acceptodds
Under review as a conference paper at ICLR 2027

Resolution Collapse in Latent-Space CEM Planning

Abstract

Planning with a learned latent world model usually means using the cross-entropy method (CEM) to search over candidate action sequences and ranking them by how close the predicted final latent state lands to the goal. We show that this ranking step fails in a specific and measurable way. Across four continuous-control environments built on a joint-embedding predictive world model, the correlation between predicted and true candidate quality collapses to chance exactly at the terminal frame CEM optimizes, even though the same predictor ranks candidates well several steps earlier. We call this resolution collapse, and we observe it on every environment we test. A set of horizon-manipulation experiments then shows that the collapse tracks CEM's own optimization target rather than a fixed prediction horizon. Stretching the planning horizon from five steps to eight steps moves the collapse point outward by exactly the same amount, and the correlation at whichever frame is currently terminal stays pinned near chance at every horizon we try. This motivates a direct fix. Ranking candidates on a frame the model can still resolve, instead of the terminal one, raises planner success rate by three to fifteen points across our environments, following an inverted-U pattern in which asking too early actively hurts and asking too late is where the original failure lives. Two independent random-cost controls, matched in magnitude to the real cost, rule out that this gain is a side effect of simply adding another term to the objective. Replacing the resolution-matched cost with noise of identical spread collapses success rate by twenty-three to ninety-four points, and an additive version of the same idea beats a matched-noise version by eleven to thirteen points at p<0.001. Finally, we give a rule that selects the reporting frame directly from the model's own resolution curve, without ever consulting success rate, and show it lands within two points of the oracle frame found by an exhaustive sweep, on every environment. Together, these results turn an invisible bottleneck in latent-space planning into something we can measure, predict, and correct at no extra test-time cost.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
latent world models, model-based reinforcement learning, cross-entropy method, planning, joint-embedding predictive architecture, model diagnostics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_XCFsOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This is an interesting paper that identifies what it calls resolution collapse in latent-space planning. The authors identify this problem as a learned world model ranking candidate action sequences reasonably well at intermediate rollout frames, however, the ranking of the candidates retained by CEM is near chance at the terminal frame used for optimization. The authors measure this in four simulated tasks; they vary the planning horizon and propose ranking candidates at an earlier frame selected from the model’s measured resolution curve. They report modest success-rate gains when comparing their method with spread-matched random-cost controls.

Critical strengths

The idea behind the paper is interesting because it questions an assumption that is often taken as given in many works: a learned world model can reliably rank candidate plans at the very horizon the planner optimizes. It's also very timely, with world models becoming a promising direction for many tasks in AI.

The paper is also well written overall.

In more detail, I believe measuring ranking quality within CEM’s elite set is a sensible diagnostic. The distinction between ranking the full candidate pool and ranking the shortlisted candidates is clear and potentially important.

Additionally, the paper includes useful controls, especially the spread-matched random-cost comparisons, and reports paired tests and detailed appendices.

Finally, the proposed intervention is simple and may be practically useful if its benefits generalize.

Critical weaknesses

The paper has a few weaknesses that would need to be addressed before acceptance.

For example, I believe the cause is not perfectly established yet through the horizon experiment . The authors interpret the collapse moving with the terminal frame as evidence that it tracks CEM’s optimization target, rather than prediction time. But doesn't changing the horizon also change the candidate sequences and the distribution CEM searches over?

Moreover, the default evaluation goal is 25 environment steps ahead; the horizon experiment extends prediction to as many as 40 steps. Isn't the longer-horizon terminal cost temporally mismatched to the goal If the goal if the embedding remains tied to 25-step? The paper should clarify exactly how the goal and true cost are defined for these longer-horizon experiments, and ideally test the target-frame explanation while holding candidate sequences and goal semantics fixed.

Additionally, as far as I understand it, the intervention changes the planning objective, not just the frame at which a reliable prediction is read. For reasons potentially beyond improved ranking resolution, can't that favor different trajectories?

