acceptodds
Withdrawn from ICLR 2027

NegativeSpace-WM: Learning to Rule Out Impossible Futures for World-Model Planning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

World models help agents choose actions by predicting their consequences, but generating physically impossible futures wastes computation. Rejecting these futures before generation requires identifying physical conflicts while preserving feasible alternatives. We propose NegativeSpace-WM, a framework that learns minimal impossibility certificates to guide this decision. Each certificate describes a physical conflict, a repair with no unnecessary edits, and the futures that the conflict rules out. A shared model trained on valid simulated events and their violating and repaired variants learns to recognize conflicts in videos and predict them from observations, actions, and brief descriptions of proposed futures. These predictions guide independent physical checks; generation is skipped only when the checks rule out every future allowed by the candidate request. Experiments show that minimal repairs improve balanced accuracy on unseen physical settings by 2.2 percentage points over equally long redundant repairs. Certificate guidance also helps the verifier reach definite conclusions for more candidates than unguided search at the same checking budget. Across four planning task families, NegativeSpace-WM uses 33.20% less total GPU time per episode than generating every candidate, with task success of 81.6% versus 82.5%.

Primary area:
applications to robotics, autonomy, planning
Keywords:
world models, physical reasoning, model-based planning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_D3YUOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces NegativeSpace-WM, which skips world-model generation for candidate futures that seem physically impossible. Each candidate comes with a symbolic description of its expected outcome called a commitment. A Cosmos3-Nano model is fine-tuned with LoRA to write a structured explanation of why each candidate might be impossible such as which objects are involved, which physical rule is broken, and when. Then these are used to decided which hand-written physical checks to run first. A candidate is skipped only if one of those checks confirms the problem. In MuJoCo environment, this saves about 33% of GPU time compared with generating every candidate while preserving task success.

Critical strengths

I like the part that the learned model does not remove a candidate on its own, and that cancellation is limited to the scope of the clear (or verified) conflict, so other outcomes of the same action are kept. The structured explanation is also nicely designed.

Critical weaknesses

I see some clear limitations: First, the method shows narrow applicability. The rule out process needs three things: 1) The future must be written as a typed symbolic claim, 2) the physical premises must be explicit (rigid bodies, closed obstcles, etc.), and 3) the object geometry and poses must be calibrated. The checks also cover only five hand-written constraint types. These conditions seem easy to meet in MuJoCo, where the simulator knows every object's shape and position and the rules of physics are fixed. However, they are hard to meet in real scenes, which is where we need learned video world models. There is no simulator, object geometry and physical properties are unknown, and candidate outcomes are not written as symbolic statements. In those cases the checks would mostly return "unknown", and the advantage becomes decreased.

Second, I don't get the practical meaning of this method. When these conditions do hold, geometry and physics are already known, so the candidate could be checked or simulated directly, probably at lower cost than video generation. The learned part of the method also does not seem to matter much. A baseline that runs the same checks in a fixed order, with no learned model, performs about the same (Table 1B). For me, the saving seems to mainly comes from the hand-written checks, not from learning.

Lastly, evaluation seems unfair to test the method's novelty or superiority. The evaluation setup was largely designed by the authors such as the violation types, the symbolic commitment that describe each candidate's expected outcome, the available checks, and the proposer. Since the proposer decides the number of impossible candidates, the amount of compute saved also depends on this author-defined proposer. With this experiment, the advantage of using the proposed method is proved narrowly. It cannot show how the things going on in various scenarios can be seend in real scene.

Questions for authors

Q1. Can the method bring advantage (or can skip the candidate) when a setting was not designed for? For example, a case with real or uncalibrated video, or with a proposer that was not tuned for this setup?

Q2. When the geometry and object poses are known, we can roll out each candidate action with a physics simulator. Or, we could check directly whether the commitment is physically possible or not (without a learned model or video generation). I think these are simpler ways compared to the use of the proposed method. Does the proposed method bring advantages in terms of cost and accuracy compared to these simpler options?

Q3. Since the learned model only changes the order of the checks, the same result could be obtained by using fixed rule-out order with a larger budget. If so, the advantage of learned model seems quite limited. Is there a case where the learned explanation allows a candidate to be ruled out that the fixed-order checks cannot rule out, even with more checks?

