Writing study: Towards Long-Horizon Interpretability: Efficient and Faithful Multi-Token Attribution for Reasoning LLMs

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Wenbo Pan, Zhichao Liu, Xianlong Wang, Yu Haining, Xiaohua Jia · Official ICML 2026 oral record

Observed writing choices, not acceptance rules. This study examines the inspected manuscript version. It does not establish why the paper was accepted or that its presentation choices should be copied unchanged.

Figure, table and section ledger

Editorial contribution type: Efficiency, inference & systems (primary).

Layout extraction, not a required paper template. Counts are distinct caption labels; a multi-panel figure counts once. Section spans are estimated from heading positions, including floats and whitespace. Method and limitations.

Canonical PDF

26 physical pages

Before references / appendix

≈ 8.38 page-equivalents

Main figures / tables

5 / 5

Attached appendix figures / tables

2 / 5

Attached appendix span

≈ 15.91 page-equivalents

Before references / appendix: ≈ 32.23%References / intervening material: ≈ 6.58%Attached appendix: ≈ 61.19%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Small-cap FlashTrace has shorter glyph boxes than ordinary section headings. Restore section4 and the following sections using source glyph positions; include the appendix contents on p.11, not just sectionA on p.13. Newly detected unnumbered impact/singular-acknowledgment statements are retained from the size-aware source extraction; authored heading corrections are unchanged.

Section-by-section lengths and source pages

The title/abstract/front matter precede the first numbered heading. A span ends at the next detected top-level heading, not at the last sentence; end-page whitespace is included. References and statements are listed separately. These are not measurements of text-only or artwork area.

Top-level heading	Starts on PDF page	Estimated span	Share of entire PDF
1. Introduction	1	≈ 0.79 pp.	3.04%
2. Related Work	2	≈ 0.31 pp.	1.19%
3. Multi-Token Attribution	2	≈ 1.36 pp.	5.23%
4. FlashTrace	4	≈ 2.46 pp.	9.46%
5. Experiments	6	≈ 2.38 pp.	9.15%
6. Conclusion	9	≈ 0.1 pp.	0.38%
Acknowledgements	9	≈ 0.08 pp.	0.31%
Impact Statement	9	≈ 0.09 pp.	0.35%
References	9	≈ 1.71 pp.	6.58%
Appendix: contents	11	≈ 2 pp.	7.69%
A. Detailed Algorithm of FlashTrace	13	≈ 2.13 pp.	8.19%
B. Complexity Analysis of FlashTrace	15	≈ 2.34 pp.	9%
C. Efficiency Experiment Details	17	≈ 0.53 pp.	2.04%
D. Baseline Methods	18	≈ 1.67 pp.	6.42%
E. Dataset Construction	19	≈ 1.61 pp.	6.19%
F. Evaluation Details	21	≈ 1.09 pp.	4.19%
G. Case Study Visualizations	22	≈ 0.11 pp.	0.42%
H. Ablation: Effect of Recursive Hops	22	≈ 0.33 pp.	1.27%
I. Pilot Study Details	22	≈ 1.2 pp.	4.62%
J. Additional Related Work	24	≈ 0.26 pp.	1%
K. Additional Faithfulness Experiments	24	≈ 0.74 pp.	2.85%
L. Memory and Runtime Trade-offs	25	≈ 0.29 pp.	1.12%
M. Alternative Span Aggregation Strategies	25	≈ 0.83 pp.	3.19%
N. Qualitative Failure Analysis	26	≈ 0.79 pp.	3.04%

All detected figure and table locations

Main text

• Figure 1 · p. 1
• Figure 2 · p. 3
• Figure 3 · p. 5
• Table 1 · p. 7
• Figure 4 · p. 7
• Table 2 · p. 8
• Figure 5 · p. 8
• Table 3 · p. 8
• Table 4 · p. 8
• Table 5 · p. 9

Attached appendix

• Table 6 · p. 17
• Figure 6 · p. 22
• Table 7 · p. 22
• Figure 7 · p. 23
• Table 8 · p. 24
• Table 9 · p. 25
• Table 10 · p. 25

Argument and organization

1. Problem and contribution

   Pairs an efficiency bottleneck with a quality failure in which intermediate reasoning absorbs attribution, motivating two corresponding mechanisms rather than speed alone.

