acceptodds
Under review as a conference paper at ICLR 2027

Coconut Tree: Adaptive Branching for Continuous Latent Reasoning

Abstract

Continuous latent reasoning replaces verbalized intermediate steps with hidden-state computation. This latent representation allows implicit representation of multiple reasoning chains in a single branch. However, we show that a single branch is a lossless implicit beam only when its state remains sufficient for every required future: when one state conflates beams requiring different futures, it cannot serve as a lossless implicit beam. To address this, we introduce Coconut Tree, a framework that determines when latent futures must remain separate and allocates reasoning width per problem. Across Coconut, CODI and CoLaR releases, the proposed approach consistently improves upon fixed-width and compute-matched baselines by several percentage points on average. To allocate reasoning width, Coconut Tree uses collision probabilities derived from the model's output-token distribution. A tight universal bound yields a stopping rule that exactly optimizes a penalised local coverage surrogate, whose only hyperparameter is a maximum branching width , without training or calibration. Predicted collision marginals track realized next-path gains, and the adaptive allocation consistently improves over the compute-matched fixed-width frontier. Live inference measurements on Coconut show that adaptive branching reduces wall-clock time by 13.6–60.0% and estimated dense-model FLOPs by 50.3–84.1% relative to using the fixed branching width (at ). The framework requires only separately continuable latent states and output-token probabilities, keeping the approach generic. Thus Coconut Tree explains both when continuous latent reasoning should branch and how much width branching justifies.

Primary area:
foundation or frontier models, including LLMs
Keywords:
latent reasoning, decoding, inference time, continuous thought

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_YxoYOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies when one continuous latent state cannot preserve the useful futures of an explicit beam, and how to allocate independent reasoning paths. It gives a conditional obstruction to lossless beam emulation and derives a collision-based marginal-coverage rule with an explicit width cap. The deployed protocol uses a deterministic first path, followed by additional independent paths; ProsQA uses answer-candidate probabilities, while open-answer tasks use a vocabulary proxy. Experiments support separate continuations and adaptive allocation at matched mean path count. Resource measurements additionally compare adaptive execution with always using the full width cap. The evaluation concerns pass@K coverage, rather than an implemented selector's final-answer accuracy.

Critical strengths
  1. The theory distinguishes the need for separate futures from the allocation of additional paths. The surrogate objective, stopping condition, and width guarantee are explicit.
  2. The matched shared-state control and fixed-width frontier make the separation and allocation results informative. Main allocation comparisons include paired intervals.
  3. Proxy ablations, architectural transfer, and measured inference overhead provide useful evidence about practical scope, including the mixed CoLaR allocation results.
Critical weaknesses
  1. W1 — The introduction moves to technical contributions before sufficiently motivating the problem. After a brief description of latent reasoning, the introduction moves directly to the method and contribution list (p.1, lines 35–53; p.2, lines 68–77). It does not sufficiently explain the distinction between implicitly representing multiple futures in one state, maintaining independent latent continuations, and allocating their width across problems before introducing unresolved futures and collision-based marginal gains. Figure 1 provides a policy comparison, and §§3–4 later formalize the concepts, but the reader must assemble much of the motivation from these later sections. For readers not already working on COCONUT-style methods, this makes the problem framing and the connection between the separation and allocation contributions harder to follow.
  2. W2 — The efficiency and allocation comparisons use different budget notions. Table 2 matches mean path count, while §5.3 and Table 4 report resource savings against full width without a joint cost–coverage frontier. Path length, batching, and the deterministic first path can affect the relationship between path count and actual cost. Tables 6–7 additionally compare stopping rules and proxies at their own mean widths. Their uplifts over the respective fixed frontiers are informative, but are not direct coverage differences at a common budget. Thus, the practical-budget advantage and the incremental value of particular allocation choices are less directly established than the individual reported comparisons (pp.7–9, 14–15, 18–19).
  3. W3 — The downloaded artifact lacks the promised low-cost verification records. Appendix A.5 states that stored path, proxy, and compute records permit regenerating the analyses without model inference (pp.13–14). The supplied archive's README instead excludes stored experiment results, and its inventory contains no such records. Recreating outputs is therefore required before checking the reported frontier, bootstrap, and resource analyses. This is a fixable artifact/documentation discrepancy; it does not establish that the reported results are wrong.
Questions for authors
  1. Q1 — Clarifying the problem and motivation (W1). Could the introduction explain, before the technical contribution list, the concrete limitation of a single implicit beam, why a fixed number of independent paths is insufficient, and how the separation and per-problem allocation questions are connected? A short running example distinguishing these settings would help readers unfamiliar with COCONUT-style work understand what the method addresses.
  2. Q2 — Common-budget quality and efficiency (W2). At matched budgets, how does the proposed policy compare in coverage with fixed-width and simpler adaptive policies? Please connect coverage with measured latency or estimated FLOPs for the practical-cost comparison, and distinguish that result from comparisons at matched mean path count. When comparing stopping rules, keep the proxy fixed so their incremental contribution can be assessed.
  3. Q3 — Availability of verification records (W3). Are the path, proxy, and compute records described in Appendix A.5 available in an omitted location? Please identify the matching records and analysis entry point, or clarify the artifact's actual reproduction scope.
3 comments