Reviewer_TT5MOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes NegativeSpace-WM, a wrapper around a frozen world model for planning. The wrapper takes three inputs: recent observations, a candidate action sequence, and a commitment, i.e., conditions on the future outcome (e.g., poses, contacts). From these inputs, a learned Gate guides pre-defined physical checks that test whether any physically possible future satisfies the commitment. If none exists, generation is skipped; otherwise, the frozen world model generates future observations. The paper reports 33.20% less GPU time than generating all candidates while achieving comparable success rates.

Critical strengths

S1. Important problem. The paper addresses the computational cost of world-model rollouts in planning, which grows with the number of candidates and the planning horizon.

S2. Thorough cost accounting. Complete-episode GPU time includes failures, timeouts, and retries.

Critical weaknesses

W1. Unclear motivation for the commitment term in the problem formulation. In addition to observations and actions, the proposed world model wrapper takes a commitment term, which specifies conditions that a candidate future must satisfy. Note that this commitment is a per-candidate assertion separate from the task goal. While the paper explains how commitments enable scoped rejection, it remains unclear to me why practical world model planning should be formulated as commitment-specific future generation requests, and what advantage this offers over rollouts conditioned only on observations and actions.

W2. The experimental setting seems artificial. The efficiency claim is supported by Table 1B and Sec. 4.5. In the fixed request pool used in Sec. 4.5, 70% of the future generation requests are infeasible (22,400 of 32,000; Appendix C.13). Table 1B’s infeasible fraction is not reported. It seems unrealistic that such a large share of infeasible generation requests would arise in practical world model use.

W3. The learned Gate offers little benefit over the pre-defined verifier alone. In Table 1B, Verifier Search runs the same pre-defined physical checks without the learned Gate. It achieves the same task success as learned Certificate Gate (81.6%) at nearly the same GPU cost (51.21 vs. 50.72 s). This suggests that the efficiency gains come mainly from the pre-defined verifier rather than from the learned gate, which is the paper's main methodological contribution.

Questions for authors

See the weaknesses

Reviewer_Su6dOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Robots can plan by using a world model to imagine what each action will lead to, but generating those imagined futures is very expensive, and many of them are physically impossible, such as an object passing through a wall. This paper rules out impossible futures before generating them. A learned model, trained on simulated physics events and their smallest fixes, guesses what is wrong and where. A separate physics checker then confirms the problem before anything is skipped, and if the evidence is unclear, the option is kept. In simulation, this saves 33% of GPU time with almost the same task success (81.6% vs. 82.5%). However, a plain checker without the learned hints saves nearly as much, so most of the benefit comes from the checker rather than the learning. The method has also only been tested in simple simulated scenes.

Critical strengths
  • I like the design that the learned model only suggests where to look, and a separate physics checker makes the final call, so a wrong guess can’t remove a good option. In addition, it rejects only the impossible version of an action (e.g., “go through the wall”) and keeps the possible ones (“bounce off”).
  • The experiments settings are fixed before testing, statistics are paired with corrections for multiple tests, and failed checks and weak results are reported in the appendix.
Critical weaknesses
  • I think one of the main weakness is that learning add little to planning. A plain checker without learned hints gets the same success (81.6%) and nearly the same GPU time. Most of the gain comes from the checker, not the learned model.
  • The efficiency gain reported in this work is measured at a fixed candidate count, where the method is cheaper than full rollout but has slightly lower task success. That comparison does not hold performance constant. The paper's own working-point analysis (Appendix C.4) asks the opposite question: what does each method cost at a matched success target? Full rollout meets the target with the default candidate budget. The proposed method falls just short and needs a larger candidate pool to close the gap, and that larger pool costs more total compute than the baseline. The reported savings therefore come from accepting a small loss in success rather than from a strict efficiency improvement. At equal task performance, the method shows no compute advantage and is in fact more expensive.
Questions for authors
  1. The main comparison is against Branch Full, which generates every commitment branch even though the world model conditions only on actions. What is the realized branch expansion factor in the main experiments? How does the method compare in cost and success against a Native Full baseline that generates action-conditioned futures without commitment branching? Appendix A.4 suggests the savings may vanish or reverse in that setting.
  2. Verifier Search reaches the same task success at statistically indistinguishable episode cost. Can you show any planning-level setting where certificate guidance beats Verifier Search, not just Verdict Gate? For example, could Search be included in the 70-second budgeted comparison?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.