acceptodds
Under review as a conference paper at ICLR 2027

CROW: Constraint-Guided Agentic Workflows for Mathematical Reasoning

Abstract

Large language models have demonstrated strong performance on mathematical reasoning tasks, yet reasoning trajectories that violate critical conditions required for a valid solution can be difficult to correct. Existing approaches primarily focus on improving reasoning generation, but provide limited mechanisms for enforcing the constraints that should govern the reasoning process. We argue that critical reasoning constraints provide explicit conditions for guiding and verifying reasoning, particularly on challenging mathematical problems. We introduce Constraint Reuse and Orchestration Workflow (CROW), a constraint-guided agentic workflow for mathematical reasoning. CROW formulates constraint utilization as a sequential decision-making problem, allowing the orchestrator to selectively retrieve, inspect, reuse, disregard, or propose constraints according to the current problem. The orchestration policy is optimized with reinforcement learning based on downstream solution correctness. During training, newly proposed constraints that contribute to successful solutions are retained for future use, progressively expanding the set of available constraints. Through reinforcement learning (RL), the orchestrator learns when to acquire and apply the retrieved constraint, and when to disregard it due to limited relevance or applicability to the current problem. Experiments on HARDMath, HARDMath2, and DeepMath show that CROW substantially improves performance on HARDMath, where critical constraints contribute to the largest gains, while RL-trained orchestration preserves strong performance across the broader benchmark suite.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Reinforcement Learning, LLM Workflow, Mathematical Reasoning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_nJogOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CROW (Constraint Reuse and Orchestration Workflow), which treats critical reasoning constraints in mathematics as knowledge objects that can be retrieved, constructed, verified, and reused. Given a problem, Qwen3-Embedding-0.6B retrieves the top-1 related problem from constraint memory. A Qwen3-4B orchestrator decides whether to read its constraint, construct a new one, directly invoke a natural-language/program-based solver, invoke the corresponding reviser, or submit an answer. Constraints can guide solving, verification, and revision. Newly constructed constraints used in successful trajectories and contributing to correct answers are added to memory. The orchestrator is trained with GRPO using final-answer correctness as a sparse terminal reward, while Qwen3-Coder-30B-A3B-Instruct remains frozen as the task model.

The main contributions are: (1) making critical mathematical conditions explicit as reasoning knowledge transferable across problems; (2) unifying decisions about when to retrieve, activate, construct, or disregard constraints, which solving modality to use, and when to revise into a sequential decision process learned through RL; (3) accumulating successful experience so that discovered constraints can be reused on later problems with similar structures; and (4) evaluation on HARDMath, HARDMath2, and the Precalculus/Calculus/Algebra subsets of DeepMath-103K, with component ablations and cost estimates.

Critical strengths
  1. Treating constraints as selectable, verifiable, and accumulable objects. This targets recurring structural mathematical errors more directly than simply increasing CoT, voting, or the number of agents. The same constraint is shared across generation, verification, and revision, and the motivation aligns with the failure cases.
  2. Relatively comprehensive component ablations showing complementary effects. Using constraints or memory alone leads to substantial degradation, while the complete system reaches 61.41, supporting the combination of reliable knowledge and selective orchestration rather than simply stacking modules.
Critical weaknesses
  1. Insufficient controls to establish the method's contributions. In Table 2, adding memory improves HARDMath accuracy from 1.95% to 30.84%, but it remains unclear whether the gains come from knowledge retrieval or from constraint representation and dynamic orchestration. Direct retrieval and fixed-workflow controls using the same memory bank are needed.

  2. Unclear memory construction and reliability. The source and number of initial seeds and the final memory size are not adequately specified. A correct final answer does not imply that the constraint or its generalized version is correct. The authors should explain the checks performed before memory insertion and assess the impact of incorrect constraints.

  3. Limited evidence of generalization. The gains are mainly concentrated on HARDMath; HARDMath2 accuracy is 5.50%, below The Conductor's 10.00%. The authors should clarify overlap in problem types between training and testing and provide a split by problem type to assess whether the method relies on reusing highly similar problems.

  4. Comparison budgets are not adequately matched. CROW uses 1,748 training problems, while search-based methods use only 100 optimization problems, and inference step limits also differ. Comparisons under matched budgets are needed, together with total cost and latency accounting for orchestration, retrieval, and program execution.