Thank you for the thoughtful review, comments and suggestions!

We are glad you see the benefit of the theory for establishing surrogate objectives, stopping conditions and width guarantees, and that you found find the separation and allocation results informative, and see useful evidence about practical scope!

We hope we address any of your remaining concerns below.

W1 — The introduction moves to technical contributions before sufficiently motivating the problem. After a brief description of latent reasoning, the introduction moves directly to the method and contribution list... It does not sufficiently explain the distinction between implicitly representing multiple futures in one state, maintaining independent latent continuations, and allocating their width across problems before introducing unresolved futures and collision-based marginal gains... the reader must assemble much of the motivation from these later sections. For readers not already working on COCONUT-style methods, this makes the problem framing and the connection between the separation and allocation contributions harder to follow. Q1 — Clarifying the problem and motivation (W1). Could the introduction explain, before the technical contribution list, the concrete limitation of a single implicit beam, why a fixed number of independent paths is insufficient, and how the separation and per-problem allocation questions are connected? A short running example distinguishing these settings would help readers unfamiliar with COCONUT-style work understand what the method addresses.

We revised the introduction to motivate separation and adaptive width before presenting the collision-based rule and listing our contributions. The new examples, marked in blue, include a running scheduling example that connects separate continuations to the choice of how many paths to explore per problem.

Q2 — Common-budget quality and efficiency (W2)..s Tbl 2 matches mean path count, while §5.3 and Tbl 4 report resource savings against full width without a joint cost–coverage frontier. Path length, batching, and the deterministic first path can affect the relationship between path count and actual cost... At matched budgets, how does the proposed policy compare in coverage with fixed-width and simpler adaptive policies? Please connect coverage with measured latency or estimated FLOPs for the practical-cost comparison, and distinguish that result from comparisons at matched mean path count. When comparing stopping rules, keep the proxy fixed so their incremental contribution can be assessed.

We have now added Fig 16, which shows the coverage comparison from Fig 4 at matched mean estimated FLOPs per example. At the eight-path cap, COCONUT TREE improves coverage over the fixed-width frontier by 0.35–3.02 percentage points across all ten settings, with paired 95% confidence intervals above zero in each setting. We have also added Fig 14(b), comparing adaptive rules at matched mean estimated FLOPs, alongside the matched mean path-count comparison in Fig 14(a). All adaptive rules use the same probability proxy. Under both budget measures, COCONUT TREE outperforms all four calibrated baselines on GSM8K, GSM-Hard, and SVAMP on both backbones at every tested cap above one. On ProsQA and MultiArith, coverage differences are smaller and vary by baseline. The low mean path counts on these tasks leave fewer extra paths to reallocate.

W3/Q3 — The downloaded artifact lacks the promised low-cost verification records... Are the path, proxy, and compute records described in Appendix A.5 available in an omitted location? Please identify the matching records and analysis entry point, or clarify the artifact's actual reproduction scope.

Sorry, this was an oversight during submission, when the system limited submissions to 50MB. We have now updated the supplementary material to include the compressed records for reproducibility, as well as refined the reproducibility statement and appendix details on this, and updated the README.md with instructions. Thank you for being so thorough with the checks!

Thank you for the comments and questions, we hope we have addressed any remaining concerns.

