acceptodds
Under review as a conference paper at ICLR 2027

Beyond the Next Step: Variable-Length Latent World Models for Long-Horizon Planning

Abstract

Recently, world models have emerged as a promising paradigm for building intelligent agents by learning predictive models that estimate future environment states conditioned on observations and actions. In particular, JEPA-style latent world models provide an efficient alternative to pixel space prediction by learning action-conditioned dynamics in compact representation spaces. However, existing latent world models typically rely on one-step prediction and must be recursively rolled out for long-horizon planning, which leads to compounding errors and a mismatch between training objectives and downstream planning tasks. To address this limitation, we propose Variable-length Latent World Models (VLWMs), a framework that learns to predict future latent states conditioned on action sequences of variable lengths. Instead of training only on one-step transitions, VLWMs directly model temporally extended dynamics, allowing the same predictor to evaluate action plans over different horizons. We further introduce a curriculum training strategy that progressively expands the action horizon, stabilizing optimization from short-range dynamics to long-range prediction. At test time, we design planning methods tailored to VLWMs to better exploit their variable-length predictive capabilities. Experiments on long-horizon control tasks show that VLWMs significantly improve latent space world models, achieving 13% average improvement over the state-of-the-art LeWM across different datasets, with especially large gains on tasks requiring extended planning. These results suggest that VLWM provides a simple yet effective paradigm for improving long-horizon prediction and planning in latent world models.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Latent world model, long-horizon planning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_PPzrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces variable-length latent world model which improves long-horizon planning by directly predicting future latent states conditioned on action sequences of varying lengths, rather than relying exclusively on recursively applied one-step predictions.

It uses a shared action-token interface to support multiple prediction horizons with one predictor, together with a progressive horizon curriculum that introduces longer prediction targets while retaining shorter-horizon supervision.

The main contributions are:

  1. a variable-length latent prediction objective and action-token architecture that support multi-horizon endpoint prediction;
  2. a cumulative horizon curriculum to stabilize training across prediction horizons
  3. five planning strategies that exploit variable-length predictions, including predefined chunk schedules, cost-adaptive chunking, and joint optimization of actions and chunk boundaries.

It evaluate VLWM on LeWorldModel (LeWM) and DINO-WM across PushT, OGBench-Cube, and TwoRoom. It report average planning-success improvements of 8.1 percentage points with the predefined long-to-short planner and 8.7 points with joint action–chunk optimization, increasing to 11.8 and 12.8 points.

Critical strengths

1.an approach to reducing long-horizon prediction error in latent world models. The paper addresses an important limitation of one-step latent prediction: recursively rolling out predictions can accumulate errors and degrade long-horizon planning. It directly predicts future latent endpoints conditioned on variable-length action sequences, using a shared action-token interface and a progressive horizon curriculum.

2.Consistent empirical gains across multiple representations and planning strategies. The authors evaluate with both jointly learned visual representations (LeWM) and frozen pretrained DINO features (DINO-WM), across three simulated control environments.

Critical weaknesses

The contribution of variable-length prediction is not fully disentangled from the effects of planning strategy and curriculum. The reported improvements combine several factors: multi-horizon latent prediction, progressive horizon curriculum, and planning strategies that vary chunk lengths or jointly optimize chunk boundaries with actions. In particular, P3 and P5 introduce different planning mechanisms that may independently contribute to the observed gains. A more systematic isolation of these components, including comparisons against equally tuned one-step baselines and alternative multi-step prediction objectives, would help establish whether the core VLWM representation and training approach are responsible for the improvements. Without clearer attribution, it is difficult to assess the method's fundamental advantage over existing planning and prediction alternatives.

Questions for authors
  1. How much of the reported improvement comes from variable-length latent prediction itself, rather than the planning strategy and horizon curriculum?

2.How robust are the results to the choice of maximum prediction horizon and the distribution of chunk lengths?

4 comments
AuthorsOpenReview ↗

Response to Reviewer PPzr (1/2)

We thank Reviewer PPzr for recognizing the empirical gains across different visual representations. We appreciate your comments on attribution and robustness and address them below.

Q1. Isolating variable-length prediction, curriculum, and planning

The contribution of variable-length prediction is not fully disentangled from planning and curriculum. How much comes from variable-length prediction itself? More systematic component isolation, including equally tuned one-step baselines and alternative multi-step objectives, would clarify whether the core representation and training approach explains the gains over existing prediction and planning alternatives.

Following your suggestions, we added action-token one-step, fixed-five-step direct, and five-step recursive training controls, alongside VLWM with and without curriculum. The non-curriculum action-token variants share the architecture, training data, update count, batch size, and global learning-rate schedule. Fixed-5 recursive is retrained with full backpropagation through five predictions and fifth-step endpoint supervision, without teacher forcing; Fixed-5 direct and VLWM-P1 both use one direct five-step prediction at evaluation.

Results below use LeWM on all three environments at Δ=100\Delta=100, with a planning horizon of five model steps, an execution budget of 200 raw steps, and 30 CEM iterations. Success rates are percentages (mean ± standard deviation over three training seeds).