Questions for authors
  1. With the same memory bank, how do direct retrieval or a fixed workflow perform compared with CROW?
  2. What are the seed sources and memory size? How are constraints validated before insertion, and how are incorrect constraints handled?
  3. How much do training and test problem types overlap? Do the gains remain under a split by problem type?
  4. Do the conclusions hold under matched budgets? Please report full costs and end-to-end latency.
5 comments
AuthorsOpenReview ↗

Response to Reviewer nJog (1/4)

W1. Insufficient controls to establish the method's contributions. In Table 2, adding memory improves HARDMath accuracy from 1.95% to 30.84%, but it remains unclear whether the gains come from knowledge retrieval or from constraint representation and dynamic orchestration. Direct retrieval and fixed-workflow controls using the same memory bank are needed.

Q1. With the same memory bank, how do direct retrieval or a fixed workflow perform compared with CROW?

We thank the reviewer for the detailed feedback and for raising these concerns. We address each point below, combining related weaknesses and questions where they can be answered by the same evidence.

In CROW, Qwen3-Embedding-0.6B always retrieves the top-1 memory entry. Our memory-without-RL ablation uses the same latest memory after CROW training, frozen during evaluation, but replaces the GRPO-trained orchestrator with the base Qwen3-4B orchestrator. It achieves 43.54% overall accuracy, compared with 61.41% for full CROW. Without learned orchestration, the model is less reliable at deciding whether the retrieved memory is applicable, which can lead it either to misuse an unsuitable constraint or to ignore a useful one. We further isolate the constraint representation using the 401 test queries for which the orchestrator chooses to use the constraint associated with the retrieved problem. For each query, we only replace this constraint with the same retrieved problem and its successful solution, all other settings remain unchanged. The results are:

DatasetConstraints (%)Demonstrations (%)
HARDMath46.2714.18
HARDMath29.767.32
Precalculus94.9286.44
Calculus77.0570.49
Algebra63.8257.45
Overall63.5948.38

Constraints outperform demonstrations on all five benchmarks, with an overall improvement of 15.21%. Analysis of cases where demonstrations fail but constraints succeed shows that demonstrations may lead the solver to reuse instance-specific assumptions or revert to its habitual reasoning, while executable programs may run successfully despite encoding an invalid formulation. Constraints instead make the required reasoning condition explicit and can directly guide generation, verification, and revision.

Together, these controls disentangle the contributions of constraint representation and learned orchestration, showing that retrieval alone does not explain CROW's gains.

To answer W1 and Q1, the same-memory ablation and matched representation control show that CROW benefits from both learned orchestration and constraint representation, rather than retrieval alone.

AuthorsOpenReview ↗

Response to Reviewer nJog (2/4)

W2. Unclear memory construction and reliability. The source and number of initial seeds and the final memory size are not adequately specified. A correct final answer does not imply that the constraint or its generalized version is correct. The authors should explain the checks performed before memory insertion and assess the impact of incorrect constraints.

Q2. What are the seed sources and memory size? How are constraints validated before insertion, and how are incorrect constraints handled?

CROW starts with 9 seed entries constructed exclusively from training data and expands to 21 entries by the end of RL training, with 12 additional entries acquired automatically during training. For seed construction, we sample training problems and first attempt to solve each problem without constraint guidance. If direct solving succeeds, no memory is created. Otherwise, Qwen3-Coder samples 10 candidate constraints, and a problem-constraint pair is retained only when constraint-guided solving succeeds. We apply the same criterion during RL training: a newly proposed constraint is added to memory only when direct solving fails but using the constraint leads to a correct solution. We agree that this does not formally prove that a stored constraint or its generalized form is universally correct. Instead, the procedure filters for constraints that provide additional reasoning value on the training problem. At test time, the constraint associated with the retrieved problem is not automatically applied, the orchestrator can reject it, construct a new constraint, or solve directly when it is not suitable for the current problem. We will clarify the initial and final memory sizes, the insertion criterion, and how non-transferable constraints are handled in the revision. To answer W2 and Q2, CROW uses training-only seed memories, grows the memory during RL training, and inserts a constraint only when it adds reasoning value beyond direct solving, non-transferable constraints can be rejected at test time.

