Organizations: Institute of Integrated Circuits and Artificial Intelligence, Qiantang Institute for Advanced Study, Hangzhou, China · Zhejiang University, Hangzhou, China · Guangdong University of Technology, Guangzhou, China
Latent world models plan towards a goal image by minimising the distance between predicted and goal latents. This distance is not geodesic: it has local minima, latent traps, behind walls, at corners and in dead ends, and memoryless receding-horizon replanning keeps the agent in them, revisiting states far from the goal. We propose Episodic Kernel Repulsion (EKR), a training-free cost term that penalises candidate plans by a top-k Gaussian kernel to a novelty-gated memory of the latents the agent has actually visited, rescaled by the current base cost so that its strength is comparable across models. Across two JEPA world models and four benchmarks, EKR raises average success from 31.8% to 46.7% (LeWM) and from 54.7% to 64.0% (H-JEPA) with the same model, planner and budget. Same-memory counterfactuals show that, in most settings, EKR pushes plans away from visited states, in navigation, often at a higher base cost; stuck episodes and revisits drop accordingly. In seven of eight settings it matches or exceeds 4× samples at 0.8--1.5× compute. Without tuning, it transfers to DINO-WM with CEM: +6.7 and +19.3 points on official and longer-goal Wall tasks, and from 44.7% to 100.0% and from 21.0% to 32.0% on long-horizon PointMaze and Push-T.
Figures & tables
Backbone
Method
AntMaze
FourRoom
Push-T
Cube
Avg.
LeWM
Baseline
19.3 ± 6.1
41.3 ± 14.7
40.0 ± 3.5
26.7 ± 1.2
31.8
+ EKR (ours)
30.0 ± 9.2 ∗
62.7 ± 5.8 ∗∗∗
54.7 ± 4.6 ∗∗
39.3 ± 7.6 ∗∗
46.7
Δ
+10.7 ( ↑ 55% )
+21.3 ( ↑ 52% )
+14.7 ( ↑ 37% )
+12.7 ( ↑ 48% )
+14.8 ( ↑ 47% )
H-JEPA
Baseline
66.0 ± 4.0
92.7 ± 4.2
10.0 ± 3.5
50.0 ± 9.2
54.7
+ EKR (ours)
73.3 ± 9.5
97.3 ± 2.3
21.3 ± 5.0 ∗∗
64.0 ± 12.0 ∗
64.0
Δ
+7.3 ( ↑ 11% )
+4.7 ( ↑ 5% )
+11.3 ( ↑ 113% )
+14.0 ( ↑ 28% )
+9.3 ( ↑ 17% )
Table 1: Planning success rate (%) on four benchmarks. Mean ± std over 3 seeds (42/43/44), 50 official evaluation tasks per seed; official planner and budget (AntMaze 350, FourRoom 210, Push-T 90, Cube 30 steps). Δ : absolute gain of EKR over the baseline (relative gain in parentheses). Exact McNemar test on 150 paired episodes: ∗p<0.05 , ∗∗p<0.01 , ∗∗∗p<0.001 .
Figure 1: A latent trap on AntMaze (LeWM, seed 42, task 13). (a) Top-down trajectories. The baseline (grey) walks into a side pocket and stays there (purple: steps labelled as trapped); EKR (orange) goes around and reaches the goal (star). (b) While trapped, the baseline’s base cost (blue) stays on a plateau at a median of 1.11× its initial value (coefficient of variation 0.09), although the true geodesic distance (black) stays at a median of 0.51 of the initial distance. EKR reaches the goal at step 207 (curves truncated at success).
Figure 2: Success vs. planning compute (log scale, relative to the 1× baseline; definition in Appendix B ). Grey line: baseline with 1 – 8× samples (shaded: 95% bootstrap CI); square: 2× iterations; star and dashed orange line: EKR at the default budget, labelled with its lead over the strongest baseline point and that point’s compute relative to EKR. Three seeds × 50 tasks, all points on the same set. Manipulation panels were chosen by their lead at 1 – 4× before the larger budgets were run; AntMaze/H-JEPA was added afterwards as the navigation example. All eight cells: Figure 5 .
Figure 3: Same-memory counterfactual. Fraction of replans in which the EKR plan ends farther from memory than the λ=0 plan solved from the same state with the same memory, inside and outside traps (dotted: 50%). Hatched: Cube/H-JEPA, selection only; LW: LeWM; HJ ℓ : H-JEPA level ℓ . Seed 42; AntMaze/LeWM 10 episodes, others 50.