Training / plannerPushTTwoRoomOGBench-Cube
Original LeWM9.3 ± 2.319.3 ± 3.142.0 ± 2.0
Action-token one-step10.0 ± 2.020.0 ± 4.042.7 ± 3.1
Fixed-5 direct11.3 ± 3.121.3 ± 4.252.0 ± 4.0
Fixed-5 recursive6.7 ± 2.311.3 ± 3.143.3 ± 4.2
VLWM, no curriculum, P110.0 ± 2.019.3 ± 1.255.3 ± 5.0
VLWM, no curriculum, P312.0 ± 2.020.7 ± 3.154.0 ± 2.0
VLWM, curriculum, P114.0 ± 2.020.7 ± 3.154.0 ± 2.0
VLWM, curriculum, P314.7 ± 3.127.3 ± 1.260.0 ± 2.0

The comparisons separate three effects:

  • Prediction training: the action-token one-step control averages 24.2%, close to original LeWM's 23.6%. Fixed-5 direct and no-curriculum VLWM-P1 both reach 28.2%, showing that direct multi-step supervision already helps; VLWM additionally supplies the range of prediction lengths used by the chunking planners.
  • Curriculum: holding P3 fixed, average success increases from 28.9% without curriculum to 34.0% with it, with gains in all three environments.
  • Planning: using the same curriculum-trained checkpoints, P3 improves over P1 from 29.6% to 34.0%, again in all three environments.

The complete VLWM-P3 configuration exceeds Fixed-5 direct by 5.8 percentage points on average. These results identify how the predictive interface, training schedule, and chunk selection work together. We include these comparisons in Section 4.3 (Tables 1–3) and Appendix H of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer PPzr (2/2)

Q2. Robustness to prediction horizon and chunk-length distribution

How robust are the results to the maximum prediction horizon and the distribution of chunk lengths?

We interpret “maximum prediction horizon” as referring to the horizon used in training and planning, and “the distribution of chunk lengths” as referring to the inference-time chunk-selection rule. We therefore added two complementary experiments on PushT with LeWM at Δ=100\Delta=100. Entries below are success rates (%, mean ± standard deviation over three training seeds).

Training and planning horizons. We trained VLWM with Kmax⁡=10K_{\max}=10 and evaluated P1–P5 with planning horizon H=10H=10, alongside the original Kmax⁡=5K_{\max}=5 setting. LeWM remains a one-step predictor and is evaluated at the corresponding planning horizon.

MethodKmax⁡=5, H=5K_{\max}=5,\ H=5Kmax⁡=10, H=10K_{\max}=10,\ H=10
LeWM9.3 ± 2.32.7 ± 1.2
LeWM+VLWM (P1)14.0 ± 2.024.7 ± 3.1
LeWM+VLWM (P2)12.0 ± 2.024.7 ± 1.2
LeWM+VLWM (P3)14.7 ± 3.130.0 ± 2.0
LeWM+VLWM (P4)15.3 ± 3.130.7 ± 3.1
LeWM+VLWM (P5)16.0 ± 2.032.7 ± 1.2

All five VLWM planners improve when the training and planning horizons are extended together. At ten model steps (50 environment steps), P3 and P5 reach 30.0% and 32.7%, respectively, compared with 2.7% for recursive LeWM. The benefit thus extends to a larger trained horizon range.

Inference-time chunk selection. Using the original Kmax⁡=5K_{\max}=5 checkpoints and fixing H=5H=5, we separately varied P3’s maximum allowed chunk length and its long-to-short bias strength β\beta. All other inference settings were held fixed within each sweep.

Maximum chunk length (β=4\beta=4)345
P3 success (%)15.3 ± 3.116.0 ± 2.014.7 ± 3.1
Bias strength β\beta (maximum chunk = 5)234
P3 success (%)12.7 ± 1.214.0 ± 2.014.7 ± 3.1

At β=0\beta=0, P3 reduces to uniform sampling over feasible chunk lengths. Thus, β=2\beta=2 stays closer to uniform sampling and imposes a weaker long-to-short preference than β=3\beta=3 or 44, consistent with its slightly lower success.

The maximum chunk length sweep spans 14.7–16.0%, and the bias sweep 12.7–14.7%. Every setting exceeds LeWM’s 9.3% at H=5H=5. These results show that the planning benefit persists across the tested maximum chunk lengths and distribution strengths. We include the horizon and chunk-distribution results in Appendix K (Tables 15–16) and clarify the default planning horizon in Section 4.1 of the revised manuscript.

Thank you again for your detailed and constructive comments. We hope our responses address your concerns and welcome any further questions.

We understand that this is a particularly busy period, and we sincerely appreciate your time and consideration.

We have provided detailed responses to all of your questions and concerns. When you have a chance, we would be very grateful if you could take a look at our responses. We hope they help clarify our work, and we would welcome any further feedback.

Thank you again, and we wish you a pleasant day!

Dear Reviewer,

With approximately 6 hours remaining before the rebuttal deadline, we would be grateful if you could take a look at our updated responses. We have provided point-by-point responses to every concern raised in your review.

We understand that this is a busy period and sincerely appreciate your time. If your schedule permits, we would greatly appreciate hearing your feedback.

Thank you again for your careful consideration. Wishing you a pleasant day!

Reviewer_YsJ5OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Variable-length Latent World Models (VLWM) to improve long-horizon planning by directly predicting future latent states conditioned on action sequences of varying lengths. The method introduces an action-token interface that supports multiple prediction horizons within a single model, together with a progressive horizon curriculum to facilitate training. Building on this formulation, the authors investigate five planning strategies that partition action sequences into variable-length chunks, including predefined schedules and joint action–boundary optimization. VLWM is implemented on both LeWM and DINO-WM and evaluated on PushT, OGBench-Cube, and TwoRoom, where it generally improves closed-loop planning performance over one-step baselines, particularly at longer goal offsets. Additional experiments on LIBERO examine goal-conditioned action ranking, while ablation studies analyze the effects of curriculum training and chunk selection.

