Pith. sign in

REVIEW 2 major objections 7 minor 22 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

When the instruction names the answer, world models copy, not perceive

2026-07-09 22:47 UTC pith:ZTRQBSB3

load-bearing objection Solid diagnostic finding with a clean fix; the independence sub-claim is the weak link the 2 major comments →

arxiv 2607.06925 v1 pith:ZTRQBSB3 submitted 2026-07-08 cs.AI

Grounding Spatial Relations in a Compact World Model: Instruction Leakage and a Goal-Free Dynamics Fix

Yufeng Wang , Lu Wei , Haibin Ling This is my paper
classification cs.AI
keywords instructionemphgoalworldleakagemodelnameswhen
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper asks when explicit reference anchors in a compact world model genuinely ground spatial relations versus when they merely appear to. The answer turns on a confound: if the language goal names the relation being scored, the model's predictor can copy the answer from the instruction rather than perceiving it from the scene. This instruction leakage inflates representation metrics (a goal-conditioned model reaches 0.90 relation-readout accuracy that collapses to 0.27 chance when the goal is withheld) and degrades control. The authors establish that leakage is governed by transcribability — whether the instruction names the scored quantity — and is essentially independent of how predictive the non-instruction inputs are, tested across a controlled tabletop environment, the BabyAI benchmark, and Language-Table. The remedy is architectural: keep the goal out of the dynamics, where it does not belong, and place it only in the planner's cost function, while supervising the perception (read) path. This recovers genuine, instruction-independent grounding at 0.88 accuracy, identical whether or not the goal is provided.

Core claim

The paper identifies and characterizes instruction leakage in goal-conditioned world models: when a language instruction directly names the spatial relation being evaluated, a goal-conditioned predictor achieves high relation-readout accuracy not by perceiving the scene but by transcribing the instruction into its predicted coordinates. Withholding the goal collapses accuracy from 0.90 to 0.27 (chance), and feeding a counterfactual instruction makes the model follow the false goal 94.5% of the time. The leakage is governed by transcribability — whether the instruction names the scored quantity — and does not depend on the predictive strength of the action or state inputs. Removing the goal 从

What carries the argument

The central mechanism is instruction leakage: a goal-conditioned dynamics predictor copies the relation named in the language instruction into its predicted anchor coordinates, bypassing scene perception entirely. The detection instrument is a pair of controls — goal-withheld (zero the goal tokens, recompute the readout) and counterfactual-goal (feed a goal naming a different relation, measure whether anchors follow the false instruction or the true scene). The fix is goal-free dynamics: the predictor never sees the goal, which enters only through the planner's cost function, combined with supervised supervision of the read (perception) path.

Load-bearing premise

The claim that leakage is independent of non-instruction input strength rests primarily on a dose-response experiment in the authors' own synthetic tabletop environment, where per-step motion is large enough for the probe to be reliable. External benchmarks like Language-Table have roughly ten times smaller motion, making their probes low-signal, so the generalization from one synthetic environment to any goal-conditioned world model depends on that environment being structur

What would settle it

Feed a goal-conditioned world model a counterfactual instruction (one naming a relation different from the true scene) and measure whether the predicted anchors follow the false instruction or the true scene. If the model perceives rather than transcribes, the anchors should follow the true scene. If it transcribes, they follow the false instruction.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 7 minor

Summary. This paper identifies and characterizes a failure mode in goal-conditioned world models: when a language instruction directly names the scored relation (e.g., 'put X left of Y'), the dynamics predictor can transcribe the instruction into predicted anchors rather than perceiving the scene. The authors demonstrate this 'instruction leakage' on a controlled 2D tabletop, the external BabyAI benchmark, and Language-Table, using two clean controls: goal-withholding (collapsing readout accuracy from 0.90 to 0.27) and counterfactual goal substitution (94.5% false-instruction following). They further show that Language-Table, whose instructions name referents but not the scored direction, does not leak until the instruction is augmented to name the direction. The proposed fix—removing the goal from the dynamics and supervising the read path—recovers instruction-independent grounding (0.88, identical with and without the goal). The paper is methodologically careful, with validated positive controls, multi-seed replication for headline claims, and honest reporting of boundaries (the fix ties rather than beats the no-goal baseline on control; grounding collapses at maximal ambiguity).

