metadata
license: cc-by-nc-4.0
pretty_name: UWM Paper Evaluation Trajectories
tags:
- world-model
- evaluation
- trajectories
- bagel
UWM paper evaluation trajectories
Complete, unselected model outputs for the UWM paper tables. 108/110 planned run archives are currently published; remaining rows are evaluations still running or not yet complete.
Each archive contains all available per-episode generated observations (gen_obs/), environment renders (renders/), rollout GIFs, verbatim generated reasoning text, actions, episode statistics, and the run summary. Text/JSON metadata is scrubbed of machine-local paths before publication.
Agent-generated observation images are required and included for CoT, forward-thinking, and branching runs. The NonCoT action-chunk policy emits no imagined image in the evaluated protocol (n_imagined=0); for those rows this field is not applicable, and the archive contains the actual actions, environment frames, GIF, and stats.
Table1
| Task | Eval set | Model | Trajectory archive | Checkpoint | Eval data | Result |
|---|---|---|---|---|---|---|
| sokoban | Ordinary Hard | k1_noncot |
download | model | dataset | 47.0% |
| sokoban | Ordinary Hard | k3_cot |
download | model | dataset | 53.0% |
| sokoban | Ordinary Hard | k3_noncot |
download | model | dataset | 10.0% |
| sokoban | Ordinary Hard | k5_cot |
download | model | dataset | 40.0% |
| sokoban | Ordinary Hard | k5_noncot |
download | model | dataset | 2.0% |
| sokoban | Ordinary Hard | k10_cot |
download | model | dataset | 15.0% |
| sokoban | Ordinary Hard | k10_noncot |
download | model | dataset | 0.0% |
| sokoban | Ordinary Hard | kinf_cot |
download | model | dataset | 19.0% |
| sokoban | Ordinary Hard | kinf_noncot |
download | model | dataset | 0.0% |
| maze2d | Ordinary Hard | k1_noncot |
download | model | dataset | 62.0% |
| maze2d | Ordinary Hard | k3_cot |
download | model | dataset | 59.0% |
| maze2d | Ordinary Hard | k3_noncot |
download | model | dataset | 54.0% |
| maze2d | Ordinary Hard | k5_cot |
download | model | dataset | 40.0% |
| maze2d | Ordinary Hard | k5_noncot |
download | model | dataset | 23.0% |
| maze2d | Ordinary Hard | k10_cot |
download | model | dataset | 21.0% |
| maze2d | Ordinary Hard | k10_noncot |
download | model | dataset | 2.0% |
| maze2d | Ordinary Hard | kinf_cot |
download | model | dataset | 3.0% |
| maze2d | Ordinary Hard | kinf_noncot |
download | model | dataset | 0.0% |
| pacman | Ordinary Hard | k1_noncot |
download | model | dataset | 84.0% |
| pacman | Ordinary Hard | k3_cot |
download | model | dataset | 38.0% |
| pacman | Ordinary Hard | k3_noncot |
download | model | dataset | 25.0% |
| pacman | Ordinary Hard | k5_cot |
download | model | dataset | 37.0% |
| pacman | Ordinary Hard | k5_noncot |
download | model | dataset | 1.0% |
| pacman | Ordinary Hard | k10_cot |
download | model | dataset | 7.0% |
| pacman | Ordinary Hard | k10_noncot |
download | model | dataset | 0.0% |
| pacman | Ordinary Hard | kinf_cot |
download | model | dataset | 6.0% |
| pacman | Ordinary Hard | kinf_noncot |
download | model | dataset | 0.0% |
| pacman | Two-Ghost | k1_noncot |
download | model | dataset | 78.0% |
| pacman | Two-Ghost | k3_cot |
download | model | dataset | 70.0% |
| pacman | Two-Ghost | k3_noncot |
download | model | dataset | 72.0% |
| pacman | Two-Ghost | k5_cot |
download | model | dataset | 62.0% |
| pacman | Two-Ghost | k5_noncot |
