PaperlayerView exact revision on arXiv

Source-linked AI summary

Reasoning Fails Where Step Flow Breaks

Xiaoyu Xu, Yulan Pan, Xiaosong Yuan, Zhihong Shen, Minghao Su, Yuanhao Su, Xiaofeng Zhang

arXiv:2604.06695v1cs.AI

TL;DR

Long reasoning traces remain difficult to interpret because token-level saliency is noisy and hard to aggregate. The paper introduces Step-Saliency to map stepwise influence and uses its observed failures to guide StepFlow, which improves accuracy across math, science, and coding benchmarks without retraining.

  • Problem

    Token-level diagnostics are noisy and difficult to aggregate across long reasoning traces, limiting readable analysis of how information flows between steps.

    Source: §1 · p. 1

  • Method

    Step-Saliency pools token-level saliency into step-to-step maps, while StepFlow applies Odds-Equal Bridge and Step Momentum Injection as a test-time intervention.

    Source: §1 · p. 2

  • Results

    StepFlow consistently improves accuracy across math, science, and coding benchmarks on multiple LRMs without retraining.

    Source: §1 · p. 2, §7 · p. 9

  • Takeaways & Limitations

    The reported gains are largest on competition-style problems requiring long reasoning chains, while StepFlow primarily corrects arithmetic carry-forward and premise-forgetting errors.

    Source: §5.2 · p. 7

    The shallow/deep layer split is backbone-specific, chosen using a small held-out split, and not selected by a fully automatic method.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Large reasoning models (LRMs) that generate long chains of thought now perform well on multi-step math, science, and coding tasks. However, their behavior is still unstable and hard to interpret, and existing analysis tools struggle with such long, structured reasoning traces. We introduce Step-Saliency, which pools attention--gradient scores into step-to-step maps along the question--thinking--summary trajectory. Across several models, Step-Saliency reveals two recurring information-flow failures: Shallow Lock-in, where shallow layers over-focus on the current step and barely use earlier context, and Deep Decay, where deep layers gradually lose saliency on the thinking segment and the summary increasingly attends to itself and the last few steps. Motivated by these patterns, we propose StepFlow, a saliency-inspired test-time intervention that adjusts shallow saliency patterns measured by Step-Saliency via Odds-Equal Bridge and adds a small step-level residual in deep layers via Step Momentum Injection. StepFlow improves accuracy on math, science, and coding tasks across multiple LRMs without retraining, indicating that repairing information flow can recover part of their missing reasoning performance.

1 Introduction

Long reasoning traces are difficult to interpret because token-level saliency is noisy and hard to aggregate. The paper introduces step-level diagnosis and a test-time intervention targeting the observed information-flow failures.

1 Introduction

  • Step-Saliency pools token-level attention–gradient saliency within reasoning steps to make long-trace step-to-step influence easier to compare.The diagnostic addresses token maps that are dense, noisy, and difficult to aggregate across positions.Source: §1 · pp. 1–2
  • StepFlow targets the two failure patterns at test time using Odds-Equal Bridge in early layers and Step Momentum Injection in late layers.Odds-Equal Bridge preserves mass on the question and earlier steps; Step Momentum Injection carries a residual summary of the previous step forward.Source: §1 · p. 2
  • StepFlow improves accuracy across the evaluated backbones and benchmarks, with its largest gains on competition-style problems requiring long reasoning chains.The evaluation covers DeepSeek-R1-Distill, GPT-OSS-20B, and QwQ-32B-Preview across six benchmarks.Source: §1 · p. 2
  • Step-Saliency identifies Shallow Lock-in and Deep Decay as depth-wise information-flow patterns that reliably distinguish incorrect from correct traces.Shallow layers suppress earlier context; in deeper layers, thinking-segment saliency fades faster on incorrect outputs while the summary becomes more self-focused.Source: §1 · p. 2

2 Related Work

Related work describes capable but unstable large reasoning models, depth-dependent reasoning functions, and information-flow analyses that struggle to explain long traces. The paper situates its approach among attention- and saliency-based diagnostics.

