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

19. AR sampled-draws eval (mean-of-samples) — screening (AR-100k draws10_t1 READ OUT 2026-08-07, all expectations met; tsens rung next; molmo2 arm waits on its endpoint)

Tag: ar-draws · idea #19 · index

Greedy decode is the AR family’s single-draw voice; the flow family’s deployment read is mean-of-10 draws. The owner’s fairness point (2026-08-06 19:15Z): when we quote flow mean-of-N, the AR models get temperature-sampled draws-N + mean-of-samples too — both sides get the same instrument or neither.

Status 2026-08-07. Instrument landed + gated 2026-08-06 (78c9f56: --ar-temperature T --sample-draws N, Gumbel-max over the grammar-masked softmax, action block only — aux value lines stay greedy; draws share one prefill via by-reference cache snapshot/restore; stable_sample_rng keying domain-separated from flow noise; _drawsN_tT provenance). Pre-registered the same night (pre-reg, T=1.0 pinned untuned, frozen reads Δ_AR vs 5.8026 / fairness vs −1.258 / family vs 5.365, q4 cost fallback). AR-100k arm running now (draws10_t1, local GPU, boundary ~13Z 2026-08-07 → frozen reads). Frozen-read script landed 2026-08-07 ~07:5xZ (draws10_t1_results.py): the pre-reg’s reads 1–5 as one command, oracle-gated on every branch before any draws data existed — degenerate self-pair → exact zeros, synthetic ×0.95/×1.005/×1.05/ ×0.75/×0.90 error effects land on the E1+E2 / null / E4-falsifier / E2-not-met / E3-overtake branches with magnitude checks, and 11 hard-abort guards (state-copy byte-match, ar_temperature/plan/count semantics, _draws10_t1 provenance, checkpoint pairing, report reproduction |d| < 5e-3). The q4 cost-fallback path is coded (index join + re-pool, flagged subset_mode) and the molmo2 arm reuses the same command with explicit paths at its endpoint. Molmo2 arm oracle-complete 2026-08-07 ~04:3xZ: the pre-reg’s mechanics were oracle-pinned on the gemma trunk only while the molmo2 arm runs the shared suffix decode over a different cache (Molmo2KVCache) — tests/test_molmo2_ar_sampling.py now pins the trunk-specific halves on the molmo2 fixture (T→0 greedy recovery, draw determinism/distinctness, snapshot/restore prefill-sharing bit-exactness, the append-only-cache contract directly, the ar_predict_sampled dispatch). Launcher prep landed 2026-08-07 ~04:4xZ (6c3cc3b): eval_box_molmo2_endpoint_draws10_t1.sh makes the endpoint read one command — greedy arm re-run only if the training launcher’s chained eval didn’t land, then the draws arm 4-GPU sharded, with the pre-registered first-~200-frames cost gate MECHANIZED (draws_rate_gate.py, 10 oracles: rank-0-shard parsing, GPU-h projection, strict >24 threshold, timeout-with-partial-progress still decides) and the q4 fallback kill+relaunch automated; babysit.toml entry prepared (commented, fill started_utc at launch). Remaining: execute at the endpoint (~2026-08-08).

Lit-sourced escalation rung (banked 2026-08-07 ~04:4xZ, NOT pre-registered). If the frozen mean-of-draws reads land small, selection-over-draws is the named next rung, two flavors from the radar: (a) MG-Select (2510.05681) — verifier-free BEST-of-N: pick the draw whose action-token distribution maximizes KL(conditional ‖ condition-MASKED reference), where the reference needs a model trained with condition dropout — AR-100k already is (state dropout 0.5, subgoal dropout 0.5) and --mask-state already computes the masked context, so this is a zero-training read over a --dump-draws npz + per-draw logit retention; (b) VLA-ATTC (2605.01194) — a trained relative action critic ranks candidates (pairwise), with uncertainty-gated test-time compute; the trained-critic alternative if (a)’s verifier-free score is noise. Both papers frame greedy as the precision bottleneck — the OPPOSITE of our expectation 2 (greedy ≈ posterior mean); the draws10 primary read adjudicates between these two priors on our own panel before any escalation.

