acceptodds
Withdrawn from ICLR 2027

Adaptive Long-Horizon Latent Action Model

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Latent action models aim to learn controllable actions from unlabeled trajectories. However, the action-induced changes between consecutive frames are hard for models to distinguish from noise. Without action labels, reconstructing observed transitions alone does not explicitly require the model to separate action effects from noise. To overcome this problem, Adaptive Long-Horizon LAM is introduced. In the real world, few controllable actions can be completed to reach a target in one frame. Also, they will exhibit strong temporal persistence and dictate a consistent state transition throughout their lifespan, while exogenous noise is stochastic and lacks dynamical consistency over time. The key idea of ALHLAM is to force the model to learn the consistent parts of changes. This constraint can be viewed as an unsupervised consistency bottleneck. In the meantime, fixed long-horizon cannot fit into all situations. These temporal supports also guide action chunk discovery. The experiments on world models show that ALHLAM is better than fixed long-horizon.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
latent action models, unsupervised representation learning, temporal abstraction, action boundary detection

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_TuXUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ALHLAM, a latent action model that infers one shared latent action from a window of 2k+1 observations and requires the forward model to explain all 2k transitions in that window with a single predicted effect. During training the radius k is sampled uniformly from 1 to 8. After training, each center selects its radius by minimizing the average reconstruction error minus λ log k, trading consistency for longer support. The selected radii are then used, with the frozen model, to discover action chunk boundaries by finding valleys in the radius sequence and confirming them with a z-scored jump in reconstruction error, and to pick one representative latent action per chunk. Experiments use a synthetic 64-dimensional additive environment with four fixed action effects, piecewise-constant chunks of 4 to 24 steps, and isotropic Gaussian noise at two amplitudes. Against a single baseline that uses the same architecture and a fixed radius of 4 but supervises only the center transition, ALHLAM improves linear-probe action accuracy from 74.7 to 84.1 percent at low noise and from 72.8 to 79.5 at high noise, and raises the fraction of selected windows containing a single action from 52 to 73 percent. Boundary detection reaches F1 of about 0.50 at one-step tolerance, chunk representatives classify the whole chunk correctly 70 percent of the time, and an exploratory extension to concurrent action factors collapses to a single factor.

Critical strengths
  1. The core question is relevant. How much temporal context a latent action should explain, and whether that extent can be chosen per sample without labels, is a real open design question in latent action modeling that existing methods fix by hand.
  2. The shared-window reconstruction objective (Eq. 4) is a clean way to force the latent to encode what is consistent across transitions, and the per-center selection rule (Eq. 6) is simple with one hyperparameter.
  3. Reusing the selected radii as a boundary cue is a reasonable idea that requires no additional model and ties the two parts of the paper together.
  4. The replacement diagnostics verify that the forward model actually uses the latent action rather than ignoring it.
Critical weaknesses
  1. The evaluation is confined to a linear toy environment. All experiments use 64-dimensional vectors with additive constant effects and isotropic Gaussian noise. There are no image observations, no nonlinear dynamics, no real distractors of the kind studied by LAOM or MaskLAM, and no standard LAM benchmark, so none of the claims about latent action models in general is tested.
  2. The abstract misstates the evidence. It says "experiments on world models show that ALHLAM is better than fixed long-horizon," but no world model is trained or evaluated.
  3. A single baseline cannot isolate the contribution. The paper itself notes that the comparison "does not separate the contribution of either individual component." Fixed-radius shared-window variants at each k from 1 to 8, a LAPO-style one-step model, and random-endpoint models like LAOM are all absent, so it is unknown whether adaptive selection adds anything over shared-window training at a good fixed radius.
  4. The toy problem has a trivial non-learned solution that is not compared. In an additive setting with piecewise-constant effects, the optimal estimate of the effect is the mean of consecutive differences within a segment, and the boundary task is classical change-point detection. A running mean with a standard change-point detector would likely match or beat the reported results, and no such baseline is included.
  5. No upper bound is reported. With known boundaries, averaging differences over a segment of 4 to 24 steps would make the four effects nearly perfectly separable at both noise levels, so the 79 to 84 percent probe accuracy suggests the method is far from the ceiling, but the reader cannot tell.
  6. The selection rule appears biased to the maximum radius. The supplied k-distribution file shows 15 percent of centers select k = 8 with only 36 percent of those windows clean, more than any other radius, which suggests the log k reward with λ = 0.1 is over-rewarding length. λ, K_max, and β are never ablated.
Questions for authors
  1. What probe accuracy and effect R² does an oracle achieve that averages Δo over the true segment?
  2. Why does k = 8 have the highest selection frequency despite the lowest clean-window rate? How do the results change with λ in {0, 0.05, 0.2} and K_max in {4, 16}?
Reviewer_NvzJOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Adaptive Long-Horizon Latent Action Model (ALHLAM), a method for learning latent actions from unlabeled observation trajectories. The key motivation is that controllable action effects often persist over multiple time steps, whereas independent transition noise lacks temporal consistency. To exploit this distinction, the authors propose a shared-window reconstruction objective that requires a single latent action to explain all transitions within a temporal window. The method further selects an adaptive temporal support for each center by balancing reconstruction consistency and window length.

The selected temporal supports are also used for unsupervised action chunk discovery without additional model training. Experiments in controlled synthetic environments with different noise levels show improvements in action classification accuracy, action-effect prediction, and temporal-support selection compared with a fixed-context, single-step reconstruction baseline. The paper also investigates concurrent latent action factorization, where the proposed extension exhibits factor collapse and fails to recover the underlying action factors reliably.

