# Josef Chen reports mixed results for learned planners on PushT [Canonical reading](https://escapemeridian.com/readings/josef-chen-reports-mixed-results-for-learned-planners-on-pusht-782739f5) Josef Chen’s paper reports mixed results for a fixed action updater on PushT. The updater solved 8 of 16 scenes reserved from configuration selection, but did not meet the task-cost criterion fixed before evaluation. The comparison planners also varied: selected Adam solved 7 tasks, adapted iCEM solved 2, and released LePlanner solved 16, with different total planning and encoding times. In a matched experiment, replaying the first ten commands and allowing the next solve 4, 8 or 12 updates yielded 9, 8 and 7 successes among 16 contexts. Extra updates did not rescue any failing task. All success/failure changes occurred in one scene cluster: extra updates lowered the costs of selected latent world-model trajectories for the updater, while tasks that had succeeded became failures. The paper includes code, weights, evaluators, execution traces and cost accounting. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Oct 27, 2026 (2026-10-27). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-13T03:05:25.818Z. First observed: 2026-10-30T06:42:57.191Z. Retrieved: 2026-10-31T07:46:43.332Z. - [When More Planning Hurts: Learned Planners for World Models](https://arxiv.org/abs/2610.37891) ## Limits The reported PushT evaluation covers 16 scenes reserved from configuration selection; the abstract does not establish how these results generalize beyond this evaluation. The fixed action updater did not satisfy the task-cost criterion fixed before evaluation. The reported changes in success and failure were confined to one scene cluster; the abstract does not explain why lower selected latent world-model trajectory costs coincided with failures. ## Open question Why did extra updates lower the costs of selected latent world-model trajectories in the affected scene cluster while previously successful tasks became failures? ## Source-stated Chen reports that the fixed action updater solved 8 of 16 PushT scenes reserved from configuration selection, but did not satisfy the task-cost criterion fixed before evaluation. > On PushT, our fixed action updater (1,505 parameters) solves 8 of 16 scenes reserved from our configuration selection. - [Supporting source](https://arxiv.org/abs/2610.37891) ## Source-stated The comparison planners had different task success counts and total planning and encoding times. > Selected Adam (7 tasks solved, total planning and encoding time 218.440s), adapted iCEM (2 tasks solved, total planning and encoding time 49.309s), released LePlanner (16 tasks solved, total planning and encoding time 5.652s). - [Supporting source](https://arxiv.org/abs/2610.37891) ## Source-stated In the matched experiment, the 4-, 8-, and 12-update budgets produced 9, 8, and 7 successes, respectively, among 16 contexts; extra updates did not rescue failing tasks. > Among all 16 contexts (8 scene clusters), the 3 budgets get 9, 8 and 7 successes, respectively. Interestingly, the extra updates did not rescue any failing task in this experiment. - [Supporting source](https://arxiv.org/abs/2610.37891) ## Source-stated All success/failure changes occurred in one scene cluster, where lower costs for selected latent world-model trajectories coincided with previously successful tasks becoming failures. > All the changes in success/failure happen in one scene cluster, where we observe that the extra updates decrease the costs of the selected latent world-model trajectories for our updater, while the tasks that used to be successes become failures. - [Supporting source](https://arxiv.org/abs/2610.37891) ## Source-stated The paper includes materials that can support examination of its results and cost accounting. > The code, weights, evaluators, execution traces and cost accounting accompany the paper. - [Supporting source](https://arxiv.org/abs/2610.37891) ## Source and contributor context Chen, Josef · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## World model An AI system's internal model of how an environment behaves, used to predict what happens next or to plan. - [Definition reference](https://arxiv.org/abs/1803.10122) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 20122097-c56d-4a2b-89b6-37fbab51e87f. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-13T03:05:25.818Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: a5f97a586dd9cc902f95a8f6c78916e779e5f15cc04efe3e5d047be50fdfddab. Source text hash: 3ede34b1736360bf18b9676156f347cf8458144ae4eb7042dd668abccd0fe4d7. ## Revision and source-check history ### Version 1 — 2026-11-13T03:05:25.818Z Josef Chen’s paper reports mixed results for a fixed action updater on PushT. The updater solved 8 of 16 scenes reserved from configuration selection, but did not meet the task-cost criterion fixed before evaluation. The comparison planners also varied: selected Adam solved 7 tasks, adapted iCEM solved 2, and released LePlanner solved 16, with different total planning and encoding times. In a matched experiment, replaying the first ten commands and allowing the next solve 4, 8 or 12 updates yielded 9, 8 and 7 successes among 16 contexts. Extra updates did not rescue any failing task. All success/failure changes occurred in one scene cluster: extra updates lowered the costs of selected latent world-model trajectories for the updater, while tasks that had succeeded became failures. The paper includes code, weights, evaluators, execution traces and cost accounting. ## Original source arXiv cs.AI · Article Original source date: Oct 27, 2026 (2026-10-27). