Writing study: Training-Free Bayesian Filtering with Generative Emulators

ICML overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICML 2026 oral writing corpus

Thomas Savary, François Rozet, Gilles Louppe · 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: Domain-grounded methods & physical 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

24 physical pages

Before references / appendix

≈ 8.21 page-equivalents

Main figures / tables

4 / 2

Attached appendix figures / tables

8 / 0

Attached appendix span

≈ 11.91 page-equivalents

Before references / appendix: ≈ 34.21%References / intervening material: ≈ 16.17%Attached appendix: ≈ 49.63%Estimated layout envelopes, not text/artwork area. Unresolved appendix boundaries remain within the undivided post-body material.

Source-bound correction: Join wrapped printed heading text using adjacent source lines; original heading start coordinates 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	≈ 0.78 pp.	3.25%
2. Preliminaries	2	≈ 1.5 pp.	6.25%
3. Method	3	≈ 1.5 pp.	6.25%
4. Experiments	5	≈ 3 pp.	12.5%
5. Related work	8	≈ 0.5 pp.	2.08%
6. Conclusion	8	≈ 0.16 pp.	0.67%
7. Limitations & Future work	8	≈ 0.34 pp.	1.42%
Acknowledgments and Disclosure of Funding	9	≈ 0.11 pp.	0.46%
Impact Statement	9	≈ 0.05 pp.	0.21%
References	9	≈ 3.88 pp.	16.17%
A. Tweedie’s formulas	13	≈ 1 pp.	4.17%
B. Training-free ensemble methods	14	≈ 2 pp.	8.33%
C. Metrics	16	≈ 1 pp.	4.17%
D. Additional results	17	≈ 5 pp.	20.83%
E. Extension to stochastic interpolants	22	≈ 2 pp.	8.33%
F. Additional details on Algorithm 1	24	≈ 0.91 pp.	3.79%

All detected figure and table locations

Main text

• Figure 1 · p. 5
• Table 1 · p. 5
• Table 2 · p. 6
• Figure 2 · p. 6
• Figure 3 · p. 7
• Figure 4 · p. 7

Attached appendix

• Figure 5 · p. 17
• Figure 6 · p. 17
• Figure 7 · p. 18
• Figure 8 · p. 19
• Figure 9 · p. 20
• Figure 10 · p. 20
• Figure 11 · p. 21
• Figure 12 · p. 21

Argument and organization

1. 1 Introduction; 2 Preliminaries

   Defines filtering and the assumed access to a pretrained generative transition model.

2. 3 Method

   Separates the adapted particle-filter algorithm, posterior proposal and weight computation.

3. 4 Experiments; 5 Related Work; 6 Conclusion; 7 Limitations

   Moves from Lorenz and fluid systems to atmospheric assimilation, then records computational and approximation limits.

Problem, claim and contribution

Training-free is explicitly conditional on an available generative emulator; the small-system experiments still train their denoisers.

Problem statement, p.1; Sections 4.1-4.2, pp.5-6 · Main paper

Argument and section order

The proposal sampler and weight estimator have separate derivations, making the ideal particle-filter construction distinguishable from approximations used to implement it.

Sections 3.1-3.3, pp.3-4 · Main paper

Figures, captions and visual explanation

A four-panel fluid example aligns truth, observation and estimates; weather maps align method rows and forecast-time columns, followed by separate wind and temperature error curves.

Figure 2, p.6; Figures 3-4, p.7 · Main paper

Tables, comparisons and evidence

The realistic weather scenario generates noisy observations from an unseen ERA5 trajectory and explicitly reports failure to reach stable skill across all variables.

Section 4.3, p.7 · Main paper

Appendix and reproducibility

The supplement lists non-exact Gaussian posterior sampling, approximate weights and inflation-induced bias rather than leaving optimal-proposal terminology unqualified.

Appendix F.2, p.24 · Main paper

Clarity and writing conventions

The main limitation section names derivative-related VRAM cost and emulator dependence, while the related-work comparison acknowledges competitive Kalman variants.

Sections 5 and 7, p.8; Appendix F.1, p.24 · Main paper

What to reuse—and what not to infer

Transferable lesson: Give the ideal Bayesian construction and its practical approximations distinct names, and make the provenance of realistic-looking observations explicit.

Boundary: This is approximate filtering with pretrained-model dependence; simulated noisy reanalysis observations and one weather trajectory are not an operational deployment evaluation.

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-oriented scan of the main argument through physical p.8, with focused physical-page text inspection on 1, 3, 4, 5, 6, 7, 8, 9. Framing, section progression, mechanism/proof explanation, displays, comparison conditions and conclusions were inspected. This is not exhaustive technical reading, proof verification or reproduction.

Appendix / supplement

Appendix organization and main-text pointers were checked; focused text inspection of physical pages 13, 17, 24. Remaining appendix proofs and details were not exhaustively read. No separately linked supplement inspected.

Visually inspected PDF pages

6, 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 proceedings PDF, PMLR volume 306, retrieved 2026-10-07 · 24 pages