APT — the seam damage is an initialization problem

Source: APT: Action Expert Pretraining Improves Instruction Generalization of Vision-Language-Action Policies (2606.12366, June 2026). Read 2026-08-07, the pre-endpoint attachment-frontier slice — the last scheduled look at the seam literature before the molmo2 stage-2 attachment decision opens (~08-08). Fed: #4 (a third explanation of why joint training hurts, one new named escalation rung — F-then-joint — and a reframe of what our F arm’s checkpoint is worth).

The theme

Every seam recipe we have banked answers “how do I stop the action expert from wrecking the trunk?” — π0.5/KI stops the gradients, AEGIS projects them, Anchor-Align leashes the trunk to its frozen self, Wall-OSS bridges them. APT asks a prior question: why are the expert’s gradients destructive in the first place? Its answer: because the expert is randomly initialized. VLA trajectories pair thousands of vision-action frames with a single instruction, so a fresh expert learns the imbalanced shortcut — act from vision, ignore language — and its noisy early gradients push the trunk the same way. On that diagnosis, the fix is not a gradient guard at all: pretrain the expert first, then decide what to unfreeze.

What the paper does

A Bayesian factorization π(a|v,ℓ) ∝ πᵖ(a|v) · L(ℓ|v,a) turned into a two-stage schedule:

1. Stage 1 — vision-action prior. Train a diffusion action expert on vision-action pairs with the language masked out entirely and the VLM backbone (Qwen3-VL) frozen. Only half the expert’s layers (N/2) exist at this stage. Vision-action pairs alone carry no language imbalance, so there is no shortcut to learn.
2. Stage 2 — language alignment. Expand to N layers (interleaved insertion, Stage-1 layers keep their weights), unmask language, and jointly finetune on the full dataset — optionally including the VLM backbone itself (“Ft VLM”). Backbone features enter through layer-wise gated fusion (sigmoid-gated residual injection of intermediate VLM features) rather than raw token insertion, explicitly to avoid stomping the Stage-1 prior.

What they ran

LIBERO-PRO (instruction/object perturbations — the benchmark where naive joint training shows its collapse: OpenVLA and π0 score 0% across the board), a four-suite pick-place generalization grid (seen/unseen objects, containers, environments), architecture transfer (π-style all-layer attention and GR00T-style final-layer cross-attention), and real-robot pick-place + compositional chaining. Headlines: 27% avg on LIBERO-PRO for APT with a finetuned VLM vs 11% for π0.5’s KI recipe; real-world APT 28/40 on the hardest suite vs π0.5’s 16/40; π0.5 “nearly collapses” on chained instructions while APT executes them without segmentation.

The decision-relevant table is the pick-place ablation grid (SO/UO/UC/UOUE):

Two readings, both load-bearing: with a random-init expert, gradient guards (KI) are what stands between you and π0’s numbers; with a pretrained expert, unfreezing everything without any gradient stopping is the best row on 3 of 4 suites — “stop-gradient is not a necessary condition.” The damage was never joint training per se; it was joint training from noise.

What transfers to us

Our #4 screen attaches a randomly initialized flow expert in both arms. APT’s grid speaks directly to that regime:

• The K arm’s design is corroborated from a new direction. Our K is phase-1 CE verbatim + stop-grad seam — the KI recipe. APT finds KI most valuable exactly when the expert starts from noise (its KI-less random-init rows are the collapsed ones). Nothing in APT argues for loosening the live screen’s seam.
• The F arm’s endpoint is APT’s Stage 1, almost verbatim. A flow expert trained to convergence against a hard-frozen trunk is a pretrained vision-action expert (ours sees language through the trunk’s representations rather than masked out — a softer prior, same structure). APT’s grid then names the follow-up we did not have: F-then-joint — warm-start a joint run (unfrozen trunk, possibly no stop-grad at all) from the F checkpoint’s expert instead of from noise. In APT’s numbers that jump is worth +8 to +26 points over the frozen row. This lands on the #4 escalation map as a named rung, after the screen reads out — the pre-reg’s two arms are untouched.
• It reframes a possible F≈K tie. If Δ_seam comes back small, the KI-vs-frozen contrast may simply be measuring two guards that both work at our scale — APT predicts the interesting contrast is then initialization (F-then-joint vs K-from-noise at matched total steps), not the seam itself.

What doesn’t transfer

• The measured axis is language generalization (unseen instructions/objects/compositions). Our panel is single-embodiment action MAE with no language-perturbation axis; APT’s largest effects live exactly where our instrument doesn’t look. The transferable content is the mechanism and the recipe ordering, not the effect sizes.
• Architecture gap: their expert is a diffusion transformer fed by sigmoid-gated per-layer fusion from Qwen3-VL; ours is a residual-tap flow expert reading 3 exported streams. Their token-insertion-vs-gated-fusion ablation (gated wins, biggest on unseen suites) is a caution for any future all-layer-reads arm (#4 external arm 1), not evidence about our current taps.
• The 0% baselines are perturbation-benchmark artifacts — LIBERO-PRO is built to break visual-shortcut policies; π0 scoring 0 there does not mean π0-class recipes score 0 anywhere else (their own pick-place grid shows π0 at 42/30/26/16).

Verdict for the stage-2 decision

Nothing here re-ranks K-vs-F before the screen — it runs as pre-registered. What APT changes is the escalation map: the F checkpoint is no longer just the conservative arm’s readout, it is free Stage-1 capital, and the strongest published recipe in its weight class (expert pretrain → unfreeze, no KI) is one warm-start run away from it. Radar note for the next slice: the initialization thread has siblings we have not read (2605.25802, VLM representation for VLA init; 2601.03309, VLM4VLA) — banked as hooks, not read.