The evidence for improved success also seems mixed. The rule improves success by 3–5 percentage points, but the gain is statistically significant only on Reacher. The abstract mentions a broader “three to fifteen points” range, which appears to include other variants so it should distinguish those results from the blind rule’s gains.

All experiments use 1 pretrained model, 1 search method, and 4 simulated environments so broad generality remains largely untested. The horizon manipulation is limited to two tasks, and the paper does not directly test the proposed mechanism on e.g.MPPI, iCEM, another model checkpoint, or real-robot data. I believe to be acceptable for ICLR, this should be extended to include at least some of these, to be able to better support better the more general claims in the paper.

Questions for authors
  • In the horizon-manipulation experiments, is the goal embedding still the frame 25 environment steps after the start, or is the goal redefined for each horizon?

  • Can the authors test the effect of changing the queried frame while evaluating the same fixed candidate sequences, to separate target-frame effects from changes in CEM’s search distribution?

  • How much offline environment interaction is needed to estimate a reliable resolution curve, and how sensitive is the selected frame to the particular candidate set used for that estimate?

1 comment
AuthorsOpenReview ↗

We have updated the revised pdf

Dear Reviewer XCFs,

Thank you again for your constructive feedback. The corresponding revisions are highlighted in red, incorporating the additional experiments and clarifications discussed in our rebuttal.

The main updates addressing your concerns can be found in:

  • Whether the goal remains the observation 25 steps ahead, and how the longer-horizon terminal relates to it. The first red paragraph in Section 5.1 states that the paired rerun holds the goal 25 steps ahead. The red paragraph in Section 4.3 and the H=8 table in the appendix section “Measurements Added During the Discussion” keep the two positions separate: at H=8, the 25-step elite Spearman on PushT and Cube recovers to 0.765 and 0.406, while the new 40-step terminal stays at 0.028 and 0.033.
  • Changing the queried frame on a fixed set of candidates. The same red paragraph in Section 4.3. On the fixed elite list of the terminal CEM, pairwise accuracy at frame 1 is 0.69, 0.89, 0.79, and 0.73, against 0.52 at the terminal frame. After a separate CEM on each frame, elite Spearman at the optimized frame lies between −0.08 and +0.09.
  • Whether the intervention changes the objective or the effective planning length. The first red paragraph in Section 5.1, with the full table in the appendix. Changing only the cost frame inside a five-step rollout raises success by 5.5, 2.5, and 3.5 points. A three-step plan reaches 72.5% and 77.0% on TwoRoom and PushT, and ties the rule frame at 67.5% on Cube. The shorter plan accounts for the larger share.
  • One planner and one world model. The red paragraph in Section 4.3, the DINO-WM table in the appendix, and the red sentence in Section 6. Terminal-elite collapse appears in the eight MPPI and iCEM cells and on DINO-WM for PushT and Wall. Closed-loop success is reported for CEM on LeWM.

These additions place the goal definition, the fixed-candidate comparison, the horizon length, and the model coverage in the corresponding parts of the paper. If this answers your questions, we would be sincerely grateful if you would consider raising your score.

Reviewer_YnLVOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

CEM plans in a learned latent space by scoring each candidate action sequence on how close the predicted final latent lands to the goal. Whether that score actually ranks candidates is what the paper tests, by re-running the elite set in the simulator and correlating predicted cost with what really happened. On the LEWM world model, in all four environments they try (TwoRoom, Reacher, PushT, Cube), the correlation at the terminal frame is chance-level, 0.05 to 0.08, yet the same predictor ranks the very same candidates well one or two frames earlier, up to 0.87. They call this resolution collapse. They pin the cause down with horizon manipulations: stretch the planning horizon from five prediction steps to eight, and the collapse point moves out to the new terminal frame while the old 25-step frame recovers to 0.44 to 0.72. Across horizons of five through eight steps the terminal-frame correlation never leaves chance, even as the step count it refers to moves from 25 to 40. What matters is which frame the search happens to be optimizing, not how far ahead it predicts. The fix costs nothing: rank candidates at a mid-rollout frame instead. A selection rule that never consults success rate picks the frame and gains 3 to 5 points, up to 15 with an additive variant called MultiHorizon, tracing an inverted-U over which frame you ask. Two spread-matched random-cost controls suggest the gain comes from what the frame tells the planner, not from simply adding another term, and the blind rule lands within 2 points of the oracle everywhere.