  • Lit slice 2026-08-07 ~06:1xZ: a THIRD selection flavor with a scaling claim — CoVer (2602.12281): scaling test-time VERIFICATION beats scaling policy pre-training for VLA-instruction alignment (contrastive verifier over rephrased-instruction × candidate-action pairs; +22% ID / +13% OOD SIMPLER, +45% real) — independent adoption evidence that the field’s compute is moving to the selection side; RoboMonkey (2506.17811) is the same bet with a VLM verifier + majority voting. Cheapest next read named: oracle best-of-10 ceiling — per-frame best draw vs ground truth from a per-draw stack bounds what ANY selector (verifier-free or trained) could buy on our panel before we build one. Retention gap found and fixed: the live AR-100k draws10 run dumps only the pooled predictions (per-draw reads there need a re-run — accepted, mean-of-samples is the registered read), but the molmo2 endpoint draws launcher now carries --dump-draws (added pre-launch, data-retention only, ~310 MB) so its selection-rung reads come free from the ~08-08 compute. The ceiling read is now a queue item (2026-08-07 ~06:5xZ, idea19-selection-ceiling-read-script, CPU: audit draws_fairness.py’s existing best-of-N against the AR draws-npz contract first, extend only the delta, oracle-gate on synthetic per-draw fixtures; exploratory read, NOT pre-registered — escalation to any actual selector needs its own pre-reg).

  • Ceiling read script LANDED 2026-08-07 ~08:1xZ (selection_ceiling_results.py): the audit found draws_fairness.py’s best-of-N is flow-probe-hardwired (panel joins + flow anchors), so the delta is a standalone sibling — the exact order-statistic best-of-K ladder for K = 1..10 (sorted per-frame draw MAEs weighted C(N−i, K−1)/C(N, K); no Monte Carlo), greedy/ensemble headroom with a paired CI on the oracle gain, first_mae mirrors, and selector diagnostics (argmin-draw uniformity, dispersion-conditioned gain quartiles — where a selector would buy most). Oracle PASS pre-data: ladder == brute-force enumeration over all K-subsets; degenerate draws=1 reproduces the 5.8026/2.1431 anchor through the ceiling path; planted best-draw pattern recovered exactly with magnitude checks; 5 abort guards (sample_draws mismatch, non-extending policy, draws=1 real mode, misaligned index, pooled-npz mean drift). Defaults = the endpoint launcher’s exact _draws.npz stems; runs the moment the ~08-08 dump lands. Follow-on queued (idea19-endpoint-fairness-es-read): the energy-score delta — the strictly-proper-scoring-rule AR-vs-flow comparison from the same npz, record-only.

  • Launcher landed 2026-08-07 ~08:3xZ (0cb8cf8, eval_ar100k_tsens_q4_draws10.sh): the pre-registered RECORD-ONLY T-sensitivity rung is one command — 3 sequential q4 rungs T ∈ {0.5, 0.7, 1.3}, draws 10, with the pre-reg’s “run ONLY if the primary lands inside the gate” clause mechanized (full-panel report

    • registered semantics + elapsed GPU-h from the babysit started_utc ≤ 24.0; five abort branches oracle-checked). --dump-draws retention per the endpoint precedent, so dispersion-vs-T and the per-T ceiling come free later. Follow-on queued (idea19-tsens-dt-read): the dT table — a T-parameterized sibling loader (the frozen-read script hard-pins T = 1.0 by design), no decision branches.

  • dT-table read script LANDED 2026-08-07 ~10:0xZ (tsens_dt_results.py): the T-parameterized sibling loader the tsens launcher’s follow-up named — registered T set {0.5, 0.7, 1.0, 1.3} only, one record-only dT table (pooled chunk/first per T on the same frozen q4 rows; the T=1.0 row re-pooled from the full-panel primary npz via the join_rows subset join), NO decision branches per the pre-reg sensitivity clause. Oracle PASS pre-data: a synthetic T=1.0 rung fixture reproduces the primary’s q4 re-pool EXACTLY; ×0.93/×0.98/×1.07 rungs land at exactly factor × re-pool; 11 guard aborts (unregistered T, wrong plan/draws/ar_temperature, policy+stem tag mismatch, rung-row disagreement, full-panel-as-rung, state-copy drift, checkpoint mismatch, report drift). Defaults = the launcher’s exact stems; the read is one command once the rungs land (which gate on the primary landing inside 24 GPU-h).

  • Energy-score read script LANDED 2026-08-07 ~09:0xZ (energy_score_results.py): the strictly-proper-scoring-rule AR-vs-flow comparison from banked data — endpoint draws ES vs the paired greedy arm as the degenerate N=1 baseline (interaction term zero by definition; ES gain + paired per-frame CI), plus the flow-side comparison via index-join to the banked drawsprobe_s7 stack (2,458 rows × 10 draws): both families get the SAME instrument on identical frames — N-draw ES, matched truth terms, paired per-frame ES delta. Audit honored: mean/best/dispersion stay in selection_ceiling_results.py; this file is ES only (draws_fairness.energy_score reused verbatim). Oracle PASS pre-data: degenerate draws=1 → interaction exactly 0 + ES == direct RMS-L2 (< 1e-12); the banked read-4 numbers (5.930763/9.882476/3.951713/8.769585) reproduced EXACTLY through this file’s own join + pooling; N=2 hand fixture exact; residual ×3 homogeneity; 5 abort guards. Defaults = the endpoint launcher’s exact stems; the tsens q4 dumps run via explicit paths (extending-policy guard is T-agnostic).

