acceptodds
Under review as a conference paper at ICLR 2027

I-Flow: Intent-Revealing Flow Optimization for Multi-Turn LLM Reasoning

Abstract

Multi-turn reasoning enables large language models (LLMs) to tackle complex problems through structured reasoning across turns. However, we uncover a critical failure mode of final-reward reinforcement learning: optimizing only for final correctness can collapse the intended hierarchy and drive diverse reasoning toward short and repetitive patterns. To address this limitation, we propose I-Flow, which structures each reasoning turn around a turn-level *intent* and introduces a Generative Flow Network (GFlowNet) optimization for the resulting intent hierarchy. Our bootstrap-free subtrajectory objective propagates terminal supervision across intent sequences, encouraging diverse high-reward reasoning structures without step-wise bootstrapping. We further replay successful intent structures to preserve effective reasoning patterns throughout training. Experiments on mathematical reasoning and code generation demonstrate consistent performance gains while effectively mitigating hierarchy collapse and reasoning simplification.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Multi-turn Reasoning, Generative Flow Networks

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_8MgPOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes I-Flow, an intent-revealing flow optimization framework for multi-turn LLM reasoning, aiming to address hierarchy collapse and reasoning simplification under terminal-reward reinforcement learning. The method introduces turn-level intents, a Bootstrap-Free Subtrajectory Balance objective, and Intent Structure Replay to preserve effective reasoning patterns. Experiments on mathematical reasoning and code generation benchmarks demonstrate consistent improvements over existing methods.

Critical strengths
  1. This paper identifies two important failure modes in multi-round LLM reasoning: hierarchy collapse and reasoning simplification. The formulation of these problems provides a compelling motivation for the proposed framework.
  2. This paper establishes a connection between intent-conditioned continuous flows and KL-regularized optimal policies. The theoretical analysis interprets intermediate flow values in terms of downstream reasoning success and derives bounds relating local flow-balance errors to the proposed training objective, providing useful theoretical justification for the method.
  3. Although I-Flow is trained on mathematical reasoning data, it also demonstrates improvements on code generation benchmarks. The learned intent-based reasoning behavior is not limited to mathematical problem-solving and may transfer to other tasks requiring structured reasoning and iterative refinement.
Critical weaknesses
  1. BF SubTB constructs multiple subtrajectory constraints anchored to the same terminal reward. Although this avoids bootstrapping from learned intermediate endpoint flows, the supervision remains highly correlated. The method does not identify which intermediate decisions contribute to success or failure, leaving credit assignment for individual intents unclear.
  2. The proposed framework relies on six predefined intent categories to represent high-level reasoning decisions. However, complex reasoning processes may involve intents that cannot be adequately captured by this fixed taxonomy, raising concerns about the adaptability of intent representations, particularly for fine-grained tasks.
Questions for authors
  1. How does I-Flow ensure that the generated high-level guidance and low-level reasoning remain semantically consistent with the selected intent?
  2. All BF SubTB residuals are anchored to the same terminal reward. How does the method distinguish intermediate intents that contribute to successful reasoning from ineffective ones? How does the proposed objective achieve more informative credit assignment than conventional trajectory-level supervision?
  3. Since I-Flow incurs a higher overall training cost than ReMA, it remains unclear whether its performance improvements primarily stem from the proposed optimization mechanism or the additional computational resources.
4 comments

We sincerely thank the reviewer for their detailed comments and questions about I-Flow’s learning mechanism. In particular, we clarify how learned flows support intent-level credit assignment and how the intent taxonomy can be adapted, and provide semantic-consistency evaluations and extended-training comparisons to further examine the effectiveness of the proposed approach.


W1/Q2 (Intent-Level Credit Assignment):

We thank the reviewer for this important question. I-Flow distinguishes intents not by the terminal reward alone, but by the flow Fϕ(st)F_\phi(s_t) learned at every post-intent state, which serves as an intermediate signal, much like learning a value function in RL instead of relying on the final return. This follows from Eq. (13), πI∗(ut∣ct)∝πrefI(ut∣ct) F∗(ct,ut)\pi^{I\ast}(u_t \mid c_t) \propto \pi^{I}_{\mathrm{ref}}(u_t \mid c_t) \thinspace F^{\ast}(c_t,u_t): since the optimal intent policy is determined by the intent-conditioned flow, an intent with higher flow at a given state is a better one, and an intent with lower flow is a worse one. Proposition 1 guarantees that the BF-SubTB conditions identify Fϕ=F∗F_\phi = F^* at every post-intent state, so the learned flow serves as this criterion.