Critical strengths
  1. The paper demonstrates that variable-length prediction can improve closed-loop planning performance across both jointly learned representations (LeWM) and frozen pretrained representations (DINO-WM). The evaluation covers three control benchmarks with different dynamics and four goal offsets, with three independent training seeds for the main experiments. The reported improvements generally become larger at longer goal offsets, supporting the motivation that direct multi-horizon prediction can alleviate error accumulation from recursive one-step rollouts. These results provide useful empirical evidence for the applicability of variable-length prediction across different latent world model architectures.

  2. The paper investigates multiple planning strategies using the same trained VLWM, helping distinguish the effects of different chunk-selection policies. In particular, Table 2 shows that the long-to-short strategy (P3) improves average planning success over fixed-length chunking (P1), suggesting that prediction partitioning can influence downstream control performance even when the world model is unchanged. Additionally, the curriculum ablation in Table 1 reports improvements across all three benchmarks, providing further evidence for the usefulness of progressively introducing longer prediction horizons.

Critical weaknesses
  1. The main contributions combine multi-horizon latent prediction, a progressive horizon curriculum, and variable-length chunk-based planning. However, related approaches such as ω-EVA (Sun et al., 2026) also predict future visual features from action prefixes at sampled horizons, although within a different policy refinement framework. While VLWM provides a distinct planning interface, its incremental contribution over existing multi-horizon prediction approaches requires stronger justification. Moreover, the proposed joint action–boundary optimization (P5) provides only a modest improvement over the simpler long-to-short strategy (P3), while the two approaches use different CEM iteration budgets (20 versus 30). Although the reported per-iteration timing is useful, a comparison under matched overall planning time would better establish whether the additional adaptive optimization is necessary and consistently beneficial.

  2. The current experiments primarily compare VLWM against conventional one-step world models, but they do not include two important alternatives: a recursive predictor explicitly trained through multi-step rollout supervision and a direct predictor trained only at the fixed planning horizon. These controls are important because the proposed improvements may arise from additional long-horizon supervision rather than the variable-length formulation itself. The P1 ablation evaluates fixed-length inference using an already multi-horizon-trained VLWM, while the direct-versus-recursive comparison in Appendix G evaluates two inference modes of the same trained model. Neither comparison determines whether variable-length training outperforms separately trained fixed-horizon or multi-step recursive models under matched architectures, supervision, and computational budgets. Without these controls, the necessity and individual contribution of the proposed training formulation remain insufficiently established.

Questions for authors
  1. Can the authors provide matched training comparisons to isolate the contribution of variable-length prediction? In particular, could the authors evaluate a recursive world model trained with explicit multi-step rollout supervision and a direct prediction model trained only at the fixed planning horizon, using comparable model capacity, representation settings, training data, and computational budgets? These comparisons would help clarify whether the reported improvements specifically require variable-length prediction or can be achieved through simpler long-horizon training objectives. Convincing results against these alternatives would substantially strengthen my assessment of the paper's methodological contribution.

  2. What is the benefit of joint action–boundary optimization under matched planning budgets? Table 2 shows that P5 achieves only a modest improvement over P3 while using fewer CEM iterations, making it difficult to isolate the value of the adaptive boundary search. Could the authors provide a success-versus-planning-time comparison by varying the CEM iteration budget for both P3 and P5 under the same evaluation settings? Such an analysis would help determine whether P5 provides a consistent advantage over simpler chunk-selection strategies and whether its additional optimization complexity is justified.

4 comments
AuthorsOpenReview ↗

Response to Reviewer YsJ5 (1/2)

We thank Reviewer YsJ5 for recognizing the usefulness of the planner and curriculum analyses. We appreciate your suggestions for clarifying the methodological contribution and address them below.

Q1. Contribution of variable-length training and planning

Related methods such as ω-EVA (Sun et al., 2026) also predict future visual features from action prefixes at sampled horizons, although within a different policy-refinement framework. While VLWM provides a distinct planning interface, its incremental contribution over existing multi-horizon prediction approaches requires stronger justification.

Can the authors provide matched training comparisons to isolate the contribution of variable-length prediction? In particular, could the authors evaluate a recursive world model trained with explicit multi-step rollout supervision and a direct prediction model trained only at the fixed planning horizon, using comparable model capacity, representation settings, training data, and computational budgets? These comparisons would help clarify whether the reported improvements specifically require variable-length prediction or can be achieved through simpler long-horizon training objectives. Convincing results against these alternatives would substantially strengthen my assessment of the paper's methodological contribution.

VLWM uses variable-horizon prediction as a composable transition interface for goal-conditioned planning. The same predictor evaluates different chunks of an action sequence, enabling both predefined partitions and joint action–boundary search. This differs from the policy proposal–prediction–refinement framework of ω\omega-EVA discussed in Section 2. Following your suggestion, we added separately trained controls to quantify what this formulation contributes beyond simpler long-horizon objectives.

We trained action-token one-step, fixed-five-step direct, and five-step recursive predictors, alongside VLWM with and without curriculum. The non-curriculum variants share the architecture, training data, and training schedule. Fixed-5 recursive uses full backpropagation through five predictions with fifth-step endpoint supervision and no teacher forcing. Fixed-5 direct and VLWM-P1 share the same single-call, five-step inference pattern. Results below use LeWM at Δ=100\Delta=100 across all three environments, with a planning horizon of five model steps, an execution budget of 200 raw steps, and 30 CEM iterations, and report mean ± standard deviation over three training seeds.

