acceptodds
Under review as a conference paper at ICLR 2027

Think Simpler to Solve Harder: Difficulty-Aware Adaptive Reasoning via Token Entropy

Abstract

Decomposing complex problems into subproblems can improve the reasoning capabilities of large language models (LLMs). However, choosing an appropriate degree of decomposition based on subproblem difficulty and limiting the impact of intermediate-answer errors on final results remain important challenges. We find that answer-token entropy, defined as the mean entropy of token distributions across positions in a generated answer, can reflect subproblem difficulty at suitable granularities. Building on this observation, we propose **DART**, a training-free algorithm for **D**ifficulty-aware **A**daptive **R**easoning through **T**oken entropy. At suitable granularities identified through one-time probing, DART uses answer-token entropy as a routing signal to adaptively choose between direct solving and further decomposition, balancing solution quality and inference overhead. We also design a confidence-aligned self-critique mechanism that jointly constrains candidate evaluation scores and score-token confidence to guide the evaluation and refinement of intermediate answers, restraining the impact of errors on subsequent reasoning. Results aggregated across three tasks show that, compared to the baseline with the lowest average error on each backbone, DART reduces average error by 22%–57% and average online inference cost by 51%–72%.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Adaptive Reasoning, Uncertainty Estimation, Efficient Inference

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_i3SNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DART, a training-free adaptive reasoning algorithm whose core idea is to use answer-token entropy, the mean entropy of token distributions across positions in a generated answer to reflect subproblem difficulty. The method has two parts: first, a difficulty-aware router, after one-time offline calibration, classifies each subproblem into solve directly / decompose directly / entropy-gated regions, keeping the answer when entropy is low and decomposing further when it is high, thereby balancing accuracy and inference cost; second, a confidence-aligned self-critique mechanism in which the critic outputs both a score and the confidence of that score token, jointly constraining them to limit the propagation of intermediate-answer errors. Across three tasks set intersection, sorting, and keyword counting on Qwen3-32B and Llama3.1-70B, DART reduces average error by 22%–57% and online inference cost by 51%–72% relative to the best baseline on each backbone, achieving the lowest error in most configurations at lower cost.

Critical strengths

1.The research problem is clearly defined and the motivation is sound. Deciding at the subproblem level whether to continue decomposition is more fine-grained than allocating a fixed budget to the entire request. The authors also demonstrate the limitations of entropy across different granularities and address these differences through a zone-partitioning strategy.

2.Several results show strong practical improvement. For example, on Qwen's 64-element sorting, the average error drops from the best baseline's 4.56 to 1.43, with clearly reduced cost as well. Results of this kind support the value of the framework on the tested tasks.

3.The ablations and supplementary experiments are fairly complete. The paper includes analyses of fixed decomposition depth, self-correction, confidence constraints, feedback history, random routing, the entropy threshold, and the number of iterations. Appendix C.7 additionally runs direct solving and forced decomposition on the same subproblems, which comes closer to validating whether the routing decision is useful than merely showing entropy–error correlation curves.

Critical weaknesses

1.Task coverage is insufficient and does not yet support a claim of general complex-reasoning contribution. All three tasks have strongly decomposable structures, and their decomposition and aggregation rules are easy to specify explicitly. These tasks are suitable for testing local mechanisms, but they do not adequately demonstrate that the method can handle reasoning tasks in which the decomposition plan itself is uncertain, subproblems are mutually dependent, and answers are hard to verify precisely.

2.Novelty positioning still need strengthening. Regarding related work, DiffAdapt and other several works has already studied entropy, difficulty, and reasoning-strategy selection, Their implementations may differ from DART, but they should be included in a more direct comparison to clarify DART's unique contribution.

Questions for authors

See weaknesses

4 comments

Thank you for your careful and thorough review of our work. We add broader reasoning tasks and direct comparisons with DiffAdapt, with the evidence below.

[W1: Evaluation beyond Predefined Structural Decomposition]

Reviewer comment (W1): Task coverage is insufficient and does not yet support a claim of general complex-reasoning contribution. All three tasks have strongly decomposable structures, and their decomposition and aggregation rules are easy to specify explicitly. These tasks are suitable for testing local mechanisms, but they do not adequately demonstrate that the method can handle reasoning tasks in which the decomposition plan itself is uncertain, subproblems are mutually dependent, and answers are hard to verify precisely.


Thank you for your suggestion. We have added two benchmarks: DeepMath [1] and ProverQA [2]. DeepMath requires models to plan mathematical derivations and build on intermediate results; ProverQA requires multistep logical reasoning in which later inferences depend on earlier conclusions. Both involve choosing decomposition steps and reasoning with dependencies between subproblems.

Reference for changes: DART versus the baseline with the highest accuracy in each column (CR, GoT, GoT, and CR, from left to right). Accuracy differences are in percentage points (pp); cost changes use the same baseline in each column.

