Writing study: Hallucination Begins Where Saliency Drops

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Xiaofeng Zhang, Yuanchao Zhu, Chaochen Gu, Xiaosong Yuan, Qiyan Zhao, Jiawei Cao, Barrett Tang, Sinan Fan, Yaomin Shen, Chen Shen, Hao Tang · 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) · Empirical analysis, diagnosis & auditing.

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

22 physical pages

Before references / appendix

≈ 9.11 page-equivalents

Main figures / tables

4 / 3

Attached appendix figures / tables

10 / 3

Attached appendix span

≈ 8.34 page-equivalents

Before references / appendix: ≈ 41.41%References / intervening material: ≈ 20.68%Attached appendix: ≈ 37.91%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	≈ 2.55 pp.	11.59%
2 ANALYSIS AND MOTIVATION	4	≈ 0.56 pp.	2.55%
3 METHODOLOGY	4	≈ 1.75 pp.	7.95%
4 EXPERIMENTS	6	≈ 2.9 pp.	13.18%
5 RELATED WORK	9	≈ 0.31 pp.	1.41%
6 CONCLUSION	9	≈ 0.14 pp.	0.64%
7 ACKNOWLEDGMENTS	9	≈ 0.24 pp.	1.09%
REFERENCES	10	≈ 4.55 pp.	20.68%
A RELATED WORK	14	≈ 3.99 pp.	18.14%
B ETHICS STATEMENT	18	≈ 0.18 pp.	0.82%
C REPRODUCIBILITY STATEMENT	18	≈ 0.37 pp.	1.68%
D MORE EXPERIMENTS IN REBUTAL	19	≈ 3.8 pp.	17.27%

All detected figure and table locations

Main text

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

Attached appendix

• Figure 6 · p. 14
• Figure 5 · p. 15
• Figure 7 · p. 16
• Figure 8 · p. 17
• Table 4 · p. 17
• Figure 9 · p. 18
• Figure 10 · p. 19
• Table 5 · p. 19
• Table 6 · p. 20
• Figure 11 · p. 20
• Figure 12 · p. 21
• Figure 13 · p. 22
• Figure 14 · p. 22

Argument and organization

1. Saliency observation

   Motivates interventions using gradient-weighted attention diagnostics.

2. SGRS and LocoRE

   Separates token rejection from forward-only local attention reinforcement.

3. Benchmark and diagnostic evidence

   Combines broad comparisons, ablations and counterexamples.

Problem, claim and contribution

The observation-to-intervention narrative is easy to follow, but the early claim of a direct causal link is stronger than the documented evidence. The conclusion's correlation wording and appendix's high-saliency failures are more appropriately bounded.

Contributions, p.3; §6, p.9; AppendixD.2, pp.20–21 · Main paper

Argument and section order

Separate subsections and algorithms distinguish backward-pass saliency rejection from a forward-only attention modification. This is important because training-free does not mean gradient-free, and the modules have different memory and latency costs.

§3 and Algorithms1–2, pp.4–6 · Main paper

Figures, captions and visual explanation

Paired attention and saliency maps illustrate the diagnostic, while Figure11 deliberately shows a hallucinated token with high saliency. That counterexample is valuable, although its adjacent traffic-cone description does not match the displayed bench example.

Figures1–2, pp.2–3; Figure11, p.20 · Main paper

Tables, comparisons and evidence

The broad tables cannot be copied without auditing: LLaVA LocoRE CHAIRI is 11.2 in Table1 but 10.2 in Table2. InternVL-13B full-method CHAIRI worsens from12.7 to14.0 despite a positive-looking improvement annotation.

Tables1–2, p.7 · Main paper

Appendix and reproducibility

Memory limitations and per-token overhead are disclosed separately from qualitative maps. Rebuttal analyses add approximately12,000 labeled tokens, but scaling a computed saliency diagnostic is not a sufficiently specified causal intervention without identifying the modified forward computation.

AppendixA.0.1, p.14; A.4, pp.16–17; D.1, pp.19–20 · Main paper

Clarity and writing conventions

Hyperparameter recommendations need consistency: the prose recommends beta1.2, while Table3 and reproducibility specify0.15 or0.20 and beta1.0 performs badly. A compact implementation checklist would prevent this contradiction from propagating to readers.

§4.5, pp.8–9; AppendixC, pp.18–19 · Main paper

What to reuse—and what not to infer

Transferable lesson: A diagnostic paper benefits from observation–intervention–failure-case structure, but causal language, implementation semantics and every reported improvement must survive a consistency audit.

Boundary: Do not present saliency as a universal hallucination detector or the full framework as gradient-free. Clarify the target loss, output-token offsets, selected layers, and whether LocoRE scales logits or normalized attention before implementation. Some table values and figure descriptions conflict.

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

pp.1–9, abstract and §§1–7; references pp.10–14

Appendix / supplement

pp.14–22, appendices A–D including efficiency, failures, reproducibility and rebuttal analyses

Visually inspected PDF pages

7, 20

• official ICLR 2026 proceedings PDF · 22 pages