Thank the authors for rebuttal and explaining. I think the blue lines indeed help me understand how you want to use multiple latent cots conceptually, but I doubt if current version has too much examples in the opening two paragraphs as a rigid academic paper. I think there are better ways for the opening.

Thank you for the follow-up. We revised the first two paragraphs of the introduction to reduce the number of examples while preserving the intuition behind separate latent trajectories. The revised opening explains why a shared latent state may lose distinctions that matter for later reasoning and how separate continuations let each trajectory proceed from its own latent state. We replaced the task-specific examples with one brief illustration of why different problems may benefit from different numbers of trajectories. This leads directly to the question of how many latent trajectories are worth exploring for each problem. We believe these revisions make the introduction easier to follow and clarify the motivation for adaptive branching, and we hope they address your concern.

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

Coconut Tree proposes a method for choosing how many independent latent reasoning paths to run for each problem. It formalizes conditions under which a single latent state cannot preserve all distinctions needed for future outcomes, then derives a stopping rule that exactly optimizes a stated coverage surrogate. After completing one deterministic path, the method uses answer-candidate probabilities or next-token probabilities to estimate the benefit of additional paths, with the threshold determined by the maximum permitted width. Experiments on Coconut, CODI, and CoLaR report improved answer coverage over fixed-width policies at matched average path counts, with mixed adaptive gains on CoLaR. The paper also reports wall-clock reductions of 13.6–60.0% compared with always running eight paths.

Critical strengths
  • The allocation mechanism is simple and inexpensive. Proposition 4 provides an exact stopping rule for the stated surrogate and a tight threshold that guarantees the width cap. The implementation reuses probabilities from the first path and requires no additional controller training or probability calibration. This makes the method straightforward to add to existing latent reasoners.

  • The comparison with fixed-width allocation is carefully designed. The paper uses the same stored path banks, compares against a fixed-width frontier at matched average path counts, and reports paired confidence intervals that account for repeated seeds. All ten primary Coconut conditions show positive adaptive gains at the largest cap, with additional support from CODI. The live runtime measurements also provide useful evidence beyond theoretical compute estimates.

Critical weaknesses
  • The advantage over other adaptive allocation rules remains unclear. Table 6 compares methods at different average widths by reporting each method’s improvement over its own fixed-width frontier. A larger improvement by this measure does not establish a better coverage-compute trade-off between adaptive methods. Moreover, rounding the Shannon effective support is different from EDEN’s published normalized-entropy mapping, so this ablation does not substitute for a faithful comparison with that approach. Since uncertainty-guided allocation already exists, a comparison between adaptive policies at matched compute is important for establishing the contribution of the proposed rule.

  • The experiments do not establish the claimed explanation of “Separation Necessity.” The shared-state control copies one stochastic proposal, whereas the independent arm starts from different proposals. This demonstrates the value of diversity at the first latent transition, but it does not directly show that a trained state has conflated future-relevant information or cannot support implicit parallel reasoning. The allocation theory also assumes that target modes and sampled continuations follow the same distribution, while the implementation combines a deterministic first path, dropout-based continuations, and an output-probability proxy. The authors acknowledge this mismatch, but the ranking diagnostics provide limited evidence that the theoretical model explains the observed gains.

Questions for authors
  • Comparison request: How does the proposed rule compare with entropy-based and confidence-based allocation when all adaptive methods use the same average compute budget? Please report coverage–compute curves, ideally using estimated FLOPs to account for differences in path length.

  • Clarification and validation request: Can you provide evidence that the output-probability proxy predicts the actual marginal gains of dropout continuations at each fixed branch index? This would help distinguish useful prediction across examples from the general decline in marginal gains as more paths are added, and would clarify the connection between the surrogate and the implemented procedure.

  • Validation request: Do the coverage gains translate into higher accuracy when the system must return one answer, using the same selection method for every policy? This matters especially for binary ProsQA, where producing both candidates guarantees coverage without identifying the correct one. Accuracy and runtime comparisons with a shared voting or reranking procedure would materially strengthen the practical case.

3 comments

Thank you for the review! We are glad you found the mechanism to be inexpensive, extensible to other latent reasoners, and noted the benefit that it requires no additional training or probability calibration.

We made the suggested additions in red to the manuscript and hope the revisions below address your concerns.

