Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

The Compression Gap: a tidy bottleneck story, measured on models 1000× smaller than the claim

Read 2026-08-09 (lit slice lit-radar-0817, priority 5: the flow-over-AR mechanism hook). Paper: 2604.03191 — “The Compression Gap: Why Discrete Tokenization Limits Vision-Language-Action Model Scaling” (Takuya Shiba, single author, 2026-04-03). Code.

The paper in plain words. Robot policies turn camera images into actions through two different kinds of “action head”: some output actions as continuous numbers (diffusion/flow heads), others compress actions into a small vocabulary of discrete tokens first (like words in a dictionary). This paper asks: if you give the policy better eyes — swap a weak vision encoder for a strong one — which head lets the improvement through? On one benchmark suite, with small models: the continuous head converts the better encoder into +21 to +26 points of success, while the discrete-codebook head gains only +4 to +10. The proposed explanation is an information bottleneck — a 1000-word action dictionary can only carry ~80 bits per motion chunk, so extra visual detail has nowhere to go. It’s a clean story, but the evidence is thin: tiny non-VLA models, one benchmark, single seeds, and the paper’s own ablations wobble in ways the bottleneck theory doesn’t explain.

What it contributes

  • A factorial claim: encoder (ResNet-18 → SigLIP/SigLIP2/ DINOv2) × action head (Diffusion Policy vs OAT’s FSQ codebook) on LIBERO-10. Encoder upgrade through DP: +21.2 (size M) / +26.0 (size L). Through the 80-bit OAT codebook: +3.6 / +10.4 (Table 1).
  • A mechanism proposal, not a measurement: the data-processing inequality — the discrete channel caps I(Z;T) at H_l·log2|V| ≈ 80 bits per 32-step chunk at OAT defaults, so once the codebook is the tightest bottleneck, encoder gains are “blocked at quantization.” No mutual information, reconstruction error, or probing is ever measured; the evidence is endpoint success deltas whose pattern is consistent with the story.
  • The one attempted causal test (grow the codebook, watch the encoder delta return) half-works: |V|=1920 → Δenc +15.2, but |V|=4375 → back down to +4.0, unexplained. Table 2 also has the codebook head getting worse under a better encoder (SigLIP2: 44.2, below the ResNet-18 baseline 53.8) — also unexplained.

The experiments it ran

Everything on LIBERO-10 (50 demos/task), with small language-free transformer policies — not VLAs: 4-layer/256-dim and 6-layer/384-dim decoders, no LLM trunk, no language conditioning. Strong encoders are frozen with pre-extracted cached features; ResNet-18 is trained end-to-end (an asymmetry inside the factorial). Metric: peak success across training, single seed per cell, 500 rollouts. The often-omitted reversal: with the weak encoder, the discrete head wins by 17.4 points (53.8 vs 36.4) — the paper’s own reading is that structured tokenization compensates for poor perception. The continuous head also saturates in the low 60s even with the best encoders.

What transfers to us

  • A weak directional prior, banked as “consistent-with,” not “predicts”: in a good-trunk regime, continuous heads extract more from trunk improvements than tight discrete codebooks do. It rhymes with our #19 finding (flow draws gains ~9× AR’s — the mean-collapse asymmetry), and it points at the vision-unfrozen adamc regime as where flow-vs-AR panel divergence would be most visible. Both of those connections are our inference — the paper contains no flow matching, no VLA, and never trains an improving encoder.
  • The bit-budget arithmetic is worth keeping as a lens: our AR decoder’s binned action vocabulary carries on the order of 1,800 bits per chunk — ~22× OAT’s 80 — so the paper’s bound plausibly never binds for our AR head. If our panels show AR attenuation anyway, the mechanism would have to be something other than raw channel capacity.

What doesn’t transfer

  • The headline numbers, wholesale. Tiny non-VLA policies, one benchmark, single seed (±5-pt noise plausible), peak-success metric, floor-effect confound (DP’s big delta starts from a 17-pt-lower baseline — more headroom), unmatched head parameter counts, and the frozen-vs-trained encoder asymmetry. The “compression gap” may be more about OAT’s specific 32×7→8-token compression ratio than about discreteness.
  • Any direct vote on #4 (frozen vs joint): the strong encoders are always frozen. Joint training opens exactly the channel the setup excludes — the encoder co-adapting to the codebook’s needs — which would shrink the gap. If adamc’s unfrozen-vision AR run beats this paper’s frozen-swap story, that’s evidence for co-adaptation, not against the paper.

Hook corrections

The banked one-liner (“encoder upgrades give >21-pt gains through continuous action heads but attenuated through discrete codebooks — mechanism-level flow-over-AR prediction that sharpens exactly when the trunk improves”) was oversold on every clause: the >21 pts is tiny language-free Diffusion Policy on LIBERO-10 at a single seed; the attenuation is one 80-bit FSQ codebook (bigger codebooks largely un-attenuate, and AR binning has orders of magnitude more bits); the “mechanism-level” evidence is arithmetic plus a non-monotonic ablation; and the paper never trains an improving trunk — the “sharpens as the trunk improves” extrapolation was ours, and it now carries an explicit conditional: with a weak encoder the discrete head won by 17 points.

Which idea/arm it fed

#19 (ar-sampled-draws) — a weak external rhyme for the mean-collapse asymmetry, filed with its conditionals (good-trunk regime only; not flow matching; our AR head’s bit budget likely escapes the bound). Watch note for the adamc k4l2 panel readout: if flow-vs-AR divergence appears under unfrozen vision, this paper is a consistent-with citation, never a predicted-by. No gate changes.