Entries report final-answer accuracy (%) / reported total cost on each evaluation set.

MethodDeepMath / QwenDeepMath / LlamaProverQA / QwenProverQA / Llama
IO48.3 / 0.6934.4 / 0.3163.4 / 0.4355.0 / 0.79
CoT63.9 / 6.1649.6 / 5.3571.7 / 2.1565.1 / 2.47
ToT75.5 / 96.1553.8 / 95.1775.5 / 35.6168.8 / 69.83
ToT269.3 / 87.1352.7 / 90.3274.9 / 32.8969.5 / 65.65
GoT73.9 / 73.6959.3‾\underline{59.3} / 63.0878.8‾\underline{78.8} / 17.8571.9 / 32.96
CR77.0‾\underline{77.0} / 41.6652.7 / 43.4973.2 / 18.4973.7‾\underline{73.7} / 25.28
AGoT72.8 / 53.4451.8 / 96.5176.5 / 29.3768.4 / 67.46
AoT70.1 / 35.7554.5 / 38.9272.3 / 14.5270.4 / 23.57
DiffAdapt75.5 / 36.3857.1 / 35.8677.6 / 15.7072.6 / 22.76
DART80.2 (↑3.2 pp) / 34.78 (↓16.5%)62.9 (↑3.6 pp) / 32.81 (↓48.0%)80.3 (↑1.5 pp) / 12.76 (↓28.5%)75.2 (↑1.5 pp) / 20.82 (↓17.6%)

Underlining marks the baseline with the highest accuracy in each column.

DART achieves the highest accuracy in all four task–model settings, at lower cost than the most accurate baseline in each. These results extend the evidence from fixed structural decomposition to mathematical and multistep logical reasoning with model-generated decomposition steps.

References

[1] Zhiwei He et al. DeepMath-103K: A Large-Scale, Challenging, Decontaminated, and Verifiable Mathematical Dataset for Advancing Reasoning. ICLR 2026.

[2] Chengwen Qi et al. Large Language Models Meet Symbolic Provers for Logical Reasoning Evaluation. ICLR 2025.

Manuscript changes: Sections 3.2 and 3.4 and Algorithm 1 clarify unified grouping and probe reuse (pp. 5–6, lines 238–294 and 318–324). Section 4.1 gives the datasets and sample sizes (pp. 7–8, lines 375–378 and 402–404); Figure 4 (p. 7, lines 357–360), Section 4.2 (p. 9, lines 448–450), and Appendix C.5, Table 13 (p. 28, lines 1459–1474), report the results. Task grouping is documented in Appendix C.2 (p. 19, lines 999–1007; Table 8, p. 20, lines 1027–1050) and Appendix C.4 (p. 22, lines 1160–1171; Figures 17–20, p. 26).


[W2: Comparison with Entropy-Driven Adaptive Methods and Contribution Positioning]

Reviewer comment (W2): Novelty positioning still need strengthening. Regarding related work, DiffAdapt and other several works has already studied entropy, difficulty, and reasoning-strategy selection, Their implementations may differ from DART, but they should be included in a more direct comparison to clarify DART's unique contribution.


Thank you for your suggestion. DART characterizes the relationship between answer-token entropy and solving performance across difficulty levels. Through clustering-based calibration, it enables subproblem-level adaptive decomposition and confidence-aligned correction within a single model, without training a routing network. The following table compares DART with related methods.

(a) Comparison of mechanisms

Related methodMain limitationsDART’s corresponding advantages
DiffAdapt [1]Requires a trained routing classifier; selects inference configurations for the whole questionNo routing-network training; clustering-based calibration selects solving or decomposition per subproblem
GlimpRouter [2]Requires switching between small and large models; lacks explicit recursive decomposition and a separate correction mechanismUses a single model; adaptive decomposition and confidence-aligned correction in both solving and aggregation
CoFiCot [3]Requires multiple candidates; main configurations rely on process/outcome reward modelsReuses the initial answer and its entropy; self-critique with the same model, without a dedicated reward model

The discussion of these methods has been added to Section 2.2 (Related Work) of the manuscript.

DART advances beyond DiffAdapt in two respects:

(1) A closer characterization of the entropy–performance relationship. DiffAdapt [1] uses U-shaped entropy patterns across difficulty levels, multiple response samples, and a trained classifier. DART uses group-level mean error and entropy–error correlation to calibrate capability regions through clustering, directly guiding solving or decomposition without training a routing network.

(2) Adaptive problem structure and correction. DiffAdapt adjusts inference configurations for the whole question without explicitly changing its structure; more reasoning may be insufficient when a problem exceeds the model’s direct-solving capability. DART recursively decomposes problems to reduce subproblem difficulty and applies self-critique during both solving and aggregation.

We further include DiffAdapt in the experiments on all five tasks and both models.