Training / plannerPushTTwoRoomOGBench-Cube
Original LeWM9.3 ± 2.319.3 ± 3.142.0 ± 2.0
Action-token one-step10.0 ± 2.020.0 ± 4.042.7 ± 3.1
Fixed-5 direct11.3 ± 3.121.3 ± 4.252.0 ± 4.0
Fixed-5 recursive6.7 ± 2.311.3 ± 3.143.3 ± 4.2
VLWM, no curriculum, P110.0 ± 2.019.3 ± 1.255.3 ± 5.0
VLWM, no curriculum, P312.0 ± 2.020.7 ± 3.154.0 ± 2.0
VLWM, curriculum, P114.0 ± 2.020.7 ± 3.154.0 ± 2.0
VLWM, curriculum, P314.7 ± 3.127.3 ± 1.260.0 ± 2.0

Fixed-5 direct training reaches 28.2% average success, confirming the value of direct long-horizon supervision. VLWM supports the full range of trained chunk lengths, attaining the same 28.2% average with uniform training and P1. With curriculum, VLWM-P1 reaches 29.6%; changing only the planner on the same checkpoints to P3 raises success to 34.0%, exceeding Fixed-5 direct in all three environments by 5.8 percentage points on average. These comparisons quantify the benefits of the training schedule and the planning flexibility enabled by the variable-length interface. The matched-time comparison below further evaluates the value of adaptive boundary search. We include all controls in Section 4.3 (Table 1) and Appendix H of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer YsJ5 (2/2)

Q2. Joint action–boundary optimization under matched planning budgets

P5 provides only a modest improvement over P3 while using a different CEM iteration budget (20 versus 30). Could you report success versus planning time by varying the CEM iteration budget for both P3 and P5 under the same evaluation settings? This would establish whether P5 offers a consistent advantage and justifies its additional optimization complexity.

We evaluated P3 and P5 using LeWM+VLWM at Δ=100\Delta=100 under four shared time budgets. For each environment, BIB_I is defined by P3's planning time with I∈{10,20,30,40}I\in\{10,20,30,40\} CEM iterations; P5's iteration count is adjusted to match each corresponding time budget. Both planners use the same checkpoints, evaluation pairs, 300 candidates, and 30 elites. Success rates are percentages (mean ± standard deviation over three training seeds); bold marks the higher mean within each environment–budget pair, including ties.

Time budget (P3 reference iterations)PlannerPushTOGBench-CubeTwoRoomAvg.
10P37.3 ± 1.246.7 ± 4.216.7 ± 3.123.6
10P59.3 ± 2.350.0 ± 2.018.7 ± 3.126.0
20P312.0 ± 2.059.3 ± 1.225.3 ± 2.332.2
20P515.3 ± 3.160.7 ± 4.229.3 ± 2.335.1
30P314.7 ± 3.160.0 ± 2.027.3 ± 1.234.0
30P516.0 ± 2.061.3 ± 3.130.0 ± 2.035.8
40P315.3 ± 4.260.7 ± 3.129.3 ± 1.235.1
40P518.0 ± 2.062.7 ± 1.230.0 ± 2.036.9

P5 achieves higher mean success in all 12 environment–budget comparisons under matched planning time. Most gains occur between B10B_{10} and B20B_{20} (+8.7 percentage points for P3 and +9.1 for P5 on average), with smaller gains at larger budgets. P5 already reaches 35.1% at B20B_{20}, compared with 34.0% for P3 at B30B_{30}. These results support a consistent benefit from joint action–boundary search across the tested planning-time budgets. We include this comparison and its budget definition in Section 4.3 (Table 4) and Appendix J (Table 12) of the revised manuscript.

Thank you again for your detailed and constructive comments. We hope our responses address your concerns and welcome any further questions.

Thank you for the additional experiments. The separately trained fixed-horizon direct and multi-step recursive baselines substantially address my concerns about attribution, and the matched-time comparisons show a modest but consistent mean improvement of P5 over P3. Variable-length training alone still does not outperform the fixed-horizon direct baseline on average under P1, so the strongest evidence supports the combined curriculum and flexible planning framework. Overall, the new results strengthen the empirical contribution, and I have raised my score from 2 to 3.

Thank you very much for your positive assessment of our rebuttal and paper, and for increasing your rating from 2 to 3. As you suggested, we have made these contributions explicit in the revised manuscript.

Thank you again for your constructive comments and for your careful evaluation of our submission.

Reviewer_Ykv8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Variable-length Latent World Models (VLWM), using action tokens and multi-horizon endpoint supervision to predict future latent states from action sequences of varying lengths, aiming to reduce error accumulation in recursive one-step prediction. Training progressively expands the prediction horizon while retaining shorter targets. For planning, the authors introduce five chunking strategies, P1–P5. Experiments with LeWM and DINO-WM show improved planning success on PushT, OGBench-Cube, and TwoRoom.

Critical strengths
  1. The cumulative horizon curriculum is well motivated and improves planning success at longer goal offsets. Training diagnostics also show reduced local fluctuations in loss and gradient norms.

  2. The method has a clear motivation and is straightforward to implement. The P1–P5 planning strategies are reasonably motivated and demonstrate downstream benefits.

Critical weaknesses
  1. Unexplained benchmark selection. Reacher was evaluated in both LeWM and DINO-WM but is omitted here without explanation. Given the claimed generality across representations and control tasks, the authors should justify this omission and, where feasible, provide Reacher results for both backbones under a consistent protocol.

  2. Insufficient transparency on efficiency and model capacity. In Table 2, P1 reduces prediction depth per candidate from 5 to 1, yet CEM iteration time decreases only from 11.58 to 11.22 seconds. P2 and P3 are slower per iteration than the baseline, while P5’s overall time savings also reflect fewer iterations. The authors should explain this discrepancy and report success under matched time or compute budgets. Despite the claim of matched capacity, encoder, predictor, total, and trainable parameter counts should be explicitly reported.

