B.6, the hinter is explicitly instructed not to reveal final answers or complete solutions, ensuring that the solver remains responsible for deriving the answer. Table 4 presents the hint-quality distributions for HintMR and AdaHintMR.

Table 4: Hint Quality Analysis Across Different Hinter Models (%)

HinterMethodCorrect/UsefulCorrect but VagueIncorrect/MisleadingAnswer Leakage
GPT-5.4HintMR72.018.59.50.0
GPT-5.4AdaHintMR68.124.57.40.0
Qwen3.8-27BHintMR33.060.96.10.0
Qwen3.8-27BAdaHintMR37.152.510.40.0

Percentages are based on GPT-6.1-sol annotations of sampled hints.

Hint quality analysis. GPT-5.4 produces substantially more correct and useful hints (68.1–72.0%) than Qwen3.8-27B (33.0–37.1%). In contrast, the local hinter produces considerably more correct but vague hints (52.5–60.9%), indicating that its guidance is often less actionable despite being mathematically valid. Incorrect or misleading hints account for 6.1–10.4% of the samples, and no explicit answer leakage was detected.

Overall, these findings suggest that the usefulness of intermediate guidance depends on hinter capability, with weaker local hinters more frequently providing vague rather than actionable hints. However, the annotations alone cannot establish whether reasoning failures are caused by incorrect hints or by the solver's inability to apply valid guidance. We will further examine unsuccessful reasoning trajectories to distinguish these failure modes and include the analysis in the revised manuscript.

Thanks for the experiments and analysis which are very insightful. I'm raising my score from 2 to 3.

Thank you for your positive feedback and for increasing the score. We appreciate your thoughtful evaluation and are glad that our additional experiments and analysis addressed your concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.