Critical strengths

1. Novel and well-motivated learning objective.

The idea of using a shared latent action to explain multiple consecutive transitions is interesting and well motivated. Unlike approaches that primarily use temporal observations as additional context, the proposed method explicitly encourages consistency across transitions within a window. This provides a meaningful way to exploit temporal persistence when learning latent actions without action supervision. I find this to be the main strength of the paper.

2. Adaptive temporal support with encouraging empirical results.

The adaptive temporal support mechanism is a natural extension of the shared-window objective, as a fixed window may include transitions caused by different actions. The experimental results are encouraging: ALHLAM consistently improves action-probe accuracy and reduces unnecessary action-boundary crossings compared with the fixed-context baseline under both noise conditions. Furthermore, reusing the learned temporal supports for action chunk discovery without additional model training is an interesting extension of the framework.

Critical weaknesses

1. Insufficient comparison with existing latent action models.

My main concern is the lack of direct comparisons with representative existing latent action learning methods. The experiments primarily compare ALHLAM against Fixed-4-OneStep, a fixed-context baseline with single-step reconstruction supervision. While this comparison demonstrates the benefits of the proposed design over this particular baseline, it remains unclear how ALHLAM performs relative to prior approaches such as CLAM or LAOM.

Given that learning latent actions from temporally extended observations has already been explored in the literature, I believe that a direct comparison with at least one representative existing method is important for establishing the empirical significance of the proposed contribution. Even one additional baseline evaluated under the same experimental setting would substantially strengthen the paper. This is my most significant concern and has a considerable impact on my overall assessment.

2. Limited ablation of the proposed components.

The main comparison changes both the reconstruction objective and the temporal support selection strategy simultaneously. Consequently, it is difficult to determine how much of the improvement comes from shared multi-transition reconstruction and how much comes from adaptive support selection. A fixed-window baseline trained with the proposed shared reconstruction objective would help isolate the contribution of the adaptive component. This distinction is particularly important because both components are presented as central elements of the proposed method.

Questions for authors

Q1. Comparison with existing methods (most important).

Could the authors provide a direct comparison with at least one representative existing latent action model, such as CLAM or LAOM, under the same synthetic experimental setting? Ideally, the comparison would use the same data splits, evaluation metrics, and comparable model capacity where applicable.

I understand that reproducing multiple existing methods may require substantial effort, but I believe that even one additional representative baseline would be highly valuable. The proposed method is interesting and shows promising results, yet this comparison is important for understanding whether its advantages extend beyond the specifically designed Fixed-4-OneStep baseline. The answer to this question would materially influence my assessment of the paper.

Q2. Contribution of individual components.

Could the authors provide an ablation that uses the proposed shared multi-transition reconstruction objective but keeps the temporal support fixed? This would help clarify whether the main performance improvements are driven primarily by shared reconstruction or by adaptive temporal support selection.

Reviewer_4m2iOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Adaptive Long-Horizon LAM, which learns a shared latent action to explain all observation transitions within a temporal window. The model is trained using randomly sampled window lengths, and the temporal support is selected after training by balancing reconstruction error against a preference for longer windows. The selected supports are also used for offline action-boundary detection and chunk representation. Experiments use synthetic additive dynamics with piecewise-constant action effects and independent Gaussian transition noise. The method improves action-probe accuracy and reduces boundary crossings relative to a fixed-context, single-transition baseline. An exploratory extension to concurrent action factors is also presented, but fails to recover the underlying factorization.

Critical strengths
  1. The paper addresses noise contamination in latent action learning with a clear motivation: using temporal persistence to estimate consistent action effects.
  2. Across two noise levels and three training seeds, the method improves action classification accuracy and reduces boundary-crossing windows relative to the evaluated baseline.
  3. The paper transparently reports limitations in boundary localization and the unsuccessful concurrent-action factorization experiment.
Critical weaknesses
  1. The comparison with Fixed-4-OneStep changes several components simultaneously: the training window lengths, the reconstruction objective, and the inference-time window selection. The reported improvements therefore do not establish how much adaptive support selection contributes beyond supervising multiple transitions. A fixed-window model trained with the same shared reconstruction objective, together with fixed versus adaptive window selection using the same trained model, would help isolate the proposed mechanism.
  2. The shared reconstruction objective has a direct connection to temporal averaging. For a fixed window, minimizing the squared error of a common prediction is equivalent to fitting the mean transition increment, up to a prediction-independent term. Under the paper’s constant-effect and independent-noise assumptions, averaging already provides a natural denoising mechanism. However, the experiments do not compare against fixed or adaptive averaging baselines, optionally followed by linear dimensionality reduction. It therefore remains unclear whether the learned representation provides substantial benefits beyond estimating a locally averaged transition effect.
  3. All main experiments use synthetic trajectories with four fixed action-effect vectors and independent, zero-mean Gaussian transition noise. This setting closely matches the assumptions that motivate the method. The experiments do not test temporally correlated disturbances, state-dependent action effects, or variation within an action segment. These cases matter because persistent disturbances may survive temporal averaging, while a single action with changing effects may appear inconsistent within a window. Real-robot experiments are not necessarily required, but controlled tests that relax these assumptions are needed to establish the method’s scope.
  4. The action segmentation results remain limited even in this controlled setting. Exact-boundary F1 is only 0.065 and 0.095 at the two noise levels; with a tolerance of two steps, it reaches 0.606 and 0.496. The comparison with “Valleys Only” evaluates the confirmation step but does not establish competitiveness against independent change-point detection methods. Segment representations also lose accuracy, particularly for the fallback cases, which achieve only about 38%. Moreover, the supplied manuscript does not fully specify the numerical settings and selection procedures for several segmentation parameters, making the pipeline difficult to reproduce and its sensitivity difficult to assess.
Questions for authors

My questions are covered in the weaknesses above.

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

Related papers

Discussion (0)

Sign in to comment.