2.1 Large Reasoning Models

2.2 Layer-wise Reasoning Behavior

2.3 Information Flow

  • Attention-based methods show which tokens models focus on but are not faithful explanations (Jain and Wallace, 2019); saliency scores are more tightly coupled to computation, yet are typically local to one prediction and often noisy.Saliency’s locality and noise make stable, sequence-level explanations of information flow through long reasoning traces difficult.Source: §2.3 · pp. 3–4

3 Motivation and Information Flow Analysis

Step-Saliency maps show distinct information-flow patterns in error traces: shallow attention becomes more local, while deep layers lose connections to earlier reasoning and the summary becomes more self-focused. The analysis aggregates token-level influence into step-to-step maps and compares these patterns across depth.

3.1 Saliency and step-level aggregation

  • Step-Saliency uses attention–gradient influence scores, normalizes them per query, and aggregates them across question, thinking-step, and summary segments into step-to-step maps.Mean pooling reduces token-level noise but loses some fine-grained signal; signed influence and separate positive/negative pooling are left for future work.Source: §3.1 · p. 4

3.2 Information flow among steps

  • Shallow Lock-in concentrates error-trace saliency on the current step and its neighbors, suppressing the question and early thinking steps; Deep Decay makes this loss of earlier-step saliency faster with depth.Correct summaries retain noticeable saliency to several earlier thinking steps, whereas error summaries focus mostly on themselves and the last few steps.Source: §3.2 · p. 4

3.3 Quantifying inter-step flow

  • Across models, error traces show higher shallow thinking self-intensity and earlier, higher summary self-intensity, alongside weaker links to earlier thinking steps.Figure 3 summarizes the layer-wise pattern across GPT-OSS-20B, R1-Distill-32B, and QwQ-32B.Source: §3.3 · p. 5, Figure 3 · p. 5

4 Method

StepFlow is a test-time intervention with two components: Odds-Equal Bridge adjusts shallow-layer attention to preserve earlier context, and Step Momentum Injection adds a residual connection across steps in deep layers.

4 Method

  • StepFlow combines Odds-Equal Bridge for shallow layers with Step Momentum Injection for deep layers as a test-time intervention guided by Step-Saliency patterns.The method reuses the paper’s step segmentation for the thinking and summary segments.Source: §4 · p. 5, §4.2 · p. 6

4.1 Odds-Equal Bridge (OEB)

  • Odds-Equal Bridge partitions past keys into current, bridge, and other groups, then shifts attention scores when bridge mass falls below its threshold to preserve a minimum share for earlier context.The projection leaves the other-group mass fixed and adjusts the group scores when the constraint is active.Source: §4.1 · pp. 5–6

4.2 Step Momentum Injection (SMI)

  • Step Momentum Injection adds a small residual from a prior thinking step to the first token of the next step in selected deep layers, maintaining a connection to earlier reasoning.The residual is added to the layer’s residual stream before the MLP block.Source: §4.2 · p. 6

5 Experiment

Across six benchmarks, StepFlow improves accuracy across the evaluated backbones, and its saliency maps show corresponding changes in information flow. Component, layer-coverage, and compute comparisons further characterize when the intervention helps.

5.1 Experimental Setup

  • Evaluation uses six math, science, and coding benchmarks, compares StepFlow with prompt-only, decode-level, internal-intervention, and unmodified-backbone baselines, and averages multiple samples for selected benchmarks.AIME24/25 and AMC23 use 16 sampled solutions per problem, GPQA-Diamond uses 8, and MATH-500 and LiveCodeBench use single samples; methods share decoding settings and run in one pass.Source: §5.1 · pp. 6–7

5.2 Main Results on Reasoning Benchmarks

  • +11.8 points on AIME25 for R1-Distill-32B (54.9→66.7), and +9.5 on LiveCodeBench for GPT-OSS-20B medium (70.0→79.5), exemplify gains across six backbones and benchmarks.Among 60 AIME24/25 problems where baselines failed but StepFlow succeeded, arithmetic carry-forward (34%) and premise forgetting (38%) account for 72% of corrections; conceptual errors (10–14%) are rarely fixed. StepFlow also runs within a single decoding run without multi-pass voting.Source: §5.2 · p. 7, Table 1 · p. 7