AuthorsOpenReview ↗

Response to Reviewer nJog (3/4)

W3. Limited evidence of generalization. The gains are mainly concentrated on HARDMath; HARDMath2 accuracy is 5.50%, below The Conductor's 10.00%. The authors should clarify overlap in problem types between training and testing and provide a split by problem type to assess whether the method relies on reusing highly similar problems.

Q3. How much do training and test problem types overlap? Do the gains remain under a split by problem type?

For HARDMath, we use the official predefined train/test split without modifying the instances. HARDMath contains six problem categories, and all six appear in both training and test sets, the split is therefore by instance rather than by problem category. We will clarify this scope in the revision. A separate problem-type split is not provided by HARDMath and would require introducing a new taxonomy outside the standard benchmark protocol. The original seed memory contained no experience from Precalculus, Calculus, or Algebra, so the orchestrator saw few successful examples of constraint use in these domains and increasingly favored direct solving, which also limited subsequent memory expansion. For each of these three domains, we add one constraint to the initial memory. We iterate through its training problems and first attempt direct solving. If direct solving succeeds, we move to the next problem, otherwise, Qwen3-Coder samples 10 candidate constraints, and the corresponding problem-constraint pair is retained once constraint-guided solving succeeds. All other settings remain unchanged, and memory continues to expand dynamically during RL training.

MethodHARDMathHARDMath2PrecalculusCalculusAlgebraOverall
Conductor3.9010.0088.9382.6163.3855.38
CROW30.845.5092.5086.9675.3563.92

Against The Conductor, our strongest baseline under the same training split and data budget, CROW achieves higher accuracy on four of five benchmarks, with particularly clear gains on HARDMath and Algebra. This provides stronger evidence that CROW’s benefit generalizes beyond HARDMath, while HARDMath2 remains the one benchmark where The Conductor performs better. To answer W3 and Q3, we follow the official HARDMath split and the broader cross-domain results show that CROW improves beyond HARDMath.

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

This paper presents CROW, an agentic workflow that uses reusable constraints to support mathematical reasoning. For each problem, an orchestrator decides whether to use a constraint from memory, create a new one, or solve the problem without one. It can call a natural language solver or a program based solver, revise failed attempts, and learn these choices from whether the final answer is correct. Experiments with a fixed Qwen3-Coder-30B model show a large improvement on HARDMath.

Critical strengths
  1. Selective constraint use. The orchestrator decides when a retrieved constraint should guide solving instead of applying it to every problem.
  2. Strong HARDMath performance. CROW reaches 30.84% accuracy, compared with 4.22% for the best baseline.
  3. Moderate inference cost. CROW uses fewer backbone calls than most multi-agent workflow baselines while achieving the highest overall accuracy.
Critical weaknesses
  1. Unclear contribution of constraint memory. CROW may simply retrieve solution patterns from similar training problems rather than learn transferable constraints. The current ablations do not compare constraints with ordinary solutions or error summaries under the same retrieval, training, and orchestration setup, so the specific contribution of the proposed constraint representation remains unclear.
  2. Limited generalization. CROW's overall gain is driven mainly by HARDMath, while improvements on the other benchmarks are small or negative. This weakens the claim that the method provides broadly reusable reasoning constraints.
  3. Limited result reliability. Results come from single runs without uncertainty estimates, while the full CROW Precalculus score differs between Tables 1 and 2 (91.43% vs. 92.43%).
Questions for authors

Under the same retrieval, memory, and orchestration setup, does CROW still outperform a baseline that stores ordinary solutions or error summaries instead of constraints, especially when HARDMath is split by problem template? Can the strongest baselines be given the same training data and retrieval memory? Please also explain the 91.43% versus 92.43% Precalculus discrepancy and report results over multiple runs.

9 comments
AuthorsOpenReview ↗

Response to Reviewer xspr (1/4)

W1. Unclear contribution of constraint memory. CROW may simply retrieve solution patterns from similar training problems rather than learn transferable constraints. The current ablations do not compare constraints with ordinary solutions or error summaries under the same retrieval, training, and orchestration setup, so the specific contribution of the proposed constraint representation remains unclear.

Q1. Under the same retrieval, memory, and orchestration setup, does CROW still outperform a baseline that stores ordinary solutions or error summaries instead of constraints, especially when HARDMath is split by problem template?