(b) Mathematical and logical reasoning tasks

Reference for changes: DiffAdapt on the same task and model. DART’s accuracy differences are in percentage points (pp), and cost changes are relative percentages.

Task / ModelDART accuracy (%)DiffAdapt accuracy (%)DART costDiffAdapt cost
DeepMath / Qwen80.2 (↑4.7 pp)75.534.78 (↓4.4%)36.38
DeepMath / Llama62.9 (↑5.8 pp)57.132.81 (↓8.5%)35.86
ProverQA / Qwen80.3 (↑2.7 pp)77.612.76 (↓18.7%)15.70
ProverQA / Llama75.2 (↑2.6 pp)72.620.82 (↓8.5%)22.76

DART improves accuracy by 2.6–5.8 percentage points while reducing cost by 4.4%–18.7% across all four settings.

(c) Structured tasks

Entries report mean final-answer error / total cost over 100 problems.

Reference for changes: DiffAdapt on the same task, input size, and model. Annotations show DART’s error and cost changes.

Task / SizeQwen: DARTQwen: DiffAdaptLlama: DARTLlama: DiffAdapt
Set intersection (N=32)0.19 (↓24.0%) / 1.96 (↓8.4%)0.25 / 2.140.03 (↓96.5%) / 2.42 (↓6.2%)0.86 / 2.58
Set intersection (N=64)1.30 (↓50.8%) / 5.66 (↓11.8%)2.64 / 6.420.83 (↓46.8%) / 8.85 (↓13.2%)1.56 / 10.20
Set intersection (N=128)7.75 (↓49.6%) / 19.40 (↑4.8%)15.39 / 18.526.43 (↓48.2%) / 27.91 (↑10.8%)12.41 / 25.19
Sorting (N=32)0.54 (↓10.0%) / 1.07 (↓3.6%)0.60 / 1.111.55 (↓20.9%) / 2.93 (↓17.5%)1.96 / 3.55
Sorting (N=64)1.43 (↓31.6%) / 5.85 (↓30.9%)2.09 / 8.474.54 (↓19.6%) / 13.06 (↓18.5%)5.65 / 16.02
Sorting (N=128)5.56 (↓43.9%) / 17.16 (↓14.9%)9.91 / 20.1714.90 (↓7.7%) / 35.36 (↓13.2%)16.15 / 40.75
Keyword counting1.80 (↓12.6%) / 5.42 (↓24.9%)2.06 / 7.221.18 (↓17.5%) / 9.65 (↓20.4%)1.43 / 12.12

DART achieves lower error in all 14 structured-task settings and lower cost in 12. The two cost exceptions are set intersection at N=128 on both models.

References

[1] Xiang Liu et al. DiffAdapt: Difficulty-Adaptive Reasoning for Token-Efficient LLM Inference. ICLR 2026.

[2] Wenhao Zeng et al. GlimpRouter: Efficient Collaborative Inference by Glimpsing One Token of Thoughts. Findings of ACL 2026, pp. 17850–17864.

[3] Dongxu Zhang et al. Not All Queries Need Deep Thought: CoFiCot for Adaptive Coarse-to-fine Stateful Refinement. UAI 2026.

Manuscript changes: Section 2.2 adds the three related methods and clarifies DART’s positioning (p. 3, lines 138–144). Appendix C.3 gives DiffAdapt implementation details (p. 22, lines 1148–1152); Sections 4.1 and 4.2 describe the baseline and comparison (p. 8, lines 415–417; p. 9, lines 438–440). DiffAdapt results appear in Figure 4 (p. 7, lines 357–360), Tables 10–12 (p. 27, lines 1411–1452), and Table 13 (p. 28, lines 1459–1474).

We hope that our responses adequately address the reviewer's concerns. We would be glad to hear any further feedback.


The author's response addressed my concerns, so I will raise my score

AuthorsOpenReview ↗

Thank you for your constructive feedback and support

Dear Reviewer i3SN,

Thank you for your careful reassessment and for raising your score. We are pleased that our responses and additional experiments have addressed your concerns.

Following your suggestions, we expanded our evaluation to DeepMath and ProverQA, added DiffAdapt comparisons across all five tasks and both models, and clarified DART’s contributions relative to other entropy-driven adaptive methods. These results and discussions have been incorporated into the revised manuscript.

We sincerely appreciate your time and constructive feedback, which helped us strengthen the empirical evaluation and clarify the positioning of our work.

Best regards,

The Authors

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

This paper presents DART, a training-free framework that adaptively decomposes problems using answer-token entropy and refines intermediate solutions through confidence-aware self-critique. Experiments on sorting, set intersection, and keyword counting report improved error–cost trade-offs over selected chain-, tree-, and graph-based reasoning baselines.

