The attachment frontier — expert memory, leashed trunks, and world-action models

Sources: AR-VLA (2603.10126, RSS 2026, read at v2), Anchor-Align (2607.13429 — “representation anchoring” in our bank), and NVIDIA’s world-action-models post (the π0.7 source). Banked across the 2026-08-05/07 lit slices; re-read at full depth for this page. Fed: #4 (two new priors on the seam screen — one for the K premise, one third recipe), #17 (the history-aware-expert direction, now with its headline ablation), #6 (a caution flag on the text-subgoal probe’s ceiling).

The theme

Our stage-2 question is how to attach an action expert to a pretrained trunk — frozen (F) vs KI-joint (K). These three sources each stretch that question along a different axis: AR-VLA gives the expert its own memory across observations; Anchor-Align trains the trunk but leashes every layer to its frozen self; the WAM tier replaces the trunk’s pretraining with video-dynamics models. None changes the live screen; all three shape what the escalation map looks like after it reads out.

1. AR-VLA — the expert that remembers (2603.10126)

What it contributes. Every VLA we run resets the action expert at each chunk: fresh prefix, no cross-observation state. AR-VLA builds the expert as a standalone AR transformer with a hybrid KV cache: a rolling FIFO buffer of proprio/action KV pairs that persists across observations (test-time length 20), plus a single-slot vision-language buffer that is replaced wholesale each time a new frame is processed. The two async streams stay coherent through Dynamic Temporal Re-anchoring: action tokens take sequential RoPE indices on the causal timeline, VL tokens take the fixed index of their capture step, and RoPE’s shift invariance makes only the relative staleness matter. Actions are continuous per-timestep regressions (not flow, not diffusion, not discrete tokens), decoded one step at a time at 5 Hz.

What they ran. SimplerEnv WidowX: 61.5% average vs CogACT 52.1, π0.5* 51.0, π0-FAST* 49.0 (starred = reproductions); 89% zero-shot on a real WidowX from BridgeV2; specialist variant beats ACT on ALOHA cube transfer (97.3 vs 86.0 scripted) though Diffusion Policy keeps PushT. Lowest action jerk of all baselines and a 28.9 ms expert pass (vs 84.3 ms for a flow expert). The ablations are the substance:

• History length 1 → 20 is worth +25 points (36.5 → 61.5; 40 slightly worse). A length-1 context is structurally our chunk-reset design — this is the cleanest published number for what cross-observation expert memory buys.
• History masking is load-bearing, not a trick: train with no masking and success is 0.0% (while validation error is the lowest — the model over-trusts its own history and collapses in closed loop); rate 0.6 is the peak at 61.5%.
• Static positional embeddings instead of DTR: 3.1% — the re-anchoring is what makes refreshable prefixes + persistent cache coherent.

Caveats. The flagship “memory” tasks (goals that become unobservable mid-episode) are figure-only — no table values; the quantitative memory evidence is the history-length ablation, which is a context-length effect, not a beyond-occlusion proof. And the authors name the failure mode themselves: an OOD action gets written into the cache as history and drags the policy further OOD — memory compounds errors as readily as it fixes them.

What transfers. Two things, one per axis. On the expert axis this is the banked #17 history-aware direction with its headline number attached — a named escalation if the attach screen leaves headroom. On the trunk axis, a sharpening of our bank: AR-VLA is not a third topology there — it freezes the VLM and blocks action gradients explicitly, writing that AR gradients degrade the trunk “similar to flow-matching experts,” necessitating knowledge insulation. An AR expert team independently converging on the K premise is evidence for the K arm’s motivation, from outside the flow family.