Significance. The paper makes a valuable methodological contribution: it provides a falsifiable characterization of when instruction leakage occurs (transcribability), a validated detection protocol (goal-withheld and counterfactual probes with an engineered-leaky positive control at 0.97), and a simple architectural remedy. The cross-environment validation (tabletop, BabyAI, Language-Table) and the dose-response ablation (Table 3) are commendable. The honesty about the control result (a tie, not a win) and the ambiguity boundary (Fig. 4) strengthens the work. The characterization that leakage is governed by transcribability and is independent of non-instruction predictor strength is the central novel claim, and the three-setting evidence for the transcribability axis is solid. The independence sub-claim (Part 2 of the central claim) is more fragile, as detailed below.

major comments (2)
  1. §5.4, Table 3 and the abstract's central claim: The independence claim ('leakage is essentially independent of how predictive the non-instruction inputs are') rests on the action-ablation dose-response, but the load-bearing positive evidence is the synthetic regime (cosine 0.975→0.986 as α:1→0). This is near-tautological in the full-transcription setting: when the instruction directly names the answer, a complete shortcut exists, so the model uses it regardless of action quality—this is what shortcuts do by definition. The informative test would be a partially transcribable instruction (e.g., instruction names the relation but not which objects are target/anchor, forcing the model to combine instruction and scene). In that regime, degrading the action or scene could plausibly shift reliance toward the instruction, and predictor-competition would predict increased leakage. This regime is未
  2. Abstract and §5.4: The claim that the protocol and remedy 'apply to any goal-conditioned world model whose instruction names the scored quantity' overreaches the evidence base. All three settings are 2D or symbolic gridworlds with discrete relations and short templated instructions. The authors acknowledge in §7(ii) that there is no 3D or real-world validation, and in §7(iii) that no released pretrained model was probed. The transcribability mechanism should generalize in principle, but the strength of the independence sub-claim and the universality of the remedy would be better supported by at least one higher-dimensional or continuous-relation setting. The authors should qualify the abstract's 'any' to match the settings tested.
minor comments (7)
  1. §5.2, Fig. 4: The across-ambiguity non-transfer result (a=2 model reads 0.86 at a=2 but falls to 0.26–0.33 at other ambiguities including easier a=0) is interesting but underexplored. A brief discussion of whether this is a memorization artifact or a fundamental limitation of per-ambiguity training would help the reader.
  2. Table 1: The 'n/a' entries for ObjToken (object latents) make it hard to compare against the anchor-based models. A brief note on why the geometric readout cannot be applied (the slot convention issue is mentioned in §7(vi) but not at the table) would clarify whether this is a limitation of the metric or the design.
  3. §3, 'PrismWM': The name appears without explanation of its etymology. A brief gloss would help.
  4. §5.3: The control result is described as 'a tie, not a win' and the authors are commended for this honesty. However, the framing that 'goal-conditioning was the thing dragging GoalDyn down' could be read as slightly circular: GoalDyn is the authors' own architecture with goal-in-dynamics, so showing it underperforms NoGoal confirms the design choice but does not independently motivate it. A sentence acknowledging that this is a consistency check on their own design, not an external validation, would be fairer.
  5. Fig. 2a: The y-axis label 'predicted-anchor accuracy (amb2)' could be misread as anchor localization accuracy rather than relation-readout accuracy. Clarifying that this is the geometric relation readout from predicted anchors would help.
  6. §5.4, Language-Table: The counterfactual cosine for the +direction regime (0.174→0.032) is acknowledged as low-signal due to ~10× smaller per-step motion (footnote 1). The footnote's reasoning that the magnitude column 'refutes' the vanishing-motion floor is somewhat terse; a one-sentence explanation of why constant magnitude rules out the floor would help the reader.
  7. References: Several citations are to 2026-dated arXiv preprints (Maes et al., Nam et al., Zhang et al., etc.). These should be verified for correctness and whether they have been published or updated since submission.

Circularity Check

0 steps flagged

No circularity found: the derivation chain is empirical with externally-defined probes

full rationale

The paper's central claim is an empirical characterization, not a formal derivation, so the self-definitional and fitted-input patterns do not apply. The leakage detection probes (goal-withheld and counterfactual goal substitution) are defined operationally and are external to the training objective — they are not quantities the model was optimized to produce, so observing their failure is not circular. The fix (GoalFree) removes the goal from the dynamics by construction; the paper explicitly acknowledges that goal-invariance of the readout is then trivially true ('by construction there is nothing to leak,' §5.3) and does not present this invariance as a finding. The actual finding is the 0.88 accuracy level itself — that the model can perceive the relation when forced to — which is an empirical result, not a tautology. The transcribability characterization is supported by contrasting three settings (tabletop/BabyAI leak, Language-Table referent-only does not, +direction token induces leak), which is independent evidence rather than a self-citation chain. No cited theorem or prior result by the same authors is load-bearing for the central claim. The dose-response ablation (Table 3) is a within-task experiment, not a fitted-parameter-renamed-as-prediction. The skeptic's concern that the independence claim is near-tautological in the full-transcription regime is a correctness/generalization concern (the informative partial-transcription regime is untested), not a circularity in the derivation chain.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The paper introduces no new physical entities or postulated forces. PrismWM and reference anchors are architectural constructs composed of known components. The free parameters are design choices for the controlled experiment, not fitted constants. The axioms are domain assumptions about the evaluation protocol, all of which are either validated within the paper (positive control) or acknowledged as limitations (render distribution, learned-anchor fairness).