Figure 4: EKR reduces trapping in navigation. (a) Fraction of episodes stuck at the end, (b) revisit rate, (c) coverage of EKR relative to the baseline, (d) outcome composition (B: baseline, E: EKR; LW/HJ/DW: LeWM/H-JEPA/DINO-WM; off./long: official/long split); for Wall, the early-trap category means trapped behind the wall, i.e. failing to find the doorway; for PointMaze it means trapped in the start corridor of the U-maze. AntMaze (LeWM, H-JEPA): n=100 paired episodes (2 seeds); FourRoom: n=150 (both backbones); DINO-WM Wall: official and long-goal task sets, 150 episodes each; DINO-WM PointMaze: our pre-registered long-horizon split (budget 150 steps, 150 episodes; on the official 75-step tasks both arms solve 100.0%, Table 3 ). Error bars: 95% episode-level bootstrap confidence intervals (2000 resamples). Manipulation tasks are in Figure 10 .
FourRoom
Push-T
Cube
AntMaze †
Variant
LeWM
H-JEPA
LeWM
H-JEPA
LeWM
H-JEPA
LeWM
Baseline
41.3
92.7
40.0
10.0
26.7
50.0
19.3
EKR (full)
62.7
97.3
54.7
21.3
39.3
64.0
30.0
What to remember
start state only
52.7
98.0
45.3
12.7
40.7
64.7
17.3
imagined (predicted) states
–
–
–
–
25.3
52.0
–
Table 2: Ablations and exploration controls (success %, 3 seeds × 50 tasks unless noted). Each row changes one component of EKR; the last block replaces the repulsion by other exploration signals at the same budget. Count-based novelty uses latent SimHash with 8 or 4 bits; we report the better of the two for each cell, which favours the control. † AntMaze/LeWM: 3 seeds × 50 tasks; remaining rows not run. The imagined-state, no- cref , no-gate, count-based and action-noise variants were run on Cube only (–: not run). AntMaze/H-JEPA results are reported in the text.
DINO-WM task (budget)
Baseline
EKR
Δ
Wall, official (75)
90.0
96.7
+ 6.7 ∗∗
Wall, long (150)
75.3
94.7
+ 19.3 ∗∗∗
PointMaze, official (75)
100.0
100.0
+ 0.0
PointMaze, long † (150)
44.7
100.0
+ 55.3 ∗∗∗
Push-T, official (75) ‡
96.0
98.0
+ 2.0
Push-T, long † (150) ‡
21.0
32.0
+ 11.0 ∗
Table 3: Generalisation with the unified configuration (success %): DINO-WM with official weights, tasks and CEM planner; Wall long-goal set and our pre-registered † long-horizon PointMaze and Push-T splits use a 150-step budget. Three planning seeds × 50 tasks ( ‡ Push-T: 1 seed official, 2 seeds long); ∗p<0.05 , ∗∗p<0.01 , ∗∗∗p<0.001 , paired exact McNemar test (EKR-only/baseline-only successes: Wall 11:1 and 29:0; PointMaze long 83:0, p=2×10−25 ; Push-T 1:0 and 16:5, p=0.027 ). LeWM/H-JEPA on OGBench PointMaze Large: Table 4 .
Model
Base
EKR unified
EKR tuned
Δtuned [95% CI]
b01:b10
Progress B → T
LeWM
9.3
14.7
18.0
+8.7∗∗∗[4.7,13.3]
13:0
0.26 → 0.34
H-JEPA
0.0
0.7
4.0
+4.0∗[1.3,7.3]
6:0
0.10 → 0.49
Table 4: LeWM/H-JEPA on OGBench PointMaze Large (success %, 3 seeds × 50 official tasks, paired). Tuned: hyperparameters selected on disjoint validation tasks ( Appendix C ); planner, budget and baseline unchanged. b01:b10 (tuned) and exact McNemar test, ∗p<0.05 , ∗∗∗p<0.001 . Progress: mean fraction of the start–goal geodesic distance covered, baseline (B) to tuned (T).
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Environment
LeWM
H-JEPA (L1/L2/L3)
AntMaze
100
100 / 100 / 400
FourRoom
200
50 / 200 / 200
Push-T
800
25 / 50 / 200
Cube
800
25 / 800 / 800
Appendix
Table 5: Official number of planner samples per replanning step (L1–L3: H-JEPA hierarchy levels).
