File size: 6,701 Bytes
e1fea06 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 | {
"schema_version": 1,
"study_id": "dial_lexical_differentiation_v1",
"title": "Differentiate Before You Integrate? When Embedding Factorization Becomes an Inductive Bias in Data-Limited Language Acquisition",
"frozen_at_utc": "2026-07-19T21:14:38Z",
"fixed_compute_deadline_utc": "2026-07-20T09:14:38Z",
"status": "prospective_before_any_dial_training_or_outcome_access",
"primary_question": "Does temporarily protecting complementary relational signals at lexical access reduce cross-talk and change acquisition before shared Transformer integration?",
"claim_boundary": {
"positive": "A temporary input-level relational scaffold changes the acquisition trajectory relative to an information-matched overlapping-support control.",
"excluded": [
"persistent cognitive modules",
"human neural equivalence",
"grounded conceptual representations",
"universal architectural optimality",
"leaderboard superiority",
"Poincare geometry"
]
},
"conditions": {
"aligned": "Real H/O/A relations in three disjoint orthogonal rank-128 input supports.",
"shared": "The exact same real H/O/A relations in three fixed dense overlapping rank-128 supports; each block is orthonormal and the concatenated condition number must be below 3.",
"shuffled": "Aligned supports with one identity permutation shared across all three relation views and restricted within frozen matching strata."
},
"operational_relations": {
"H": "Hierarchical-distributional Brown radial-angular targets.",
"O": "Order-sensitive directional relative-position PPMI at offsets -3,-2,-1,+1,+2,+3.",
"A": "Associative-distributional symmetric PPMI."
},
"architecture": {
"backbone": "dense PRISM GPT-BERT",
"hidden_size": 384,
"rank_allocation": [128, 128, 128],
"embedding_tables": 1,
"transformer_streams": 1,
"shared_transformer_from_layer": 1,
"input_output_embeddings_tied": false,
"final_factor_readout": false,
"parser": false,
"attention_masks_from_priors": false,
"block_specific_normalization": false
},
"scaffold": {
"loss": "L_LM + lambda(t) * sum_k alpha_k L_k",
"lambda_schedule": "lambda0 * max(0, 1 - words_seen / 10000000)",
"withdrawal_words": 10000000,
"calibration_minibatch": 0,
"alpha_rule": "equalize initial H/O/A gradient norms on the input embedding table",
"lambda0_rule": "aggregate auxiliary gradient norm divided by LM gradient norm equals 0.25",
"outcome_tuning": false,
"orthogonality_penalty": false,
"cross_covariance_role": "diagnostic_only",
"structured_initialization": "0.75 fixed relational solution plus 0.25 paired isotropic initialization, renormalized to the registered initializer RMS"
},
"design": {
"paired_seeds": [0, 1, 2, 3, 4, 5],
"conditions_per_seed": 3,
"fresh_models": 18,
"terminal_exposure_words": 40000000,
"learning_rate_schedule_words": 100000000,
"checkpoints_words": [1000000, 2000000, 3000000, 4000000, 5000000, 6000000, 7000000, 8000000, 9000000, 10000000, 20000000, 30000000, 40000000],
"same_within_pair": [
"non-embedding initialization",
"document and batch order",
"dropout stream",
"edge batches",
"optimizer",
"gradient clipping",
"corpus",
"tokenizer",
"relation arrays",
"auxiliary schedule"
],
"intended_difference": "support overlap or the frozen shared identity permutation only"
},
"primary_endpoint": {
"battery": ["BLiMP", "COMPS", "EWoK", "Entity Tracking"],
"model_outputs": "native LM probabilities without fine-tuning or probes",
"trajectory": "equal-weight composite AUC over log exposure from 1M through terminal exposure",
"contrast": "aligned minus shared",
"falsification_gate": "aligned greater than shuffled",
"normalization": "independent frozen baseline only; otherwise equal-accuracy mean"
},
"validity_gates": {
"lm_loss_noninferiority_nats_per_token": 0.02,
"minimum_shuffle_direction": "aligned greater than shuffled",
"random_basis_controls": 32,
"random_basis_matching": ["rank", "energy"],
"maximum_failed_seeds_per_condition": 1,
"systematic_throughput_or_clipping_difference_allowed": false
},
"mechanism": {
"checkpoint_ages_words": [1000000, 10000000, 40000000],
"intervention": "input-block donor swaps matched by frequency, subword length, norm, coverage and degree",
"controls": "32 Haar-random equal-rank equal-energy bases",
"joint_interventions": "all 2^3 H/O/A swaps",
"interaction": "non-additive HxOxA contrast",
"gradient_audit_minibatches": 64,
"gradient_audit": "LM versus H/O/A cosine and conflict rates",
"imposed_not_result": "between-support orthogonality in aligned"
},
"cognitive_bridge": {
"primary": "AoA prediction with frequency, length, contextual diversity and lexical class controls, including within-frequency-bin effects",
"secondary": "cross-validated log1p self-paced reading time, grouped by the 205 sentences, with frequency, length, position, spillover and paired shared-condition surprisal controls",
"exploratory": "ERP only; no post-hoc mapping from subspaces to components",
"human_causality_claimed": false
},
"statistics": {
"unit": "paired training seed",
"primary_test": "two-sided exact paired sign-flip",
"minimum_attainable_two_sided_p": 0.03125,
"report": ["every seed", "paired effect", "confidence interval"],
"no_checkpoint_pseudoreplication": true
},
"compute_fallback": {
"decision_input": "preflight throughput only, before outcome access",
"if_predicted_wall_time_to_40m_exceeds_hours": 8.5,
"fallback_terminal_exposure_words": 30000000,
"node_failure": "retain complete balanced pairs only, first five then first four seeds",
"forbidden": "asymmetric condition truncation or metric-driven adaptation"
},
"priority_order": [
"primary acquisition composite",
"AoA",
"input swaps",
"gradient conflict",
"reading time",
"full zero-shot for one preregistered model",
"SuperGLUE",
"ERP",
"embedding visualization"
],
"artifact_policy": {
"model_checkpoints": "all registered ages plus 40M main and recovery on Hugging Face",
"priors_tokenizer_corpus": "pinned Hugging Face revisions",
"study_archive": "data and machine-readable provenance only",
"adaptive_result_access_before_all_lineages_sealed": false
},
"retrospective_context_only": [
"FACTORIZED tournament",
"FACTORIZED mechanism study",
"factorized resource substitution study",
"STRUCTURED-DS tournament"
]
}
|