The advantage over other adaptive allocation rules remains unclear. Tbl 6 compares methods at different average widths. How does the proposed rule compare? Ideally using estimated FLOPs

We added coverage–compute curves in Fig 14 to compare COCONUT TREE directly with four adaptive baselines, including entropy-based and confidence-based allocation. At each cap, we match the baselines to COCONUT TREE’s mean path count in Fig 14a and its mean estimated FLOPs in Fig 14b. The FLOP estimates account for path-length differences.

At matched FLOPs, COCONUT TREE achieves higher coverage than all four baselines on GSM8K, GSM-Hard, and SVAMP on both backbones for Bmax⁡>1B_{\max}>1. The pointwise paired 95% CIs for these differences are above zero. The same pattern holds at matched mean path counts. The methods achieve similar coverage on ProsQA and MultiArith, where low average path counts leave few extra paths to reallocate. On ProsQA with Llama-1B, for example, the shared first path already achieves 97.6% coverage, and COCONUT TREE averages just 1.118 paths per example even at the 8-path cap.

Rounding the Shannon effective support is different from EDENs normalized-entropy, so this ablation does not substitute for a faithful comparison

Tbl 6 and Fig 14 now compare COCONUT TREE with EDEN’s normalized-entropy mapping, which we adapted to allocate complete paths (App C3). At the largest shared budget in estimated FLOPs, COCONUT TREE achieves 0.69–2.27 percentage points higher coverage on GSM8K, GSM-Hard, and SVAMP across both backbones, with paired 95% CI's above zero. EDEN’s only lead with a paired 95% CI above zero is 0.27 percentage points on ProsQA with GPT-2. The intervals include zero in the other three settings. The same pattern holds at matched mean path counts (App C6).

The experiments do not establish the claimed explanation of “Separation Necessity”

We renamed “Separation Necessity” to “Value of distinct proposal states.” Sec 5.1 now clarifies that the separate arm starts from independently sampled state–cache pairs and the shared arm copies one pair. Both use fresh dropout during continuation. We limit the empirical claim to the coverage benefit of retaining and independently continuing distinct initial proposals.

Theory assumes that target modes and sampled continuations follow the same distribution, while the implementation combines a deterministic first path, dropout-based continuations, and an output-probability proxy.

We added a clarification after Algo 1 explaining how the common-distribution assumption is used to derive the surrogate and its stopping rule. At inference, we compute the score from the deterministic first path’s output probabilities and add a dropout continuation when the score exceeds the threshold. The practical premise, tested below, is that higher scores predict larger gains across examples at the same path count.

Provide evidence that the output-probability proxy predicts the actual marginal gains of dropout continuations at each fixed branch index? This would help distinguish useful prediction from the general decline in marginal gains as more paths are added

We revised Fig 7 to compare score-based allocation with matched random allocation at each fixed path count. At every path count, score-based allocation yields 0.92–1.36 percentage points more coverage per added path than matched random allocation. Both policies allocate the same number of extra paths at each bb, so the comparison shows that the proxy-derived score directs dropout continuations to examples with larger realized gains while accounting for the overall decline in marginal gains as more paths are added.

Fig 12 further shows that at each fixed path count, the mean coverage gain from one additional dropout path increases across score quartiles, further supporting score-based allocation.

Do coverage gains translate into higher accuracy when the system must return one answer, using the same selection method for every policy?

We added a selection experiment in App C7, using the same LLM judge and prompt across policies. With GPT-OSS-20B, COCONUT TREE outperforms the fixed-width frontier in nine of ten settings at the eight-path cap and matched mean path counts, with paired 95% CI above zero in eight. On binary ProsQA, accuracy rises from 80.95% to 82.27% with GPT-2, and both policies reach about 97.6% with Llama-1B. With Gemma 4 12B, COCONUT TREE improves accuracy in all eight arithmetic settings. The coverage gains can translate into higher final-answer accuracy, with the size of the improvement depending on the selector.

Thank you for the comments. We hope we have addressed your concerns

I thank the authors for the thorough response. The new matched-compute comparisons address my main concern about the advantage over other adaptive allocation rules, and narrowing the separation claim together with the fixed-path-count validation adequately addresses the gap between theory and implementation. I have raised my score to 3.

Thank you for the response and review, we are glad to have addressed your concerns!