Critical strengths
  1. The paper considers both solution quality and inference cost, addressing a meaningful trade-off in allocating computation across subproblems.
  2. The component ablations and paired direct-solving versus decomposition experiments provide useful evidence for the proposed routing mechanism within the evaluated settings.
Critical weaknesses
  1. Problem decomposition, iterative correction, and difficulty-adaptive computation are established ideas. The main additions appear to be entropy-based routing and score-token confidence thresholds, whose advantages over simpler alternatives are not sufficiently established.
  2. Sorting, set intersection, and keyword counting are narrowly structured tasks with predefined decomposition and aggregation procedures, closely matching the design of earlier GoT-style methods. Success on these benchmarks provides limited evidence of progress on contemporary reasoning problems involving nontrivial dependencies, planning, or open-ended decomposition.
  3. DART requires reference-answer-based offline probing to determine routing regions and entropy thresholds, introducing additional cost beyond the reported online inference cost. The paper does not establish whether this calibration transfers across tasks or distributions, making the cost and effort of adapting the framework unclear.
  4. The reported AGoT configuration differs from its original experimental settings, and the AoT description leaves its iterative execution unclear. Moreover, performance differences are reported without uncertainty estimates, limiting confidence in smaller improvements.
Questions for authors
  1. Can the authors demonstrate gains on a task requiring nontrivial decomposition and dependencies, including a comparison against the backbone’s strong native reasoning configuration under a matched inference budget?
  2. How many calibration samples are necessary, and can the calibrated regions and thresholds transfer to unseen tasks, input sizes, or distributions? Please report total costs including calibration and the request volume needed to recover that overhead.
  3. Can the authors separate the effects of task-specific verification prompts, decomposition policy, and entropy routing, particularly given that sorting uses no instance-level entropy gating?
10 comments

[W1: Comparison with Simpler Routing and Confidence Schemes]

W1: Problem decomposition, iterative correction, and diffi◊culty-adaptive computation are established ideas. The main additions appear to be entropy-based routing and score-token confidence thresholds, whose advantages over simpler alternatives are not sufficiently established.


Thank you for raising this point. The following comparisons address this concern. Mean error refers to the final answer; total cost is in USD.

(1) Does entropy routing outperform simpler policies?

(a) Entropy routing versus fixed decomposition granularity

Fixed C1/C2/A corresponds to 8/4/2 sentences for keyword counting and 128×32, 128×16, and 128×8 for set intersection. Only the routing policy changes across methods.

Reference for changes: Full DART, with routing and self-critique enabled. Each variant’s error and cost changes are measured against this reference.

TaskModelMetricDARTFixed C1Fixed C2Fixed A
Keyword countingQwenError ↓1.802.77 (↑53.9%)1.89 (↑5.0%)1.54 (↓14.4%)
Cost ↓5.423.99 (↓26.4%)5.84 (↑7.7%)8.31 (↑53.3%)
Keyword countingLlamaError ↓1.181.83 (↑55.1%)1.43 (↑21.2%)0.95 (↓19.5%)
Cost ↓9.658.13 (↓15.8%)10.63 (↑10.2%)14.93 (↑54.7%)
Set intersection (N=128)QwenError ↓7.7512.57 (↑62.2%)8.80 (↑13.5%)6.74 (↓13.0%)
Cost ↓19.4013.38 (↓31.0%)19.90 (↑2.6%)28.78 (↑48.4%)
Set intersection (N=128)LlamaError ↓6.439.37 (↑45.7%)7.81 (↑21.5%)5.61 (↓12.8%)
Cost ↓27.9121.87 (↓21.6%)28.38 (↑1.7%)42.59 (↑52.6%)

Relative to DART, the error increase under fixed C1 (45.7%–62.2%) substantially exceeds its cost reduction (15.8%–31.0%). The cost increase under fixed A (48.4%–54.7%) substantially exceeds its error reduction (12.8%–19.5%). Fixed C2 increases both error and cost. Thus, entropy routing achieves a more favorable error–cost trade-off.

Manuscript changes: Table 1 (p. 8, lines 379–389) and the experimental scope and two-part analysis in Section 4.3 (p. 9, lines 450–464) have been updated.

(b) Entropy routing versus random routing

Only the Zone C decision changes; calibrated regions, prompts, and self-critique remain unchanged.

Reference for changes: DART on the same task and model. Annotations show the error and cost changes under random routing.

TaskModelMetricDARTRandom, p=0.25Random, p=0.50Random, p=0.75
Keyword countingQwenError ↓1.802.55 (↑41.7%)2.32 (↑28.9%)1.88 (↑4.4%)
Cost ↓5.425.21 (↓3.9%)5.61 (↑3.5%)7.00 (↑29.2%)
Keyword countingLlamaError ↓1.181.81 (↑53.4%)1.48 (↑25.4%)0.94 (↓20.3%)
Cost ↓9.659.54 (↓1.1%)10.27 (↑6.4%)12.71 (↑31.7%)
Set intersection (N=64)QwenError ↓1.301.61 (↑23.8%)1.69 (↑30.0%)1.31 (↑0.8%)
Cost ↓5.665.35 (↓5.5%)5.96 (↑5.3%)6.82 (↑20.5%)
Set intersection (N=64)LlamaError ↓0.830.97 (↑16.9%)0.92 (↑10.8%)0.89 (↑7.2%)
Cost ↓8.858.81 (↓0.5%)9.24 (↑4.4%)9.94 (↑12.3%)

The results show that entropy routing achieves the best performance–cost trade-off.

