- 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
- 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.
- 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.
- 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.
- The replacement diagnostics verify that the forward model actually uses the latent action rather than ignoring it.
Critical weaknesses
- 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.
- 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.
- 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.
- 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.
- 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.
- 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
- What probe accuracy and effect R² does an oracle achieve that averages Δo over the true segment?
- 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}?