2. Anchor-Align — train the trunk, leash it to its frozen self (2607.13429)

What it contributes. Between F (frozen trunk: nothing forgotten, nothing adapted) and K (stop-grad joint training) there is a third recipe: fine-tune the trunk with a distillation leash. Anchor-Align adds two losses to standard BC. Anchoring: keep a frozen copy of the pretrained VLM and penalize the trainee’s hidden-state drift from it (vision and text tokens, every decoder layer, per-layer Frobenius norm). Language-action alignment: label each chunk with a programmatic 6-way motion direction (up / down / left / right / forward / backward), and train the trunk to predict that word through its frozen pretrained LM head — so the language space itself is forced to encode intended motion. Our bank had only the first half; the second is co-equal (worth +4.9 / +3.5 on their two suites alone) and drives their flashiest result — after alignment training, a policy told to grab the pink mug grabs it 100% of the time where the BC baseline grabs the training-biased green one 90% of trials.

What they ran. LIBERO-PRO (semantic OOD): 71.9 mean vs 61.0 naive-BC, and — the row that matters for #4 — 43.8 for a Co-training+KI baseline and 43.1 for full-freeze: on their benchmark, both of our screen’s endpoints lose to the leash by ~28 points. On position-swap every baseline sits at 0.0–2.3% and Anchor-Align is the only method off the floor (22.6%). Real xArm7: 28.3 → 54.2 and (on a flow-head architecture) 36.7 → 60.0. Representation preservation is measured directly: naive BC loses 94% of GQA VQA accuracy within 10k fine-tuning steps; anchoring retains ~70%. Shuffled-label controls collapse to baseline, so the alignment gain is not generic regularization.

Caveats. Primary backbone is 0.5B — small; no layer-subset ablation (is every-layer anchoring necessary?); λ values live in a truncated appendix; the KI baseline is their reimplementation on their benchmark, not π0.5’s recipe at π0.5’s scale. Grasp-and-drop errors slightly increase (12 → 13) — more grasps survive to the drop stage.

What transfers. A named third recipe for the seam screen’s escalation map: if K wins Δ_seam but shows the named cost (VQA/semantic drift), anchoring is the published repair — and for us the teacher is free, since the frozen copy is exactly our F trunk. The cheap probe to steal regardless: measure VQA retention on the Molmo2 trunk before/after stage-2 training — the 94%-in-10k figure says drift is fast when it happens.

3. The world-action-model tier — π0.7 and the video prior (NVIDIA post)

What it says. The tier above VLA trunks: policies built on pretrained video models, which already encode language→ visual-change dynamics. Two families — inverse dynamics (generate future video, recover actions: LingBot-VA on Wan 2.2-5B, 16k hours of robot pretraining, two-way video↔action conditioning) and joint denoising (DreamZero: one DiT from Wan 2.1-14B denoises video and actions together; RoboArena 1750 vs π0.5’s 1622 from DROID-only action tuning). Being-H0.7 bridges latently (V-JEPA2.1-encoded future observations as a posterior plan, dropped at test time). And π0.7: high-level policy emits a subtask, a BAGEL world model renders it as a subgoal image, the action expert conditions on observation + subgoal — reported necessary for dataset-bias-breaking tasks where no-subgoal variants fail, and faster to train because action prediction becomes near-inverse- dynamics. All banked systems and numbers verified (one date touch-up: GR-1 is 2023/ICLR-24). The costs the post is honest about: WAM inference runs 590–800 ms per chunk vs ~190 ms for π0.5, and the full video-pretrain stacks sit at ~10× the compute of a VLA fine-tune.

The addition worth banking: Fast-WAM — a representation- only WAM that skips test-time video generation, runs 3–4× faster, and reportedly matches LingBot-VA in sim without the 16k-hour robot pretraining. That is the strongest single argument that the video prior, not test-time generation, carries the value — and it points at the reachable version for us: a video-capable trunk (Molmo2 — already ours) conditioning the expert on predictive features rather than rendered frames.

What transfers — and a flag for #6. The reachable-scale ladder stands as banked: text-subgoal probe → subgoal-image conditioning → video-prior features. But π0.7 is also a caution for the self-subgoal probe’s ceiling: they found text subtasks insufficient for the bias-breaking tasks and needed rendered images. If our rung-(a) text probe reads null, that is consistent with the field’s experience, not evidence the hierarchy thesis is dead — the pre-registered read should say so in advance rather than after.