Writing study: Optimistic Task Inference for Behavior Foundation Models

ICLR overviewCorpus atlasBy contribution typeVisual galleryRules & templateWriting patternsPaper studies

← ICLR 2026 oral writing corpus

Thomas Rupf, Marco Bagatella, Marin Vlastelica, Andreas Krause · 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

≈ 9.35 page-equivalents

Main figures / tables

6 / 0

Attached appendix figures / tables

14 / 2

Attached appendix span

≈ 21.89 page-equivalents

Before references / appendix: ≈ 27.5%References / intervening material: ≈ 8.12%Attached appendix: ≈ 64.38%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	≈ 1.05 pp.	3.09%
2 PRELIMINARIES	2	≈ 0.95 pp.	2.79%
3 OPTIMISTIC TASK INFERENCE FOR BEHAVIOR FOUNDATION MODELS	3	≈ 2.19 pp.	6.44%
4 PRACTICAL ALGORITHM	5	≈ 0.88 pp.	2.59%
5 EXPERIMENTS	6	≈ 2.25 pp.	6.62%
6 RELATED WORK	8	≈ 0.89 pp.	2.62%
7 CONCLUSION	9	≈ 0.5 pp.	1.47%
ACKNOWLEDGEMENTS	10	≈ 0.11 pp.	0.32%
REFERENCES	10	≈ 2.75 pp.	8.09%
A PROOFS	13	≈ 10.3 pp.	30.29%
B ADDITIONAL EXPERIMENTS	23	≈ 4.7 pp.	13.82%
C FULL EXPERIMENTAL RESULTS	28	≈ 0.51 pp.	1.5%
D TRAINING PROTOCOL	28	≈ 3.1 pp.	9.12%
E EXPERIMENT PROTOCOL	31	≈ 0.93 pp.	2.74%
F INFERENCE HYPER-PARAMETERS	32	≈ 2.36 pp.	6.94%

All detected figure and table locations

Main text

• Figure 1 · p. 2
• Figure 2 · p. 7
• Figure 3 · p. 7
• Figure 4 · p. 8
• Figure 5 · p. 8
• Figure 6 · p. 9

Attached appendix

• Table 1 · p. 23
• Figure 7 · p. 24
• Figure 8 · p. 24
• Figure 9 · p. 25
• Figure 10 · p. 25
• Figure 11 · p. 26
• Figure 12 · p. 27
• Figure 13 · p. 27
• Figure 14 · p. 28
• Figure 15 · p. 29
• Figure 16 · p. 30
• Figure 17 · p. 31
• Figure 18 · p. 32
• Figure 19 · p. 33
• Figure 20 · p. 34
• Table 2 · p. 34

Argument and organization

1. Offline limitation → online task inference

   Turns a practical labeling problem into a precise inference setting before introducing optimism.

2. Preliminaries → regret analysis → practical method

   Separates idealized successor-feature guarantees from approximate deployment choices.

3. Question-led experiments

   Tests learning speed, informativeness, update frequency and changing tasks separately.

Problem, claim and contribution

The opening contrasts collecting labels for an existing dataset with actively requesting rewards during deployment; this identifies the changed information-access assumption, not merely an improved return.

Sections 1–2; Figure 1, p. 2 · Main paper

Argument and section order

The mathematical statement names perfect successor features, linear rewards and exact optimization; practical random shooting and per-step updates are introduced afterward rather than silently folded into the theorem.

Sections 3–4; Appendix A.1, pp. 13–14 · Main paper

Figures, captions and visual explanation

The first diagram shares one pretrained model across offline and online branches, making the moving confidence set the visual difference. Paired learning curves later distinguish return during collection from the usefulness of collected labels.

Figure 1, p. 2; Figures 2–3, p. 7 · Main paper

Tables, comparisons and evidence

Relative performance uses a same-model oracle inferred from 50,000 labeled pretraining samples, not the environment optimum. Shaded ranges are min–max over three training seeds, not confidence intervals.

Section 5, pp. 6–7; Appendix E, p. 31 · Main paper

Appendix and reproducibility

Additional studies probe misspecified successor features, nonlinear and noisy rewards, explicit label requests and inference cost; full per-task curves make aggregate improvements inspectable.

Appendices B–F, pp. 23–34 · Main paper

Clarity and writing conventions

Experiment headings are concrete questions answered immediately by a figure and a qualified conclusion. The theory/practice gap is revisited with episode-only versus per-step updates.

Sections 5.1–5.5, pp. 7–9 · Main paper

What to reuse—and what not to infer

Transferable lesson: Organize evaluation around the information assumptions your method changes, and separate theorem conditions, practical approximations and measured behavior.

Boundary: Near-oracle performance is relative to the pretrained behavior model; the episodic regret result does not directly certify the implemented per-step heuristic. Label-saving experiments also trade off additional environment interactions.

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 1–10, Sections 1–7, including assumptions, algorithm and all experiments.

Appendix / supplement

Pages 13–34, Appendices A–F: complete proofs, additional experiments, full per-task results, training and inference protocols.

Visually inspected PDF pages

2, 7

• ICLR 2026 official proceedings PDF, retrieved 2026-09-19 · 34 pages