Source-linked AI summary
Order-Consistent Latent Geometry for Reliable Short-Horizon Planning in World Models
Junlin Wu, Wen Fan, Dandan Zhang
TL;DR
Terminal Euclidean latent-goal scores may reverse true transition progress even when rollouts are predicted exactly. The paper identifies a pathwise ordering condition, analyzes polygonal rollouts, and derives acute-bend candidate reversals; a six-branch study reports misranking and success differences across embeddings.
Problem
Terminal Euclidean latent-goal scores need not preserve true transition progress, so exact rollout prediction alone does not ensure correct candidate ordering.
Source: §1 · pp. 1–2Method
The paper identifies pathwise transition-order consistency with self-approachingness and analyzes discrete latent rollouts through polygonal interpolation.
Source: §1 · p. 2, §2.2 · p. 3Results
At H = 0.05, 2D and padded embeddings yield 8.0% strict misranking and 60.0% success; cross-polytope yields zero strict misranking and 100% success.
Source: §3 · p. 5Takeaways & Limitations
Terminal-distance guidance needs no global isometry, but the guarantees are pathwise for graph-like point bottlenecks with locally reversible branches.
Source: §4 · p. 6The branch-capacity result assumes locally distinct branches share a state and every pair may form a transition-shortest two-segment rollout.
Source: §2.3 · p. 4
Abstract
from arXiv · showhide
Latent planners often rank candidate action sequences by terminal Euclidean distance to a goal latent. Exact agreement with transition distance would make this score ideal, but is stronger than planning requires: along a transition-shortest rollout, the essential requirement is that latent distance does not reverse true transition progress. We show that this pathwise no-reversal requirement is precisely the classical self-approaching condition. To the best of our knowledge, we are the first to identify self-approachingness as a planner-facing compatibility condition for terminal Euclidean latent planning and to connect its violation to an exact finite-horizon candidate-order reversal. Because visual world models predict discrete latent sequences, we analyze their polygonal interpolation. We prove that an acute bend creates a nonempty region in which terminal Euclidean distance ranks a transition-worse candidate ahead of one that makes genuine progress, even under exact rollouts. We further interpret latent dimension as angular capacity through a classical spherical-code bound. A controlled six-branch experiment, including a zero-padded three-dimensional baseline, validates the predicted ordering failure and separates angular arrangement from ambient dimensionality.
1 Introduction
Latent planners can predict candidate rollouts exactly yet choose actions that move in the wrong direction because terminal Euclidean scoring may reverse true transition progress. The paper identifies self-approachingness as the weaker ordering condition needed for this score and connects acute bends to finite-horizon misranking.
1 Introduction
- Latent world models plan over predicted candidate consequences, and goal-conditioned systems can optimize toward a latent image goal without decoding future pixels.The paradigm spans latent control, differentiable planning, and online model predictive control; related goal-conditioned and reconstruction-free systems are cited as [7] [8].Source: §1 · p. 1
- Exact rollout prediction can still pair with a misleading objective when terminal Euclidean distance ranks a transition-inferior candidate ahead of one making progress.The paper frames reliable planning as depending on the ordering induced by latent geometry, not predictive accuracy alone.Source: §1 · p. 1
- Pathwise transition-order consistency, equivalent to self-approachingness, requires preserving progress ordering without reproducing transition distances numerically.Exact metric fidelity would force every shortest latent rollout to be straight, while transition structures may contain bends, obstacles, merges, and branches.Source: §1 · p. 2
- Prior work studied planning-aware representations and geometric routing, while this paper distinguishes its focus as horizon-truncated terminal objectives and exact candidate-order reversal.The paper does not claim that raw latent proximity can be insufficient is a new general observation; it cites prior work on structured transition models, learned action distances, curvature regularization, and reachability-aware objectives [11] [12] [13].Source: §1 · p. 2
- The paper derives a sharp non-acute criterion and a finite-horizon wrong-way optimum, then validates the mechanism in a six-branch stress test with a padded-dimension control.The stated angular-capacity bound is n ≤ 2D.Source: §1 · p. 2
2 Theory
The theory replaces exact latent-distance fidelity with a pathwise ordering condition and analyzes discrete predicted rollouts through polygonal interpolation. It shows how acute bends can reverse finite-horizon candidate rankings and bounds how many branches can remain pairwise non-acute in a given latent dimension.
