AHR v2 ARC-AGI-3 research checkpoints
Action balancing and corrected world gradients: no baseline win (19 October 2026)
The complete 329,350,583-parameter AHR v2 remains intact: Queen transformer, Worker and Scout EBMs, literal Lorentz geometry, observed memory, learned spatial world model and bounded two-step planning. Informative actions are sampled uniformly, with one forced action per eight examples. Conditional coordinate supervision is separated from branch action weights; world heads use different successor bootstraps and changed-cell supervision.
A first action-balanced trial completed 3,446 updates / 55,136 sampled examples. Its calibrated policy recovered 34/68 optional clicks, but its raw policy recovered 0/68. It also exposed a concrete implementation error: subtracting a detached current state from a forecast containing that same identity base created an unwanted encoder gradient. The corrected loss differentiates both terms. Identity-limit and nonzero-residual finite-difference tests verify the correction; initial forward predictions remain bitwise identical.
The corrected trial completed 2,516 updates / 40,256 sampled examples in 26.64 minutes, including 1,812 joint updates. All three actual stage recovery gates passed. It used the same parent, seed, data, sampler and learning rates, with a separately registered shorter allowance. This makes terminal cross-trial comparisons descriptive; matched TRAIN updates are an additional mechanism diagnostic.
| Same 512 development observations | Action agreement | Joint NLL, lower is better | Mean click distance |
|---|---|---|---|
| TRAIN-fitted baseline | 43.75% | 4.4300 | 23.19px |
| Parent 4,385 | 35.16% | 4.9727 | 23.67px |
| First action-balanced 3,446 | 40.43% | 4.9127 | 22.93px |
| Corrected terminal 2,516 | 35.74% | 4.9493 | 23.26px |
| Preregistered TRAIN-selected 704 | 36.52% | 4.9797 | 23.61px |
Baseline cells use metric-specific references: mask-table action probabilities with uniform legal coordinate density for joint NLL, and the TRAIN-fitted coordinate median for click distance. They are not one predictor yielding every cell.
Neither corrected endpoint beats the unchanged baseline criterion. Optional clicks remain 0/68, raw and calibrated. The terminal raw joint NLL worsens from 5.0000 to 5.1244; TRAIN-fitted calibration lowers it to 4.9493. Loss reduction after calibration does not demonstrate that the raw policy learned useful click decisions.
Grounding is more stable: corrected world forecast MSE is 0.08603, versus 0.21326 in the first action-balanced trial and 0.08911 at initialization. Encoder alignment is 0.06875, versus 0.19342 and 0.07334 respectively. At matched TRAIN update 2,204, world MSE is 0.08472 versus 0.12296. However, the corrected terminal forecast still loses to neural-current persistence 0.08382 and fixed-descriptor persistence 0.02211. Useful action-dependent dynamics beyond persistence are not established. Distinct world heads alone do not prove calibrated epistemic uncertainty.
Corrected final/best weights and benchmark/restore metadata · First diagnostic trial.
Checkpoint preservation verified: both corrected final/best payloads total 2,653,937,277 bytes. All 22 payload/metadata assets and the separate restore receipt passed complete anonymous byte-count/SHA256 verification at 2026-10-19T22:33:44.406082+00:00. Latest two resumable checkpoints plus best remain on the training server. The temporary upload credential has been removed. GitHub releases remain mutable; verify full hashes on restore.
These are offline human-playthrough imitation and one-step descriptor diagnostics, using the same 512 observations from 44 recordings in four reused development families. This is not an ARC-AGI-3 gameplay score, game-completion result or new-game generalization test. The TRAIN-family selector has only four optional-click labels; ancestors were exposed to its families. One seed, reused development data and recording-bootstrap intervals conditional on these families limit conclusions. Inference controls establish component activity, not a causal architecture advantage.