  3. Model and planning contributions are not sufficiently separated. The full P1–P5 comparison is provided only at the largest goal offset, where low success rates—including those of the baselines—raise questions about performance relative to a random policy. Results across all goal offsets are needed, particularly for P1, which uses the maximum fixed prediction span. These would establish whether VLWM consistently improves over the one-step baseline without sophisticated chunking, and quantify the additional gains of P2–P5 over P1.

  4. Insufficient curriculum ablation. Table 1 compares the cumulative curriculum with uniform sampling over the full horizon range, but this changes both training order and cumulative sampling frequencies. A reverse, long-to-short curriculum, with other variables controlled, would help determine whether the gains come from curriculum direction or horizon weighting.

Questions for authors
  1. Evaluation coverage: Why was Reacher omitted? Could you explain the benchmark selection and provide results for both backbones under a consistent protocol?

  2. Efficiency and capacity: What computational bottlenecks explain the limited runtime reduction despite fewer prediction steps? Could you report encoder, predictor, total, and trainable parameter counts for the baselines and VLWM?

  3. Model versus planning contributions: Could you provide P1–P5 results across all goal offsets for both backbones, especially P1, to establish its gains over the one-step baseline and the additional benefits of P2–P5? Could you also report random-policy success under the same execution budget,especially the largest goal offset.

  4. Curriculum direction: How does a long-to-short curriculum perform when cumulative sampling counts for each horizon, training budget, and learning-rate schedule are matched? Does the short-to-long ordering remain beneficial under these controls?

5 comments
AuthorsOpenReview ↗

Response to Reviewer Ykv8 (1/3)

We thank Reviewer Ykv8 for recognizing the motivation and implementation simplicity of our method. We appreciate your detailed suggestions and address each concern below.

Q1. Benchmark selection and Reacher coverage

Reacher was evaluated in both LeWM and DINO-WM but is omitted here without explanation. Please justify the benchmark selection and, where feasible, provide Reacher results for both backbones under a consistent protocol.

Thank you for the suggestion. We initially prioritized benchmarks where increasing goal offset more strongly challenges planning. Reacher involves controlling a two-joint planar arm without contact interactions. Its goal offset measures temporal separation along a recorded trajectory, which need not translate into a longer required control sequence.

Following your suggestion, we now report Reacher results for both backbones at all four offsets, using the same execution-budget protocol, B=2ΔB=2\Delta. Entries are success rates (%, mean ± standard deviation over three seeds; independent action seeds for the random policy). Bold indicates the best result within each backbone.

MethodΔ=25\Delta=25Δ=50\Delta=50Δ=75\Delta=75Δ=100\Delta=100
Random policy13.3 ± 1.210.7 ± 1.28.7 ± 1.25.3 ± 1.2
LeWM88.0 ± 4.088.0 ± 2.088.7 ± 3.184.0 ± 2.0
LeWM+VLWM (P3)80.7 ± 4.290.7 ± 4.288.7 ± 4.291.3 ± 3.1
LeWM+VLWM (P5)76.0 ± 7.288.0 ± 2.087.3 ± 3.185.3 ± 3.1
DINO-WM77.3 ± 3.180.0 ± 2.083.3 ± 3.188.0 ± 2.0
DINO-WM+VLWM (P3)78.7 ± 3.182.7 ± 3.186.0 ± 2.085.3 ± 4.2
DINO-WM+VLWM (P5)74.0 ± 2.082.0 ± 2.085.3 ± 3.184.7 ± 3.1

The baselines maintain high success across offsets, consistent with Reacher’s weaker sensitivity to this difficulty parameter under our protocol. VLWM keeps comparable with the basline and provides additional gains in several settings: P3 improves LeWM from 84.0% to 91.3% at Δ=100\Delta=100, and DINO-WM from 83.3% to 86.0% at Δ=75\Delta=75.

Q2. Runtime bottlenecks and matched planning budgets

In Table 2, what computational bottlenecks explain the limited runtime reduction despite fewer prediction steps? Please explain the discrepancy and report success under matched time or compute budgets.

Thank you for raising this point. Table 2's prediction depth counts sequential predictor calls, whose costs differ across methods. Longer chunks contain more action tokens, which increases per-call Transformer work since the attention term scales quadratically with token length. Complete CEM timing also includes sampling, elite updates, and batching overhead, so fewer prediction calls may not yield a proportional runtime reduction.

We additionally evaluated all planners at Δ=100\Delta=100, using LeWM's runtime with 30 CEM iterations as the target planning-time budget. Each planner's iteration count was calibrated from the existing timing measurements, keeping the checkpoints, evaluation protocol, 300 candidates, and 30 elites fixed. Success rates are percentages (mean ± standard deviation over three training seeds); iteration triples follow the environment order in the table.

MethodPushTOGBench-CubeTwoRoomAvg.CEM iterations
LeWM9.3 ± 2.342.0 ± 2.019.3 ± 3.123.630 / 30 / 30
LeWM+VLWM (P1)14.7 ± 1.254.0 ± 2.021.3 ± 2.330.031 / 30 / 32
LeWM+VLWM (P2)11.3 ± 2.352.0 ± 5.322.7 ± 1.228.726 / 28 / 31
LeWM+VLWM (P3)15.3 ± 2.360.0 ± 2.028.0 ± 2.034.426 / 30 / 31
LeWM+VLWM (P4)16.0 ± 4.058.0 ± 2.026.7 ± 3.133.635 / 30 / 31
LeWM+VLWM (P5)16.7 ± 4.261.3 ± 1.230.0 ± 2.036.035 / 31 / 32

Average success changes by at most 0.9 percentage points from Table 2. P3 and P5 achieve 34.4% and 36.0%, compared with 23.6% for LeWM, and P5 has the highest mean success in all three environments. The planning gains therefore persist after time-based iteration calibration. We include these results in Appendix J (Table 13) and explain the runtime interpretation in Section 4.3 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer Ykv8 (2/3)

Q3. Model capacity and parameter counts

Despite the claim of matched capacity, encoder, predictor, total, and trainable parameter counts are not explicitly reported. Could you provide these counts for the baselines and VLWM?