In contrast, the terminal-reward-only TB variant also evaluated in Section 5.4 imposes only a single constraint from the initial state to the terminal state, so no intermediate flow is learned and no such criterion is available. While TB constrains only the sum of local balance residuals, where positive and negative errors can cancel, I-Flow bounds their sum of squares by its training loss up to a constant factor (Proposition 2), thereby providing finer intermediate supervision with the same terminal reward. Empirically, Figure. 9 in Appendix D.4 shows that, from the same reasoning prefix, selecting the highest-flow intent yields a 64.2% success rate, compared with 55.8% for random and 46.1% for lowest-flow selection, confirming that the flow indeed distinguishes intents that contribute to success. We thank the reviewer again for this insightful question.


W2 (Adaptability of the Intent Taxonomy):

We thank the reviewer for raising this concern. Our taxonomy consists of five core reasoning intents (EXP, VER, BCK, CON, and FIN) reflecting reasoning operations identified in prior work on hierarchical reasoning for mathematical and coding tasks [R.1, R.2]. Alongside these five core intents, the taxonomy already includes IDL for general progress, accommodating steps outside them. For different domains, categories can be added, merged, or refined without modifying I-Flow’s optimization objective, with corresponding updates to SFT intent annotations. For example, in document-based question answering, behaviors assigned to IDL could be refined into SELECT and SYNTHESIZE for evidence identification and integration, respectively. Thus, the intent taxonomy can be adapted to different reasoning requirements while preserving I-Flow's core optimization framework. We have clarified the taxonomy's rationale in Section 5.1 and expanded Appendix A with this adaptation example and a discussion of its scope.

[R.1] Wan, Ziyu, et al. “ReMA: Learning to Meta-Think for LLMs with Multi-agent Reinforcement Learning.” NeurIPS, 2025.

[R.2] Shahariar, G. M., et al. “Modeling Hierarchical Thinking in Large Reasoning Models.” ICML, 2026.

Q1 (Semantic Consistency):

We thank the reviewer for asking whether the selected intent, high-level guidance, and low-level reasoning remain semantically consistent. We directly assessed semantic consistency on I-Flow’s MATH500 rollouts and added the analysis to RQ2 in Section 5.3, with judging prompts in Appendix C.5. GPT-4o-mini evaluates whether (1) the high-level guidance matches the selected intent and (2) the low-level reasoning follows that guidance. As shown in Table R.4 below, Both consistency rates exceed 92% across all six intents, providing empirical evidence that the intended relationships among intent, guidance, and reasoning are maintained. This likely reflects how BF-SubTB supports exploration of diverse successful intent structures and ISR re-executes those structures with freshly generated guidance and reasoning. Repeated learning from successful trajectories under these structures may help maintain semantic alignment among the selected intent, high-level guidance, and low-level reasoning.

Table R.4. Semantic consistency (%) of I-Flow by intent type, reproduced from Table 5 in Section 5.3.

Intent(1) Intent–Guidance(2) Guidance–Reasoning
EXP92.2496.75
VER98.5594.24
BCK95.8092.43
CON94.8893.24
FIN96.0993.16
IDL93.1395.99

Q3 (Explaining Performance Gains):

We thank the reviewer for asking whether I-Flow’s performance gains can be explained by additional training compute alone. To examine this, we compare I-Flow and ReMA under approximately matched total training times in Appendix E.1 and analyze the corresponding trajectory lengths and hierarchy-collapse rates in Section 5.4. Under identical maximum token and turn limits, I-Flow takes approximately 54% longer per training step than ReMA, while flow-network training itself accounts for only about 0.11%. The additional time primarily reflects generating and optimizing longer trajectories that sustain hierarchical multi-turn reasoning, whereas ReMA’s trajectories shorten alongside hierarchy collapse and reasoning simplification.