Spatial world repair: no baseline win (19 October 2026)
The separate repair keeps the entire 329,334,183-parameter core, adds a learned 16,400-parameter spatial readout, and retains the Queen, both EBMs, literal Lorentz event geometry, observed memory and bounded two-step planning: 329,350,583 parameters total. It corrects the previous input/target coordinate mismatch using grounded neural spatial states, identity-initialized simplex residual forecasts, an existing-head spatial click pointer and train-only calibrated outcome/availability/error heads. The observed semantic context remains fixed during imagination.
All three actual stage-boundary exact-recovery gates passed, including full policy-only component gradients, inference effects and teacher-target isolation. The run completed 4,385 updates / 70,160 examples in 46 minutes 30 seconds, including 3,617 full joint updates / 57,872 examples, within the original one-hour allowance. No historical weights were deleted. GPU0 was shared; the other experiments were untouched.
The fixed cohort is the same 512 observations, 44 eligible recordings and four reused development families as the full pilot below. Both old and repaired point errors now use floating conditional means, registered before this repair's outcomes. Exact joint NLL remains the proper two-dimensional branch-mixture likelihood.
| Fixed development measurement | Old full model | Repair initial | Repair final | Baseline |
|---|---|---|---|---|
| Human action agreement | 41.60% | 43.75% | 35.16% | 43.75% |
| Agreement, 384 multi-legal observations | 22.14% | 25.00% | 13.54% | 25.00% |
| Exact joint NLL, lower is better | 5.2245 | 4.4616 | 4.9727 | 4.4300 |
| Mean click distance, 196 clicks | 24.37 px | 24.24 px | 23.67 px | 23.19 px |
| Common fixed-descriptor world MSE | 0.46159 | 0.75462 | 0.08911 | 0.02211 |
The final model does not beat the combined baseline criterion. Joint NLL improves over the old model by 0.2517 nats (paired recording 95% interval 0.1959–0.3074), but deteriorates against the repaired initialization by 0.5111 nats (0.3879–0.6185). Action agreement is 8.59 percentage points below the mask baseline (5.66–11.70 pp below). The comparison algorithms are a training-fitted mask action prior, its product with uniform coordinates, a training-fitted constant median click, and exact current-descriptor persistence. Exact clicks remain 0/196. The historical 31.15-pixel result below used marginal modes; it is not the matched old mean-decoder result in this table.
Grounding improves substantially (old encoder alignment MSE 0.43259 → 0.07334), but world prediction still loses to copying its own neural state (0.08911 versus 0.08810) and to raw-descriptor persistence (0.02211). In learned successor coordinates, forecast MSE is 0.016300, versus persistence 0.016046. Wrong-action alignment and state-shuffle intervals do not establish useful state-specific generalization. Planning now helps same-weight joint NLL by only 0.00382 (0.00225–0.00548), without changing action agreement. Removing EBM refinement improves NLL by 0.01545; the Lorentz-versus-Euclidean interval crosses zero. These are inference interventions, not separately trained causal architecture comparisons.
The most concrete remaining shortcut is click-action suppression: on 68 target-click observations with other legal actions, agreement falls from 68/68 initially to 0/68 finally. The final policy selects action 2 on 28 and action 4 on 40 of these rows; the 128 forced-click rows remain correct. This is an exploratory post-run slice of the fixed cohort, not a new independent test.
The new repair excludes 21 complete TRAIN recordings (7,289 rows) from all new gradient phases, leaving 66,833 optimizer rows. The inherited parent used the earlier full pool, so this is a current-run holdout rather than globally unseen data. Calibration and checkpoint selection use only these reserved TRAIN recordings; the unchanged benchmark priors retain all 74,122 original training rows. Fixed TRAIN-selection NLL improves 4.4533 → 3.9472, while development NLL worsens 4.4616 → 4.9727, indicating poor transfer/overfitting. The final checkpoint is also best by the registered TRAIN-selection metric; this does not turn the development failure into a win.