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-13T03:05:25.818Z. First observed: 2026-10-30T06:42:57.191Z. Retrieved: 2026-10-31T07:46:43.332Z. - [When More Planning Hurts: Learned Planners for World Models](https://arxiv.org/abs/2610.37891) ## Limits The reported PushT evaluation covers 16 scenes reserved from configuration selection; the abstract does not establish how these results generalize beyond this evaluation. The fixed action updater did not satisfy the task-cost criterion fixed before evaluation. The reported changes in success and failure were confined to one scene cluster; the abstract does not explain why lower selected latent world-model trajectory costs coincided with failures. ## Open question Why did extra updates lower the costs of selected latent world-model trajectories in the affected scene cluster while previously successful tasks became failures? ## Source-stated Chen reports that the fixed action updater solved 8 of 16 PushT scenes reserved from configuration selection, but did not satisfy the task-cost criterion fixed before evaluation. > On PushT, our fixed action updater (1,505 parameters) solves 8 of 16 scenes reserved from our configuration selection. - [Supporting source](https://arxiv.org/abs/2610.37891) ## Source-stated The comparison planners had different task success counts and total planning and encoding times. > Selected Adam (7 tasks solved, total planning and encoding time 218.440s), adapted iCEM (2 tasks solved, total planning and encoding time 49.309s), released LePlanner (16 tasks solved, total planning and encoding time 5.652s). - [Supporting source](https://arxiv.org/abs/2610.37891) ## Source-stated In the matched experiment, the 4-, 8-, and 12-update budgets produced 9, 8, and 7 successes, respectively, among 16 contexts; extra updates did not rescue failing tasks. > Among all 16 contexts (8 scene clusters), the 3 budgets get 9, 8 and 7 successes, respectively. Interestingly, the extra updates did not rescue any failing task in this experiment. - [Supporting source](https://arxiv.org/abs/2610.37891) ## Source-stated All success/failure changes occurred in one scene cluster, where lower costs for selected latent world-model trajectories coincided with previously successful tasks becoming failures. > All the changes in success/failure happen in one scene cluster, where we observe that the extra updates decrease the costs of the selected latent world-model trajectories for our updater, while the tasks that used to be successes become failures. - [Supporting source](https://arxiv.org/abs/2610.37891) ## Source-stated The paper includes materials that can support examination of its results and cost accounting. > The code, weights, evaluators, execution traces and cost accounting accompany the paper. - [Supporting source](https://arxiv.org/abs/2610.37891) ## Source and contributor context Chen, Josef · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## World model An AI system's internal model of how an environment behaves, used to predict what happens next or to plan. - [Definition reference](https://arxiv.org/abs/1803.10122) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: 20122097-c56d-4a2b-89b6-37fbab51e87f. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-13T03:05:25.818Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: a5f97a586dd9cc902f95a8f6c78916e779e5f15cc04efe3e5d047be50fdfddab. Source text hash: 3ede34b1736360bf18b9676156f347cf8458144ae4eb7042dd668abccd0fe4d7. ### uncertain — 2026-11-13T03:05:25.818Z — anthropic/claude-opus-5.5 Publication: 20122097-c56d-4a2b-89b6-37fbab51e87f attribution — pass: The page lists Josef Chen as the sole author. The draft attributes the results to Chen's paper and does not present them as established fact. > Authors:Josef Chen summary — uncertain: The numbers are accurate, but the summary is hard to follow without specialist background. PushT, fixed action updater, Adam, iCEM, LePlanner, task-cost criterion and latent world-model trajectories are all unexplained, and the summary never says what a world model or planner does. It also buries the paper's central finding, signalled by its title: more updates lowered the model's own trajectory costs while real successes turned into failures. It drops the condition that subsequent solves all got 8 updates. Smallest repair: lead with that finding, add a plain-language gloss or tooltips for the core terms, and restore the condition. > (subsequent solves all get 8 updates) date — pass: The originalPublished date of 2026-10-27 matches the v1 submission date. The feed date of Oct 31 is later and is correctly not used. > \[Submitted on 27 Oct 2026\] support — pass: Each claim is source-stated and matches the abstract. The limitations properly note the narrow 16-scene scope and the unexplained mechanism. The question is a genuine open issue that follows from the reported cluster result, not a leading claim. The draft adds no invented benefits or costs. > Note that our action updater does not satisfy the task-cost criterion we fixed before evaluation. media — pass: The selected media is the verified arXiv abstract page for this paper, which is the direct original source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/782739f5c9bc522b62b1c60855382737fe76553a8e802ff3900920a8bc9f28b4/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/josef-chen-reports-mixed-results-for-learned-planners-on-pusht-782739f5.json)