download | model | dataset | 69.0% |
| pacman | Two-Ghost | k10_cot |
download | model | dataset | 52.0% |
| pacman | Two-Ghost | k10_noncot |
download | model | dataset | 57.0% |
| pacman | Two-Ghost | kinf_cot |
download | model | dataset | 57.0% |
| pacman | Two-Ghost | kinf_noncot |
download | model | dataset | 48.0% |
| pusht | Mid (left-up) | k1_noncot |
download | model | dataset | 60.0% |
| pusht | Far (four corners) | k1_noncot |
download | model | dataset | 31.0% |
| pusht | Mid (left-up) | k3_cot |
download | model | dataset | 19.0% |
| pusht | Far (four corners) | k3_cot |
download | model | dataset | 9.0% |
| pusht | Mid (left-up) | k3_noncot |
download | model | dataset | 7.0% |
| pusht | Far (four corners) | k3_noncot |
download | model | dataset | 7.0% |
| pusht | Mid (left-up) | k5_cot |
download | model | dataset | 18.0% |
| pusht | Far (four corners) | k5_cot |
download | model | dataset | 9.0% |
| pusht | Mid (left-up) | k5_noncot |
download | model | dataset | 9.0% |
| pusht | Far (four corners) | k5_noncot |
download | model | dataset | 0.0% |
| pusht | Mid (left-up) | k10_cot |
download | model | dataset | 6.0% |
| pusht | Far (four corners) | k10_cot |
download | model | dataset | 3.0% |
| pusht | Mid (left-up) | k10_noncot |
download | model | dataset | 2.0% |
| pusht | Far (four corners) | k10_noncot |
download | model | dataset | 0.0% |
| pusht | Mid (left-up) | kinf_cot |
download | model | dataset | 1.0% |
| pusht | Far (four corners) | kinf_cot |
download | model | dataset | 2.0% |
| pusht | Mid (left-up) | kinf_noncot |
download | model | dataset | 0.0% |
| pusht | Far (four corners) | kinf_noncot |
download | model | dataset | 3.0% |
Table2
| Task | Eval set | Model | Trajectory archive | Checkpoint | Eval data | Result |
|---|---|---|---|---|---|---|
| sokoban | 2-box Deadlock | van5k |
download | model | dataset | 22.0% |
| sokoban | 2-box Deadlock | always5k |
download | model | dataset | 60.0% |
| sokoban | 2-box Deadlock | bra5k |
download | model | dataset | 30.0% |
| sokoban | 3-box Deadlock | van5k |
download | model | dataset | 39.0% |
| sokoban | 3-box Deadlock | always5k |
download | model | dataset | 48.0% |
| sokoban | 3-box Deadlock | bra5k |
download | model | dataset | 39.0% |
| sokoban | 4-box Deadlock | van5k |
download | model | dataset | 27.0% |
| sokoban | 4-box Deadlock | always5k |
download | model | dataset | 24.0% |
| sokoban | 4-box Deadlock | bra5k |
download | model | dataset | 32.0% |
| sokoban | 2-box Structural Fork | van5k |
download | model | dataset | 85.0% |
| sokoban | 2-box Structural Fork | always5k |
download | model | dataset | 84.0% |
| sokoban | 2-box Structural Fork | bra5k |
download | model | dataset | 79.0% |
| sokoban | 3-box Structural Fork | van5k |
download | model | dataset | 34.0% |
| sokoban | 3-box Structural Fork | always5k |
download | model | dataset | 46.0% |
| sokoban | 3-box Structural Fork | bra5k |
download | model | dataset | 39.0% |
| sokoban | 4-box Structural Fork | van5k |
download | model | dataset | 22.0% |
| sokoban | 4-box Structural Fork | always5k |
download | model | dataset | 22.0% |
| sokoban | 4-box Structural Fork | bra5k |
download | model | dataset | 32.0% |
| pacman | Ordinary Hard | pacman_forward_v1 |
download | model | dataset | 6.0% |
| pacman | Two-Ghost | pacman_forward_v1 |
download | model | dataset | 54.0% |
| pacman | Forward Stress G8 | pacman_vanilla_kinf |
download | model | dataset | 78.0% |
