acceptodds
Withdrawn from ICLR 2027

Coupling Sampling Width and Reasoning Depth for Efficient LLM Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Inference-time scaling improves LLM reasoning, but its accuracy gains come with a rapidly growing token budget. Existing adaptive methods optimize a single axis: width-based approaches stop sampling once answers agree, risking premature commitment to an agreed-upon hallucination, while depth-based approaches prune trajectories in isolation, truncating hard but recoverable reasoning. Treating the two axes as independent leaves the budget-quality trade-off unresolved. We propose Dual-Dimensional Consistency (DDC), which couples both axes through one shared path-quality signal computed during generation. On the width axis, a confidence-weighted Bayesian rule terminates sampling only when the leading answer holds a high-confidence absolute majority; on the depth axis, trend-aware stratified pruning distinguishes persistent degradation from transient uncertainty via a spectral analysis of the confidence trajectory, gated per query without tuned constants. The budget freed by early termination is thereby redirected to the trajectories still worth extending. Across five reasoning benchmarks and five LLMs from 1.7B to 32B, DDC cuts token consumption by over while matching or exceeding strong adaptive baselines.

Primary area:
foundation or frontier models, including LLMs
Keywords:
inference-time scaling, large language models, LLM reasoning, self-consistency, efficient inference

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_bc4jOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors address the inefficiency of existing inference-time scaling methods for Large Language Model (LLM) reasoning, where width-based and depth-based adaptive strategies are optimized independently, leading to either premature commitment to consensus hallucinations or erroneous pruning of recoverable reasoning paths. To resolve this limitation, the authors propose Dual-Dimensional Consistency (DDC), a unified framework that couples both axes through a shared path-quality signal computed during generation. Specifically, on the width axis, a confidence-weighted Bayesian sequential decision rule terminates sampling only when the leading answer achieves a high-confidence absolute majority; on the depth axis, a trend-aware stratified pruning mechanism distinguishes persistent degradation from transient uncertainty via spectral analysis of the confidence trajectory, gated per query without tuned constants. The budget freed by early termination is redirected to paths still worth extending. Extensive evaluations across five reasoning benchmarks and five LLMs ranging from 1.7B to 32B parameters demonstrate that DDC reduces token consumption by over 10× while matching or exceeding strong adaptive baselines.

Critical strengths
  1. The paper proposes a novel unified framework that couples inter-path sampling width and intra-path reasoning depth through a shared confidence signal, addressing the fundamental limitation of existing adaptive inference methods that optimize these two dimensions independently.

  2. The proposed DDC method achieves substantial efficiency gains, reducing token consumption by over 10× across five reasoning benchmarks and five LLMs ranging from 1.7B to 32B parameters, while matching or surpassing strong adaptive baselines and demonstrating a highly favorable accuracy-cost trade-off.

Critical weaknesses
  1. DDC heavily relies on token confidence to guide termination and pruning. However, confidence may be poorly calibrated, especially in RL-trained reasoning models, which raises concerns about the robustness of the method.

  2. The comparison with RASC is conducted only on Qwen3-4B, and since RASC achieves comparable results on GPQA-diamond, the additional benefit of intra-path pruning is not sufficiently demonstrated on other models.

Questions for authors
  1. How sensitive is DDC to confidence miscalibration, particularly in RL-trained reasoning models?

  2. Can the authors clarify the advantage of DDC over strong width-only baselines such as RASC under comparable compute budgets?

  3. Would the reported speedups hold in optimized serving frameworks such as vLLM or SGLang?

  4. Does DDC generalize to other tasks, such as code generation, and larger models (e.g., 70B+)?

Reviewer_HPW5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Dual-Dimensional Consistency (DDC), a training-free adaptive inference-time scaling framework that jointly controls the number of sampled reasoning trajectories (width) and the generation length of individual trajectories (depth).

The method consists of three main components:

  1. Multi-granular confidence estimation: Token-level confidence is aggregated into sliding-window and path-level confidence signals.
  2. Confidence-Weighted Bayesian Termination: A Beta-distribution-based sequential stopping rule accumulates confidence-weighted evidence for the leading answer and terminates sampling when the posterior probability of absolute-majority support exceeds a threshold.
  3. Trend-Aware Stratified Pruning: A hierarchy of confidence-based gates and a spectral trend score identify trajectories that should be retained or terminated.
Critical strengths

S1. Important and practically relevant problem. Efficient allocation of inference-time compute is increasingly important for reasoning models. The paper correctly identifies that adaptive sampling and trajectory-level pruning should not necessarily be treated as independent decisions.

S2. A coherent training-free framework. The integration of path-quality estimation, sequential termination, and dynamic pruning is conceptually appealing. It can potentially be incorporated into existing inference engines without additional model training.

S3. Strong empirical improvements in several settings. On Qwen3-4B/AIME25, the paper reports an accuracy improvement from 66.5% to 82.1%, while token consumption decreases from 41.0 to 1.5 units of 10710^7. This is a particularly compelling result.

Critical weaknesses

The method defines the accumulated evidence for candidate answer and constructs a Beta distribution using these confidence-weighted counts.

However, although fractional evidence updates can be interpreted within a generalized Bayesian framework, the paper does not justify why the unnormalized confidence scores should correspond to likelihood contributions of the stated magnitude. In particular, the posterior concentration depends on the absolute scale of wiw_i, not only on the relative amount of supporting evidence. Because the path-confidence proxy is not normalized to [0,1][0,1], this issue directly affects the interpretation and behavior of the stopping rule. I think this is the biggest issue of the paper. If authors can clarify this, i may change my score.

Questions for authors

What probabilistic model justifies mapping path confidence to effective sample counts? How sensitive are stopping decisions to the scale of the confidence weights?

Reviewer_imXTOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DDC, an efficient training-free inference-time scaling method that jointly adapts sampling width and reasoning depth using a shared path-confidence signal. A confidence-weighted Beta-posterior rule stops sampling once the leading answer holds a confident absolute majority. Trend-aware pruning, using percentile gates based on per-query Tukey thresholds, discards low-quality paths during generation. On five reasoning benchmarks and models from 1.7B to 32B, DDC reports about 8.5× fewer tokens than Self-Consistency with comparable or higher accuracy.

Critical strengths
  1. This paper proposed a training-free method for improving the reasoning efficiency of inference-time scaling methods, called DDC.
  2. Easy-to-follow motivation. The paper clearly explains why optimizing width or depth alone is insufficient: width-only methods can stop early on a wrong consensus, and depth-only methods can prune hard but recoverable paths. That makes the case for coupling the two intuitive.
Critical weaknesses
  1. The experimental evaluation is insufficiently comprehensive. Table 1 uses a budget of B=512, but the work does not provide systematic comparisons under smaller budgets such as (B=8,16), which are more practical and common settings. Since self-consistency often exhibits diminishing returns as the number of paths increases, using SC@512 as the primary reference may overstate the reported efficiency gains. Moreover, the comparison of final aggregation strategies is unfair. DDC uses confidence-weighted voting, but the same strategy could also be applied to SC.
  2. The end-to-end inference latency should be reported. Since DDC may require multiple sampling rounds, token savings may lead to higher user-perceived waiting time.
  3. Several important baselines are missing. In particular, Difficulty-Adaptive Self-Consistency [1] and CISC [2] should be included, as they closely relate to adaptive sampling.
    1. Make Every Penny Count: Difficulty-Adaptive Self-Consistency for Cost-Efficient Reasoning
    1. Confidence Improves Self-Consistency in LLMs
Questions for authors

N/A

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

Related papers

Discussion (0)

Sign in to comment.