Model
Env
λ
K
levels
σ mode
Mmax
α
where applied
LeWM
AntMaze
1.2
3
–
obs ( ×2 )
256
0.1
gradient + selection
LeWM
FourRoom
2.0
3
–
obs ( ×2 )
256
0.1
gradient + selection
LeWM
Push-T
1.0
5
–
obs ( ×2 )
256
0.1
gradient + selection
LeWM
Cube
0.5
1
–
endpoint ( τσ=1.5 )
128
0.1
gradient + selection
H-JEPA
AntMaze
1.0
3
L3
endpoint ( τσ=1 )
128
0.1
gradient + selection
H-JEPA
FourRoom
1.5
3
L1–L3
endpoint ( τσ=1 )
128
0.1
gradient + selection
Appendix
Table 6: EKR configuration per cell. Mmax : memory capacity; α : novelty coefficient of Equation 6 . “obs”: observation calibration σ=st (twice the median real-trajectory spacing, β=2 ); “endpoint”: endpoint calibration ( 10 ). All cells use W=64 and β=2 (Cube: β=1 ) in Equation 5 .
Model
Env
Baseline (seeds 42/43/44)
EKR (seeds 42/43/44)
wins / losses ( p )
LeWM
AntMaze
7 / 9 / 13
20 / 14 / 11
33 / 17 (0.033)
LeWM
FourRoom
15 / 29 / 18
28 / 33 / 33
41 / 9 ( 5.6×10−6 )
LeWM
Push-T
21 / 21 / 18
26 / 30 / 26
38 / 16 (0.0038)
LeWM
Cube
13 / 14 / 13
18 / 17 / 24
34 / 15 (0.0094)
H-JEPA
AntMaze
33 / 35 / 31
42 / 35 / 33
33 / 22 (0.18)
H-JEPA
FourRoom
44 / 48 / 47
48 / 50 / 48
11 / 4 (0.12)
Appendix
Table 7: Per-seed successes (out of 50) and paired win/loss counts (EKR succeeds and baseline fails / vice versa) over 150 episodes.
Official tasks
Held-out tasks (seed 10042)
Env.
Backbone
Baseline
EKR
Δ [95% CI]
Baseline
EKR
Δ [95% CI]
AntMaze
LeWM
19.3 ± 6.1
30.0 ± 9.2
+10.7 [1.3, 20.0]
25.3 ± 4.2
38.7 ± 3.1
+13.3 [4.7, 21.3]
AntMaze
H-JEPA
66.0 ± 4.0
73.3 ± 9.5
+7.3 [-2.7, 16.7]
70.0 ± 8.7
72.0 ± 12.5
+2.0 [-6.7, 11.3]
FourRoom
LeWM
41.3 ± 14.7
62.7 ± 5.8
+21.3 [12.7, 30.0]
34.7 ± 4.2
56.7 ± 4.2
+22.0 [13.3, 31.3]
FourRoom
H-JEPA
92.7 ± 4.2
97.3 ± 2.3
+4.7 [0.0, 9.3]
95.3 ± 3.1
99.3 ± 1.2
+4.0 [0.7, 8.0]
Push-T
LeWM
40.0 ± 3.5
54.7 ± 4.6
+14.7 [6.0, 24.0]
34.7 ± 8.1
46.0 ± 3.5
+11.3 [3.3, 20.0]
Appendix
Table 8: Held-out task set vs. official tasks (success %, mean ± sd over 3 seeds × 50 tasks). Held-out tasks are drawn from the official distribution with a new task seed (10042), with the same success criterion and budget; configurations are frozen from Table 6 , and baseline and EKR are run on the same tasks. Δ : paired difference EKR − baseline in points with bootstrap 95% CI over matched (seed, task) pairs.
Model
Candidate
Hyperparameters
Val. success
Val. progress
LeWM
U
λ=1.5 , K=3
2/10
0.28
LeWM
C
λ=1.5 , K=3 , σ×4
2/10
0.32
LeWM
D ⋆
λ=2 , K=5
3/10
0.35
H-JEPA
U
λ=1.5 , K=3
2/10
0.51
H-JEPA
E
λ=3 , K=5
3/10
0.53
H-JEPA
F ⋆
λ=3 , K=5 , gate ×4
3/10
0.64
Appendix
Table 9: Validation results of all EKR candidates on PointMaze Large ( ⋆ : selected; U: unified configuration; σ×c : LeWM observation-mode bandwidth multiplier c instead of 2; gate ×c : H-JEPA novelty-gate scale c instead of 2, bandwidth unchanged).