2. Argument sequence

   A diagnostic pilot precedes mechanism definitions, then experiments separate runtime, evidence recovery and perturbation-based faithfulness before comparing an exhaustive attribution control.

3. Evidence and verification

   The exhaustive-rollout control exposes an approximation trade-off, and later checks relax correctness filtering. Main comparisons use selected correct generations; appendix hardware, baseline granularity and cache assumptions qualify the 130-fold speed claim.

Problem, claim and contribution

Pairs an efficiency bottleneck with a quality failure in which intermediate reasoning absorbs attribution, motivating two corresponding mechanisms rather than speed alone.

Introduction and diagnostic setup, pp. 1-4 · Main paper

Argument and section order

A diagnostic pilot precedes mechanism definitions, then experiments separate runtime, evidence recovery and perturbation-based faithfulness before comparing an exhaustive attribution control.

Sections 3-5, pp. 3-8 · Main paper

Figures, captions and visual explanation

Hop heatmaps include a signed difference map, while scaling panels vary input and generation length separately and mark OOM. The adjacent table retains metric-specific losses instead of displaying only a speed multiplier.

Figure 3, p. 5; Figure 4 and Table 1, p. 7; both PNGs inspected · Main paper

Tables, comparisons and evidence

The exhaustive-rollout control exposes an approximation trade-off, and later checks relax correctness filtering. Main comparisons use selected correct generations; appendix hardware, baseline granularity and cache assumptions qualify the 130-fold speed claim.

Table 3, p. 8; Appendices C-E, pp. 17-21; K, p. 24 · Main paper

Appendix and reproducibility

The full cost retains a scalar target-by-context term and excludes explicit attention maps from working-memory notation; a cached-versus-recomputed experiment makes the practical memory-latency trade-off visible.

Appendix B, pp. 15-17; L and Table 9, p. 25 · Main paper

Clarity and writing conventions

Section 4.1 carefully calls the score informational contribution rather than direct causal effect, and the final failure analysis restricts interpretation to coarse localization for diffuse evidence. Some captions overstate: Table 3 says hours-to-seconds despite displayed 11.2-to-0.72 seconds.

Section 4.1, p. 4; Table 3, p. 8; Appendix N, p. 26 · Main paper

What to reuse—and what not to infer

Transferable lesson: Connect each efficiency mechanism to a distinct bottleneck, then report approximation loss, working-memory exclusions and actual hardware alongside the headline acceleration.

Boundary: Attribution faithfulness under deletion is not complete causal understanding; O(N) shorthand omits the retained scalar aggregation term and cache costs, and OOM methods plotted without measured faithfulness should not be interpreted as valid quality comparisons.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-10-07

Reading status

Writing-focused manuscript scan + selected visuals inspected

Main text

Writing-focused distributed manuscript scan, not exhaustive technical/proof validation. Sections 1-6, pp. 1-9: long-reasoning attribution problem, pilot diagnostic, span aggregation and recursion, quality/efficiency evaluation, approximation controls and conclusion.

Appendix / supplement

Writing-focused appendix scan; code, data and external assets not executed. Appendix cover/contents, pp. 11-12, and Appendices A-N, pp. 13-26: algorithm, full complexity and cache accounting, hardware/baseline adaptations, data/metrics, hop controls, unfiltered checks, memory trade-offs and failure analysis.

Visually inspected PDF pages

5, 7

This is a writing-focused scan of the manuscript and the stated supplementary scope: framing, organization, visual explanation, evidence presentation and wording. It is not a line-by-line proof audit, reproduction of results or exhaustive scientific review.

• Published ICML 2026 proceedings PDF, PMLR volume 306; retrieved 2026-10-07. Public OpenReview decision notes were not independently reverified. · 26 pages