| { |
| "schema_version": 1, |
| "study_id": "factorized_resource_substitution_v1", |
| "status": "preregistered_pending_execution", |
| "title": "Geometry as a Cognitive Resource: Can Structured Representations Substitute for Attention Depth?", |
| "scope": "compute_only_non_benchmark", |
| "primary_question": "Does an orthogonally factorized representational geometry reduce the attention depth required to acquire the same frozen linguistic relations?", |
| "primary_claim": "Factorized geometry substitutes for attention depth when reducing a model from 12 to 8 layers causes less loss of frozen relational competence under orthogonal factorization than under the matched dense control.", |
| "design": { |
| "factors": { |
| "geometry": ["dense", "orthogonal"], |
| "attention_layers": [8, 12] |
| }, |
| "seeds": [47, 53, 59], |
| "fresh_seeds_unseen_by_selection": true, |
| "models": [ |
| {"variant": "resource_dense_l12_s47", "geometry": "dense", "attention_layers": 12, "seed": 47}, |
| {"variant": "resource_dense_l12_s53", "geometry": "dense", "attention_layers": 12, "seed": 53}, |
| {"variant": "resource_dense_l12_s59", "geometry": "dense", "attention_layers": 12, "seed": 59}, |
| {"variant": "resource_dense_l8_s47", "geometry": "dense", "attention_layers": 8, "seed": 47}, |
| {"variant": "resource_dense_l8_s53", "geometry": "dense", "attention_layers": 8, "seed": 53}, |
| {"variant": "resource_dense_l8_s59", "geometry": "dense", "attention_layers": 8, "seed": 59}, |
| {"variant": "resource_orthogonal_l12_s47", "geometry": "orthogonal", "attention_layers": 12, "seed": 47}, |
| {"variant": "resource_orthogonal_l12_s53", "geometry": "orthogonal", "attention_layers": 12, "seed": 53}, |
| {"variant": "resource_orthogonal_l12_s59", "geometry": "orthogonal", "attention_layers": 12, "seed": 59}, |
| {"variant": "resource_orthogonal_l8_s47", "geometry": "orthogonal", "attention_layers": 8, "seed": 47}, |
| {"variant": "resource_orthogonal_l8_s53", "geometry": "orthogonal", "attention_layers": 8, "seed": 53}, |
| {"variant": "resource_orthogonal_l8_s59", "geometry": "orthogonal", "attention_layers": 8, "seed": 59} |
| ] |
| }, |
| "training": { |
| "corpus_words": 10000000, |
| "exposure_words": 100000000, |
| "strict_small": true, |
| "same_corpus": true, |
| "same_tokenizer": true, |
| "same_width": true, |
| "same_optimizer": true, |
| "same_checkpoint_schedule": true, |
| "saved_checkpoint_words": [1000000, 2000000, 3000000, 4000000, 5000000, 6000000, 7000000, 8000000, 9000000, 10000000, 20000000, 30000000, 40000000, 50000000, 60000000, 70000000, 80000000, 90000000, 100000000] |
| }, |
| "assay": { |
| "source_study": "factorized_mechanism_v2", |
| "source_items_sha256": "0517ba2a8b7966698f08f7d2fb7742b40bbb9b2855f35111cb69f3533a7a5b25", |
| "source_item_spec_sha256": "0b33a75dc26136438d7bfe6463afc502efabd3464e1ded0d706a84a7b032c8b5", |
| "selection_seed": 20260717, |
| "selection_reads_new_models": false, |
| "items_per_assay": 1024, |
| "crossfit_folds": 5, |
| "regularization_c": 1.0, |
| "evaluation_words": [1000000, 5000000, 10000000, 20000000, 40000000, 60000000, 80000000, 100000000], |
| "primary_words": 40000000, |
| "asymptotic_words": 100000000, |
| "holdout_words": 100000, |
| "encoding_batch_size": 32 |
| }, |
| "estimands": { |
| "primary": "(orthogonal_8_minus_orthogonal_12) - (dense_8_minus_dense_12) at 40M; positive means factorization protects competence under depth reduction", |
| "secondary": [ |
| "paired-seed interaction in normalized competence AUC from 1M to 40M", |
| "paired-seed interaction at 100M", |
| "first registered age reaching the competence criterion", |
| "held-out language-model loss interaction", |
| "competence per parameter and estimated attention-layer compute" |
| ], |
| "competence": "equal-weight mean of six cross-fitted direct-assay component accuracies", |
| "criterion": { |
| "minimum_component_accuracy": 0.65, |
| "required_components": 5 |
| } |
| }, |
| "inference": { |
| "seed_is_independent_training_unit": true, |
| "paired_by_seed": true, |
| "bootstrap_seed_resamples": 10000, |
| "bootstrap_seed": 20260718, |
| "directional_primary_hypothesis": true, |
| "report_all_components": true, |
| "no_checkpoint_selection": true, |
| "no_leaderboard_selection": true |
| }, |
| "claim_boundary": "The study tests computational resource substitution in frozen relational competence under one matched BabyLM training recipe. It does not claim human equivalence, universal architectural optimality, or leaderboard superiority. The direct assays measure representational competence, while held-out loss is a general-performance control." |
| } |
|
|
