acceptodds
Under review as a conference paper at ICLR 2027

Shared Accuracy–Length Patterns in Reasoning Compression

Abstract

Reinforcement learning with verifiable rewards enables strong reasoning, and reasoning-compression methods seek shorter correct responses through different training objectives and length controls. We place their independently trained policies in a common observable coordinate defined by dataset accuracy and mean response length. Across the observed task, model, and controller extensions, low-parameter functions predict held-out policy endpoints in this accuracy–length space. One shared endpoint direction learned from one group of prompts captures most transferable variation on a disjoint group. Its endpoint values are approximately monotone in realized response length with nonlinear spacing. Simple exponential and hyperbolic one-parameter shapes capture this spacing with little predictive loss and remain predictive after correctness is averaged over prompts. Together, these results connect the dataset-level accuracy–length structure of completed compression sweeps to compact transferable structure already shared across individual prompts.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Reasoning Compression, Reinforcement Learning with Verifiable Rewards, Large Language Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6QSfOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies the trade-off of accuracy and reasoning length from several reasoning-compression methods. The authors show that the average response length and accuracy of each trained model can be fit by a low-dimensional saturating curve. They also further report that correctness matrices across problems and compressed models are dominated by a shared direction that transfers to held-out problems, and that this direction varies monotonically but nonlinearly with response length.

Critical strengths
  1. conduct a broad and careful comparison across methods, models and task types.
  2. The low-dimensional regularities are presented clearly and is useful.
Critical weaknesses
  1. The topic itself is very old, has been explored much, and this paper doesn't raise something much insightful.
  2. The main regularity is not surprising relative to the experimental construction. If generating each model family just by sweeping one scalar length, it will do have a dominant one-dimensional direction.
  3. Do not have much practical use. Need 3-9 training point according to the paper to rebuild the curve.
Questions for authors
  1. When two or more interventioon axes are mixed, still mainly focus on one direction?
  2. How does the proposed fit compare with linear interpolation, isotonic regression, and standard saturating-curve baselines?
  3. Does the different setting of max length of training and eval matters? Why the code task (LiveCodeBench) show weaker structure?
  4. What decision (controller choice, choice of setting) would a practitioner make differently because of these findings? Can you demonstrate one?
3 comments

We have incorporated the corresponding clarifications and analyses from our rebuttal into the revised PDF. The updated passages and added results are marked in blue.

  • Novelty and scalar sweeps: We clarified why a reward setting does not prescribe the resulting accuracy–length curve, and why nonlinear prompt responses need not share a linear direction (§1, lines 40–44; §2, lines 108–110; §5.2, lines 376–377).
  • Practical decisions: Section 4.3, lines 288–300 now explains how marginal accuracy cost guides compression exploration and how curves compare controllers at equal realized cost. Section 6.3, lines 500–504 and Appendix I.4 add prompt sensitivity sis_i as a potential training signal, with offline calibration and training validation left to future work.
  • Mixed interventions: Appendix I.2 pools 42 distinct DSDT-1.5B policies from five controllers. Adding the first shared direction lowers held-out question-level RMSE on all six math datasets by a median of 6.88 percentage points. This is preliminary cross-design evidence; joint multi-factor training remains untested.
  • Baselines: Table 2 and Appendix I.1 compare nested saturation-family selection with linear interpolation, isotonic regression followed by linear interpolation, and shape-preserving cubic interpolation. Across six DSDT-1.5B math datasets, LP’s median RMSE is 1.656 versus 2.371 percentage points for linear interpolation.
  • Length caps and code: We clarified the shared cap protocol and its untested independent effect (§7, lines 525–527), and added the math-to-code transfer context for code results (§4.2, lines 267–269).

We would be grateful for a brief follow-up on whether these updates resolve your concerns or whether further clarification is needed.