| pacman | Forward Stress G8 | pacman_forward_v1 |
download | model | dataset | 78.0% |
| pacman | Forward Stress G8 | pacman_forward_v2 |
pending | pending model publication | dataset | |
| pacman | Forward Stress G12 | pacman_vanilla_kinf |
download | model | dataset | 9.0% |
| pacman | Forward Stress G12 | pacman_forward_v1 |
download | model | dataset | 4.0% |
| pacman | Forward Stress G12 | pacman_forward_v2 |
pending | pending model publication | dataset |
Table3
| Task | Eval set | Model | Trajectory archive | Checkpoint | Eval data | Result |
|---|---|---|---|---|---|---|
| sokoban | 2-box Deadlock | van5k_rl90 |
download | model | dataset | 21.0% |
| sokoban | 2-box Deadlock | fwd3k_sft |
download | model | dataset | 41.0% |
| sokoban | 2-box Deadlock | fwd3k_rl90 |
download | model | dataset | 19.0% |
| sokoban | 2-box Deadlock | bra3k_sft |
download | model | dataset | 37.0% |
| sokoban | 2-box Deadlock | bra3k_rl90 |
download | model | dataset | 36.0% |
| sokoban | 3-box Deadlock | van5k_rl90 |
download | model | dataset | 42.0% |
| sokoban | 3-box Deadlock | fwd3k_sft |
download | model | dataset | 38.0% |
| sokoban | 3-box Deadlock | fwd3k_rl90 |
download | model | dataset | 38.0% |
| sokoban | 3-box Deadlock | bra3k_sft |
download | model | dataset | 35.0% |
| sokoban | 3-box Deadlock | bra3k_rl90 |
download | model | dataset | 44.0% |
| sokoban | 4-box Deadlock | van5k_rl90 |
download | model | dataset | 37.0% |
| sokoban | 4-box Deadlock | fwd3k_sft |
download | model | dataset | 25.0% |
| sokoban | 4-box Deadlock | fwd3k_rl90 |
download | model | dataset | 34.0% |
| sokoban | 4-box Deadlock | bra3k_sft |
download | model | dataset | 30.0% |
| sokoban | 4-box Deadlock | bra3k_rl90 |
download | model | dataset | 41.0% |
| sokoban | 2-box Structural Fork | van5k_rl90 |
download | model | dataset | 85.0% |
| sokoban | 2-box Structural Fork | fwd3k_sft |
download | model | dataset | 78.0% |
| sokoban | 2-box Structural Fork | fwd3k_rl90 |
download | model | dataset | 66.0% |
| sokoban | 2-box Structural Fork | bra3k_sft |
download | model | dataset | 84.0% |
| sokoban | 2-box Structural Fork | bra3k_rl90 |
download | model | dataset | 85.0% |
| sokoban | 3-box Structural Fork | van5k_rl90 |
download | model | dataset | 42.0% |
| sokoban | 3-box Structural Fork | fwd3k_sft |
download | model | dataset | 35.0% |
| sokoban | 3-box Structural Fork | fwd3k_rl90 |
download | model | dataset | 41.0% |
| sokoban | 3-box Structural Fork | bra3k_sft |
download | model | dataset | 33.0% |
| sokoban | 3-box Structural Fork | bra3k_rl90 |
download | model | dataset | 44.0% |
| sokoban | 4-box Structural Fork | van5k_rl90 |
download | model | dataset | 30.0% |
| sokoban | 4-box Structural Fork | fwd3k_sft |
download | model | dataset | 23.0% |
| sokoban | 4-box Structural Fork | fwd3k_rl90 |
download | model | dataset | 28.0% |
| sokoban | 4-box Structural Fork | bra3k_sft |
download | model | dataset | 22.0% |
| sokoban | 4-box Structural Fork | bra3k_rl90 |
download | model | dataset | 28.0% |
Integrity
manifest.json records the source run identifier, summary metrics, remote archive path, byte size, file count, and SHA-256 digest when the archive was packaged by this release tool.
Table 2 uses the frozen six-set, 100-episode, max-40 execution contract. Table 3 currently tracks the matched pre-RL/RL90 max-40 evaluation; completed rows appear automatically when their run summary exists.