  • Lit slice 2026-08-07 ~08:3xZ — a SIXTH selection flavor, the cheapest trained one: What Frozen VLAs Already Know About Success (2605.28527) — LINEAR probes on frozen VLA features (OpenVLA, pi0.5) recover value-like success structure their imitation objective never asked for (~92% pairwise success-ordering on LIBERO-Goal, beating progress/time-to-go/ proprioception baselines), and the probe used as a selector over sampled action prefixes lifts push-plate success 26.7% → 44.3% (gains not universal, costs inference compute). Slots between MG-Select (verifier-free) and VLA-ATTC/CoVer (trained critics): a one-linear-layer trained selector over representations we already compute. Same gate as flavors 1–5: the oracle best-of-10 ceiling read decides if ANY of this is worth building on our panel. Also independent evidence for the #6/#17 prior that frozen trunks carry more task structure than their action head uses. Papers-page re-read 2026-08-07 (page) caveat the hook missed: the 26.7→44.3 selector result is NOT probe-only — each candidate prefix is rolled out in the simulator from a snapshot before the probe breaks ties (probe adds value on top of rollout screening; probing R²/pairwise numbers stand clean).

  • draws10_t1 READOUT (2026-08-07 ~12:2xZ, frozen reads via draws10_t1_results.py → reports/analysis__draws10_t1_ar100k_k4l2.json): ALL THREE PRE-REGISTERED EXPECTATIONS MET. E1: Δ_AR (draws10 − greedy) = −0.14505, CI95 [−0.182, −0.109], excludes zero. E2: |Δ_AR| is ~9× smaller than the flow draws gain 1.258 — the pre-registered mean-collapse shape (greedy AR decode already sits near the predictive mean). E3: draws10_t1 5.6515 does not overtake the flow draws10 band 5.365. Falsifier (Δ_AR > +0.1) NOT tripped; oracles clean (row pairing full byte-match, both report arms reproduced |d| < 5e-3). Cost ~12.7 GPU-h — inside the 24 GPU-h gate by ~2×, so the q4 fallback stays closed. Next: the T-sensitivity q4 rung (eval_ar100k_tsens_q4_draws10.sh, record-only), then the molmo2 arm at its endpoint.

  • Lit slice 2026-08-07 (decode-temperature page, 5 sources): a directional prior for the dT read, written down BEFORE the rungs land — the multimodal-failure analysis (2605.22493: deterministic beats every generative variant on near-unimodal tasks; coverage ≠ success) plus MARS (2605.29766: stochasticity pays only in genuinely diverse phases) predict a near-flat dT table with mild asymmetry against T=1.3 on our unimodal-dominated panel. Also banked: BOKBO (2605.30660) measures policy-internal confidence correlating poorly with violations — the SECOND independent strike (after the rollout caveat) that any post-ceiling selector needs trained scoring, not a free confidence readout; and 2603.20538 gives the q-token + CE trunk its sample-complexity-optimality citation. Hook parked, not armed: DDVLA’s temperature schedule (decay 1→0 beats both fixed extremes, 97.4 vs 96.4/96.2 LIBERO-Goal) is an unexplored axis in AR-VLA decoding — nothing opens unless the dT table shows real sensitivity.

  • T-SENSITIVITY dT TABLE BANKED (2026-08-07 23:09Z, record-only per the pre-reg sensitivity clause — reports/analysis__tsens_dt_ar100k_q4.json via tsens_dt_results.py, all guards green, T=1.0 re-pooled from the full-panel primary npz onto the same 4301 q4 rows): chunk MAE 6.5004 / 6.5668 / 6.7812 / 7.1843 at T = 0.5 / 0.7 / 1.0 / 1.3 (dChunk −0.2808 / −0.2144 / 0 / +0.4032; first-MAE mirrors −0.1710 / −0.1357 / 0 / +0.3654). Against the banked decode-temperature prior (“near-flat with mild asymmetry against T=1.3”): the T=1.3 asymmetry CONFIRMED (+0.40, the largest entry); the low side is not flat but mildly monotone toward T=0.5 — mean-of-10 at cooler temperatures sits closer to the greedy mean, the same mean-collapse shape as the primary read (draws add noise that averaging removes; cooling adds less to remove). Total spread ~0.68 on a 6.78 base over T ∈ [0.5, 1.3]. Per pre-reg this is quoted as a dT diagnostic, never a headline, never a license to re-pick T — primary stays T=1.0. DDVLA’s temperature-schedule hook stays parked: the table is its recorded input, and the primary read already bounds draws-at-any-T as a small-effect axis for this AR family. Rungs cost ~7.2 GPU-h ≤ 12 gate (t0.5 ~2.4 h, t0.7 ~2.3 h, t1.3 ~2.4 h). Follow-up read item still open: T-guard delta via the q4 subset join. Local GPU confirmed free 23:09Z (transient unit exited, 0 MiB) — the selfsubgoal probe window is open.