Manuscript location: The existing random-routing comparison is in Section 4.4 (p. 9, lines 475–485) and Table 4 (p. 10, lines 487–493).

(2) Does confidence alignment improve on score-only acceptance?

(c) Score-only acceptance

We remove the confidence constraint on the score token while retaining critic scores and all other settings.

Reference for changes: “Score + confidence” in the same row. Annotations show changes after removing the confidence constraint.

TaskModelMetricScore + confidenceScore only
Keyword countingQwenError ↓1.801.97 (↑9.4%)
Cost ↓5.425.49 (↑1.3%)
Keyword countingLlamaError ↓1.181.43 (↑21.2%)
Cost ↓9.659.60 (↓0.5%)
Set intersection (N=64)QwenError ↓1.301.66 (↑27.7%)
Cost ↓5.665.78 (↑2.1%)
Set intersection (N=64)LlamaError ↓0.831.04 (↑25.3%)
Cost ↓8.858.98 (↑1.5%)
Sorting (N=64)QwenError ↓1.431.68 (↑17.5%)
Cost ↓5.855.97 (↑2.1%)
Sorting (N=64)LlamaError ↓4.544.79 (↑5.5%)
Cost ↓13.0613.14 (↑0.6%)

Removing confidence alignment increases error by 5.5%–27.7% across all six settings, while cost changes by only −0.5% to +2.1%. This shows that adding a confidence constraint beyond the score threshold improves answer quality.

Manuscript location: The existing confidence ablation is in Table 2 (p. 8, lines 392–400) and Section 4.3 (p. 9, lines 470–473).


[Q3: Ablations of Entropy Routing, Decomposition Policy, and Verification Prompts]

Q3: Can the authors separate the effects of task-specific verification prompts, decomposition policy, and entropy routing, particularly given that sorting uses no instance-level entropy gating?


Thank you for your question. Sorting uses entropy-informed offline A/B calibration, with no Zone C. GMM+BIC combines group-level error and entropy–error correlation to route 64-number inputs to decomposition and 32-number subproblems to solving. The resulting online procedure for sorting (N=64) uses a fixed 32 + 32 split, followed by solving and aggregation.

All other settings match full DART; each control makes only the changes specified below.

  • w/o entropy routing: bypass entropy-guided region decisions and default to directly solving all 64 numbers.
  • w/o decomposition policy: disable the 32 + 32 decomposition and directly solve all 64 numbers. These two interventions yield the same execution path on sorting (N=64) and share one direct-solving control.
  • w/o task-specific verification prompts: use GPT-6-Astra to generate P1/P2 from the task description, algorithm description, and prompt framework, then replace the original verification prompt.

Reference: Original DART for the same model. Percentages use unrounded results for direct solving and displayed values for P1/P2, rounded to one decimal place.

ModelMethod (ablation)ConfigurationError ↓Error changeCost ↓Cost change
QwenFull methodOriginal DART (paper): 32 + 321.43—5.85—
Qwenw/o entropy routing / decomposition policyDirect solving: 64 (shared control)7.46+421.4%2.50-57.2%
Qwenw/o task-specific verification promptsGeneric verification prompt P11.40-2.1%6.07+3.8%
Qwenw/o task-specific verification promptsGeneric verification prompt P21.49+4.2%5.90+0.9%
LlamaFull methodOriginal DART (paper): 32 + 324.54—13.06—
Llamaw/o entropy routing / decomposition policyDirect solving: 64 (shared control)7.25+59.7%4.43-66.1%
Llamaw/o task-specific verification promptsGeneric verification prompt P14.81+5.9%13.09+0.2%
Llamaw/o task-specific verification promptsGeneric verification prompt P24.540.0%12.10-7.4%

Entropy-guided partitioning and decomposition. Their shared direct-solving control increases Qwen/Llama error by 421.4% / 59.7%, with 57.2% / 66.1% lower cost. This supports the overall contribution of calibrated partitioning and the decomposition policy it triggers.

Task-specific verification prompts. Replacing the original prompts with generic P1/P2 changes error by −2.1% to +5.9% and cost by −7.4% to +3.8%, showing that DART accommodates LLM-generated verification prompts while maintaining good solution quality.

Manuscript changes and locations: Appendix C.10 and Table 19 present the shared direct-solving control and verification-prompt ablations (p. 32, lines 1674–1713; Table 19, lines 1688–1706). Sorting's A/B routing is described in Appendix C.3 (p. 22, lines 1153–1157).