To assess whether additional training can compensate for this difference, we extended ReMA from 250 to 400 steps, approximately matching I-Flow’s total training time at 250 steps. Even with this extended training, ReMA’s turn and token counts continued to decrease, and its hierarchy collapse rate remained around 60%, compared with 3.21% for I-Flow (Section 5.4). I-Flow also achieved higher scores than ReMA’s 300- and 400-step checkpoints on all seven mathematical reasoning benchmarks (Appendix E.1). These results suggest that additional training alone has limited effects on the observed hierarchy collapse and reasoning simplification. Importantly, preserving distinct high-level guidance and low-level execution enables verification, revision, and adaptive reasoning across turns, which are central benefits of hierarchical multi-turn reasoning. Thus, I-Flow’s additional computational cost largely reflects its ability to sustain these capabilities throughout training.


The theoretical clarification and empirical analyses support the usefulness of learned flows as intent-level learning signals and show high semantic consistency between selected intents, guidance, and execution in the evaluated outputs. We sincerely thank the reviewer for their valuable feedback, which helped us clarify the learning mechanism and strengthen the evidence for its effectiveness.

Thank you for the detailed response and additional experimental analyses. The authors' clarifications have addressed most of my concerns. I will maintain my original score of 3 (Weak Acceptance).

Thank you for your prompt response and for taking the time to consider our rebuttal. We appreciate your thoughtful feedback.

Authors

Reviewer_r6ZgOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes I-Flow, a framework for improving multi-turn LLM reasoning through explicit intent modeling and flow-based optimization. The method introduces turn-level reasoning intents, uses a GFlowNet-based objective to optimize reasoning trajectories, and reuses successful intent sequences during training. Experiments on mathematical reasoning and code generation show consistent improvements over existing baselines.

Critical strengths
  1. The proposed flow-based optimization provides a meaningful technical contribution by introducing intermediate supervision while preserving terminal reward signals.

  2. The experimental results are strong, with consistent improvements across multiple benchmarks and ablation studies supporting the proposed components.

Critical weaknesses
  1. The method relies on a predefined set of reasoning intents, which may not generalize well to tasks requiring different reasoning strategies.

  2. The evaluation focuses mainly on mathematical reasoning, with code generation used for transfer evaluation. It remains unclear how well the approach works on more open-ended reasoning tasks.

  3. The proposed method introduces additional training complexity and computational cost. Its practical efficiency compared with simpler multi-turn reasoning methods is not fully established.

Questions for authors

How does the additional training cost compare with the performance improvements achieved by I-Flow?

How did you come up with the reasoning intent types? How do you justify them?

Overall I think this is a solid work, but reasoning tuning has been fully studied, how do you distinguish your work from the mainstream? Does pre-defined types really bring difference?

2 comments

We sincerely thank the reviewer for their insightful comments on the applicability and practical value of I-Flow. In response, we provide additional evaluations on open-ended tasks and extended-training comparisons, and clarify the rationale for the intent taxonomy, its adaptability, and I-Flow’s distinction from existing reasoning tuning.


W1/Q2 (Selection and Adaptability of the Intent Taxonomy):

We thank the reviewer for raising this concern. Our taxonomy consists of five core reasoning intents EXP, VER, BCK, CON, and FIN reflecting reasoning operations identified in prior work on hierarchical reasoning for mathematical and coding tasks [R.1, R.2]. Alongside these five core intents, the taxonomy already includes IDL for general progress, accommodating steps outside them. For different domains, categories can be added, merged, or refined without modifying I-Flow’s optimization objective, with corresponding updates to SFT intent annotations. For example, in document-based question answering, behaviors assigned to IDL could be refined into SELECT and SYNTHESIZE for evidence identification and integration, respectively. We have clarified the taxonomy's rationale in Section 5.1 and expanded Appendix A with this adaptation example and a discussion of its scope.

[R.1] Wan, Ziyu, et al. “ReMA: Learning to Meta-Think for LLMs with Multi-agent Reinforcement Learning.” NeurIPS, 2025.

[R.2] Shahariar, G. M., et al. “Modeling Hierarchical Thinking in Large Reasoning Models.” ICML, 2026.