Critical strengths
  1. The causal design. In addition to constructing a mechanistic model, expanding the field of vision across four time spans for intervention and dose-response is a good way to distinguish between the two explanations. The two random-cost controls are the right falsification tests, and they are reported with exact p-values.
Critical weaknesses
  1. The claims outrun the evidence. "On every environment we test" means four tasks on one model, and within that base only one environment shows a significant gain. The paper also never connects its per-frame account to the per-iteration story (candidates crowd together as CEM converges). Are these the same phenomenon? Would plain early stopping recover terminal-frame resolution? The paper leaves both open.
  2. Consistency. The same configuration shows 63.5% in Table 8 and 55.5% in Table 9 on TwoRoom, while Appendix G pins the run-to-run floor on PushT and Cube and calls TwoRoom bit-exact. One of these statements needs a correction. And the abstract's range mixes arms with different baselines, which flatters the result.
Questions for authors
  1. Reconcile the TwoRoom baselines (63.5% vs 55.5%). Different episode sets, L2 calibrations, process invocations? Given the bit-exactness claim, what differs? And does the +15.0 MultiHorizon gain survive against the Table 8 baseline?
  2. Can ρ_elite(t) be split into between-candidate true-cost spread and per-candidate prediction error, as functions of frame and CEM iteration? That would tell us why the terminal frame dies and whether "candidates become near-tied" is the whole story.
  3. Any transfer result, e.g. DINO-WM on PushT, or MPPI/iCEM? One would do a lot for the generality claim.
  4. Would early stopping alone (halt when candidate spread drops below model resolution) recover terminal-frame performance? Worth a run.
1 comment
AuthorsOpenReview ↗

We have updated the revised pdf

Dear Reviewer YnLV,

Thank you again for your constructive feedback. The corresponding revisions are highlighted in red, incorporating the additional experiments and clarifications discussed in our rebuttal.

The main updates addressing your concerns can be found in:

  • The 63.5% versus 55.5% baselines, and the +15 in the abstract. The red paragraph on MultiHorizon in Section 5.1. The two tables share their episodes. The TwoRoom terminal arm in Table 9 uses λ=1.441, and Table 8 uses λ=1.265. The +15 is 70.5−55.5. Against the Table 8 terminal arm, MultiHorizon gains +7.0, +7.0, +2.0, and +2.5 points, and Reacher moves from 91.0% to 98.0% (p=0.0026). All four environments are positive.
  • The per-iteration crowding, and early stopping. The second half of the first red paragraph in Section 5.1. Stopping once the predicted-score spread falls below the final error scale cuts a 30-iteration search to a median iteration of 11, 11, and 3, with success at 61.5%, 69.0%, and 67.5% against 58.0%, 70.0%, and 64.0% for the full terminal search.
  • A second planner and a second world model. The red paragraph in Section 4.3 after the causal-horizon figure, the tables in the appendix section “Measurements Added During the Discussion,” and the red sentence in Section 6, which states the new coverage. Terminal-elite Spearman lies in [−0.10, +0.11] in all eight MPPI and iCEM cells. On DINO-WM, the top-30 terminal values are −0.01 on PushT and +0.02 on Wall. The horizon stretch now includes PushT and Cube: the 25-step elite Spearman rises from 0.052 and 0.072 at H=5 to 0.765 and 0.406 at H=8, while the 40-step terminal stays at 0.028 and 0.033.

These revisions assign the abstract’s range to the comparison it belongs to, and they extend the diagnosis to three planners and a second world model. If this answers your questions, we would be sincerely grateful if you would consider raising your score.