5.3 Effect on Information Flow

  • In representative traces and aggregated AIME25 error cases, StepFlow reduces shallow thinking and deep summary self-intensity while strengthening question-to-thinking and summary-to-thinking links.Figure 4 shows reduced shallow near-diagonal self-loops and broader deep links from the summary into thinking; Figure 5 reports these shifts across backbones.Source: §5.3 · p. 8, Figure 4 · p. 8

5.4 Ablation Analysis

  • On GPT-OSS-20B, OEB alone helps on all three benchmarks, SMI alone helps somewhat more—especially on GPQA-D and LiveCodeBench—and combining them gives the largest gains.The component pattern matches the diagnostic: shallow layers need less current-step lock-in, while deep layers need a stronger thinking-to-summary link.Source: §5.4 · p. 8

5.5 Layer Coverage and Decoding Cost

  • +13.8 points on medium and +14.2 points on hard LiveCodeBench problems accompany smaller gains on easy items near ceiling; bottom/top-quarter layer activation gives the best gains.Broader coverage over one third or one half of layers slightly reduces gains but remains above baseline; decoding efficiency is reported separately in Appendix C.2.Source: §5.4 · p. 8, §5.5 · pp. 8–9, Table 3 · p. 9

5.6 Compute-Normalized Comparison

  • At matched compute of approximately 1.35×, StepFlow achieves 5.7× the gain of longer generation; matching its accuracy requires self-consistency with k≈8 at 8.0× compute.StepFlow plus self-consistency (k=2) at approximately 2.7× compute outperforms self-consistency (k=4) at 4.0×, suggesting the approaches address orthogonal failure modes.Source: §5.6 · p. 9

6 Discussion

The discussion treats the causal link between diagnosed saliency patterns and improved accuracy as suggestive rather than proven. It also leaves open whether propagation failures are best classified as memory or reasoning errors.

Causal status of the diagnostic

  • The causal link between diagnosed saliency patterns and StepFlow’s improvements remains suggestive rather than formally proved.The authors cite component profiles, layer-band specificity, and selective correction patterns as evidence against a generic regularization account, while stopping short of causal proof.Source: §6 · p. 9

Memory versus reasoning

  • Whether failures to propagate correct intermediate results are memory or reasoning errors remains open, and the distinction warrants further investigation.The cited attempts to disentangle the categories rely on external annotations rather than model-internal definitions.Source: §6 · p. 9

7 Conclusion

The conclusion presents Step-Saliency as a step-level diagnostic of depth-wise information-flow failures and StepFlow as a test-time intervention guided by those patterns. Across multiple LRMs, StepFlow improves accuracy on math, science, and coding benchmarks without retraining.

7 Conclusion

  • Step-Saliency maps question–thinking–summary influence and reveals Shallow Lock-in and Deep Decay; guided by these patterns, StepFlow improves benchmark accuracy across multiple LRMs without retraining.The conclusion characterizes StepFlow as a lightweight test-time intervention that repairs information flow.Source: §7 · p. 9

Limitations

StepFlow’s layer ranges are backbone-specific and selected on a small held-out split, while its diagnostic aggregates token influence row by row and averages maps across layers or depth bands.

Limitations

  • StepFlow uses one concrete intervention design and does not explore alternatives such as head-level steering or value-space projections.The authors note that these alternatives could be analyzed within the same framework.Source: Limitations · p. 9
  • Its error taxonomy primarily covers information-propagation errors (72%), while finer-grained category breakdowns and benchmarks beyond AIME remain future work.Conceptual errors account for 10–14%.Source: Limitations · pp. 9–10

A.1 Computation Protocol

  • For each query position, the influence matrix is computed from gradients with respect to that attention row, and row results are stacked into a T × T matrix.Depth-collapsed maps average across all layers; layer-wise analysis averages the bottom or top quarter.Source: §A.1 · p. 12

A.2 Segmentation Robustness

The appendix tests segmentation and hyperparameter robustness, records how the interventions are configured, surveys comparison baselines, and quantifies decoding overhead.