free parameters (4)
  • Number of objects N = 4
    Fixed design choice for the tabletop environment; not fitted but constrains all results.
  • Ambiguity levels a = 0-4
    Discrete knob controlling duplicate distractors; chosen by design to probe referential ambiguity.
  • Action scaling alpha = 1.0, 0.5, 0.0
    Dose-response ablation parameter in Table 3; chosen to test predictor-competition hypothesis.
  • JEPA latent dimension = not stated
    Architecture hyperparameter not specified; affects capacity but not load-bearing for the leakage claim.
axioms (4)
  • domain assumption The encode and read path never uses the goal (its readout is identical with and without the goal for every model).
    Stated in Section 4: this is what makes the observed-anchor accuracy a clean perception measure. If false, the separation between perception and transcription breaks down.
  • domain assumption Leakage-controlled readouts must run on the training render distribution.
    Stated in Section 4: a fresh-scene variant under-counted a render-sensitive encoder (0.35 vs 0.84). All leakage probes use dataset frames.
  • domain assumption The geometric readout (observed-anchor accuracy) is a fair measure of supervised anchors.
    Section 7 acknowledges this is NOT fair for learned (unsupervised) anchors, which are flagged rather than interpreted.
  • domain assumption The counterfactual probe is a valid instrument (validated by positive control at 0.97).
    The engineered-leaky model fires the probe at 0.97 vs 0.03 for no-goal baseline, establishing the probe is not dead.
invented entities (2)
  • PrismWM independent evidence
    purpose: The factored world model architecture (JEPA latent + metric point anchors + goal) used for the study.
    It is the authors' own architecture but is a composition of known components (JEPA, ViT, point anchors). The leakage finding does not depend on PrismWM being novel; it is the testbed.
  • Reference anchors (metric point anchors p) independent evidence
    purpose: Sparse explicit coordinates read from patch features to localize referents.
    Point-based representations are established in prior work (cited: Huang et al., Kim et al.). The contribution is the evaluation, not the entity.

pith-pipeline@v1.1.0-glm · 17545 in / 3218 out tokens · 370736 ms · 2026-07-09T22:47:43.181050+00:00 · methodology

0 comments
read the original abstract

Compact world models that condition on a language goal promise to ground relations such as ``put the red block left of the blue block'' using a sparse set of explicit \emph{reference anchors}. We ask when such references actually ground a relation, and identify a trap: a goal-conditioned predictor reaches a striking $0.90$ relation-readout accuracy, yet this is \emph{instruction transcription}, not perception. Withholding the goal collapses it to chance ($0.90\!\to\!0.27$, three seeds) and a counterfactual instruction makes the predicted anchors follow the \emph{false} instruction $94.5\%$ of the time (true scene $2.3\%$; $N{=}256$). Tested across three settings and a within-task ablation, our central claim characterizes the confound: \textbf{instruction leakage occurs when the scored quantity is transcribable from the instruction (when the instruction names the answer) and is essentially independent of how predictive the non-instruction inputs are.} Our tabletop and the external BabyAI benchmark leak, whereas a Language-Table forward-dynamics world model whose instruction names \emph{referents} does not, until the instruction is augmented to name the direction; and degrading the action never increases leakage, the opposite of what predictor-competition predicts. The diagnosis prescribes the fix: keep the goal out of the dynamics (it belongs to the planner's cost) and supervise the \emph{read} path, recovering genuine, instruction-independent grounding ($0.88$, identical with and without the goal). The detection protocol and remedy apply to any goal-conditioned world model whose instruction names the scored quantity.

Figures

Figures reproduced from arXiv: 2607.06925 by Haibin Ling, Lu wei, Yufeng Wang.