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

The paper studies the resolution collapse in latent-space planning with the CEM. It represents that a learned world model loses its ability to accurately rank elite candidate trajectories at the terminal prediction step. The author proposes resolution-matched planning, which evaluates candidate trajectories at an intermediate frame where ranking accuracy remains reliable. Experiments across four continuous control environments show improved planning success rates across all four environments, supported by random-cost control experiments.

Critical strengths

The author raise a fair point of a world model's ability to predict future states does not necessarily translate into reliable ranking of candidate action sequences.

I like the idea of changing the planning horizon, which provides evidence that resolution collapse is associated with CEM's optimization target rather than simply increasing prediction distance.

Critical weaknesses

The observed resolution collapse may largely result from optimization-induced selection bias. Since CEM selects candidates based on predicted terminal costs, the low ranking correlation within the elite set could reflect an optimizer's curse rather than a distinct failure of the world model. The experiments do not fully disentangle these explanations.

The resolution curve is measured using elites selected by terminal-cost CEM. My concern is that when the objective changes to an intermediate frame, the candidate distribution also changes, potentially causing resolution collapse at the newly optimized frame.

Also, I believe that optimizing an intermediate frame could shorten the planning horizon. The improvements may therefore come from shorter-horizon planning rather than resolution-aware selection. A matched shorter-horizon CEM baseline is needed.

Questions for authors

See weaknesses

1 comment
AuthorsOpenReview ↗

We have updated the revised pdf

Dear Reviewer Cxoa,

Thank you again for your constructive feedback. The corresponding revisions are highlighted in red, incorporating the additional experiments and clarifications discussed in our rebuttal.

The main updates addressing your concerns can be found in:

  • Whether the collapse follows the newly optimized frame. The red paragraph in Section 4.3, immediately after the causal-horizon figure. With the horizon held at five, a separate CEM on each of frames 1 to 5 (four environments, 50 episodes per cell) leaves elite Spearman between −0.08 and +0.09 at the frame just optimized. The frame the diagnosis points to is the frame the search is optimizing.
  • Whether the same candidates remain ordered at an earlier frame. The same red paragraph. On the fixed terminal elite list, pairwise accuracy at frame 1 is 0.69, 0.89, 0.79, and 0.73, against 0.52 at the terminal frame. A null places the observed elite correlations between the 0.53 and 0.58 quantiles of chance. The earlier frame still supplies an ordering of those candidates.
  • Whether the gain comes from changing the cost frame or from a shorter plan. The first red paragraph in Section 5.1, with the full numbers in the appendix section “Measurements Added During the Discussion.” On one shared list of 200 episodes, with the goal held 25 steps ahead, changing only the cost frame inside a five-step rollout raises success by 5.5, 2.5, and 3.5 points on TwoRoom, PushT, and Cube. A three-step plan reaches 72.5% and 77.0% on TwoRoom and PushT, above the rule frame, and ties the rule frame at 67.5% on Cube. The shorter plan accounts for the larger share, and changing only the cost frame still raises success in the same direction.

These experiments put both explanations you raised where they can be checked, and they say which frame the search should stop on. If this answers your questions, we would be sincerely grateful if you would consider raising your score.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Added experiments summary for now

Thank you all for reading the paper so carefully. The rebuttal period is now about halfway through, so we are gathering here the experiments added during the discussion, for the convenience of all reviewers. They give the paper several strengths that can be checked on their own.

1. The collapse follows the frame being optimized, so the diagnosis says where the search should not be spent. We ran a full CEM on each of frames 1 to 5, in all four environments, 50 episodes per cell, 20 cells in total. At the optimized frame, mean elite Spearman lies between −0.08 and +0.09. Inside that same elite the true cost is still spread out: at the terminal frame its standard deviation is 31.6, 94.3, 8.67 and 41.4 (TwoRoom, Reacher, PushT, Cube), while the predicted score has standard deviation 3.2, 1.58, 0.373 and 0.454. Under a null that rebuilds the true scores from the full-list correlation, the observed elite Spearman sits between the 0.53 and 0.58 quantiles of the null. After the horizon is stretched to 8, the old 25-step position recovers to 0.440, 0.765, 0.716 and 0.406, and the new 40-step terminal remains at 0.035, 0.028, 0.032 and 0.033. The score the search fits last has no resolution inside its own elite. That is a concrete place to improve the planner.