[W2: Mathematical Decomposition and Logical Dependencies]

W2: Sorting, set intersection, and keyword counting are narrowly structured tasks with predefined decomposition and aggregation procedures, closely matching the design of earlier GoT-style methods. Success on these benchmarks provides limited evidence of progress on contemporary reasoning problems involving nontrivial dependencies, planning, or open-ended decomposition.


Thank you for your suggestion. We have added two benchmarks: DeepMath [1] and ProverQA [2]. DeepMath requires models to plan mathematical derivations and build on intermediate results; ProverQA requires multistep logical reasoning in which later inferences depend on earlier conclusions. Both involve choosing decomposition steps and reasoning with dependencies between subproblems.

Reference for changes: The highest-accuracy baseline for each task–model setting: CR for DeepMath/Qwen and ProverQA/Llama; GoT for the other two. Accuracy changes are in percentage points (pp); cost changes are relative to the same baseline.

We report final-answer accuracy (%) and total cost (USD) on each evaluation set.

(a) Task: DeepMath

MethodQwen: accuracy (%) ↑Qwen: cost ↓Llama: accuracy (%) ↑Llama: cost ↓
IO48.30.6934.40.31
CoT63.96.1649.65.35
ToT75.596.1553.895.17
ToT269.387.1352.790.32
GoT73.973.6959.3‾\underline{59.3}63.08
CR77.0‾\underline{77.0}41.6652.743.49
AGoT72.853.4451.896.51
AoT70.135.7554.538.92
DiffAdapt75.536.3857.135.86
DART80.2 (↑3.2 pp)34.78 (↓16.5%)62.9 (↑3.6 pp)32.81 (↓48.0%)

(b) Task: ProverQA

MethodQwen: accuracy (%) ↑Qwen: cost ↓Llama: accuracy (%) ↑Llama: cost ↓
IO63.40.4355.00.79
CoT71.72.1565.12.47
ToT75.535.6168.869.83
ToT274.932.8969.565.65
GoT78.8‾\underline{78.8}17.8571.932.96
CR73.218.4973.7‾\underline{73.7}25.28
AGoT76.529.3768.467.46
AoT72.314.5270.423.57
DiffAdapt77.615.7072.622.76
DART80.3 (↑1.5 pp)12.76 (↓28.5%)75.2 (↑1.5 pp)20.82 (↓17.6%)

Underlining marks the highest baseline accuracy for each task and model.

DART achieves the highest accuracy in all four task–model settings, at lower cost than the most accurate baseline in each. These results extend the evidence from fixed structural decomposition to mathematical and multistep logical reasoning with model-generated decomposition steps.

References

[1] Zhiwei He et al. DeepMath-103K: A Large-Scale, Challenging, Decontaminated, and Verifiable Mathematical Dataset for Advancing Reasoning. ICLR 2026.

[2] Chengwen Qi et al. Large Language Models Meet Symbolic Provers for Logical Reasoning Evaluation. ICLR 2025.

Manuscript changes: Sections 3.2 and 3.4 and Algorithm 1 clarify unified grouping and probe reuse (pp. 5–6, lines 238–294 and 318–324). Section 4.1 gives the datasets and sample sizes (pp. 7–8, lines 375–378 and 402–404); Figure 4 (p. 7, lines 357–360), Section 4.2 (p. 9, lines 448–450), and Appendix C.5, Table 13 (p. 28, lines 1459–1474), report the results. Task grouping is documented in Appendix C.2 (p. 19, lines 999–1007; Table 8, p. 20, lines 1027–1050) and Appendix C.4 (p. 22, lines 1160–1171; Figures 17–20, p. 26).


Reviewer_1tfxOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces a new method to be more balanced between cost and error when deciding how much further sub-problem decomposition is needed. The motivation came from a deep analysis of how the answer token entropy correlated with errors at different granularity levels, and then adaptive reasoning is conducted to perform more decomposition when it's beneficial, so the reasoning depth is not fixed but is really adaptive, and therefore cost in $ and error rate are better balanced. Of course, it is not just based on the answer token entropy; the self-critique is also important to avoid any wrong token with high entropy from being encouraged.

Critical strengths

Token entropy-based adaptive reasoning is not novel; however, how the adaptiveness is conducted matters. Figure 3 does a great job of explaining the relationship between error and entropy across different levels of complexity and granularity. The methodology based on this observation is solid and interesting for communication. Empirically, the zone-based routing makes a lot of sense given the analysis, especially since we don't want to waste decomposition on simple tasks or do zero decomposition on super hard problems. This method gives a practical solution to a better tradeoff. The results are also quite encouraging. I like both Figure 3 and Figure 4; they are quite informative.