The counts below correspond to the models used in our experiments, with PushT shown here. Each entry uses (trainable/all), with frozen parameters included only in all.

MethodEncoderPredictorOtherTotal
LeWM(5,501,376/5,501,376)(10,791,360/10,791,360)(1,741,742/1,741,742)(18,034,478/18,034,478)
LeWM+VLWM(5,501,376/5,501,376)(10,792,128/10,792,128)(1,741,742/1,741,742)(18,035,246/18,035,246)
DINO-WM(0/21,585,792)(19,261,824/19,261,824)(1,590/1,590)(19,263,414/40,849,206)
DINO-WM+VLWM(0/21,585,792)(24,234,240/24,234,240)(1,590/1,590)(24,235,830/45,821,622)

For LeWM, VLWM adds only 768 parameters (0.0043%), while retaining the same encoder and auxiliary modules. For DINO-WM, both variants use the same frozen DINOv2-S/14 encoder at 168×168 resolution and the same action/proprio encoders. Replacing the predictor adds 4,972,416 trainable parameters (25.8%) for encoding action tokens. The VLWM predictor preserves the 384-dimensional patch representation and shares one prediction core across patches, with the conditioning adapter included in the Predictor count. “Other” comprises the action encoder and two projection modules for LeWM, and the action/proprio encoders for DINO-WM.

We include the complete per-environment parameter counts in Appendix M (Tables 20–21), with backbone-specific predictor configurations in Appendices C and M of the revised manuscript.

Q4. P1–P5 across goal offsets and backbones

The full P1–P5 comparison is provided only at the largest goal offset. Could you provide results across all goal offsets for both backbones, especially P1 with its maximum fixed prediction span? These comparisons would establish whether VLWM consistently improves over the one-step baseline without sophisticated chunking and quantify the additional gains of P2–P5 over P1.

Thank you for requesting this comparison. We now report P1–P5 on PushT for both LeWM and DINO-WM at all four goal offsets. P1 uses the maximum fixed chunk length, k=5k=5. For each backbone, P1–P5 share the trained VLWM checkpoints and use the execution budget B=2ΔB=2\Delta. Entries are success rates (%, mean ± standard deviation over three training seeds); bold marks the best mean within each backbone and offset, including ties.

MethodΔ=25\Delta=25Δ=50\Delta=50Δ=75\Delta=75Δ=100\Delta=100
LeWM90.0 ± 8.050.7 ± 8.114.7 ± 3.19.3 ± 2.3
LeWM+VLWM (P1)94.0 ± 2.059.3 ± 3.118.0 ± 2.014.0 ± 2.0
LeWM+VLWM (P2)93.3 ± 1.258.0 ± 6.018.0 ± 5.312.0 ± 2.0
LeWM+VLWM (P3)94.7 ± 4.259.3 ± 1.222.7 ± 3.114.7 ± 3.1
LeWM+VLWM (P4)93.3 ± 3.161.3 ± 3.122.0 ± 2.015.3 ± 3.1
LeWM+VLWM (P5)94.7 ± 1.260.7 ± 3.123.3 ± 3.115.3 ± 3.1
DINO-WM65.3 ± 2.314.7 ± 1.29.3 ± 1.22.7 ± 1.2
DINO-WM+VLWM (P1)68.0 ± 2.022.0 ± 2.014.7 ± 3.18.0 ± 2.0
DINO-WM+VLWM (P2)68.7 ± 4.223.3 ± 3.114.7 ± 1.29.3 ± 4.2
DINO-WM+VLWM (P3)70.0 ± 2.024.0 ± 4.015.3 ± 2.39.3 ± 2.3
DINO-WM+VLWM (P4)67.3 ± 4.625.3 ± 4.616.0 ± 5.39.3 ± 4.2
DINO-WM+VLWM (P5)66.7 ± 1.224.7 ± 3.114.7 ± 3.18.7 ± 3.1

P1 improves over the one-step baseline at every offset on both backbones, with an average gain of 5.2 percentage points for each backbone. Thus, the planning benefit is already present with fixed-length chunking.

Averaged over both backbones and all four offsets, P2, P3, P4, and P5 change success relative to P1 by −0.1, +1.5, +1.5, and +1.3 percentage points, respectively. P2 is therefore comparable to P1 overall, while P3–P5 provide additional average gains.

We include this complete PushT comparison in Appendix I (Table 11) of the revised manuscript, with a reference in Section 4.2.

AuthorsOpenReview ↗

Response to Reviewer Ykv8 (3/3)

Q5. Random-policy success

Low success rates at the largest goal offset, including those of the baselines, raise questions about performance relative to a random policy. Could you report random-policy success under the same execution budget, especially at the largest goal offset?

