Grounding & conditioning placement — where the trunk should feed the expert
Papers: IVRA (2601.16207, read at v2), FLOWER (2509.04996, CoRL 2025), SCALE (2602.04208, ICML 2026 spotlight, read at v2), and SmolVLA (2506.01844). Banked across the 2026-08-05/06 lit slices; re-read at full-text depth for this page. Fed: #11 (the grounding front — all four triangulate the acuity probe’s story) and the architecture batch (arm B’s full-residual conditioning has its published head-to-head baseline here).
The theme
Our acuity probe found position information sharpest at the vision-tower output and degraded through the LM stack; our panel’s first_mae sits barely ahead of state-copy. These four papers are the field’s answers to the same diagnosis — where in the stack the action head should read (FLOWER, SmolVLA), whether lost visual geometry can be re-injected without training (IVRA), and whether the vision encoder can be modulated at test time (SCALE). Three of the four needed corrections against our skim-depth bank; the corrected versions actually agree with each other better.
1. IVRA — re-injecting the tower’s geometry, training-free (2601.16207)
What it contributes. VLAs flatten image patches into a 1-D token sequence, and the LM gradually discards the 2-D structure. IVRA computes a patch-affinity matrix (cosine similarities from an intermediate layer of the frozen vision encoder) and uses it to smooth the visual tokens at one mid-LM layer: each visual token is replaced by a convex blend of itself and its affinity-weighted average over the other patches (λ ≈ 0.2–0.3). No training, no new parameters, +3% latency. Note the mechanism precisely — it is token-feature mixing along tower affinities, not attention-score editing (our bank’s “injects affinity signals” was loose).
What they ran. The headline is the low-data regime: VIMA with LLaRA at 12% data, average 53.9 → 58.1 (+4.2), shrinking to +1.5 at full data. On LIBERO: OpenVLA +1.1, FLOWER +0.8 (from a 96.3 base). One banked correction: we had “consistent LIBERO gains across LLaRA/OpenVLA/FLOWER” — LLaRA is never evaluated on LIBERO; its results are VIMA + a small real-robot study. The ablations are where the value is for us: injection at layer 20 of 32 (~62% depth) is best, any layer above 19 works, one layer beats two or five — and injecting at the LM input (the projector) is catastrophic (0.0–6.9% across tasks, vs 22.5–73.1 at layer 20). Early re-injection destroys the very computation that needs the signal later.
What transfers. Direct support for the acuity story, plus a zero-training A/B we could run on the frozen teacher: affinity-mix our soft visual tokens at one Gemma layer near the kv9–kv14 zone and read panel first_mae. Banked to #11 as the cheap rung behind the img280 read — more tokens and better-used tokens are the same front from opposite ends.
2. FLOWER — cut the top of the trunk, feed the middle (2509.04996)
What it contributes. A 947M-parameter flow VLA built on a deliberately truncated VLM: prune the top of the LM and cross-attend the flow transformer into the truncation-point hidden states, reallocating the saved capacity to a bigger action head (339M flow transformer vs 205M of kept LM). Plus Global-AdaLN — one shared modulation across all flow layers (−20% head params at identical performance, 4.43 vs 4.44 CALVIN).
What they ran. CALVIN ABC→D 4.53 average rollout length (previous SoTA 4.28, π0 4.01), LIBERO-Long 94.9 (π0 85.2), a 20-task real-kitchen suite at 61.0 vs OpenVLA’s 31.0 — from 200 H100-hours of pretraining on ~250k trajectories. The fusion-placement ablation is the sharpest published version of our question: on CALVIN, early fusion 57.1, intermediate 89.5, late 71.2. Reading top-of-stack features is 18 points worse than mid-stack; feeding the expert too early is 32 points worse. Honest scope: SIMPLER Google-Robot loses to RT-1X (31.9 vs 42.4).
The corrections. Two precision fixes to our bank. The “prunes up to 50%” headline is the encoder-decoder (Florence-2) config, where 50% = deleting the decoder; for decoder-only trunks (our case) their own sweep says 30% pruning is the optimum and 50% degrades (72.1/70.7 → 66.4/62.5). And “conditions below mid-stack” is wrong: the truncation point sits at ~50% depth for Florence-2 but ~70% depth for decoder-only — the accurate phrase is at or somewhat above mid-stack, never the top.