We thank the reviewer for the detailed feedback and for raising these concerns. We address each point below, combining related weaknesses and questions where they share the same evidence.

To isolate the contribution of constraint memory from the effects of retrieving similar training problems, we conduct a controlled comparison between constraints and solution demonstrations while keeping the retrieval results and orchestration decisions fixed. We consider the 401 test queries for which the orchestrator chooses to use the constraint associated with the retrieved problem. For each query, we replace only that constraint with the same retrieved problem and its successful solution, all other settings remain unchanged. The results are:

DatasetConstraints (%)Demonstrations (%)
HARDMath46.2714.18
HARDMath29.767.32
Precalculus94.9286.44
Calculus77.0570.49
Algebra63.8257.45
Overall63.5948.38

Constraints outperform demonstrations on all five benchmarks, with an overall improvement of 15.21%. We further analyze cases where demonstrations fail but constraints succeed. In natural-language reasoning, the solver may revert to its habitual reasoning or transfer assumptions that do not apply. In program-based reasoning, code may execute successfully while encoding an invalid assumption or solving a different formulation. A constraint instead makes the required condition explicit and can directly guide generation, verification, and revision. For HARDMath, we use the predefined train/test split provided by the dataset authors without modifying the instances. The split is by instance rather than category: all six HARDMath categories appear in both training and test sets. We will clarify that the official HARDMath split does not evaluate generalization to entirely unseen problem categories. To answer W1 and Q1, the controlled comparison shows that constraints outperform ordinary problem-solution demonstrations under the same retrieval and orchestration setup.

AuthorsOpenReview ↗

Response to Reviewer xspr (2/4)

W2. Limited generalization. CROW's overall gain is driven mainly by HARDMath, while improvements on the other benchmarks are small or negative. This weakens the claim that the method provides broadly reusable reasoning constraints.

The initial memory contained no seed experience from Precalculus, Calculus, or Algebra, so the orchestrator observed few successful examples of constraint use in these domains and increasingly favored direct solving. For each of these three domains, we add one constraint to the initial memory. We iterate through training problems and first attempt direct solving, if it fails, Qwen3-Coder samples 10 candidate constraints, and the first problem-constraint pair that enables a correct solution is retained. All other settings remain unchanged, and memory continues to expand dynamically during RL training. The results against our strongest baseline, The Conductor, are:

MethodHARDMathHARDMath2PrecalculusCalculusAlgebraOverall
Conductor3.9010.0088.9382.6163.3855.38
CROW30.845.5092.5086.9675.3563.92

CROW outperforms The Conductor on four of five benchmarks, with particularly large gains on HARDMath and Algebra, while also improving Precalculus and Calculus. This provides stronger evidence that the benefit of CROW generalizes beyond HARDMath. To answer W2, the broader-seed experiment shows that CROW outperforms The Conductor on four of five benchmarks, with clear gains outside HARDMath.

AuthorsOpenReview ↗

Response to Reviewer xspr (3/4)

Can the strongest baselines be given the same training data and retrieval memory?

The strongest baseline, The Conductor is memory-free, so adding persistent retrieval would change the method itself. For search-based baselines, we use an optimization budget of approximately 2 days. Because workflow search repeatedly evaluates candidate workflows, these methods use a shared 100-example optimization set. To match data exposure, we additionally train CROW on the same 100 examples for 10 epochs, requiring approximately 1.5 days:

MethodHARDMathHARDMath2PrecalculusCalculusAlgebraOverall
AFlow4.225.5089.1177.7263.3854.23
ADAS2.605.5083.2176.0954.9350.43
MASS3.575.5090.1877.7261.9754.38
EvoMAS1.954.5073.3957.0764.0844.62
CROW29.225.5091.9683.1569.0162.20

With the same 100-example budget, CROW outperforms the strongest search-based baseline, MASS, by 7.82%. This demonstrates strong data efficiency under matched training exposure. To answer Q2, The Conductor already uses the same training split and data budget as CROW, while the matched 100-example experiment shows strong data efficiency for the search-based comparison and W1 isolates retrieval and memory representation directly.

Reviewer_kY3AOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an agentic workflow that uses a memory of mathematical constraints to improve mathematical reasoning, with particularly strong results on HARDMath. The authors use GRPO to train a small orchestrator to decide whether to access stored constraints, generate new constraints, use natural-language or program-based solving, and revise previous attempts. The ablation experiments show that constraints, memory, and learned orchestration work best together, while adding constraints alone may reduce accuracy.