W2 (Generalization to Open-ended Reasoning Tasks):

We thank the reviewer for highlighting the need to evaluate generalization to open-ended tasks. Following the reviewer’s suggestion, we added an evaluation of the mathematics-trained models on WildBench’s 1,024 open-ended tasks using GPT-4o as the evaluator in Appendix D.3, and discussed potential extensions to open-ended tasks in Appendix A. As in code generation, we directly transfer the mathematics-trained models without task-specific SFT data or reward models. As shown in Table R.3 below, I-Flow achieves the highest score (−1.11), outperforming Qwen2.5-7B-Intent (−17.75), ReMA (−9.67), and Dr.MAS (−11.64). In the examined examples, I-Flow preserves distinct hierarchical roles and performs intent-guided reasoning and verification, whereas the baselines exhibit hierarchy collapse and reasoning simplification. These results demonstrate I-Flow's potential for generalization to open-ended reasoning tasks without additional training or taxonomy adaptation.

Table R.3. Results on WildBench, reproduced from Table D.3 in Appendix D.3.

MethodWildBench Score
Qwen2.5-7B-Intent−17.75
ReMA−9.67
Dr.MAS−11.64
I-Flow (Ours)−1.11

W3/Q1 (Training Cost and Practical Efficiency):

We thank the reviewer for asking whether I-Flow’s performance gains can be explained by additional training compute alone. To examine this, we compare I-Flow and ReMA under approximately matched total training times in Appendix E.1 and analyze the corresponding trajectory lengths and hierarchy-collapse rates in Section 5.4. Under identical maximum token and turn limits, I-Flow takes approximately 54% longer per training step than ReMA, while flow-network training itself accounts for only about 0.11%. The additional time primarily reflects generating and optimizing longer trajectories that sustain hierarchical multi-turn reasoning, whereas ReMA’s trajectories shorten alongside hierarchy collapse and reasoning simplification.

To assess whether additional training can compensate for this difference, we extended ReMA from 250 to 400 steps, approximately matching I-Flow’s total training time at 250 steps. Even with this extended training, ReMA’s turn and token counts continued to decrease, and its hierarchy collapse rate remained around 60%, compared with 3.21% for I-Flow (Section 5.4). I-Flow also achieved higher scores than ReMA’s 300- and 400-step checkpoints on all seven mathematical reasoning benchmarks (Appendix E.1). These results suggest that additional training alone has limited effects on the observed hierarchy collapse and reasoning simplification. Importantly, preserving distinct high-level guidance and low-level execution enables verification, revision, and adaptive reasoning across turns, which are central benefits of hierarchical multi-turn reasoning. Thus, I-Flow’s additional computational cost largely reflects its ability to sustain these capabilities throughout training.


Q3 (Distinction from Reasoning Tuning):

We thank the reviewer for this question. We agree that reasoning tuning has been extensively studied. However, our focus is on preserving the intended separation between high-level guidance and low-level execution during multi-turn optimization, which enables meaningful verification and revision across turns. Unlike conventional outcome-based reasoning tuning, which primarily optimizes final correctness, I-Flow explicitly represents high-level decisions through turn-level intents, making reasoning structures identifiable and learnable. These intents are not merely predefined labels but serve as intermediate decision points for flow-based supervision.

Specifically, GFlowNet-based optimization learns intent-conditioned flows that distinguish promising reasoning continuations, while BF-SubTB anchors the intermediate constraints to verified terminal outcomes. Intent Structure Replay further preserves successful reasoning structures throughout training. Regarding whether predefined intent types make a difference, removing explicit intents consistently degrades performance (Table 6), showing that explicit intents themselves help. At the same time, adding the same intents to the ReMA objective still falls well short of I-Flow (GRPO+Intent, Table E.4), indicating that intent labels alone are not sufficient without our flow-based optimization. The ablation results in Section 5.4 show that removing explicit intents degrades performance, supporting their contribution beyond predefined categorization. Moreover, the case studies in Appendix D.5 illustrate how I-Flow maintains distinct guidance and execution through verification and revision. Thus, I-Flow differs from conventional reasoning tuning by explicitly learning and preserving effective hierarchical reasoning structures, rather than optimizing final correctness alone. We have clarified this distinction in the Conclusion and provided supporting analyses in Section 5.4 and Appendix D.5.