  • 2026-08-09 — seventh selection flavor, with the first direct evidence on WHY external critics fail best-of-K (ForesightFlow page, 2606.04968, deep-read same session the sweep banked it): the same flow network generates a per-step success-potential track beside each action chunk (~1K extra params); best-of-K picks the highest-mean track. The load-bearing table is the K-sweep: IDQL’s separate 500M critic ranks the policy’s own candidates at chance (39.0 → 38.4 SR from K=1→5) while the jointly-generated scorer climbs +5.0 — selector shape beats selector size, and that’s the third strike on post-hoc probe selectors against our banked best-of-10 ceiling. Training = decoupled advantage-weighted FM (weights on action velocities only — coupled weighting lets an overconfident scorer mask its own corrective gradient). Carried as a directional prior, not a plan: needs stage labels + mixed-quality rollouts, which the offline panel doesn’t have. Bonus instrument: 1-NFE endpoint preview ranks candidates at Kendall τ ≈ 0.83 vs full integration (~97% of the selection gain) — score-before- integrate is available to any future selector rung.

  • Production dT sighting (2026-08-09, Qwen-VLA page): RL rollouts sample at τ=1.0, deployment sharpens to τ=0.6 — a production stack independently landing on the cool side, the same direction as our record-only dT table’s mild monotone toward T=0.5. Banked beside the table as a sighting, not evidence (their mechanism is exploration-vs-exploitation, not mean-collapse); primary stays T=1.0 per pre-reg.

  • 2026-08-09 ~14:3xZ — trained-critic pole placed and parked (Robot Critics page, 2606.21572): fine-tuned VLM critic (pairwise success/failure supervision from policy rollouts) + action-conditioned video outcome model → +11% real / +5.9% sim. Consistent with RoVer and the free-scorer arc: learning the judge is what makes judging work. Parked at its stated price — needs rollout labels (#16) and a video model, and our ceiling reads cap the payoff on our decodes (AR small, flow ~null). No rank change.

  • External prior on the head’s trained shape (2026-08-09, FAN page, 2604.01570): the training-side push toward exactly the unimodal-around-the-mode distribution our decode reads measured. FAN regularizes a discrete-token VLA head toward a Gaussian bump centered on its own argmax (CVPR26; OOD gains +5–6 pts on OpenVLA/-OFT). Read for #19: a head trained this way should widen the greedy-vs-sampled gap (more mass adjacent to the mode → sampling averages over near-mode neighbors), i.e. an external vote that mean-collapse is a property of well-trained manipulation heads, not a defect of ours. Record-only; no rank change, family decodes stand.

  • Lit 0816 2026-08-09 — the cheap-draws hook died on verification (ActionCache page, 2607.06370): the banked “changes #19’s cheap-draws cost model” claim is corrected — ActionCache’s retrieval is top-1 and returns ONE deterministic cached chunk (it collapses the draw distribution rather than amortizing N draws), and it accelerates only the flow head while the trunk — our dominant cost — runs every tick to produce the cache key. Our draws economics are unchanged (the 1191 ms 10-draw figure is the AR decoder anyway). A top-k retrieval variant would be a cheap-diverse-draws mechanism, but that is our extrapolation, not the paper. Record-only; family decodes stand.

  • 2026-08-09 lit 0817 — the draws cost model splits cleanly, and the mean-collapse asymmetry gets a weak external rhyme (Reflex 2607.14695 + Compression Gap 2604.03191): Reflex’s timestep-invariance observation (trunk KV exactly valid across all ODE steps — true of our frozen trunk by construction) means K draws share ONE trunk prefill: marginal draw cost = expert-only FLOPs. Combined with ActionCache’s per-decision anchor (~102 ms, VLM ≈ 22 ms), draw economics are better than the 0816 correction alone implied — the fixed trunk cost is per-decision, not per-draw. Compression Gap files as consistent-with, never predicted-by: continuous heads convert encoder upgrades to +21–26 pts where an 80-bit FSQ codebook passes +4–10 — but tiny non-VLA single-seed models, frozen encoders, mechanism asserted not measured, and our AR head’s ~1,800-bit budget plausibly escapes the bound entirely; with a WEAK encoder discrete wins by 17 pts. Watch note for the adamc k4l2 readout: flow-vs-AR divergence under unfrozen vision may cite it as a rhyme only.