Critical strengths
  • Learning to decide whether a mathematical constraint is relevant and worth using addresses a concrete problem with practical value.
  • The experiments are thorough, cover a comprehensive set of baselines, and provide convincing results.
  • The method shows clear improvements on benchmarks such as HARDMath.
  • The method has novelty.
Critical weaknesses
  • The ablations show that the full combination of components is effective, but they do not directly establish why the gains require mathematical constraints as the memory representation. Ordinary problem-solving experience might work just as well with the same orchestration setup. This makes the novelty of the constraint-based representation less convincing.
  • The gains are mainly concentrated on HARDMath, while the method does not consistently outperform the strongest baseline on the other test sets. These results do not provide enough evidence to conclude that the method works broadly and reliably.
Questions for authors

None

5 comments
AuthorsOpenReview ↗

Response to Reviewer kY3A (1/2)

W1. The ablations show that the full combination of components is effective, but they do not directly establish why the gains require mathematical constraints as the memory representation. Ordinary problem-solving experience might work just as well with the same orchestration setup. This makes the novelty of the constraint-based representation less convincing.

We thank the reviewer for recognizing the novelty of CROW, the practical value of learning when constraint knowledge should be used, and the strength of our experimental evaluation. We address the two concerns below.

We consider the 401 test queries for which the trained orchestrator chooses to use the constraint associated with the retrieved problem. For each query, we only replace that constraint with the corresponding retrieved example problem and solution, all other settings remain unchanged. The results are:

DatasetConstraints (%)Demonstrations (%)
HARDMath46.2714.18
HARDMath29.767.32
Precalculus94.9286.44
Calculus77.0570.49
Algebra63.8257.45
Overall63.5948.38

Constraints outperform demonstrations on all five benchmarks, with an overall improvement of 15.21%. We further analyze cases where demonstrations fail but constraints succeed. In natural-language reasoning, the solver may recognize the analogous example but still revert to its habitual reasoning pattern or transfer assumptions that do not apply to the current problem. In program-based reasoning, generated code may execute successfully while encoding an invalid assumption or solving a different formulation. The key distinction is that a demonstration provides an instance-specific solution and leaves the solver to infer what should transfer, whereas a constraint explicitly isolates the condition required for correctness. This allows the same condition to directly guide generation, verification, and revision. To answer W1, the controlled comparison shows that constraints provide a consistent advantage over problem-solution demonstrations under the same retrieval and orchestration setup.

AuthorsOpenReview ↗

Response to Reviewer kY3A (2/2)

W2. The gains are mainly concentrated on HARDMath, while the method does not consistently outperform the strongest baseline on the other test sets. These results do not provide enough evidence to conclude that the method works broadly and reliably.

The initial memory contained no seed experience from Precalculus, Calculus, or Algebra, so the orchestrator observed few successful examples of constraint use in these domains and increasingly favored direct solving. For each of these three domains, we add one constraint to the initial memory. We iterate through its training problems and first attempt direct solving. If direct solving succeeds, we move to the next problem. Otherwise, Qwen3-Coder samples 10 candidate constraints, and the corresponding problem-constraint pair is retained once constraint-guided solving succeeds. All other settings remain unchanged, and the memory continues to expand dynamically during RL training. For a single-run comparison with our strongest baseline, The Conductor:

MethodHARDMathHARDMath2PrecalculusCalculusAlgebraOverall
Conductor3.9010.0088.9382.6163.3855.38
CROW30.845.5092.5086.9675.3563.92

CROW achieves higher accuracy on four of five benchmarks, with particularly large gains on HARDMath and Algebra, while also improving Precalculus and Calculus. This provides stronger evidence that the benefit of CROW generalizes beyond HARDMath. We also repeated held-out evaluation three times because Qwen3-Coder-30B shows slight output variation even with temperature 0:

MethodHARDMathHARDMath2PrecalculusCalculusAlgebraOverall
Conductor4.00 ± 0.5010.17 ± 0.7689.23 ± 1.3682.79 ± 2.9964.32 ± 0.8155.67 ± 0.29
CROW28.57 ± 2.035.17 ± 1.0492.32 ± 2.0687.14 ± 3.5475.12 ± 1.7763.30 ± 0.54

CROW maintains a 7.63% overall advantage and higher mean accuracy on four of five benchmarks, with particularly clear gains on HARDMath and Algebra. To answer W2, the broader-seed and repeated-evaluation results show that CROW outperforms our strongest baseline on four of five benchmarks, with clear gains beyond HARDMath.