Thanks for your reply. Yeah I think you solve part of my problems, proving that the pattern exists in a wide field and perhaps are not restricted to training settings. But I still didn't see the innovation here as it is common sense that there exists best length for each problem and certainly moderate compression may remove redundant or error-prone reasoning and improve accuracy, whereas excessive compression may eventually hurt it, and I didn't know what you want to tell. In LEASH or LASER, they do such kind of training because they realize this kind of problem and thus find a way to get the optimized length for each problem, so if you base on these and then do different length, I do not see what else is revealed. Actually I share same concern with other reviewers, I do not believe the remaining contribution currently meets the ICLR acceptance bar so I will keep my score.

Thanks for spelling this out. We agree that the basic intuition is familiar: moderate compression can help, while excessive compression can hurt. What we want to add is a way to measure this trade-off and understand where it comes from at the question level.

First, knowing that an optimal length exists does not tell us how much accuracy we pay for the next token saved, or which method is better when they end up at different lengths. LEASH still leaves target-length selection to the user (§6 of the LEASH paper). Our fit gives a differentiable curve and predicts held-out points more accurately than common interpolation methods. For DSDT-1.5B with LP, median RMSE across the six math datasets is 1.66 percentage points, versus 2.37 for linear interpolation and 2.46 for PCHIP. This lets us estimate where further compression becomes costly and compare methods at the same actual length (§4.3).

Second, our question-level results offer another way to think about adaptive compression. The approximate shared shape helps explain why the average curve stays simple, even though questions have different accuracy levels and response amplitudes. LASER-D groups questions by current accuracy and gives shorter targets to easier groups. Current accuracy describes performance at one setting. On the other hand, with the shared shape and its scale fixed, our amplitude sis_i describes how question ii’s accuracy changes across trained compression policies. It is therefore closer to the compression response we want to measure. It provides a different perspective: considering sensitivity to compression alongside current difficulty.

Reviewer_21eYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper places independently trained reasoning-compression policies in a common coordinate of dataset accuracy and mean response length, and finds that low-parameter functions predict held-out policy endpoints. On the main math sweeps, a dominant endpoint direction learned on one prompt set transfers to disjoint prompts. This direction is roughly monotone but nonlinearly spaced in mean length, and simple exponential or hyperbolic shapes predict about as well. The paper links these prompt-level findings to aggregate accuracy–length curves and tests extensions across models, controllers, and domains.

Critical strengths
  1. An interesting, potentially novel empirical regularity.

The finding is not just that accuracy varies with length, but that prompt-level correctness across compression policies has a compact shared description that connects back to aggregate accuracy–length curves, shown across several controller sweeps.

  1. Validation beyond in-sample fitting.

Source/target prompt splits, independent rollout halves, held-out prediction, and a randomized smoothness check give real evidence of transferable structure, and the shape-transfer test excludes the held-out endpoint. The limitations are stated clearly.

Critical weaknesses
  1. The analysis does not yet explain what drives the shared pattern.

    The results are mostly descriptive. Does the dominant direction reflect reasoning compression specifically, general policy quality, or simply varying one scalar control within a narrow family of training procedures? The smoothness check does not distinguish these. And since length is a dataset-wide mean, its relation to per-prompt reasoning effort is unclear.

    A full mechanistic theory is not required, but a more discriminating analysis would help. For example, the authors could test whether length explains residual variation beyond a simple source-prompt quality statistic, or use the matched-length controller checks to identify failure conditions.

  2. The statistical evidence leaves the pattern's stability and complexity uncertain.

    Each sweep has only 7–16 endpoints, with no repeated runs per controller setting, so residual error cannot be split into training variability, evaluation noise, and misspecification. RQ1 also picks the best function family using the same cross-validation scores it reports. This introduces selection optimism even with pre-specified families.

Questions for authors
  1. Can you separate a length-related shared direction from a generic policy-quality direction? Does realized length still add explanatory or predictive value after accounting for it?
  2. What do the LP-C/LP-D matched-length checks and the weaker code results say about when the shared-shape approximation holds or fails?
  3. Are repeated runs available for any controller setting, and does the family comparison hold when selection is separated from final evaluation?
3 comments

We have integrated the clarifications and additional checks from our rebuttal into the revised PDF, with all changes marked in blue.