Our revisions clarify the motivation and scope of I-Flow and extend its evaluation to broader reasoning tasks. We are grateful for the reviewer’s insightful comments, which helped us better communicate the method’s practical value and distinction from existing approaches.

Reviewer_xK9TOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

I-Flow introduces explicit intents for each reasoning turn and combines a flow-based training objective with replay of successful intent sequences. The paper reports improvements on math and code benchmarks while preserving more varied reasoning structures and a clearer separation between planning and execution.

Critical strengths

The paper tackles a clear problem: during training, the high-level controller can stop providing useful guidance, and reasoning patterns become repetitive.

The method is well motivated, and the theoretical analysis helps explain what the intermediate flow estimates represent.

The experiments cover several benchmarks and include useful ablations, another model backbone, and transfer from math to code.

Critical weaknesses

W1 It is hard to tell how much of the improvement comes from additional computation. I-Flow takes about 54% longer per training step than ReMA, so comparisons with equal training and inference budgets would strengthen the results.

W2 The structure-transfer experiment is promising, but the comparison favors structures taken from successful I-Flow runs over structures from failed SFT runs. Comparing against fresh SFT samples and alternative intent sequences of the same length would make the evidence more convincing.

Questions for authors

reply to weaknesses

Are the results consistent across training seeds, and do human judgments agree with the automatic hierarchy-collapse evaluation?

the authors should more cleanly discriminate the novel findings - subtb also include trajectories with terminal states already.

4 comments

We sincerely thank the reviewer for their thoughtful comments and constructive suggestions. In particular, we provide extended-training comparisons and evaluations across three training seeds to address comparison fairness and robustness, and clarify the structure-transfer evaluation and the distinction between BF-SubTB and standard SubTB.


W1 & Q1 (Computational Cost Comparison):

We thank the reviewer for the comments on training cost and comparison settings. We have added cost comparisons under approximately matched total training time in Appendix E.1 and trajectory and hierarchy collapse analyses in Section 5.4. All main comparisons use identical training steps and common token and turn limits. The training analysis in Section 5.4 (Figure 7) shows that ReMA’s turn and token counts decrease alongside hierarchy collapse and reasoning simplification, whereas I-Flow maintains longer trajectories with distinct high-level guidance and low-level execution. The additional training cost arises because I-Flow generates and optimizes longer trajectories (Figure 7), rather than from training the flow network itself (Appendix E.1). Regarding the inference budget, all compared methods are evaluated under the same maximum numbers of turns and tokens. Matching the budget to the actual usage of the baselines would instead confine I-Flow to the short trajectories that result from hierarchy collapse, restricting the verification and revision it is designed to preserve.

To examine whether training time alone explains the gap, we extended ReMA’s training to approximately match I-Flow’s total training time. Additional training showed limited improvement in the baseline’s hierarchical structure, while I-Flow maintained higher math accuracy across all seven benchmarks and less hierarchy collapse on MATH500. Maintaining the separation between high-level guidance and low-level execution enables verification, revision, and adaptive reasoning across turns, which are central benefits of hierarchical multi-turn reasoning. These findings suggest that I-Flow’s additional computational cost supports sustaining these reasoning capabilities.


W2 (Understanding Intent Structure Transfer):

We thank the reviewer for this comment. The purpose of Figure 6 is to examine whether novel and successful intent structures discovered by I-Flow remain effective when transferred to the original SFT model, rather than to compare the overall quality of intent structures generated by different policies. Since I-Flow jointly optimizes intent selection, high-level guidance, and low-level reasoning, we isolate the contribution of the discovered structures by transferring only the intents while keeping the SFT model’s guidance and reasoning policies fixed.

We focus on problems where I-Flow succeeds but SFT fails under Pass@1 to examine whether the discovered structures can help the original model solve previously unsuccessful problems. Under the same Pass@16 budget, the fixed SFT model achieves 13.9% success with its original structures and 56.9% when following I-Flow’s novel structures. As shown in Table R.1, the two conditions use comparable numbers of turns (6.85 vs. 6.79 on average), indicating that the gain does not come from longer intent sequences. This result demonstrates that the discovered structures retain useful reasoning guidance even when executed by the unchanged SFT model. We have clarified this experimental objective and interpretation in Section 5.3, RQ3.

