Writing study: TD-JEPA: Latent-predictive Representations for Zero-Shot Reinforcement Learning

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Marco Bagatella, Matteo Pirotta, Ahmed Touati, Alessandro Lazaric, Andrea Tirinzoni · 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) · Theory & formal foundations.

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

34 physical pages

Before references / appendix

≈ 10.11 page-equivalents

Main figures / tables

4 / 1

Attached appendix figures / tables

9 / 5

Attached appendix span

≈ 19.89 page-equivalents

Before references / appendix: ≈ 29.74%References / intervening material: ≈ 11.76%Attached appendix: ≈ 58.5%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	≈ 1.22 pp.	3.59%
2 PRELIMINARIES	2	≈ 0.97 pp.	2.85%
3 LATENT-PREDICTIVE TEMPORAL-DIFFERENCE REPRESENTATIONS	3	≈ 1.83 pp.	5.38%
4 THEORETICAL ANALYSIS	5	≈ 1.64 pp.	4.82%
5 RELATED WORK	7	≈ 1.23 pp.	3.62%
6 EXPERIMENTS	8	≈ 2.22 pp.	6.53%
7 CONCLUSION	10	≈ 0.41 pp.	1.21%
REFERENCES	11	≈ 4 pp.	11.76%
A EXTENDED RELATED WORK	15	≈ 1 pp.	2.94%
B PROOFS	16	≈ 3.63 pp.	10.68%
C THEORETICAL ANALYSIS UNDER RELAXED ASSUMPTIONS	19	≈ 2.77 pp.	8.15%
D ADDITIONAL RESULTS	22	≈ 7.84 pp.	23.06%
E IMPLEMENTATION DETAILS	30	≈ 4.66 pp.	13.71%

All detected figure and table locations

Main text

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

Attached appendix

• Table 2 · p. 23
• Figure 5 · p. 24
• Figure 6 · p. 25
• Figure 7 · p. 26
• Figure 8 · p. 26
• Figure 9 · p. 26
• Figure 10 · p. 27
• Table 3 · p. 28
• Table 4 · p. 28
• Figure 11 · p. 29
• Figure 12 · p. 29
• Figure 13 · p. 30
• Table 5 · p. 31
• Table 6 · p. 33

Argument and organization

1. Bridge two vocabularies

   Preliminaries distinguish state representations, task representations and successor features before combining them.

2. Derive before specializing

   Moves from a Monte Carlo latent-prediction objective to an offline temporal-difference loss, then separates the encoders and constructs zero-shot policies.

3. Separate guarantees and empirical questions

   An explicitly idealized analysis precedes experiments organized around performance, prediction horizon, encoder asymmetry and adaptation.

Problem, claim and contribution

The introduction locates the gap along three independent axes: single-step prediction, single-policy learning and dependence on on-policy data. It then names the components that address them. This is a useful way to explain a combination of established ideas without pretending each component is new.

Sections 1–3, pages 1–5 · Main paper

Argument and section order

The theory labels its tabular, linear-predictor idealization before presenting results. Appendix C explains that removing symmetry needs changed sampling dynamics, not merely a stronger claim for the implemented algorithm. Keeping this boundary visible helps readers avoid transferring a theorem to a different method.

Section 4; Appendix C.2–C.3, pages 21–22 · Main paper

Figures, captions and visual explanation

Figure 1 links three views of the same idea: the encoder/predictor graph, trajectories in a maze, and colored latent trajectories whose stars denote predicted barycenters. Matching colors and symbols carry the explanation across panels; the caption explicitly defines the stars, circles and arrows.

Figure 1, page 2 · Main paper

Tables, comparisons and evidence

The main result table retains domain-level results instead of only an aggregate win. A separate probability-of-improvement heatmap addresses consistency, and the text acknowledges that BYOL-γ is slightly better on OGBench RGB. Experimental question headings connect each ablation to a design decision.

Table 1 and Figures 2–3; Section 6, pages 8–9 · Main paper

Appendix and reproducibility

The appendix is an operational companion: proofs, additional baseline controls, speed on common hardware, low-coverage data, architecture tables, tuning grids, evaluation seeds and pseudocode. Essential exceptions such as flow-based behavior regularization are there, so the main evaluation should point readers to the exact subsection.

Appendices B–E; Tables 4–6; E.6–E.8 · Main paper

Clarity and writing conventions

Question-led experiment paragraphs make a long evaluation navigable. However, a final consistency pass remains necessary: simplified Listing 1 adds a positive actor objective to a minimized loss, whereas Algorithm 1 minimizes its negative; some appendix figure references also remain section-like placeholders.

Section 6; Algorithm 1, page 5; Listing 1, page 34; Appendix D.5 and D.8 · Main paper

What to reuse—and what not to infer

Transferable lesson: Build a bridge between neighboring research communities by defining their objects before combining them. Present an intuition figure, an executable algorithm, qualified theory and question-led experiments as complementary views of one contribution.

Boundary: Zero-shot means a downstream reward is encoded from rewarded samples without policy retraining, not that the system receives no task information. Theoretical guarantees concern idealized dynamics; the practical offline correction and adapted BYOL baselines must remain visible. Do not copy the simplified actor-loss sign without checking the actual implementation.

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

Sections 1–7, including all definitions, four main theorems, algorithm and experiments, pages 1–10.

Appendix / supplement

Appendices A–E, pages 15–34: related work, proofs, relaxed assumptions, additional experiments and implementation/pseudocode.

Visually inspected PDF pages

2, 9

• Official ICLR 2026 proceedings manuscript with integrated appendices A–E · 34 pages