AuthorsOpenReview ↗

Additional Updates

To further address your concern in W1, we trained a Qwen3-4B orchestrator using the same CROW training pipeline, with retrieved problem-solution demonstrations instead of constraints. Both experiments use the same training data, retrieval procedure, same training settings, and evaluate on full test set. The result is as follows:

MethodHARDMathHARDMath2PrecalculusCalculusAlgebraOverall
Demonstration-based orchestrator6.825.0086.6172.2860.5652.73
CROW30.845.5092.5086.9675.3563.92

CROW remains superior in overall accuracy by 11.19% and performs better on all five benchmarks under matched training conditions.

If these new results and our previous clarifications address your concerns, we would sincerely request you to adjust the score at your convenience. Thank you.

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

The paper proposes CROW, a workflow for math reasoning that stores useful “constraints” from solved problems and reuses them later. A small orchestrator, trained with RL, decides at each step whether to use a retrieved constraint, write a new one, solve directly with text or code, or revise. A frozen larger model does the actual solving. CROW gets the best overall results on HARDMath, HARDMath2, and three DeepMath subsets, and most of that gain comes from HARDMath. The ablations suggest the constraints, memory, and RL work together.

Critical strengths

S1. Retrieving a constraint is kept separate from deciding to use it, and the ablation supports this design. Constraints used without memory or RL hurt accuracy (29.41% vs. 51.87%), and learned selective use recovers performance. This is a useful finding.

S2. The HARDMath gain is large (≤4.22% → 30.84%). It suggests that reusing structural knowledge helps on applied-math problems where models repeat the same mistakes.

Critical weaknesses

W1. Almost all of the gain comes from HARDMath. On the other benchmarks, CROW is behind The Conductor on HARDMath2 and Calculus, and only slightly ahead on Precalculus and Algebra. None of the baselines use retrieval, so the improvement may come from retrieving similar training problems rather than from the constraints themselves.

W2. Constraints without memory hurt performance (29.41 vs. 51.87), and the policy uses them less over training. Memory seems to be doing most of the work. The paper also doesn’t say where the seed constraints come from.

W3. Search-based baselines are optimized on 100 problems, while CROW trains on 1,748. Results are from a single run with no variance.

W4. Some details are missing or inconsistent. The RL algorithm isn’t given, the no-RL memory setup is unclear, Precalculus is 91.43 in Table 1 but 92.43 in Table 2, the ablation text credits “revision” even though Table 2 has no revision column, and Self-Consistency is listed as a baseline but missing from Table 1.

Questions for authors

Q1. How does CROW compare to a simple baseline that retrieves the same top-1 training problem and its solution as a one-shot example?

Q2. Was the HARDMath train/test split done by template or by instance? How similar are test problems to the problems retrieved for them?

Q3. How is memory built in the no-RL ablation? Is memory frozen at test time?

Q4. How often are constraints actually used at test time on each benchmark?

Q5. Which RL algorithm did you use, and do the results hold across multiple seeds?

9 comments
AuthorsOpenReview ↗

Response to Reviewer zGec (1/6)

W1. Almost all of the gain comes from HARDMath. On the other benchmarks, CROW is behind The Conductor on HARDMath2 and Calculus, and only slightly ahead on Precalculus and Algebra. None of the baselines use retrieval, so the improvement may come from retrieving similar training problems rather than from the constraints themselves.

Q1. How does CROW compare to a simple baseline that retrieves the same top-1 training problem and its solution as a one-shot example?

We thank the reviewer for the constructive feedback and for raising these concerns. We address each point below, combining related weaknesses and questions where they can be answered by the same evidence.

