Writing study: p-less Sampling: A Robust Hyperparameter-Free Approach for LLM Decoding

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Runyan Tan, Shuang Wu, Phillip Howard · Official ICLR 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: Inference, algorithms & systems efficiency (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

44 physical pages

Before references / appendix

≈ 10.27 page-equivalents

Main figures / tables

4 / 4

Attached appendix figures / tables

15 / 11

Attached appendix span

≈ 30.89 page-equivalents

Before references / appendix: ≈ 23.34%References / intervening material: ≈ 6.45%Attached appendix: ≈ 70.2%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

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	≈ 1.19 pp.	2.7%
2 RELATED WORK	2	≈ 0.96 pp.	2.18%
3 METHODOLOGY	3	≈ 2.93 pp.	6.66%
4 EXPERIMENTS	6	≈ 1.83 pp.	4.16%
5 ANALYSIS	8	≈ 2.2 pp.	5%
6 CONCLUSION	10	≈ 0.36 pp.	0.82%
7 REPRODUCIBILITY STATEMENT	11	≈ 0.16 pp.	0.36%
REFERENCES	11	≈ 2.84 pp.	6.45%
A HUMAN EVALUATION	14	≈ 1 pp.	2.27%
B ADDITIONAL PROPERTIES OF p-LESS AND p-LESSNORM SAMPLING	15	≈ 5.65 pp.	12.84%
C ADDITIONAL EXPERIMENTAL DETAILS AND RESULTS	20	≈ 24.25 pp.	55.11%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Table 1 · p. 7
• Figure 2 · p. 7
• Table 2 · p. 8
• Table 3 · p. 8
• Table 4 · p. 9
• Figure 3 · p. 9
• Figure 4 · p. 10

Attached appendix

• Figure 5 · p. 19
• Figure 6 · p. 20
• Figure 7 · p. 21
• Figure 8 · p. 22
• Table 5 · p. 23
• Figure 9 · p. 23
• Figure 10 · p. 24
• Table 6 · p. 24
• Figure 11 · p. 25
• Table 7 · p. 26
• Table 8 · p. 26
• Table 9 · p. 27
• Table 10 · p. 28
• Table 11 · p. 28
• Table 12 · p. 29
• Table 13 · p. 30
• Figure 12 · p. 30
• Table 14 · p. 30
• Figure 13 · p. 31
• Figure 14 · p. 31
• Figure 15 · p. 32
• Figure 16 · p. 32
• Figure 17 · p. 33
• Table 15 · p. 33
• Figure 18 · p. 43
• Figure 19 · p. 43

Argument and organization

1. Sampling problem and collision-probability threshold

   Motivate and define the small algorithm before alternative interpretations

2. Temperature sweeps, efficiency and diversity

   Test robustness rather than only an optimal operating point

3. Proofs, protocols and generated examples

   Expose implementation and qualitative failure details

Problem, claim and contribution

The title's hyperparameter-free claim concerns the truncation threshold, computed as the sum of squared token probabilities. Temperature still changes the distribution and remains an explicit experimental and deployment choice.

Sections 3.1 and 4.1, pages 3–6 · Main paper

Argument and section order

A three-step threshold/filter/renormalize recipe precedes Rényi-entropy interpretations. This makes the mechanism reproducible without requiring readers to accept every theoretical motivation; the executable snippet arrives much later in the appendix.

Page 4, equations 2–4; page 32, Figure 15 · Main paper

Figures, captions and visual explanation

Figure 1 holds logits fixed and varies temperature, with probability bars, cumulative probability and threshold lines. Figure 16 complements this conceptual figure with binned CPU-time distributions rather than only a headline speed percentage.

Page 2, Figure 1; page 32, Figure 16 · Main paper

Tables, comparisons and evidence

Table 1's AUC integrates accuracy over temperature; it is not ROC-AUC or single-temperature accuracy. Table 3's 0.01942 versus 0.02497 seconds measures token sampling, while Appendix C.11 profiles CPU time separately; neither establishes a universal end-to-end speedup.

Sections 4.2 and 5.1; Tables 1, 3 and 15 · Main paper

Appendix and reproducibility

The appendix adds baseline hyperparameter sweeps, deterministic-decoding comparisons and explicit wrong-answer traces. Seed counts differ: Llama-2 uses three seeds, while Mistral and Llama-3 use one. Human evaluation includes author annotators, which is transparently disclosed.

Appendices A, C.3–C.6 and C.13 · Main paper

Clarity and writing conventions

Probability-of-correctness language depends on treating model probabilities as the unknown target distribution. Entropy intuitions should not be overstated: an entropy bound alone does not prove universal monotonicity across unrelated distributions, and confidence does not ensure a correct answer.

Sections 3.1–3.4; Appendix C.13, pages 42–44 · Main paper

What to reuse—and what not to infer

Transferable lesson: For a compact inference method, pair an implementable recipe with controlled operating-range plots, qualitative failures and explicit definitions of every aggregate metric.

Boundary: Do not transfer the headline to temperature-free decoding, guaranteed correctness, or optimized-serving speed. Theoretical interpretations and author-involved human preferences need their stated assumptions and limitations.

Apply this lesson using the argument and evidence worksheet.

Inspection record and sources

Checked

2026-09-19

Reading status

Full text + selected visuals inspected

Main text

Pages 1–11, Sections 1–7; references 11–13

Appendix / supplement

Pages 14–44, Appendices A–C including bounds, all results, prompts, examples and failure cases

Visually inspected PDF pages

2, 32

• official ICLR 2026 proceedings PDF · 44 pages