Model
Env
Base
S2
S3
S4
S6
S8
I2
EKR
LeWM
FourRoom
41.3
51.3
55.3
60.0
–
–
–
62.7
LeWM
Push-T
40.0
46.7
48.7
46.0
47.3
51.3
–
54.7
LeWM
Cube
26.7
27.3
28.0
30.0
32.7
34.0
28.0
39.3
H-JEPA
Cube
50.0
48.7
56.7
58.7
56.7
54.7
45.3
64.0
LeWM
AntMaze
19.3
–
–
28.7
–
–
–
30.0
H-JEPA
AntMaze
66.0
72.0
67.3
68.0
64.7
73.3
–
73.3
Appendix
Table 10: Baseline with larger search budgets (success %, 3 seeds × 50 tasks; –: not run on all three seeds). S k : k× samples; I2: 2× iterations.
Figure 5: Success vs. planning compute, all eight cells. As Figure 2 . Labels give EKR’s lead over the strongest baseline point. Every point is measured on the same 150 (seed, task) pairs; budgets whose runs do not cover all three seeds are omitted and noted in the panel. At ≤1.5× compute EKR matches or exceeds 4× samples in 7 of 8 cells.
Figure 6: Optimisation versus exploration on Cube (LeWM (LW) and H-JEPA (HJ), 150 episodes each). First-plan base cost, executed iterations and success for the baseline, the baseline with 2× iterations (I2) and EKR. Error bars: 95% bootstrap confidence intervals.
Figure 7: EKR versus a random cost of the same scale (baseline, EKR and random cost evaluated on the same tasks; cells with ≥20 such episodes; AntMaze uses seeds 42–43, n=100 per backbone; over all three seeds AntMaze/LeWM success is 19.3/30.0/21.3% for baseline/EKR/random cost; other cells 3 seeds × 50). Success, stuck at end, revisit rate, coverage relative to the baseline, and early-trap fraction, with the definitions of Appendix B . LW/HJ: LeWM/H-JEPA; FR: FourRoom. Error bars: 95% bootstrap confidence intervals.
Figure 8: Executed trajectories in the frozen latent space. (a–d) Latent states actually visited by the baseline (grey, light to dark with episode time) and EKR (orange) on the same task (seed 42). Horizontal axis: projection of zt−z0 onto the start–goal direction, scaled so that z0↦0 and zg↦1 ; vertical axis: first principal component of the orthogonal residual. The dashed hull marks the baseline’s last quarter of the episode. Examples (a–d) are selected among the seed-42 tasks where only EKR succeeds (FourRoom/LeWM: 15 of 50; AntMaze/LeWM: 18 of 50). (e) Latent recurrence rate: fraction of steps t with mins≤t−W∥zt−zs∥<τ , where W is the trap window (FourRoom 30, AntMaze 50, Push-T 20, Cube 10 steps) and τ twice the median one-step latent displacement of the baseline–EKR pair, computed on the lowest-level latent of each backbone (level 0 for H-JEPA); the three conditions are truncated to a common horizon, and tasks whose common horizon is at most W are excluded (Cube/LeWM: 11 of 150, all fast successes). Mean and 95% bootstrap CI over paired tasks (seeds 42–44; AntMaze 42–43). ∗ EKR below baseline, † EKR below the random cost (paired CI excludes zero); FR: FourRoom, LW: LeWM, HJ: H-JEPA. Projections are for visualisation only; recurrence is computed in the full latent space.
Figure 9: Full-episode memory versus start-only, against replanning steps per episode. Paired success gap between full EKR and the start-only variant (95% bootstrap CI). On Cube (about four replans per episode) the gap is about −1 point; on AntMaze/LeWM it is +12.7 points [4.0,21.3] over 150 paired episodes; on AntMaze/H-JEPA (seed 42) +6 points with a CI that includes zero. The FourRoom and Push-T gaps are given in Section 4.5 .
Figure 11: Gain of EKR on DINO-WM Wall for the official tasks (budget 75) and the long-goal tasks (budget 150); goal distance and budget change together. The same pattern holds on DINO-WM PointMaze: no gain at the ceiling on the official tasks and +55.3 points on the long-horizon split ( Table 3 ).