Fewer layers than you think: prune the twins, then finetune

Lit slice 2026-08-09 (work session 12:15Z, third read — sweep hook cleared same session it was banked). “Finetuning Vision-Language- Action Models Requires Fewer Layers Than You Think” (2606.20246, June 2026; the method is called CLP, CKA-guided Layer Pruning). Fed #17 (the trunk-redundancy ledger gets its first VLA-specific measurement) and the throughput accounting (a fourth lever class: fewer FLOPs per step, training AND inference).

The paper in plain words

Big robot-policy models inherit their size from the language models inside them — but how much of that depth does the control task actually use? This paper measures how similar each transformer layer’s output is to the next one’s (using CKA, a scale-invariant similarity score), finds long runs of near-identical “twin” layers, and simply deletes the twins — before finetuning, with no distillation, no extra parameters, from one forward pass over a calibration set. Finetuning then heals the seams: the remaining layers reorganize to restore the original representation manifold. Up to half the depth can go. The surprise is the low-data regime: with only 10% of the demos, the pruned π₀ beats the full model by +6.9 points — less capacity means less overfitting. Training gets ~28–31% faster, inference ~28–30% faster, and this holds across LIBERO, RoboCasa, SimplerEnv, and 10 real-robot tasks.

Contribution

Method: one forward pass over calibration data → CKA between consecutive layers → cluster runs with similarity ≥ τ → within each run keep the first layer, delete the rest (up to κ). Prune first, then finetune normally. Training-free selection; the finetune is the healing.
Where the redundancy lives: π₀ loses 6 of 18 layers (positions 1–9 range); GR00T-N1.5 loses 7 of 12 VLM layers (3–9) and — the eyebrow-raiser — 8 of 16 DiT action-head layers. Early-to-mid trunk and half the action expert are twins; the “major feature transformations concentrate in a few distinct transitions.”
Why it can gain: implicit regularization — capacity removal discourages overfitting to task noise in low-data finetunes; PCA shows the pruned network’s manifold re-aligning with the original during finetuning.

Experiments

LIBERO at 10% data (~30% params removed): π₀ 77.7 → 84.6 (+6.9); GR00T-N1.5 93.9 → 93.0 (−0.9 — at a strong baseline the prune is ~free, not a gain).
RoboCasa (30 demos): π₀ 15.6 → 18.0; GR00T flat out to 50% pruning ratio.
SimplerEnv (GR00T): 16.6 → 20.0, training 22.9 → 15.7 h.
Real world (10 tasks, GR00T): 73.5 → 75.9 avg, single tasks up to +20.
Speed: training −27.8% (π₀) / −30.8% (GR00T); inference 211 → 152 ms (π₀) / 121 → 85 ms (GR00T); FLOPs ×1.39–1.42 down.
Ablations: CKA beats MSE/cosine layer selection (those “distort representations into isolated subspaces”), and beats random or drop-last-k badly — which layers go matters more than how many.

What transfers to us, what doesn’t

A fourth lever class in the throughput accounting. The owner’s thread now has: step-time kernels (perf pass-1 — measured dead on the true recipe), memory→batch (#20 actckpt — the live lever), fewer-steps-to-quality (Spatial Forcing), and now fewer-layers — the only one claiming training AND inference wins at once, with a train-time mechanism (fewer FLOPs) that can’t be falsified by kernel-scheduling artifacts the way pass-1’s microbenches were. The catch for any pre-reg: their gains are measured on finetunes from a pretrained VLA in low-data regimes; our stage-2 attach trains an expert from a 50-60k-step trunk on the full recipe — closer to their GR00T full-data cell (−0.9, cost-neutral) than their +6.9 cell. “Same quality, 30% cheaper” is the honest expectation, and it competes head-on with actckpt’s memory→batch route for the same wall-clock budget.
Half the action expert was twins — a capacity datapoint for the #4/#17 expert-sizing conversation. GR00T’s 16-layer DiT head pruned to 8 with no loss. Reads with HyperVLA (0.1M generated policies suffice per episode) and against QDepth’s scaffold effect (extra expert capacity carrying wins): expert capacity is cheap to add and cheap to remove, so measured sizing beats inherited sizing. Our 18-layer-class experts have never had a CKA pass; the diagnostic is one forward pass and an afternoon.
The CKA map is a free diagnostic even with zero pruning intent. One forward pass over a panel batch gives the twin-block map of our Molmo2 trunk + expert at any checkpoint — worth having in the ledger the next time trunk-depth or expert-depth choices come up (#17’s fractional-depth mount discussion cites depth ratios with no redundancy evidence behind them). Cross-link: FLOWER’s deep-layer pruning (grounding-conditioning) pruned for placement reasons; this gives the general instrument.
Doesn’t transfer: the +6.9 headline (10%-data regularization regime, not ours); absolute latencies (their serving stack); any claim about pretraining-stage pruning (explicitly untested); and note their π₀ baseline (77.7 LIBERO single-view) is the same open-π₀ number QDepth-VLA starts from — these low baselines inflate every 2026 delta on that table.

Where it lands

#17: trunk-redundancy ledger opens with real numbers — 33–50% of finetuned-VLA depth is twins (CKA-selected, healing-finetune recipe); the CKA map banked as a one-forward-pass diagnostic for our own trunk/expert checkpoints; expert-sizing datapoint (16→8 DiT layers free) filed beside HyperVLA.
Throughput accounting: fourth lever class (fewer layers, train+inference), honest expectation “cost-neutral quality at ~30% savings” in our regime; competes with #20 for the same budget, and unlike pass-1’s kernels the mechanism is FLOP-count, not scheduler-dependent.
#4 (context): any future attach-screen sequel could prune the trunk before attaching — their prune-then-finetune order is exactly our attach shape; parked as a named sequel arm, needs its own pre-reg.