What transfers. This is arm B’s published head-to-head baseline: a single truncation-point cross-attention feed at 60–70% depth captures most of what a multi-stream design chases — if full-residual res0..res14 nulls, the corrected FLOWER read says the follow-on is a single deep-but-not-top tap (near our layer-14 stream), not more streams and not maximally-early ones. The frozen-VLM row (3.42 vs 4.44 trained) is also worth carrying: at sub-1B scale, freezing the trunk costs a full CALVIN point.
3. SCALE — temperature, not tokens (2602.04208)
What it contributes. Our bank said “token budget spent adaptively rather than uniformly” — there is no token budget in this paper; the correction is the content. SCALE is training-free and verifier-free, for AR/FAST-token VLAs, at constant single-pass compute. It computes a per-token “self-uncertainty” (a dual-KL score against both a one-hot and a uniform reference — capturing spread and decisiveness) and uses it twice: as the action sampling temperature (near-greedy when confident, explorative when not), and — through its step-to-step change — as the vision-encoder attention temperature (attention flattens to explore when uncertainty rises above its recent average, sharpens to focus when it falls).
What they ran. OpenVLA on LIBERO 75.7 → 81.5 (+5.8, where naive sampling and top-k/p buy nothing); π0-FAST 91.2 → 93.0 (+1.8, where naive sampling degrades to 84–88); π0-FAST on SIMPLER-WidowX +14.6; real UR10e +19.5/+13.9 in-distribution. It beats the trained test-time-selection methods it compares against (MG-Select by +10.7 average) at 1× compute — and sits within 1.3 points of a double-pass oracle that uses the exact current uncertainty. Component ablation: decoding alone +5.3, visual attention alone +3.3, both +10.6 (super-additive). The row that matters most for us: modulating the vision encoder beats modulating cross-modal attention in the VLA (63.3 vs 57.4).
What transfers. Directly pluggable on our AR/FAST decode path — their π0-FAST config (2048-vocab, T0=0.3, κ=2) is the template, and it stacks conceptually with the #19 selection program (it is a draw-shaping method at N=1, orthogonal to best-of-N). The vision-encoder-beats-cross-modal result is yet another independent pointer that the tower is the highest-leverage intervention site. Limitation for us: nothing here applies to the continuous flow head — this is an AR-side tool only.
4. SmolVLA — the half-depth cut is cheap, not better (2506.01844)
What it contributes. The 450M community-data VLA: SmolVLM-2 (SigLIP + SmolLM2) with the flow expert cross-attending to features from the first 16 of 32 LM layers — the upper half discarded entirely — interleaved cross/self-attention in the expert, and an async inference stack that decouples acting from predicting. Also the closest published analogue of our data situation: pretrained on 481 community datasets, 22.9k episodes — an order of magnitude below the big VLAs — with VLM-regenerated instructions.
What they ran. LIBERO average 87.3 (above π0-3.3B’s 86.0 at 7× the size), Meta-World 57.3 (π0 47.9), real SO100 78.3 vs π0 61.7; async inference cuts task time ~30% and doubles fixed-window throughput (19 vs 9 cycles/60s). Community pretraining is worth +26.6 points on their real suite (51.7 → 78.3 multitask). Our banked claim (L/2 conditioning, 450M, SigLIP) is confirmed — with the nuance their own Table 8 adds: the full 32-layer stack is slightly better (80.3 vs 78.5); N=16 was chosen as a compute/performance tradeoff. Several shipped choices are likewise efficiency-driven rather than ablation-best (expert width 0.75× ships, 1.0× ablates better; chunk 50 ships, 10 ablates better) — quote them as tradeoffs, not optima.
What transfers. SmolVLA is the softer datapoint in the mid-stack story: it says the top half of a small VLM is nearly free to discard, not that discarding helps. Its real lessons for us are elsewhere — external validation that community-corpus pretraining transfers to a real rig (the north-star protocol, #16), the interleaved-CA/SA expert ablation (85.5 vs 79.0 CA-only / 74.5 SA-only), and the async execution stack as deployment reference for the rig.
The triangulation
After corrections, the four papers plus our acuity probe agree on a sharper story than the one we banked. Grounding signal lives in the vision tower and decays up the LM stack — but the optimum tap is not maximally early: IVRA’s best injection is ~62% depth and its projector-level injection is catastrophic; FLOWER’s decoder-only optimum keeps 70% of the stack and early fusion collapses; SmolVLA’s half-depth cut costs a little rather than gaining. The trunk’s early layers do work the expert needs done — the failure is only at the top, where features over-specialize for next-token prediction. For arm B that reframes the follow-on before its data lands: if full-residual nulls, the literature’s bet is one tap at 60–70% depth, and if it wins, the early streams are the ones to suspect of contributing least.