Three successor heads are retained, but equal initialization and their shared objective preserve symmetry. Independent ensemble diversity is unestablished; uncertainty is a learned conditional error scale with TRAIN-only calibration, not demonstrated epistemic or out-of-distribution uncertainty. One seed, correlated recordings, reused development families and uncorrected multiple comparisons limit inference. No ARC gameplay, game completion, competition score, or new-game generalization is measured.
Public final/best storage verified: the world-repair release contains one deduplicated final/best step-4,385 payload (1,326,910,741 bytes), 14 metadata files and a public restore receipt. Every complete anonymous payload/metadata download and the restore receipt passed byte-count/SHA256 checks at 2026-10-19T20:34:38.438244+00:00. The checkpoint remains on the training server; the temporary upload credential was removed.
Full AHR pilot: no baseline win (19 October 2026)
The new full-path run retained 329,334,183 parameters and activated differentiable Worker/Scout EBM refinement, literal 1+3 Minkowski/Lorentz event calculations, observed two-frame memory and eight-root/two-step learned-world planning. All 17 exact-recovery and activation checks passed before training. This is a new initialization: the world and geometry modules were freshly initialized, all LayerNorm affine parameters reset, and the verified parent encoder/Queen/Worker/Scout tensors outside LayerNorm preserved. Exact continuation applies within this new run.
It completed 5,500 updates / 22,000 examples in 47 minutes 1 second, including smoke checks and evaluation, within the original one-hour allowance on shared GPU0. This is less than one pass through the 74,122-example pool. The frozen development cohort has 512 observations, 384 multi-legal rows and 196 click rows, balanced across four reused families and covering all 44 eligible recordings. It differs from the historical 4,096-row cohort below.
| Fixed development metric | New-run initial | Final step 5,500 | Baseline |
|---|---|---|---|
| Human action agreement | 38.67% | 41.60% | 43.75% |
| Agreement with multiple legal actions | 18.23% | 22.14% | 25.00% |
| Proper joint action/click NLL, lower is better | 12.9404 | 5.2245 | 4.4300 |
| Mean click distance, lower is better | 28.42 px | 31.15 px | 23.19 px |
| Fixed-descriptor world MSE, lower is better | 1.8328 | 0.4616 | 0.0221 |
The action baseline is a training-fitted legal-mask prior; joint NLL uses that prior with uniform coordinates; click distance uses a training-fitted constant median; world MSE uses the fixed current-frame descriptor as its successor prediction. Final exact clicks were 0/196. The final action difference from the mask baseline is -2.15 percentage points (exploratory 95% interval -5.46 to +0.98 pp), so superiority is not established. The best checkpoint by development joint NLL is step 2,500, also below the baseline criterion.
Real observations outperform shuffled observation bundles on multi-legal action agreement by 3.13 pp (95% interval +0.52 to +5.80 pp), but not on joint NLL. This is limited evidence of useful observation dependence. Disabling planning slightly improves joint NLL by 0.0069 nats (paired interval 0.0044 to 0.0093); the Lorentz-versus-Euclidean difference includes zero. These are same-checkpoint inference controls, not separately trained architecture ablations. Intervals resample recordings within fixed reused families, with one training seed and no multiple-comparison correction.
World prediction and encoder anchoring improve from their new-run initial states, but final world MSE remains 20.9 times the copy-current error (19.7 times on descriptor-changed rows). The coarse fixed auxiliary target is 16-colour occupancy in 8x8 spatial cells; it is not a direct policy input or full visual reconstruction. The planner uses approximate future transitions and unverified future masks. Computation stays at fixed depth two; persistent event DAG/pheromone memory and useful adaptive halting are not demonstrated. No ARC gameplay, completion rate or competition score was measured.
Public final/best storage verified: the full-path release contains final step 5,500, best step 2,500, 11 selected metadata files and a public restore receipt. Both payloads total 2,652,910,016 bytes. Every complete anonymous payload download and the restore receipt passed byte-count/SHA256 checks. The weights also remain on the training server; the temporary upload credential was removed.