  • Interpretation: We clarified that mean length is a policy cost and that overall quality may itself change because of compression training. We do not claim extra length prediction after controlling source accuracy (§3.2, lines 142–144; §6.3, lines 496–499; §7, lines 521–522).
  • Beyond one controller: The PDF now includes the pooled analysis of 42 distinct DSDT-1.5B policies from five controllers. Adding the first shared direction lowers held-out question-level RMSE on all six math datasets by a median of 6.88 percentage points (Appendix I.2, Table 21).
  • Selection and testing: Table 2 and Appendix I.1 now report nested family selection against interpolation. Across six DSDT-1.5B math datasets, LP’s median RMSE is 1.656 versus 2.371 percentage points for linear interpolation.
  • Stability: Appendix I.3 reports endpoint subsampling; adding the first direction on seven-policy subsets lowers RMSE by a median of 6.01 percentage points. Section 7, lines 515–517 explicitly states that fixed-setting training repeats are unavailable.
  • Scope: We added interpretations of LP-C/LP-D matched-length checks (Appendix F.3, lines 1459–1461 and 1487–1489) and the math-to-code transfer context (§4.2, lines 267–269).

Could you let us know whether these updates address your concerns, or which points still need clarification?

Thank you for the detailed response. The nested family selection, the pooled analysis across five controllers, and the subsampling check address my statistical concerns. However, my main concern about depth remains. The response confirms that the shared direction may reflect general policy quality, and that length adds no predictive value beyond source accuracy. The result is therefore a well-supported description of shared variation, but it does not yet explain what drives it or what is specific to reasoning compression. I will keep my rating of 2.

Thanks for the follow-up! General policy quality could explain the shared direction. What makes it useful for compression is that it lets us estimate which questions lose more accuracy when we train for shorter outputs. For example, LASER-D groups questions by their current solve rate and uses sampled response lengths to set a target for each group. Two questions with the same solve rate can still respond differently to stronger compression. With the shared shape fixed, sis_i captures that difference: for the same move to a shorter policy, a larger sis_i means a larger predicted accuracy drop for question ii (§6.3, Eq. 10). That gives us something concrete to test: would weaker length penalties on more sensitive questions preserve more accuracy at the same average length? Estimating sis_i takes offline evaluations across several trained policies, and we have yet to test whether it helps in a new training run (Appendix I.4).

Reviewer_3dG3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates the accuracy–length relationship in reasoning-compressed RLVR model training. The authors study the accuracy–length tradeoffs by investigating different length-control mechanisms. They present three key findings: (1) low-parameter functions can predict held-out accuracy–length endpoints; (2) a single shared direction in the prompt–policy correctness matrix captures most transferable variation across prompts; and (3) this direction is approximately monotonic but nonlinearly in response length and can be approximated using simple one-parameter functions.

Critical strengths

S1: The paper investigates whether independently trained reasoning-compression policies exhibit common accuracy–length patterns. Understanding such relationships could be useful for analyzing the efficiency of reasoning models.

S2: The reported results show that three-parameter functions generally fit the accuracy–length relationships well, and that one shared endpoint direction improves held-out prediction on mathematical reasoning tasks.

Critical weaknesses

W1: The main findings are largely empirical observations based on curve fitting and standard low-rank matrix decomposition. Although the observations are interesting, it's still unclear whether the findings constitute a sufficiently substantial methodological or conceptual advance.

W2: The paper does not demonstrate how the fitted relationships can improve model reasoning efficiency or guide model training policies in real-world scenarios.

Questions for authors
  1. How can the identified accuracy–length relationships be used to improve reasoning compression in practice?

  2. How well do the fitted functions predict the accuracy of policies whose response lengths fall outside the observed training sweep?

  3. Why is the observed structure substantially weaker on coding tasks?

2 comments

We have incorporated the clarifications from our rebuttal into the revised PDF. All additions and revisions are marked in blue.