2. The same diagnosis predicts two changes that both raise success, and shortening the plan raises it by more. Both tables use the same 200 episodes. The goal image is the observation 25 steps ahead in every row. Reacher's absolute success stays the Table 8 number.

With the prediction length fixed at 5, only the cost frame changes, from the terminal frame to the rule frame (frame 4 on TwoRoom and PushT, frame 3 on Cube):

EnvTerminal frameRule frameGain
TwoRoom58.0%63.5%+5.5
PushT70.0%72.5%+2.5
Cube64.0%67.5%+3.5

Shortening the prediction itself, with the cost on the terminal frame of that shorter rollout:

Env1 step2345Rule frame
TwoRoom57.569.072.569.056.563.5
PushT58.573.077.075.567.572.5
Cube61.063.567.563.566.067.5

Horizons 3 and 4 are above the rule frame on TwoRoom and PushT, and on Cube horizon 3 ties the rule frame at 67.5%. The curve is an inverted U, the same shape as the frame sweep. The effective planning length accounts for the larger share, and changing only the objective frame still moves success in the same direction. Both are what the diagnosis predicts: the search should avoid the frame whose score no longer separates the elite.

3. The frame is chosen from the offline resolution curve alone, the direction agrees on all four environments, and the gain comes from what that frame contains. The blind rule never consults success. In Table 8 the gains are +4.5, +5.0, +3.0 and +4.0, and Reacher goes from 91.0% to 96.0% (p=0.013). On the same 200 episodes, TwoRoom, PushT and Cube go from 58.0% to 63.5%, 70.0% to 72.5%, and 64.0% to 67.5%. With spread-matched noise in place of that frame, success is 13.0%, 2.5%, 0.5% and 45.5%. Fifty offline episodes are enough to fix the frame.

4. Adding one intermediate frame, using only the model's own scores, is higher on every environment. The TwoRoom terminal arm in Table 9 uses the MultiHorizon weight λ=1.441, and the Table 8 terminal arm uses 1.265, on the same episodes. The +15 in the abstract is 70.5−55.5. Against the Table 8 terminal arm, MultiHorizon is:

EnvTable 8 terminalMultiHorizonGain
TwoRoom63.5%70.5%+7.0
Reacher91.0%98.0%+7.0
PushT69.5%71.5%+2.0
Cube64.5%67.0%+2.5

The paired test on Reacher gives p=0.0026. All four environments are positive.

5. Early stopping cuts a 30-iteration search down to a few iterations, and success stays with the terminal CEM. The standard deviation of the elite's predicted score falls below the final error scale at iteration 20, 9, 15 and 4 (TwoRoom, PushT, Reacher, Cube). Stopping at that threshold gives:

EnvTerminal CEMEarly stopMedian halt step
TwoRoom58.0%61.5%11
PushT70.0%69.0%11
Cube64.0%67.5%3

The original search runs for 30 iterations. On TwoRoom and Cube the success rate is a little higher as well.

6. Terminal-elite collapse appears under three planners, and on a second world model. CEM's terminal elite Spearman lies between 0.047 and 0.072. The eight MPPI and iCEM cells lie between −0.10 and +0.11. On the fixed elite list of the H=5 terminal CEM, pairwise accuracy at frame 1 is 0.69, 0.89, 0.79 and 0.73 (TwoRoom, Reacher, PushT, Cube), against 0.52 at the terminal frame.

