Writing study: Mean Flow Policy with Instantaneous Velocity Constraint for One-step Action Generation

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Guojian Zhan, Letian Tao, Pengcheng Wang, Yixiao Wang, Yuxin Chen, Yiheng Li, Hongyang Li, Masayoshi Tomizuka, Shengbo Li · 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: Learning & optimization algorithms (primary) · Inference, algorithms & systems efficiency.

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

23 physical pages

Before references / appendix

≈ 9.83 page-equivalents

Main figures / tables

5 / 3

Attached appendix figures / tables

3 / 3

Attached appendix span

≈ 9.89 page-equivalents

Before references / appendix: ≈ 42.74%References / intervening material: ≈ 14.26%Attached appendix: ≈ 43%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	≈ 1.06 pp.	4.61%
2 PRELIMINARIES	2	≈ 0.89 pp.	3.87%
3 METHOD	3	≈ 3.31 pp.	14.39%
4 EXPERIMENTS	6	≈ 3.01 pp.	13.09%
5 RELATED WORK	9	≈ 0.74 pp.	3.22%
6 CONCLUSION	10	≈ 0.19 pp.	0.83%
7 ACKNOWLEDGMENT	10	≈ 0.08 pp.	0.35%
REFERENCES	10	≈ 3.27 pp.	14.22%
A THEORETICAL ANALYSIS ON THE MEAN VELOCITY POLICY IMPROVEMENT	14	≈ 4.12 pp.	17.91%
B SUPPLEMENTARY RESULTS	18	≈ 0.29 pp.	1.26%
C ENVIRONMENTS DESCRIPTION	18	≈ 1.59 pp.	6.91%
D VISUALIZATIONS	20	≈ 1.71 pp.	7.43%
E REPRODUCIBILITY STATEMENT	21	≈ 1.29 pp.	5.61%
F LLM USAGE DISCLOSURE	23	≈ 0.89 pp.	3.87%

All detected figure and table locations

Main text

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

Attached appendix

• Table 4 · p. 18
• Table 5 · p. 18
• Figure 6 · p. 19
• Figure 7 · p. 20
• Figure 8 · p. 21
• Table 6 · p. 22

Argument and organization

1. Introduction / Preliminaries

   Connects repeated action-generation cost to flow and RL definitions.

2. Mean velocity policy / Instantaneous constraint

   Separates imitating a target from selecting a useful target, then motivates a boundary constraint.

3. Mean flow RL / Experiments

   Combines the pieces in pseudocode before testing success, ablations and timing.

4. Appendices A–F

   Provides assumptions, numerical ablations, task descriptions, training settings and disclosure.

Problem, claim and contribution

The opening converts a practical control bottleneck into a specific question: whether expressive action distributions require repeated numerical integration. The proposed answer concerns a mean-velocity policy, rather than treating all generative policies as inherently slow.

§1, pp.1–2 · Main paper

Argument and section order

Two explicitly numbered questions separate how to imitate a target action from how to obtain that target. This prevents the reader from confusing the flow-fitting objective with the critic-guided best-of-N improvement procedure.

§3.1, pp.3–4 · Main paper

Figures, captions and visual explanation

Figure 2 uses interval arrows and tangent arrows to distinguish mean from instantaneous velocity. The training plots then retain common axes and mark offline versus online phases, helping readers connect algorithm phases to observed learning.

Figure 2, p.3; Figure 3, p.7 · Main paper

Tables, comparisons and evidence

The experiments isolate the added constraint, compare naive one-step variants and report success alongside training/inference cost. Table 3 importantly shows FQL slightly faster at inference, qualifying the prose's broad fastest-policy language.

§4.2, pp.8–9; Tables 2–3, p.9 · Main paper

Appendix and reproducibility

The proof explicitly states critic-error, smoothness and matching-error assumptions. The reproducibility table exposes best-of-N counts, one-million offline and online updates, and evaluation episodes, which are necessary to interpret the apparently one-step policy.

Appendix A, pp.14–18; Table 6, p.22 · Main paper

Clarity and writing conventions

The conclusion names JVP memory as a limitation and leaves physical-robot validation to future work. However, words such as guaranteed and significant need tighter qualification by approximation error and overlapping seed variability.

§3.3, p.6; §4.1, p.7; §6, p.10 · Main paper

What to reuse—and what not to infer

Transferable lesson: Separate target construction from target fitting; align each claimed mechanism with its own ablation, and report policy quality and deployment cost as independent outcomes.

Boundary: One-step generation still samples 16 or 32 candidates and scores them with a critic. Theorem 1 permits a negative improvement bound when fitting errors dominate. CPU timing disables JAX JIT and is not universal deployment latency. Main text reports the partial-IVC result as 0.45±0.15, whereas Table 4 gives 0.44±0.08; do not silently combine these. The main loss also reverses the stated time-order subscript.

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

PDF pp.1–10, §§1–7; references pp.10–13; page ranges verified with page-specific extraction because the first figure creates spurious form-feed characters

Appendix / supplement

PDF pp.14–23, Appendices A–F, including policy-improvement proof, all ablations, environment definitions, visual trajectories and hyperparameters

Visually inspected PDF pages

3, 8, 9

• official ICLR 2026 proceedings PDF · 23 pages