2.1 Pathwise Transition-Order Consistency
- A differentiable shortest rollout is order-consistent exactly when future points stay in the forward tangent half-space, so nonzero curvature alone is not failure.The condition is self-approachingness and preserves true transition-progress ordering without requiring numerical distance agreement.Source: §2.1 · pp. 2–3
2.2 Polygonal Rollouts and Acute-Bend Reversal
- Polygonal interpolation joins consecutive predicted latent states with straight segments, providing a continuous geometric analysis of discrete rollouts without assuming globally linear dynamics.Analyzing the interpolation makes ordering reversals independent of temporal sampling resolution.Source: §2.2 · p. 3
- A two-segment shortest rollout preserves ordering for every source and goal radius if and only if its outward directions are non-acute; acute angles create strict reversals.For cross-segment points, squared latent distance has derivative 2(r − b u⊤v), which is nonnegative for all positive radii exactly when u⊤v ≤ 0.Source: §2.2 · p. 3
- When c > 0 and r < bc, terminal Euclidean scoring ranks a transition-worse backward candidate ahead of an inward candidate that reduces remaining transition distance.This is a strict pairwise inversion and does not require those candidates to be the planner’s only options.Source: §2.2 · pp. 3–4
- On the studied same-branch reachable interval, the global Euclidean-loss minimizer moves away from the turning state when c > 0 and r < bc.The result assumes the short-horizon reachable set stays on the same branch; the preceding ordering results assume exact candidate rollouts, so the failure comes from scoring geometry rather than prediction error.Source: §2.2 · p. 4
2.3 Latent Dimension as Angular Capacity
- If every pair among n branches meeting at a state must be non-acute in R^D, then n ≤ 2D, and the bound is tight for subsets of ±e1, . . . , ±eD.The proof uses Rankin’s spherical-code bound A(D, π/2) = 2D and the cross-polytope construction.Source: §2.3 · p. 4
- Increasing latent dimension allows more transition-distinct branches to remain geometrically separated without acute configurations that reverse pathwise ordering.The branch-direction constraint applies when every pair of branches may form a transition-shortest two-segment rollout.Source: §2.3 · p. 4
3 Controlled Six-Branch Study
The six-branch study tests how latent branch geometry affects terminal-distance planning. Acute hexagon embeddings produce misrankings and lower success, while a cross-polytope avoids them; zero-padding shows that angular arrangement, not dimension alone, matters.
3 Controlled Six-Branch Study
- The 2D hexagon and its zero-padded 3D copy have six acute branch pairs and acute loss 1.5; the cross-polytope has no acute pairs and zero acute loss.Zero-padding changes ambient dimension while preserving pairwise angles.Source: §3 · p. 4, Figure 1 · p. 5
- At H=0.05, 2D/padded hexagons had 8.0% strict misranking and 60.0% success; at H=0.30, 3.6% misranking and 79.6% success, while cross-polytope stayed at zero and 100%.The 2D and zero-padded 3D hexagons overlap at every horizon; the cross-polytope removes the acute conflicts, misrankings, and associated planning failures.Source: Figure 2 · p. 5, §3 · pp. 5–6
- The planner discretizes branches at 0.05, enumerates terminal states within budget H, minimizes terminal Euclidean score, executes one graph step, and replans.The protocol evaluates all ordered branch pairs and start/goal radii from 0.1 to 1.0, with 3,000 queries per horizon.Source: §3 · p. 5
- Acute-embedding errors vanish once the horizon reaches beyond the local trap, while the padded control shows that the cross-polytope's gain comes from angular arrangement, not coordinate count alone.The 2D and zero-padded 3D conditions are identical at every horizon.Source: §3 · pp. 5–6, Figure 1 · p. 5
4 Scope and Conclusion
Terminal-distance planning requires scores not to reverse progress along acted-on rollout prefixes, rather than a global isometry. Acute mandatory junctions can create wrong-way optima; the guarantees are pathwise and limited to graph-like point bottlenecks with locally reversible branches.
4 Scope and Conclusion
- The guarantees cover pathwise ordering on graph-like point bottlenecks with locally reversible branches, not global ranking across unrelated paths or an impossibility result for continuous-action tasks.Acute mandatory junctions create exact wrong-way optima.Source: §4 · p. 6