Critical weaknesses

The analysis shows that answer-token entropy correlates with direct-solving error at selected granularities. However, the routing decision requires evidence about the benefit of further decomposition relative to retaining the initial answer. Moreover, a reduction in error does not necessarily translate into a fully correct answer or improved end-to-end task success. It remains unclear how decomposition benefit is defined and evaluated: as partial error reduction, increased task success, or improvement relative to additional inference cost.

Questions for authors

How do we identify the subproblem size in practice, and understand its easy or hard label? I wonder how much of the gain comes from those, as opposed to the answer-token entropy?

Can you also demonstrate that entropy predicts this benefit, beyond simply predicting direct-solving error?

5 comments

Thank you for your careful review and your positive assessment of our work and figures. We address your concerns individually below.

[W1 & Q2: Definition and Evidence of Routing Benefit]

Reviewer comment (W1): The analysis shows that answer-token entropy correlates with direct-solving error at selected granularities. However, the routing decision requires evidence about the benefit of further decomposition relative to retaining the initial answer.

Reviewer comment (W1): It remains unclear how decomposition benefit is defined and evaluated: as partial error reduction, increased task success, or improvement relative to additional inference cost.

Reviewer comment (Q2): Can you also demonstrate that entropy predicts this benefit, beyond simply predicting direct-solving error?


Thank you for raising this point. The benefit of routing is to allocate additional computation to subproblems that benefit more from decomposition, achieving a larger reduction in final-answer error for a smaller increase in cost and thereby balancing solution quality and computational expense. We evaluate this benefit through the following paired experiment on the original problems.

For each original problem, we compare direct solving and further decomposition at the same Zone C node (the intermediate-difficulty region, where entropy determines whether to solve or decompose), then complete the reasoning for the original problem. Other settings remain unchanged. We report mean final-answer error and the total cost for the n problems in each group.

Reference for changes: Direct solving in the same row. Parentheses show the error and cost changes resulting from decomposition.

Task / ModelRoute decisionnDirect: errorDecompose: errorDirect: costDecompose: cost
Set intersection (N=64) / QwenDecompose321.750.97 (↓44.6%)1.592.01 (↑26.4%)
Set intersection (N=64) / QwenSolve681.561.24 (↓20.5%)3.165.26 (↑66.5%)
Set intersection (N=64) / LlamaDecompose542.280.96 (↓57.9%)3.895.11 (↑31.4%)
Set intersection (N=64) / LlamaSolve460.910.93 (↑2.2%)3.084.09 (↑32.8%)
Keyword counting / QwenDecompose512.941.49 (↓49.3%)2.353.08 (↑31.1%)
Keyword counting / QwenSolve492.371.84 (↓22.4%)2.133.78 (↑77.5%)
Keyword counting / LlamaDecompose502.040.92 (↓54.9%)4.235.35 (↑26.5%)
Keyword counting / LlamaSolve501.421.18 (↓16.9%)4.215.95 (↑41.3%)

When routing recommends decomposition, decomposing reduces final-answer error by 44.6%–57.9% at a 26.4%–31.4% increase in cost; the proportional error reduction consistently exceeds the cost increase. When routing recommends solving, forced decomposition reduces error by only 16.9%–22.4% in three settings and increases it by 2.2% in the fourth, while raising cost by 32.8%–77.5%. Thus, in all four settings, the groups selected for decomposition achieve larger error reductions at smaller cost increases. This shows that entropy predicts the benefit of the routing decision—decomposition versus direct solving (Q2).

Manuscript changes: Appendix C.9 adds the paired whole-problem experiment and analysis (p. 31, lines 1620–1652); Table 18 is at lines 1630–1643.


[W2: Evaluation Target and Metric Definition]

Reviewer comment (W2): Moreover, a reduction in error does not necessarily translate into a fully correct answer or improved end-to-end task success.


Thank you for raising this point. For set intersection, sorting, and keyword counting, our metric is the mean number of errors in the final answer to the original problem: errors are counted in each final answer and then averaged over all test instances, as defined in Appendix C.1.1. Unless explicitly identified as subproblem-level analyses, the reported error metrics concern the final answer to the original problem.

Manuscript location: The mean-error definitions are in Appendix C.1.1 (pp. 17–18, lines 898–924).


[Q1: Identifying Statistical Groups and the Roles of Grouping and Entropy]

Reviewer comment (Q1): How do we identify the subproblem size in practice, and understand its easy or hard label? I wonder how much of the gain comes from those, as opposed to the answer-token entropy?


Thank you for your question. For general tasks, the procedure is: sample subproblems across difficulty levels offline and record mean answer-token entropy and error from one solving attempt → sort probes by entropy and form groups → compute each group’s mean error and entropy–error correlation → use GMM+BIC to identify regions A (easy), B (hard), and C (intermediate, with entropy-threshold gating) and freeze the configuration → use one online entropy probe to assign the group and look up its region → reuse the probe answer for subsequent DART reasoning. For structured tasks, group membership can instead be determined directly from the input structure, such as the sizes of the two sets in set intersection.