Thank you for suggesting this comparison. We evaluated a random-action policy at all four offsets using the same execution budget, B=2ΔB=2\Delta environment steps. Random-policy success rates (%, mean ± standard deviation over three independent runs) are:

EnvironmentΔ=25\Delta=25Δ=50\Delta=50Δ=75\Delta=75Δ=100\Delta=100
PushT2.7 ± 3.12.7 ± 1.20.0 ± 0.00.0 ± 0.0
TwoRoom28.0 ± 4.08.0 ± 2.04.7 ± 1.20.7 ± 1.2
OGBench-Cube49.3 ± 4.232.0 ± 5.336.7 ± 3.133.3 ± 4.2
Reacher13.3 ± 1.210.7 ± 1.28.7 ± 1.25.3 ± 1.2

At the largest offset, Δ=100\Delta=100, the direct comparison with our reported model-based results is:

EnvironmentRandomLeWMLeWM+VLWM (P3)LeWM+VLWM (P5)
PushT0.0 ± 0.09.3 ± 2.314.7 ± 3.115.3 ± 3.1
TwoRoom0.7 ± 1.219.3 ± 3.127.3 ± 1.229.3 ± 2.3
OGBench-Cube33.3 ± 4.242.0 ± 2.060.0 ± 2.060.7 ± 4.2

Both LeWM and LeWM+VLWM outperform the random policy largely across all evaluated offsets on these three benchmarks. At the largest offset specifically, VLWM exceeds random-policy success by 14.7–28.7 percentage points, while also improving over LeWM in every environment. These results demonstrate effective goal-directed planning even at the most challenging offset, with further benefits from variable-length prediction.

Q6. Curriculum direction versus horizon weighting

How does a reverse, long-to-short curriculum perform when cumulative sampling counts for each horizon, training budget, and learning-rate schedule are matched? Does the short-to-long ordering remain beneficial under these controls, separating curriculum direction from horizon weighting?

Following your suggestions, we evaluated LeWM+VLWM with P3 at Δ=100\Delta=100 on all three control benchmarks. The reverse condition reverses the forward schedule’s sampled horizon sequence while preserving the cumulative sampling count for each horizon, training budget, and global learning-rate schedule. No curriculum uses uniform sampling over the full horizon range throughout training. Results are success rates (%, mean ± standard deviation over three training seeds):

Training schedulePushTOGBench-CubeTwoRoom
No curriculum12.0 ± 2.054.0 ± 2.020.7 ± 3.1
Forward curriculum14.7 ± 3.160.0 ± 2.027.3 ± 1.2
Reverse curriculum9.3 ± 1.248.7 ± 1.218.7 ± 3.1

Forward curriculum performs best in all three environments, while reverse curriculum performs worse than no curriculum. Starting with long horizons and dropping them later may impair early optimization and long-horizon retention. The matched-count comparison supports the benefit of short-to-long curriculum ordering.

Thank you again for your detailed and constructive comments. We hope our responses address your concerns and welcome any further questions.

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

VLWM predicts a future latent directly from a variable-length action sequence, rather than repeatedly applying a one-step model. It combines action tokens, a horizon curriculum, and several ways to partition a plan into prediction chunks. Results on LeWM and DINO-WM show improved goal-reaching performance on three environments, with additional goal-conditioned action-ranking results on LIBERO.

Critical strengths
  • The improvement transfers to both jointly learned and frozen visual representations. Across three environments and four goal offsets, P3 and P5 improve success by 8.1 and 8.7 percentage points, respectively. Three training seeds and shared evaluation pairs strengthen the cross-backbone evidence.
  • One predictor supports direct prediction and chunked composition, and the simpler P3 captures most of the benefit. Appendix G also provides a useful within-model comparison: direct multi-step prediction has lower endpoint error than recursive one-step prediction at each tested multi-step horizon.
Critical weaknesses
  • The main control results need a narrower interpretation. Appendices B–D keep the actual planning window at five model steps, or 25 environment steps, even as the goal offset reaches 100. Larger offsets therefore test repeated short-window control rather than longer imagined rollouts. Appendix G supports reduced recursive prediction error within the trained range, but does not establish that increasing imagination length explains the widening control gains or that chunk composition remains effective beyond that range.
  • The mechanism behind the planning gains is not fully isolated. VLWM changes action conditioning, prediction targets, and the training schedule; a matched one-step action-token control would help separate these changes. Moreover, the diagnostic in Section 4.3 and Appendix F ranks candidates by prediction error against each candidate’s own realized endpoint, rather than by distance to a shared goal. A better reference rank under this diagnostic does not directly establish better action selection. The LIBERO shared-goal ranking experiment addresses this distinction, but only in that setting.
Questions for authors
  1. At fixed goal offset and execution budget, how does the performance gap change with the actual planning horizon, including a horizon beyond the maximum trained chunk length? This would test whether the benefit extends to longer composed predictions rather than only improved local control.
  2. How much of the control gain remains against a one-step action-token predictor with matched training settings? For the control benchmarks, can you also report agreement between shared-goal predicted costs and realized goal-reaching outcomes, rather than ranking candidates by their individual prediction errors?
2 comments
AuthorsOpenReview ↗

Response to Reviewer gxcS (1/2)

We thank Reviewer gxcS for recognizing the cross-backbone improvements. We appreciate your careful assessment and address your concerns point by point below.

Q1. Actual planning horizon versus goal offset

Appendices B–D keep the planning window at five model steps (25 environment steps), even when the goal offset reaches 100. At fixed goal offset and execution budget, how does the performance gap change with the actual planning horizon, including horizons beyond the maximum trained chunk length?

Thank you for the suggestion. We now directly compare H=5H=5 and H=10H=10 (25 and 50 environment steps) on all three environments, fixing the goal offset at Δ=100\Delta=100. All the other settings are fixed.

VLWM retains Kmax⁡=5K_{\max}=5. P3 and P5's partitions obey the same chunk-length constraint. Results are success rates (%, mean ± standard deviation over three training seeds); bold marks the best mean within each environment and horizon, including ties.