Table R.1. Average number of turns (mean ± standard deviation) in the structure-transfer experiment.

SFTSFT + I-Flow’s Intent
Number of turns6.85 ± 2.546.79 ± 2.01

Q2 (Consistency across Training Seeds and Manual Validation):

To address the reviewer’s concern within the rebuttal period, we focused our multi-seed evaluation on I-Flow and ReMA, training and evaluating each with three seeds. Table R.2 shows higher mean Avg@16 and Pass@16 for I-Flow on all seven benchmarks, consistent with the performance trends reported in the paper. In particular, the mean ± standard deviation ranges are clearly separated on AIME25 and AMC23 for both metrics. We plan to extend the multi-seed evaluation to the remaining baselines for the final version.

For hierarchy collapse analysis, we manually reviewed generated trajectories and present representative cases in Appendix D.5. In ReMA, high-level outputs often directly provide answers that low-level outputs merely repeat, leaving essential reasoning unresolved, whereas I-Flow maintains distinct guidance and execution roles through revision and verification. These behaviors are apparent on manual inspection and are consistent with the automatic hierarchy-collapse labels in the examined cases, providing supporting evidence for the behaviors captured by the evaluator.

Table R.2. Mean ± standard deviation of Avg@16 and Pass@16 over three training seeds. The higher mean for each metric and benchmark is bolded.

MethodMetricGSM8KMATH500Gaokao23MinervaOlympiadAIME25AMC23
I-FlowAvg@1692.60±0.1778.27±0.4566.33±1.0336.20±0.1742.13±0.2510.47±0.5764.93±0.64
ReMAAvg@1691.03±0.7171.50±0.2060.00±2.1435.17±0.7538.42±1.178.13±0.4058.37±1.36
I-FlowPass@1698.17±0.2393.73±0.1279.73±0.4656.50±0.8266.90±0.6238.87±1.9690.83±1.44
ReMAPass@1696.57±0.1288.63±0.9176.40±0.6650.83±0.7658.43±1.1026.67±3.3580.00±2.50

Q3 (Further Clarification of BF-SubTB):

We thank the reviewer for this observation and agree that standard SubTB already includes terminal-ending subtrajectories. The key distinction of BF-SubTB is not the inclusion of such subtrajectories, but the exclusive use of terminal-anchored constraints. Standard SubTB imposes balance over arbitrary subtrajectories, including those with two nonterminal endpoints whose flows are learned through function approximation. Errors in these estimates may propagate through coupled balance constraints, particularly in long-horizon LLM reasoning with sparse terminal rewards.

In contrast, BF-SubTB restricts its constraints to suffixes from each post-intent state (sts_t) to the terminal state (sT+1s_{T+1}): F(st)PF(st:T+1)=F(sT+1),∀tF(s_t)P_F(s_{t:T+1})=F(s_{T+1}),\quad \forall t where (PB=1P_B=1) in our autoregressive state space. The terminal flow F(sT+1)F(s_{T+1}) is determined by the verified reward and fixed reference policy πref(τ∣x)exp⁡(β−1r(x,τ)))\pi_{\mathrm{ref}}(\boldsymbol{\tau}\mid\mathbf{x})\exp(\beta^{-1}r(\mathbf{x},\boldsymbol{\tau}))). Thus, while the starting-state flow remains learned, every intermediate constraint is directly anchored to a fixed terminal target rather than a learned nonterminal endpoint flow. This retains intermediate supervision without bootstrapping through learned endpoint estimates. The objective ablation in Section 5.4 further shows that BF-SubTB outperforms standard SubTB, supporting the effectiveness of this design. We have revised Section 4.2 to clarify this distinction and its motivation.


We thank the reviewer for their careful evaluation and constructive feedback. We hope that the additional comparisons and clarifications address the concerns about experimental fairness and robustness and make our evaluation more transparent.

thank you for addressing my concerns, in that detail.

  • the addition R2 is most valuable addition to me, so training budget is not an issue. for inference it still remains a bit unclear to me.
  • fig 6 comparison with freshly sampled SFT structures remains open

given that my main concerns were addressed, and also seeing that for the other reviews, i'll raise it to accept i would appreciate clear & precise separation of findings going into the paper

We sincerely thank the reviewer for raising the score and for the thoughtful follow-up. We are grateful that our responses have addressed your main concerns. In the final version, we will extend our turn and token analysis to cover inference and add a comparison with freshly sampled SFT intent structures to the structure-transfer experiment in Figure 6. We will also clearly and precisely separate the findings from the rebuttal discussion as they are incorporated into the paper. We sincerely appreciate your time and constructive feedback throughout the review process.