Figure 1
Figure 1. Figure 1: PrismWM and the leak path. State factors into a JEPA latent z, sparse metric point anchors p, and a goal g. In the goal-conditioned model the goal enters the dynamics (red), letting the predictor transcribe the instruction into anchors. Our fix (GoalFree) deletes that edge: the goal enters only the planner cost (green), where it is used for control anyway. axis of the target relative to the anchor. Referen… view at source ↗
Figure 2
Figure 2. Figure 2: Instruction transcription. (a) Predicted-anchor relation accuracy at medium ambiguity with the goal given (grey) versus withheld (color). The goal-conditioned model collapses to chance when the goal is withheld (leak￾age); adding observed-supervision partly de-leaks the predictor; goal-free dynamics (GoalFree, ours) is unchanged because it never uses the goal. (b) Under a counterfactual goal, GoalDyn’s pre… view at source ↗
Figure 3
Figure 3. Figure 3: Transcription, visualized. One tabletop scene (true relation: target below anchor) with GoalDyn’s predicted anchors overlaid (target ◦, anchor □) under three instructions. (a) True goal: the predicted target satisfies the named relation. (b) Goal withheld: the placement is uninformative. (c) Counterfactual goal (left of): the predicted target moves to satisfy the false instruction, not the scene; the model… view at source ↗
Figure 4
Figure 4. Figure 4: The ambiguity boundary of perception. Goal-independent observed-anchor accuracy versus scene ambiguity. Supervised read paths (GoalDyn+Read, GoalFree) ground the relation at low and medium ambiguity and collapse at a=4; without observed-supervision (GoalDyn) or without anchors the read path is near chance throughout. For control, the clean result is that goal-conditioning hurts, not that our fix wins. We a… view at source ↗
Figure 5
Figure 5. Figure 5: Goal-involvement gradient (medium ambiguity). [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages · 15 internal anchors

  1. [1]

    Don't Just Assume; Look and Answer: Overcoming Priors for Visual Question Answering

    Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don’t just assume; look and answer: Overcoming priors for visual question answering, june 2018.arXiv preprint arXiv:1712.00377. Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael G. Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from i...

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  2. [2]

    V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning

    URLhttps://api.semanticscholar.org/CorpusID:255999752. Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, Artem Zholus, et al. V-jepa 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Planning with Reasoning using Vision Language World Model

    Delong Chen, Theo Moutakanni, Willy Chung, Yejin Bang, Ziwei Ji, Allen Bolourchi, and Pascale Fung. Planning with reasoning using vision language world model.arXiv preprint arXiv:2509.02722, 2025a. Jiaming Chen, Wentao Zhao, Ziyu Meng, Donghui Mao, Ran Song, Wei Pan, and Wei Zhang. Vision- language model predictive control for manipulation planning and tr...

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R Bowman, and Noah A Smith

    URLhttps://api.semanticscholar.org/CorpusID:8081284. Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel R Bowman, and Noah A Smith. Annotation artifacts in natural language inference data. InProceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, ...

    work page 2018

  5. [5]

    Learning Latent Dynamics for Planning from Pixels

    URLhttps://api.semanticscholar.org/CorpusID:52171619. Danijar Hafner, Timothy P. Lillicrap, Ian S. Fischer, Ruben Villegas, David R Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels.ArXiv, abs/1811.04551,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    Dream to Control: Learning Behaviors by Latent Imagination

    URLhttps: //api.semanticscholar.org/CorpusID:53280207. Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination.arXiv preprint arXiv:1912.01603,

    work page internal anchor Pith review Pith/arXiv arXiv 1912

  7. [7]

    Td-mpc2: Scalable, robust world models for continuous control

    Nick Hansen, Hao Su, and Xiaolong Wang. Td-mpc2: Scalable, robust world models for continuous control. InInternational Conference on Learning Representations, volume 2024, pp. 47376–47405,

    work page 2024

  8. [8]

    ParticleFormer: A 3D Point Cloud World Model for Multi-Object, Multi-Material Robotic Manipulation

    SuningHuang, QianzhongChen, XiaohanZhang, JiankaiSun, andMacSchwager. Particleformer: A3dpoint cloud world model for multi-object, multi-material robotic manipulation.arXiv preprint arXiv:2506.23126, 2025a. Yuanhui Huang, Weiliang Chen, Wenzhao Zheng, Xin Tao, Pengfei Wan, Jie Zhou, and Jiwen Lu. Terra: Explorable native 3d world model with point latents....