HH (model steps)MethodPushTOGBench-CubeTwoRoomAvg.
5LeWM9.3 ± 2.342.0 ± 2.019.3 ± 3.123.6
5LeWM+VLWM (P3)14.7 ± 3.160.0 ± 2.027.3 ± 1.234.0
5LeWM+VLWM (P5)16.0 ± 2.061.3 ± 3.130.0 ± 2.035.8
10LeWM2.7 ± 1.233.3 ± 6.15.3 ± 4.213.8
10LeWM+VLWM (P3)10.0 ± 2.054.7 ± 3.115.3 ± 3.126.7
10LeWM+VLWM (P5)11.3 ± 3.150.0 ± 4.020.0 ± 2.027.1

At H=10H=10, both VLWM planners outperform LeWM in every environment. P3's average advantage increases from 10.4 to 12.9 percentage points, and P5's from 12.2 to 13.3 percentage points. Absolute success is lower at H=10H=10 with the CEM settings held fixed, while VLWM retains its comparative benefit. This directly supports composing trained-range predictions into imagined rollouts beyond the maximum trained chunk length.

We include these results in Section 4.3 (Table 5) and Appendix K (Table 14) of the revised manuscript; Section 4.1 defines the goal offset, planning horizon and replanning interval.

Q2. Matched one-step action-token control

How much of the control gain remains against a one-step action-token predictor with matched training settings?

Following your advice, we trained this control using the VLWM action-token architecture with only one-step supervision. It shares the architecture, training data, update count, batch size, and global learning-rate schedule with the no-curriculum VLWM. During planning it recursively predicts five steps; VLWM-P1 predicts the same endpoint in one call. The no-curriculum VLWM samples training horizons uniformly from k=1,…,5k=1,\ldots,5.

Results below use LeWM on all three environments at Δ=100\Delta=100, with a planning horizon of five model steps, an execution budget of 200 raw steps, and 30 CEM iterations. Success rates are percentages (mean ± standard deviation over three training seeds).

Training / plannerPushTTwoRoomOGBench-Cube
Original LeWM9.3 ± 2.319.3 ± 3.142.0 ± 2.0
Action-token one-step10.0 ± 2.020.0 ± 4.042.7 ± 3.1
VLWM, no curriculum, P110.0 ± 2.019.3 ± 1.255.3 ± 5.0
VLWM, curriculum, P114.0 ± 2.020.7 ± 3.154.0 ± 2.0
VLWM, curriculum, P314.7 ± 3.127.3 ± 1.260.0 ± 2.0

The action-token one-step control remains close to original LeWM: their average success rates are 24.2% and 23.6%, respectively. Without curriculum or variable chunk selection, VLWM-P1 reaches 28.2%, with the clearest gain on Cube (42.7% → 55.3%). Curriculum-trained VLWM reaches 29.6% with P1 and 34.0% with P3. These comparisons locate the improvement beyond the action-token interface and shows the benefit of curriculum and multi-step prediction.

We include these controls in Section 4.3 (Table 1) and their training configurations in Appendix H of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer gxcS (2/2)

Q3. Prediction-error ranking versus goal-directed action selection

Section 4.3 and Appendix F rank candidates by prediction error against each candidate's own realized endpoint, rather than distance to a shared goal. For the control benchmarks, can you report agreement between shared-goal predicted costs and realized goal-reaching outcomes instead of ranking candidates by individual prediction errors?

Thank you for this suggestion. We evaluated agreement between shared-goal predicted costs and simulator-realized outcomes on PushT, OGBench-Cube, and TwoRoom, using 50 queries with 300 candidates each at h=1,…,5h=1,\ldots,5 and one checkpoint per method. Both methods score the same candidate pools against a shared goal. Outcomes are computed from physical simulator states using the task’s goal criteria, independently of either encoder.

We report two complementary measures: the predicted-cost ranking of simulator-optimal candidates, and the realized goal-distance gap of the model-selected candidate:

Δd=dreal ⁣(arg⁡min⁡icpred(i))−min⁡idreal(i).\Delta d=d_{\mathrm{real}}\!\left(\arg\min_i c_{\mathrm{pred}}(i)\right)-\min_i d_{\mathrm{real}}(i).

This gap measures the additional goal distance incurred by selecting according to the model’s predicted cost. Results at h=4h=4 and h=5h=5 are:

EnvironmenthhLeWM Top-1 (%) ↑+VLWM Top-1 (%) ↑LeWM Δd\Delta d ↓+VLWM Δd\Delta d ↓
PushT474.078.00.2520.222
PushT576.080.00.3650.243
OGBench-Cube452.360.31.293 mm1.179 mm
OGBench-Cube564.170.20.831 mm0.732 mm
TwoRoom470.074.00.755 pixels0.702 pixels
TwoRoom570.074.00.784 pixels0.621 pixels

VLWM uses P3. Top-1 inclusion averages uniformly over tied simulator optima. PushT distances use goal-criterion units: max⁡(Δposition/20,Δangle/(π/9))\max(\Delta\mathrm{position}/20,\Delta\mathrm{angle}/(\pi/9)), with wrapped angular differences.

At both h=4h=4 and h=5h=5, VLWM achieves higher Top-1 inclusion and a lower selected-candidate distance gap in all three environments. At h=5h=5, the gap reductions are 33.4% on PushT, 11.9% on OGBench-Cube, and 20.8% on TwoRoom. VLWM also reduces the distance gap across all five PushT horizons. These measurements directly connect shared-goal model scores to the physical outcomes of the selected actions.

We include this analysis in Section 4.3 and the complete ranking and realized-distance-gap results for all five horizons in Appendix L (Tables 17–19) of the revised manuscript.

Thank you again for your detailed and constructive comments. We hope our responses address your concerns and welcome any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.