Comments on the paper

2 comments

General Response to All Reviewers

We sincerely thank the Area Chair and all reviewers for their constructive feedback. We have strengthened the manuscript with additional experiments, analyses, and clarifications, and uploaded a revised version with all changes highlighted in blue. The major updates are summarized below.


(i) Clarification of I-Flow’s motivation and distinctions. (Reviewers r6Zg, xK9T, 8MgP; Sections 4.2 and 5.4, Appendix D.4) To address the reviewers’ questions about I-Flow’s motivation and distinction from existing reasoning tuning, we clarified that its contribution is not predefined intent labels alone, but preserving the separation between high-level guidance and low-level execution during optimization. Maintaining this separation supports sustained verification, revision, and adaptive reasoning across turns. Section 4.2 explains how learned post-intent flows provide intermediate supervision and how BF-SubTB differs from standard SubTB through its exclusive use of terminal-anchored constraints. The objective ablations in Section 5.4 and flow-guided intent-selection analysis in Appendix D.4 support these distinctions and the usefulness of learned flows for identifying promising reasoning decisions.

(ii) Computational cost and reasoning behavior. (Reviewers xK9T, r6Zg, 8MgP; Section 5.4, Appendix E.1) Since all reviewers raised concerns about computational cost, we addressed questions about comparison settings, practical efficiency, and the contribution of additional computation through training-cost comparisons in Appendix E.1 and trajectory analyses in Section 5.4. The additional training time is primarily associated with generating and optimizing longer trajectories, while flow-network training accounts for only about 0.11% of the step time. Extending the baseline’s training to 400 steps to approximately match I-Flow’s total training time appeared to have a limited effect on preserving its hierarchical reasoning, and I-Flow achieved higher scores across all seven mathematical benchmarks. These findings suggest that I-Flow’s additional computation supports sustained guidance, verification, and revision across turns.

(iii) Intent taxonomy extensibility and semantic consistency. (Reviewers r6Zg, 8MgP; Sections 5.1 and 5.3, Appendices A and C.5) For r6Zg’s and 8MgP’s questions about the taxonomy’s rationale and adaptability, we clarified the grounding of the first five intents in established reasoning operations and IDL’s role in accommodating additional behaviors, with adaptation examples in Section 5.1 and Appendix A. For 8MgP’s semantic-consistency question, the new analysis in Section 5.3, RQ2, reports over 92% consistency for both intent–guidance and guidance–reasoning relationships across all six intents. The judging prompts are provided in Appendix C.5. These results provide empirical evidence that the selected intents correspond to actual reasoning behavior.

(iv) Open-ended transfer and further validation. (Reviewers r6Zg, xK9T; Section 5.3, Appendices D.3 and D.5, rebuttal Table R.1 and R.2) In response to r6Zg’s question about open-ended tasks, WildBench results in Appendix D.3 show that mathematics-trained I-Flow achieves higher scores than the evaluated baselines without task-specific fine-tuning. For xK9T, we clarified the purpose and controlled settings of the fixed-model intent-structure transfer ablation in Section 5.3, RQ3, and report in rebuttal Table R.1 that the two conditions use comparable numbers of turns. We also expanded qualitative case analyses in Appendix D.5 and reported three-seed comparisons in rebuttal Table R.2 to address questions about automatic evaluation and performance consistency.


We believe these revisions address the main concerns raised during review and improve the clarity and completeness of the paper. We are grateful for the reviewers’ guidance, which materially enhanced the manuscript.

Dear Reviewers,

We appreciate the time and effort you have devoted to evaluating our work. Your feedback has helped us strengthen the manuscript through additional experiments and clarifications on computational cost, intent taxonomy, and the learning mechanism.

With the discussion period approaching its conclusion, we would welcome any remaining questions or concerns regarding our responses and revisions. We would be happy to address them.

Thank you again for your consideration.

Best regards, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.