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C

    URLhttps://api.semanticscholar.org/ CorpusID:260844398. Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C. Lawrence Zitnick, and Ross B. Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning.2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1988–1997,

    work page 2017

  10. [10]

    JisooKim, JungbinCho, SanghyeokChu, AnanyaBal, JinhyungKim, GunheeLee, SihaengLee, SeungHwan Kim, BohyungHan, HyunminLee, etal

    URL https://api.semanticscholar.org/CorpusID:15458100. JisooKim, JungbinCho, SanghyeokChu, AnanyaBal, JinhyungKim, GunheeLee, SihaengLee, SeungHwan Kim, BohyungHan, HyunminLee, etal. Pri4r: Learningworlddynamicsforvision-language-actionmodels with privileged 4d representation.arXiv preprint arXiv:2603.01549,

    work page arXiv

  11. [11]

    A path towards autonomous machine intelligence version 0.9

    Yann LeCun et al. A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27.Open Review, 62(1):1–62,

    work page 2022

  12. [12]

    Object-Centric Learning with Slot Attention

    Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention.ArXiv, abs/2006.15055,

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  13. [13]

    Understanding-in-generation: Reinforcing generative capability of unified model via infusing under- standing into generation.arXiv preprint arXiv:2509.18639,

    Yuanhuiyi Lyu, Chi Kit Wong, Chenfei Liao, Lutao Jiang, Xu Zheng, Zexin Lu, Linfeng Zhang, and Xuming Hu. Understanding-in-generation: Reinforcing generative capability of unified model via infusing under- standing into generation.arXiv preprint arXiv:2509.18639,

    work page arXiv

  14. [14]

    LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels

    Lucas Maes, Quentin Le Lidec, Damien Scieur, Yann LeCun, and Randall Balestriero. Leworldmodel: Stable end-to-end joint-embedding predictive architecture from pixels.arXiv preprint arXiv:2603.19312,

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    SOLD: Slot Object-Centric Latent Dynamics Models for Relational Manipulation Learning from Pixels

    Malte Mosbach, Jan Niklas Ewertz, Angel Villar-Corrales, and Sven Behnke. Sold: Slot object-centric latent dynamics models for relational manipulation learning from pixels.arXiv preprint arXiv:2410.08822,

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    Causal-JEPA: Learning World Models through Object-Level Latent Masking

    12 Heejeong Nam, Quentin Le Lidec, Lucas Maes, Yann LeCun, and Randall Balestriero. Causal-jepa: Learning world models through object-level latent interventions.arXiv preprint arXiv:2602.11389,

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    The Pitfalls of Simplicity Bias in Neural Networks

    Harshay Shah, Kaustav Tamuly, Aditi Raghunathan, Prateek Jain, and Praneeth Netrapalli. The pitfalls of simplicity bias in neural networks.ArXiv, abs/2006.07710,

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  18. [18]

    CLIPort: What and Where Pathways for Robotic Manipulation

    URLhttps://api.semanticscholar. org/CorpusID:219687117. Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Cliport: What and where pathways for robotic manipula- tion.ArXiv, abs/2109.12098,

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    Jingwen Sun, Wenyao Zhang, Zekun Qi, Shaojie Ren, Zezhi Liu, Hanxin Zhu, Guangzhong Sun, Xin Jin, and Zhibo Chen

    URLhttps://api.semanticscholar.org/CorpusID:237396838. Jingwen Sun, Wenyao Zhang, Zekun Qi, Shaojie Ren, Zezhi Liu, Hanxin Zhu, Guangzhong Sun, Xin Jin, and Zhibo Chen. Vla-jepa: Enhancing vision-language-action model with latent world model.arXiv preprint arXiv:2602.10098,

    work page arXiv

  20. [20]

    Factored Latent Action World Models

    Zizhao Wang, Chang Shi, Jiaheng Hu, Kevin Rohling, Roberto Martín-Martín, Amy Zhang, and Peter Stone. Factored latent action world models.arXiv preprint arXiv:2602.16229, 2026a. Zizhao Wang, Kaixin Wang, Li Zhao, Peter Stone, and Jiang Bian. Dyn-o: Building structured world mod- els with object-centric representations.Advances in Neural Information Proces...

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [21]

    ThinkJEPA: Empowering Latent World Models with Large Vision-Language Reasoning Model

    Haichao Zhang, Yijiang Li, Shwai He, Tushar Nagarajan, Mingfei Chen, Jianglin Lu, Ang Li, and Yun Fu. Thinkjepa: Empowering latent world models with large vision-language reasoning model.arXiv preprint arXiv:2603.22281,

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning

    Gaoyue Zhou, Hengkai Pan, Yann LeCun, and Lerrel Pinto. Dino-wm: World models on pre-trained visual features enable zero-shot planning.arXiv preprint arXiv:2411.04983,

    work page internal anchor Pith review Pith/arXiv arXiv