Reviewer_D6JDOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates an interesting idea of deciding how many independent latent reasoning paths to run for each input. It first runs one deterministic trajectory and uses its output probabilities to decide how many additional paths to run. For tasks with explicit answer candidates, it uses the probabilities over those candidates. For open-answer tasks, it uses the vocabulary distribution at a selected token position in the first answer. It treats this distribution as a proxy for possible reasoning outcomes, estimates the gain from adding another path, and stops when the estimated gain falls below a threshold determined by the maximum width. The additional paths are generated independently using dropout.

The theoretical part gives general arguments for keeping different reasoning states separate and derives the stopping rule under assumptions about the outcome distribution. These arguments do not directly establish why the selected token probabilities should represent latent-path outcomes. The experiments compare separate-state and shared-state continuations, and adaptive allocation against fixed-width policies at matched average path counts. The main metric is coverage, meaning that at least one generated answer is correct.

Critical strengths

The method is simple and does not require additional model training. Using information from the first run to decide whether further computation is worthwhile is a practical idea. The allocation rule also avoids separately tuning a stopping threshold for each model or dataset.

I appreciate the comparisons at matched average path counts, which make the benefit of adaptive allocation easier to see. The experiments cover several latent reasoning models and tasks, with ablations on both the allocation rule and the probability proxy. The paper also measures actual wall-clock time and includes the allocation overhead, giving a clearer picture of its practical cost.

Critical weaknesses

My main concern is why the probabilities at one token position should describe the possible results of different latent reasoning paths. A token can be uncertain because of wording, even when the reasoning is settled. The score-validation experiment is useful, but it does not fully explain why this distribution is a good basis for deciding how many paths to run.

Propositions 1 and 2 seem fairly elementary to me, and their connection to the actual method is unclear. The separation experiment shows that different initial states are useful, but does not directly establish the information loss discussed in these propositions. Proposition 4 gives a valid stopping rule under the assumed outcome distribution, but does not establish why the resulting choice of B should be good in practice.

The novelty also seems limited, since multi-path latent sampling and uncertainty-based branching have already been explored in LatentTTS and EDEN. The main addition is the rule for choosing B. Table 6 includes some allocation baselines, but compares their gains at different average widths. A direct comparison at the same budget would better show whether the new rule is preferable to simpler alternatives.

Finally, the main results measure whether at least one path gives the correct answer. It is unclear whether the gains remain when the system has to select one final answer, for example through voting or an existing scorer.

Questions for authors

Why should the token probabilities from the first answer tell us how useful additional latent paths will be, and can the proposed rule be tested against a simple entropy-based rule at the same average number of paths?

2 comments

Thank you for the detailed review. We’re glad you found the idea interesting and the method simple and practical. We hope our responses address your concerns. Changes are marked in red in the manuscript.

My main concern is why the probabilities at one token position should describe the possible results of different latent reasoning paths.

Why the token score can help. For tasks with fixed answer choices (e.g., ProsQA), we use the first path’s probability distribution over candidate answers directly. For open-answer tasks, we use its next-token distribution at the eligible position with the largest Rényi-2 effective support. Because token uncertainty may reflect uncertainty about the answer itself, we use the score to estimate how much another latent path might help. Empirically, in the revised Fig. 7, we show allocating by score yields 0.92–1.36 percentage points more coverage per added path than budget-matched random allocation (App. C.2).

Semantics As you point out, token uncertainty can also reflect wording or formatting. Inspired by semantic entropy, we compare to grouping completions by answer and using their group frequencies to compute an answer-level collision marginal (App C5). At matched mean path counts, the token-based rule still improves coverage over this probe by 0.91–2.13 percentage points (Tbl 8), motivating the chosen rule. This is an interesting area for future research.

Props 1 and 2 seem fairly elementary to me their connection to the actual method is unclear. Prop 4 does not establish why the resulting choice of B should be good in practice.

Prop 1 motivates retaining each proposal as a distinct state–cache pair to preserve distinctions that can affect outcomes. Prop 2 motivates continuing each pair independently so later computation can respond to those distinctions. We now make both links explicit in the discussion of Algorithm 1. We also reframed Section 5.1 around the coverage benefit of distinct initial proposals at matched compute.