The second model is the DINO-WM baseline released with facebook/jepa-wms: a frozen DINOv2 ViT-S/14 plus a predictor trained apart from it. Planning follows that release, for 50 episodes on PushT and 50 on Wall: CEM, 300 samples, 10 elites, 30 iterations, horizon 6, frameskip 5. Mean Spearman of the top 30, from frame 1 to the terminal frame:

Env123456 (terminal)
PushT+0.340+0.350+0.354+0.259+0.147−0.010
Wall+0.718+0.517+0.336+0.303+0.336+0.018

The top 10, which is the set this CEM refits, ends at −0.012 on PushT and +0.027 on Wall. Among the top 30, the terminal predicted-score standard deviation is 0.0020 and 0.00058, while the true cost still spreads by 0.031 and 0.045. The earlier frames still separate the same candidates.

Success rates are reported for CEM on LeWM. On the three planners and on DINO-WM, what we report is the resolution: the terminal elite collapses at the frame being optimized.

Overall, these additions make the use of the paper concrete. It identifies which frame in latent-space planning still ranks candidates, and that position follows the optimization target. Following the diagnosis, both a shorter plan and an intermediate-frame cost raise success, and the shorter plan raises it by more. The frame is read off an offline curve, with no retraining. The same phenomenon appears under CEM, MPPI and iCEM, and on DINO-WM for PushT and Wall. We hope these checks show what the work offers: a low-cost, reusable way to decide, from offline logs alone, which frame the search should be spent on.

If these results are helpful to your assessment, we would be sincerely grateful if you would consider raising your score. We are glad to add any further comparison you would like to see.

AuthorsOpenReview ↗

Paper Revision

Thank you all for reading. We have just updated a revised pdf. Every sentence added in this revision is printed in red, and the full tables are in the appendix section “Measurements Added During the Discussion.” The additions support the following conclusions.

  1. With the horizon held at five, a separate CEM on each of frames 1 to 5 leaves elite Spearman between −0.08 and +0.09 at the frame just optimized. On that same fixed elite list, pairwise accuracy at frame 1 is 0.69, 0.89, 0.79, and 0.73 (TwoRoom, Reacher, PushT, Cube), against 0.52 at the terminal frame. The diagnosis says where the search should not be spent.
  2. Terminal-elite collapse appears in all eight MPPI and iCEM cells, with Spearman in [−0.10, +0.11], and on the released DINO-WM baseline. There the top-30 terminal Spearman is −0.01 on PushT and +0.02 on Wall, the top-10 terminals are −0.012 and +0.027, and the predicted-score standard deviations shrink to 0.0020 and 0.00058 while the true cost still spreads by 0.031 and 0.045.
  3. Stretching the horizon to eight steps restores the 25-step elite Spearman on PushT and Cube from 0.052 and 0.072 to 0.765 and 0.406, while the new 40-step terminal stays at 0.028 and 0.033. The collapse follows the optimization target.
  4. On one shared list of 200 episodes, changing only the cost frame inside a five-step rollout raises success from 58.0%, 70.0%, and 64.0% to 63.5%, 72.5%, and 67.5% on TwoRoom, PushT, and Cube. A three-step plan reaches 72.5% and 77.0% on TwoRoom and PushT, above the rule frame, and ties the rule frame at 67.5% on Cube. The shorter plan accounts for the larger share of the gain.
  5. Stopping CEM when the predicted-score spread falls below the final error scale cuts a 30-iteration search to a median iteration of 11, 11, and 3, with success at 61.5%, 69.0%, and 67.5% against 58.0%, 70.0%, and 64.0% for the full terminal search.
  6. MultiHorizon, using only the model’s own scores, gains +7.0, +7.0, +2.0, and +2.5 points against the Table 8 terminal arm, and Reacher moves from 91.0% to 98.0%. All four environments are positive.

These results show the contribution of the paper: which frame still ranks candidates, how that position follows the optimization target, and how a shorter plan and an intermediate-frame cost raise success once the diagnosis is followed. The frame is read off an offline curve, with no retraining. The same phenomenon now has support from three planners and from a second world model.

If this update is helpful to your assessment, we would be sincerely grateful if you would consider raising your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.