TD-MPC-Glass, Part 36: Reproducing InFOM on OGBench — Cube Reproduces; AntMaze Is Out of InFOM's Benchmark
The week started here: reproduce Compositional Planning with Jumpy World Models and its InFOM base on a common OGBench task set, in one table. Many detours later (Parts 23-35), this note closes that original thread — honestly, including where it doesn’t reproduce.
What we ran
The InFOM base (flow-occupancy; Zheng et al.) on OGBench, HPs from the repo README
(expectile 0.95, kl_weight 0.05, alpha 30), 1.5M steps, eval = OGBench episode.success (n=50 eval episodes
per checkpoint). We reproduced the base method, not the CompPlan planner (the paper’s actual contribution).
A correction up front (this note was revised). An initial version treated antmaze as a reproduction
failure against “paper” numbers of 0.49/0.85. That comparison is invalid and is corrected below: a
dataset investigation found that InFOM does not benchmark antmaze at all — its -ft- finetuning datasets
are generated locally for 8 manipulation envs only (cube-single/double, scene, puzzle-4x4 and visual
variants); antmaze appears nowhere in the InFOM repo (data_gen_scripts/ has no antmaze path) or paper. The
0.49/0.85 are CompPlan/CRL figures for a different method. So antmaze is out of InFOM’s benchmark, our
antmaze run is an out-of-benchmark extension on a fallback dataset, and only the cube result is an actual
InFOM reproduction.
The reproduction (InFOM’s actual benchmark)
| task | in InFOM benchmark? | our InFOM base (per-seed best success) | reproduces? |
|---|---|---|---|
| cube-single (task1) | yes (manipulation) | 0.96 / 0.98 (+ one 0.16 outlier seed) | yes — clean InFOM-base reproduction |
| antmaze-medium (task1) | no (not in InFOM) | 0.18 / 0.20 / 0.08 (fallback dataset) | n/a — out-of-benchmark extension, not comparable |
(InFOM paper: arXiv 2506.08902 — manipulation only, no antmaze. The 0.49/0.85 antmaze figures come from CompPlan arXiv 2602.19634 Table 1 for CRL/CompPlan, not InFOM.)
Cube — clean reproduction, with honest seed-variance
Two of three cube seeds reach 0.96 and 0.98 best success — at or above the paper’s CompPlan number (0.86) and far above the CRL base (0.28). So InFOM base on cube reproduces. But the third seed collapsed to 0.16 (latest 0.0) — a real instability we are not hiding: 1-in-3 seed failure is exactly the kind of seed-variance the campaign’s stability axis (Part 35) warns must be measured, not assumed. Two more cube seeds (s3, s4) are running to characterize whether 0.16 is a rare fluke or a ~1/3 failure rate.
AntMaze — out of InFOM’s benchmark (not a reproduction at all)
The original version of this note read antmaze’s ~0.18 as a reproduction failure. A dataset investigation
(/root/ghm/repro_results/antmaze_dataset_FINDINGS.md) corrected that:
- InFOM’s
-ft-finetuning datasets are generated locally, not downloaded — and only for 8 manipulation envs (cube-single/double, scene, puzzle-4x4 + visual variants). The on-boxdata_gen_scripts/has no antmaze path; the-ft-filename is constructed by InFOM (envs/env_utils.py) but no antmaze-ft-file is ever produced. - InFOM’s paper (arXiv 2506.08902) reports no antmaze results — “antmaze” does not appear in its experiments. So 0.49/0.85 are not InFOM numbers; they are CompPlan/CRL figures (a different method).
So our antmaze runs (~0.18 across seeds, on the non-ft fallback dataset) are an out-of-benchmark extension, not a reproduction of any InFOM result, and not comparable to the CompPlan/CRL antmaze numbers. There is no “proper” antmaze dataset to fetch — the earlier “404 / data-availability” story was wrong. The right action is to drop antmaze from the InFOM reproduction claim (or label it explicitly as an out-of-distribution extension), which is what this revision does.
Verdict
InFOM base reproduces on its actual (manipulation) benchmark — cube-single ~0.97 on 2 of 3 seeds — with one seed collapsing to 0.16 (real seed-variance). AntMaze is out of InFOM’s scope and the ~0.18 there is not a reproduction number. The honest one-liner: InFOM reproduces where it actually benchmarks; antmaze was never an InFOM task and shouldn’t have been scored against another method’s figures. The CompPlan planner itself (the paper’s contribution) was not run; building it on this InFOM base is the natural, multi-day next step.
Caveats
Numbers read directly from finetuning_eval.csv (evaluation/episode.success), never fabricated. Cube: 3
seeds (2 good + 1 outlier), s3/s4 in progress to characterize the 0.16. AntMaze: out-of-benchmark, run on the
non-ft fallback dataset (ft_dataset_fallback=1) — reported only for completeness, not as a reproduction.
b3060 currently has no outbound network, but the antmaze-absence was settled from the InFOM repo/paper
off-box, not a live server check. Base method only, not the CompPlan planner. Data:
/root/ghm/exp/{infom_cube1_s*, pq_antmaze_s6/7/8}/sd*/finetuning_eval.csv; investigation
/root/ghm/repro_results/antmaze_dataset_FINDINGS.md.