In the original submission, the seed memory contained no initial experience from Precalculus, Calculus, or Algebra. Consequently, the orchestrator observed few successful examples of constraint utilization in these domains and increasingly favored direct solving, which also limited subsequent memory expansion. For each of these three domains, we add one constraint to the initial memory. We iterate through training problems and first attempt direct solving. If direct solving succeeds, we move to the next problem. Otherwise, Qwen3-Coder samples 10 candidate constraints, and the corresponding problem-constraint pair is retained once constraint-guided solving succeeds. All other settings remain unchanged, and memory continues to expand dynamically during RL training. For a single-run comparison under the same evaluation protocol with our strongest baseline, The Conductor:

MethodHARDMathHARDMath2PrecalculusCalculusAlgebraOverall
The Conductor3.9010.0088.9382.6163.3855.38
CROW30.845.5092.5086.9675.3563.92

CROW outperforms The Conductor on four of five benchmarks and improves overall accuracy by 8.54%, with particularly large gains on HARDMath and Algebra. These results show that, with broader seed coverage, CROW’s gains are not concentrated on HARDMath. We further compare constraints with problem-solution demonstrations under the same retrieval and orchestration setup. We consider the 401 test queries for which the trained orchestrator chooses to use the constraint associated with the retrieved problem. For each query, we only replace that constraint with the corresponding retrieved problem and successful solution, all other settings remain unchanged. The resulting accuracies are:

DatasetConstraints (%)Demonstrations (%)
HARDMath46.2714.18
HARDMath29.767.32
Precalculus94.9286.44
Calculus77.0570.49
Algebra63.8257.45
Overall63.5948.38

Across the same 401 queries, constraint guidance improves accuracy by 15.21% over demonstrations and performs better on all five benchmarks. We further analyze cases where demonstrations fail but constraints succeed. In these cases, the solver may revert to its habitual reasoning pattern or transfer assumptions that do not apply to the current problem, while the constraint isolates the critical condition required for correctness and can guide generation, verification, and revision. To answer W1 and Q1, the broader-seed and matched demonstration controls show that CROW generalizes beyond HARDMath and that constraint representation provides a benefit beyond retrieving the same solved problem.

AuthorsOpenReview ↗

Response to Reviewer zGec (2/6)

W2. Constraints without memory hurt performance (29.41 vs. 51.87), and the policy uses them less over training. Memory seems to be doing most of the work. The paper also doesn’t say where the seed constraints come from.

Constraints generated directly for a new query may be unreliable, while memory retains constraints that have previously contributed to successful solutions. However, a constraint associated with the retrieved problem may still be unsuitable for a new problem, so RL learns whether it should be used. This complementarity is reflected in the ablation: constraints without memory or RL achieve 29.41% overall accuracy, memory without RL reaches 43.54%, and the full system reaches 61.41%. For the construction of seed constraints, we randomly sample a small subset of training problems as candidates and filter out those that can already be solved directly. For each remaining problem, Qwen3-Coder samples 10 candidate constraints, and a problem-constraint pair is retained only if constraint-guided solving succeeds. To answer W2, memory is important, but it is complementary to constraint construction and RL rather than replacing them.

AuthorsOpenReview ↗

Response to Reviewer zGec (3/6)

W3. Search-based baselines are optimized on 100 problems, while CROW trains on 1,748. Results are from a single run with no variance.

We thank the reviewer for raising the concern about unequal training-data exposure between CROW and the search-based baselines.

For search-based methods, we use a common optimization budget of approximately 2 days. Since workflow search repeatedly evaluates candidate workflows, these methods use a shared 100-example optimization set. To control for data exposure, we additionally train CROW on the same 100 examples for 10 epochs, requiring approximately 1.5 days:

MethodHARDMathHARDMath2PrecalculusCalculusAlgebraOverall
AFlow4.225.5089.1177.7263.3854.23
ADAS2.605.5083.2176.0954.9350.43
MASS3.575.5090.1877.7261.9754.38
EvoMAS1.954.5073.3957.0764.0844.62
CROW29.225.5091.9683.1569.0162.20

With the same 100-example budget, CROW still substantially outperforms all search-based baselines, demonstrating strong data efficiency under matched training exposure. To answer W3, CROW remains substantially stronger than the search-based baselines under matched training exposure, demonstrating strong data efficiency.

Comments on the paper

1 comment

Dear Reviewers kY3A, nJog, and zGec,

As we are approaching the end of the discussion period, could you read the authors' rebuttal and see if it has addressed your concerns and questions? Thanks again for your time!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.