We distinguish these roles through the following two comparisons. KC denotes keyword counting and SI denotes set intersection.

(a) Adaptive routing versus fixed decomposition granularity

After identifying subproblem size, we replace adaptive routing with fixed decomposition that stops at C1, C2, or A. These correspond to 8/4/2 sentences for KC and 128×32, 128×16, and 128×8 for SI, respectively.

Reference for changes: Full DART, with routing and self-critique enabled. Each variant’s error and cost changes are measured against this reference.

Task / ModelMetricDART (reference)Fixed C1Fixed C2Fixed A
Keyword counting / QwenError ↓1.802.77 (↑53.9%)1.89 (↑5.0%)1.54 (↓14.4%)
Cost ↓5.423.99 (↓26.4%)5.84 (↑7.7%)8.31 (↑53.3%)
Keyword counting / LlamaError ↓1.181.83 (↑55.1%)1.43 (↑21.2%)0.95 (↓19.5%)
Cost ↓9.658.13 (↓15.8%)10.63 (↑10.2%)14.93 (↑54.7%)
Set intersection (N=128) / QwenError ↓7.7512.57 (↑62.2%)8.80 (↑13.5%)6.74 (↓13.0%)
Cost ↓19.4013.38 (↓31.0%)19.90 (↑2.6%)28.78 (↑48.4%)
Set intersection (N=128) / LlamaError ↓6.439.37 (↑45.7%)7.81 (↑21.5%)5.61 (↓12.8%)
Cost ↓27.9121.87 (↓21.6%)28.38 (↑1.7%)42.59 (↑52.6%)

Knowing the size still leaves the decomposition granularity to be chosen. Among the tested fixed policies, finer decomposition reduces error but increases cost. Relative to DART, fixed C1 costs less but has higher error, while fixed C2 has both higher error and higher cost. Fixed A reduces error by 12.8%–19.5% but raises cost by 48.4%–54.7%. DART uses entropy to select subproblems worth decomposing further, balancing quality and cost.

(b) Entropy routing versus random routing with fixed region assignments

We retain the calibrated regions and replace only the Zone C decision with random decomposition at probability p.

Reference for changes: DART on the same task and model. Annotations show the error and cost changes under random routing.

Task / ModelMetricDART (reference)Random p=0.25Random p=0.50Random p=0.75
Keyword counting / QwenError ↓1.802.55 (↑41.7%)2.32 (↑28.9%)1.88 (↑4.4%)
Cost ↓5.425.21 (↓3.9%)5.61 (↑3.5%)7.00 (↑29.2%)
Keyword counting / LlamaError ↓1.181.81 (↑53.4%)1.48 (↑25.4%)0.94 (↓20.3%)
Cost ↓9.659.54 (↓1.1%)10.27 (↑6.4%)12.71 (↑31.7%)
Set intersection (N=64) / QwenError ↓1.301.61 (↑23.8%)1.69 (↑30.0%)1.31 (↑0.8%)
Cost ↓5.665.35 (↓5.5%)5.96 (↑5.3%)6.82 (↑20.5%)
Set intersection (N=64) / LlamaError ↓0.830.97 (↑16.9%)0.92 (↑10.8%)0.89 (↑7.2%)
Cost ↓8.858.81 (↓0.5%)9.24 (↑4.4%)9.94 (↑12.3%)

These results show that entropy routing achieves a more favorable error–cost trade-off.

Manuscript changes and locations: Sections 3.2 and 3.4 and Algorithm 1 clarify unified grouping and probe reuse (pp. 5–6, lines 238–294 and 318–324). Table 1 (p. 8, lines 379–390) and the fixed-policy analysis in Section 4.3 (p. 9, lines 452–468) have been updated. The existing random-routing evidence is in Section 4.4 (pp. 9–10, lines 479–499) and Table 4 (p. 10, lines 487–495).

We hope that our responses adequately address the reviewer's concerns. We would be glad to hear any further feedback.


AuthorsOpenReview ↗

A Friendly Reminder to Review Our Responses

Dear Reviewer 1tfx,

Thank you again for your careful evaluation and constructive feedback. To make the correspondence clear, we have organized our responses under [W1 & Q2], [W2], and [Q1], with supporting evidence for each concern.

In [W1 & Q2], we defined routing benefit and added paired experiments measuring how direct solving versus decomposition affects the original problem's final-answer error and cost. In [W2], we clarified the final-answer error metric. In [Q1], we explained group assignment and compared entropy routing with fixed and random policies. The supporting evidence is included directly in our responses, with corresponding manuscript locations indicated.

When convenient, we would be grateful if you could review our responses and let us know whether any concerns remain unclear or insufficiently addressed. If you find these clarifications satisfactory, we would sincerely appreciate your consideration of updating your assessment.

Thank you again for your time and thoughtful suggestions.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.