Public research checkpoints are hosted as GitHub release assets. The completed longer human-playthrough run, human step-200 parent, and historical diagnostics are separate releases. On 19 October 2026, published payloads, selected metadata and public restore receipts passed full anonymous download and SHA256 verification.
Earlier policy-only continuation: no baseline win
The continuation added 10,766 optimizer updates, reaching global step 10,966 in 50 minutes 53 seconds, on a shared A100. All nine exact checkpoint-recovery smoke checks passed before training. The registered one-hour lease was not renewed, and the run completed at its wall-time cutoff.
Evaluation uses a fixed 4,096-row development cohort from four held-out game families and all 44 recordings with eligible labels out of 48 raw development recordings. Four recordings contain no eligible imitation labels. These development families have been reused; this is not an untouched test set.
| Metric | Initial step 200 | Final step 10,966 | Train-only comparison |
|---|---|---|---|
| Action agreement | 36.28% | 37.87% | Mask-conditioned prior: 47.68% |
| Agreement with multiple legal actions | 15.04% | 17.15% | Mask-conditioned prior: 30.24% |
| Mean click distance, lower is better | 21.95 px | 25.36 px | Global mean coordinate: 23.32 px |
| Joint action/click NLL, lower is better | 4.5785 | 5.7308 | Global action prior with uniform coordinates: 4.4622 |
Exact click coordinates remained 0/1,604 at the final endpoint. The action improvement is small but supported by a paired recording bootstrap within these fixed families; the final action score is 9.81 percentage points below the mask-conditioned baseline (exploratory 95% interval: 8.10–11.76 points below). Click error and joint likelihood worsened. Shuffling pixels changed 4,093/4,096 grids and provided no reliable evidence of useful state-specific pixel dependence. The best checkpoint by the declared development joint NLL remains the initial step-200 checkpoint.
The release includes the machine-readable learning report and terminal checkpoint verification. It does not establish ARC game completion, a competition score, new-game generalization, or a causal architecture advantage.
Earlier policy-only architecture scope
The full model contains 329,334,183 parameters. That earlier human-imitation path gives gradients to 211,749,274 parameters in the current-frame encoder and Queen/Worker/Scout policy paths. It mean-pools spatial features, uses independent x/y click heads, and supplies the current latent in the nominal history/memory slots. World prediction, energy-based refinement, Lorentz geometry and planning are inactive in this experiment. The result is evidence about this implemented imitation path, not a full test of those proposed reasoning mechanisms.
Data
Human playthroughs come from the existing public dataset magic-sword/arc_agi_3_public_demo_human_testing, revision 455b7c58eaa9d5ea05f91cd1b5b603121599e758. The prepared mirror contains 340 raw recordings across 25 public game families, yielding 178,387 validated transitions and 86,306 eligible imitation examples. Whole families are separated into 21 training and four development families. The continuation uses all 74,122 eligible training examples; the development pool contains 12,184 eligible examples. Raw recordings are not republished here.
Earlier 200-update check
A separate earlier 512-row development sample gave 42.58% → 40.04% action agreement, below a 48.05% global-prior baseline. On its 207 click rows, error improved from 40.26 to 20.20 pixels, with 0/207 exact coordinates. This smaller sample differs from the fixed 4,096-row continuation cohort; these numbers should not be mixed.
Recovery
These are PyTorch recovery dictionaries, not a Transformers AutoModel package. Matching architecture notebooks, runtime and dataset identities are required; executable notebooks are not published in these repositories. Each release includes a public restore receipt with original paths, asset IDs, byte counts and SHA256 hashes. GitHub releases are mutable: verify downloaded bytes before loading. Final and best payloads are also retained on the training server. Historical diagnostic snapshots are labelled separately from human-playthrough checkpoints.