# VLA-MBPO (ICML 2026) — independent reproduction bundle

Reproduction of **"Towards Practical World Model-based Reinforcement Learning for
Vision-Language-Action Models"** (VLA-MBPO), ICML 2026, OpenReview `yKQ8GrwEhr`,
arXiv:2603.20607.

No official code/checkpoints were released; this is an independent reproduction
from the paper text. Trackio logbook is the primary record.

## Contents

| Path | What it is |
|---|---|
| `paper/paper.txt` | Extracted full text of the arXiv PDF (27 pages) |
| `paper/vla_mbpo_arxiv.pdf` | The paper PDF |
| `claim3_value_gap.py` | Claim 3 (theory): recomputes the value-gap constants, verifies Thm 4.2 from Lemma A.4 with sympy, and validates both bounds on 200 tabular MDPs |
| `claim3_figure.py` | Plotly figure for Claim 3 |
| `claims45_audit.py` | Claims 4 & 5: internal-consistency audit of Tables 2/3/5 |
| `claim2_world_model.py` | Claim 2/1: reduced-scale multi-view world model (IVD vs parallel, pretrained vs random, joint vs separate reward) on LIBERO-Object |
| `claim2_world_model_job.py` | Same script with a PEP 723 header, for `hf jobs uv run` |
| `outputs/claim3/` | Claim 3 results JSON + figure |
| `outputs/claims45/` | Claims 4 & 5 audit JSON |
| `outputs/claim2/` | Claim 2 results (populated by the GPU job) |

## Reproduce

```bash
# Claim 3 — theory (seconds, CPU)
python claim3_value_gap.py --seeds 200
python claim3_figure.py

# Claims 4 & 5 — table audit (instant, CPU)
python claims45_audit.py

# Claim 2 — world model (GPU; ~45 min, ~$0.60 on one L4)
hf jobs uv run --flavor l4x1 --timeout 2h --secrets HF_TOKEN \
  -e HF_HUB_ENABLE_HF_TRANSFER=1 -d claim2_world_model_job.py -- \
  --steps 8000 --batch-size 128 --dim 384 --layers 6 \
  --reward-weight 0.5 --workers 8 --push-to srilearns/vla-mbpo-repro
```

Data: `physical-intelligence/libero` (LIBERO-Object suite, tasks 20–29).
Base UMM used at full scale by the paper (not by this reduced-scale run):
`ByteDance-Seed/BAGEL-7B-MoT`.

## Findings (summary)

- **Claim 3 — reproduced.** All four case-study constants match exactly
  (4183.32, 18916.58, 1710.78, 400.00); Thm 4.2 is a correct consequence of
  Lemma A.4; both bounds hold on 200/200 random MDPs. Caveat: the headline
  "47×" model-error tightening is a step-vs-chunk **unit mismatch** — ~4.7× in
  matched units.
- **Claim 5 — refuted as literally stated.** Table 5 shows 2 of 12
  hyperparameters vary (Sample size 512→1280, Update-to-data 20→50 on
  LIBERO-Long); the paper's own appendix admits this. True in spirit (10/12
  constant across 5 suites and 2 platforms).
- **Claim 4 — not reproduced (infeasible); numbers internally consistent.** One
  typo found (Goal 92.8 should be 92.6); it does not affect any headline claim.
- **Claims 1 & 2 — mechanism code validated end-to-end; GPU run pending.**
- **Claim 6 — not verifiable** (no robots, no data, no numbers, no video).