aryan1219/hopfield-vla-checkpoints

Checkpoints: Hopfield vs. HAMLET memory-augmented VLA

Trained checkpoints from a controlled comparison of a Modern Hopfield associative-memory module vs. a paper-faithful HAMLET (sequential) memory [Koo et al., ICLR 2026, arXiv:2510.00695], wrapping two frozen-backbone VLAs — SmolVLA and pi0 — evaluated on four LIBERO suites (Object, Spatial, Goal, Long-horizon).

Full method, code, and paper: hopfield-vla on GitHub

Results summary

3 seeds × 10 episodes/task. mem = memory-on (last RWR round); base = same checkpoint, memory disabled at inference (matched control); Δ = mem − base.

SmolVLA backbone

Suite	Off-shelf	Hop mem / base / Δ	HAM mem / base / Δ
Object	84.0	94.0 / 83.0 / +11.0	88.0 / 43.3 / +44.7
Spatial	54.0	83.0 / 41.8 / +41.2	63.7 / 30.3 / +33.4
Goal	84.0	86.0 / 75.7 / +10.3	82.1 / 73.6 / +8.5
Long	35.0	64.7 / 33.3 / +31.4	53.3 / 41.0 / +12.3
Avg	64.3	81.9 / 58.5 / +23.4	71.8 / 47.1 / +24.7

pi0 backbone (cross-backbone check)

Suite	Hop mem / base / Δ	HAM mem / base / Δ
Object	94.7 / 79.3 / +15.4	89.0 / 77.8 / +11.2
Spatial	79.0 / 63.3 / +15.7	61.0 / 41.3 / +19.7
Goal	94.0 / 77.3 / +16.7	83.0 / 72.7 / +10.3
Long	61.2 / 29.1 / +32.1	50.2 / 32.3 / +17.9
Avg	82.2 / 62.3 / +19.9	70.8 / 56.0 / +14.8

Hopfield achieves the higher absolute success rate on every suite, on both backbones. See the GitHub repo for best@3 figures, latency benchmarks, and the full evaluation protocol/caveats.

SmolVLA (smolvla_*.pt)

• smolvla_hopfield_{goal,long}_t{N}_grpo_round{R}.pt — Hopfield-memory arm, per-task RWR checkpoint (plateau-stopped round) for task t{N} of the given suite.
• smolvla_hamlet_{goal,long,object,spatial}_final.pt — HAMLET-memory arm, BC-trained checkpoint (one shared checkpoint per suite, used for every task/pool at eval time).

pi0 (pi0_*.pt)

• pi0_hopfield_{object,spatial,goal,long}_t{N}_round{R}.pt — Hopfield-memory arm, per-task RWR peak checkpoint.
• pi0_hamlet_{object,spatial,goal,long}_step_from_step_{N}.pt — HAMLET-memory arm, BC-phase checkpoint (shared across tasks within a suite).

Why some tasks/suites have more than one checkpoint file. Each pi0 task is evaluated across 3 independent eval pools, and each pool's reported number comes from whichever checkpoint (round for Hopfield, BC step for HAMLET) scored best for that pool. So a task can legitimately have up to 3 round/step files behind it — these are not redundant duplicates, they're each the exact checkpoint needed to reproduce one pool's result. For HAMLET specifically, the checkpoint is a single shared BC run per suite (not per-task), so the handful of step files per suite are already deduplicated at the physical-checkpoint level across all (task, pool) pairs that happened to peak at the same step.

Methodology

• SmolVLA: BC warm-start → reward-weighted regression (RWR), evaluated with paired memory-on/memory-off controls at the plateau-stopped final RWR round.
• pi0: same BC → RWR protocol, backbone frozen, memory module (and where applicable the action expert) trained.

Full results tables, ablations, and evaluation protocol: see the project README.

Citation

@misc{associative_vs_sequential_memory_vla_2026,
  title   = {Associative vs. Sequential Memory in Pre-Trained Vision-Language-Action Models},
  year    = {2026},
  note    = {Preprint}
}