  • Contribution: We clarified why reward settings do not guarantee the observed accuracy–length relation or shared prompt-level structure (§1, lines 40–44; §2, lines 108–110).
  • Practical guidance: We expanded how the derivative estimates local accuracy loss per token saved and guides training exploration, and how curves compare controllers at the same realized output cost (§4.3, lines 288–300).
  • Potential training insight: The discussion now includes prompt sensitivity sis_i as a possible allocation signal distinct from current solve rate. Offline estimation and training validation remain future work (§6.3, lines 500–504; Appendix I.4).
  • Sweep coverage: We clarified how the intended compression range can be covered directly, so these decisions need not rely on extrapolation (§4.3, lines 302–307).
  • Code results: We added the math-only training and cross-domain transfer context for the weaker code evidence (§4.2, lines 267–269).

We would appreciate your feedback on whether these revisions address your concerns and whether any points need further clarification.

Thank you for the detailed response and revisions. The clarifications help explain the potential value of the observed accuracy–length relationships. However, my main concerns remain. The contributions are still primarily empirical observations based on standard techniques, and the practical benefits have not been sufficiently demonstrated. Therefore, I will maintain my original rating.

Comments on the paper

1 comment

Dear Area Chair, Senior Area Chair, and Program Chairs,

Thank you for handling our submission. We appreciate the follow-ups from all three reviewers.

Discussion summary

  • 21eY said the new analyses addressed their statistical concerns. They still ask what drives the shared direction and whether it reflects general policy quality.
  • 6QSf acknowledged the broader evidence across training settings and still questions what the findings add beyond prior compression work.
  • 3dG3 said the revisions clarified the potential value of the accuracy–length relationship. They still see the contribution as primarily empirical and seek stronger evidence of practical benefits.

What the findings add

Agreed. The nested evaluation separates curve selection from final testing (Table 2 and Appendix I.1). Adding the shared direction improves held-out prediction on all six math datasets after pooling policies from five controllers, and the benefit persists with fewer policy endpoints (Appendices I.2–I.3).

No Response on Specific Applications. The follow-ups discuss practical value in general terms. Our responses described two concrete applications.

  • Dataset level. The fitted slope estimates the accuracy cost per token saved. Practitioners can use it to identify where further compression is worth exploring and compare controllers at the same realized mean response length (§4.3, 3dG3 W2/Q1 and 6QSf W3/Q4). The fits had lower median prediction error than interpolation on held-out math results (Table 2, 6QSf Q2), and Appendix H evaluates predicted accuracy changes between policies (Table 16).
  • Prompt level. On math, the shared shape learned from individual questions also predicts held-out aggregate accuracy (§6.3, 3dG3 W1 and 6QSf W1). Previous compression algorithms like LASER-D groups questions by current solve rate, yet questions with similar solve rates can respond differently to compression. For the same move to a shorter policy, a larger fitted sensitivity sis_i predicts a larger accuracy loss for question ii. As discussed with 21eY, this suggests estimating sis_i offline across trained policies and testing whether weaker length penalties for more sensitive questions preserve more accuracy at the same mean response length (Appendix I.4).

Changes in the revision

  • 21eY. We separated curve family selection from final evaluation (W2/Q3, Table 2 and Appendix I.1), pooled five controllers to test whether the shared direction extends beyond one sweep (W1, Appendix I.2), and checked stability with fewer endpoints (W2, Appendix I.3). We clarified the roles of mean length and policy quality (W1/Q1, §§3.2 and 6.3) and the scope of the matched-length and coding results (Q2, §4.2 and Appendix F.3).
  • 6QSf. We clarified why a single parameter sweep can yield different prompt responses (W1/W2, §§1–2 and 5.2), added the pooled analysis and interpolation comparisons (W2/Q1 and Q2, Table 2 and Appendices I.1–I.2), clarified length caps and coding evidence (Q3, §§4.2 and 7), and expanded guidance on choosing controllers and settings (W3/Q4, §4.3).
  • 3dG3. We clarified the link between aggregate and prompt-level results (W1, §§1–2), expanded guidance from curve slopes and matched-length comparisons (W2/Q1, §4.3), and explained how to cover the target length range (Q2, §4.3) and interpret the weaker coding evidence after math-only training (Q3, §4.2).

Changes are marked in blue.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.