A.2 Segmentation Robustness

  • Manual inspection found 99% boundary-detection accuracy, and even extreme boundary perturbations reduced accuracy by at most 2.0 points while StepFlow remained above baseline.The few observed detection errors involved repeated delimiters.Source: §A.2 · pp. 12–13, Table 6 · p. 13

B.1 Hyperparameter Selection

  • OEB improves over baseline across τmax ∈ [0.1, 0.5] with under 1 point of peak-to-trough variation; SMI accuracy spreads are typically 1–2 points within viable α intervals.AIME24-selected α⋆ is reused across other benchmarks without further tuning.Source: §B.1 · p. 13, Figure 6 · p. 13, Table 7 · p. 14

B.2 Reproducibility Summary

  • OEB groups question, thinking, and summary keys into current, bridge, and remaining regions; SMI mean-pools the prior step’s value states and injects a residual at the next step’s start.SMI triggers at detected thinking-step boundaries in selected deep layers.Source: §B.2 · p. 14, Table 8 · p. 14

C.1 Baselines

  • The baselines span prompt-only methods, decode-level budget control, and internal interventions, alongside the proposed StepFlow.Examples include PS+ and Hint-Infer, Budget Forcing, and attention or activation steering.Source: §C.1 · p. 14

C.2 Decoding Efficiency

  • On AIME25, StepFlow gains +11.8 versus +3.2 for two-sample self-consistency at equivalent compute, with estimated overhead of 30–37%.The paper frames the overhead as worthwhile for long-chain, high-accuracy tasks and single-pass correctness settings.Source: §C.2 · p. 15

C.3 Experiment Settings

The experiments use shared decoding settings and uncertainty tests, and compare StepFlow with longer generation and self-consistency at equivalent compute across three backbones.

C.3 Experiment Settings

  • All models use temperature 0.6, top-p 0.95, and up to 32,768 new tokens; sample counts vary by benchmark, with shared random seeds for baseline and StepFlow.AIME24, AIME25, and AMC23 use 16 samples; GPQA-Diamond uses 8; MATH-500 and LiveCodeBench use one.Source: §C.3 · p. 15

C.5 Full Compute-Normalized Comparison

  • At equivalent compute, StepFlow substantially outperforms longer generation and self-consistency across all three backbones, and the methods compose well.The comparison covers averaged AIME 24/25 accuracy.Source: Table 10 · p. 15, §C.5 · p. 15

C.4 Uncertainty and Significance

  • The authors report 95% bootstrap confidence intervals from 10,000 resamples and use paired bootstrap tests, plus McNemar’s test for multiple-choice datasets.Improvements are marked at p < 0.05.Source: §C.4 · p. 15

D Error Taxonomy

StepFlow corrections mainly involve cross-step information propagation, with examples showing retained intermediate results and counterexamples showing that better flow cannot fix a wrong premise.

D Error Taxonomy

  • Information-propagation errors account for 72% of corrections, while conceptual errors account for 10–14% and remain outside StepFlow’s scope.The taxonomy covers 60 AIME 24/25 problems where baselines failed but StepFlow succeeded.Source: D Error Taxonomy · pp. 15–16, Table 12 · p. 16

E.1 StepFlow Repair Example (GPQA-Diamond)

  • In the GPQA-Diamond example, the baseline drops an exponent contraction, whereas StepFlow combines the relations before substituting the period values.The example attributes this to earlier derivation steps remaining visible rather than lock-in on the latest computation.Source: §E.1 · p. 16

E.2 Arithmetic Carry-forward (AIME 24 P-7)

  • In the AIME carry-forward example, the baseline omits a previously derived +36 sin θ term, while StepFlow retains it in the aggregation.The omitted term changes the sine coefficient from −432 sin θ to −468 sin θ.Source: §E.2 · p. 16

E.3 Bad Case Analysis

  • Bad cases show that StepFlow can preserve a wrong premise or propagated wrong result; it repairs information flow, not the content generated.One example starts from an incorrect geometric restriction, and another drops an absolute value before later steps build on it.Source: §E.3 · pp. 16–17

Related papers

Loading 2604.06695v1…