Prop 4 derives the optimal stopping rule for the coverage surrogate. After Algorithm 1, we now distinguish the theoretical outcome distribution from the first-path proxy used to choose BB. Our practical assumption is that, at a fixed path count, a higher proxy score predicts a larger gain from another dropout path. At Bmax⁡=8B_{\max}=8, the resulting policy improves coverage by 0.35–3.03 percentage points over the fixed-width frontier at matched mean path counts across all ten settings (Tbl 2). Together with the fixed-path-count results below, this supports using the surrogate-derived rule in practice.

The novelty also seems limited, since multi-path latent sampling and uncertainty-based branching have already been explored in LatentTTS and EDEN.

LatentTTS evaluates preset path counts and finds diminishing coverage gains (Fig 2 of their paper). EDEN uses token entropy to decide how many next-token alternatives to expand at each decoding step. COCONUT TREE estimates the coverage gain of adding another path to answer a question left open by both approaches. How many complete latent reasoning paths should we run for each problem?

A direct comparison at the same budget would better show whether the new rule is preferable to simpler alternatives.

We added a direct Shannon-entropy baseline to Tbl 6. Using the same probability proxy across rules, we compared them at matched mean path counts (Fig 14a, App C3) and at matched mean estimated FLOPs to account for path length (Fig 14b, App C6). Under both budget measures, COCONUT TREE outperforms all four calibrated baselines on GSM8K, GSM-Hard, and SVAMP on both backbones at every tested cap above one. At matched mean path counts, the pointwise paired 95% confidence intervals for these differences are above zero. Results on ProsQA and MultiArith are smaller and mixed, with fewer extra paths to reallocate. For instance, on Llama-1B ProsQA, COCONUT TREE averages 1.118 paths at the eight-path cap.

It is unclear whether the gains remain when the system has to select one final answer, for example through voting or an existing scorer.

We added a final-answer selection experiment in App C7 using LLM-as-a-judge.

Under selection, COCONUT TREE achieves higher accuracy than the fixed-width frontier in most settings. With GPT-OSS-20B, it leads in nine of ten settings at the eight-path cap and matched mean path counts, with paired 95% confidence intervals for the difference above zero in eight. On binary ProsQA, accuracy rises from 80.95% to 82.27% on GPT2 and is essentially tied at 97.6% on Llama-1B. With Gemma 4 12B, COCONUT TREE improves accuracy in all eight arithmetic settings.

Coverage gains can carry over to final-answer accuracy, and how much carries over depends on the selector. Developing stronger selectors, including improved reward models, is complementary to the path-allocation approach proposed here.

Thank you for the questions. We hope this addresses your concerns.

Thank you for the review! We hope we have successfully addressed your concerns during the rebuttal. With the rebuttal period coming to its final day, please do let us know if you have any additional questions. Thank you

Comments on the paper

1 comment

With the rebuttal period coming to an end in 2 days, we would like to thank all the reviewers for their questions and suggestions. We have updated the manuscript, with changes marked in red or blue. We are including a summary below of these main changes for convenience.

  • Equated FLOP comparisons, not just equated average path comparison. COCONUT Tree still consistently outperforms all comparisons, even when equating for FLOPs.
  • Additional baseline comparisons, including a direct Shannon entropy version and path-level EDEN version. COCONUT Tree convincingly outperforms both.
  • Clarified connections between propositions and the method itself, to make the connection with theory cleaner. Specifically, we explain how Propositions 1 and 2 motivate the design of Algorithm 1, and clarify the assumptions behind the stopping rule.
  • Added supporting experimental evidence for the allocation score to Figure 7, motivating the proposed score-based allocation.
  • Reworded the experimental claim from "Separation necessity" to "Value of distinct proposal states"
  • Added accuracy comparisons under single path selection, rather than just pass@k/coverage. COCONUT Tree still outperforms the comparisons under final answer selection.
  • Added an ablation of semantic grouping rather than token-level variability (Appendix C.5), showing the token-level rule still outperforms this semantic grouping.
  • Improved the introduction and clarified with explicit examples for improved readability.
  • Updated the supplementary material to include the stored paths for replaying the methods without needing to rerun

We believe the revised manuscript has significantly improved as a result, so we thank the reviewers and hope we have addressed all concerns. We are happy to make any additional clarifications or answer any further questions over the remainder of the rebuttal period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.