scan.json (16000B)
1 { 2 "paper": { 3 "title": "Optimal scaling laws in learning hierarchical multi-index models", 4 "authors": ["Leonardo Defilippis", "Florent Krzakala", "Bruno Loureiro", "Antoine Maillard"], 5 "year": 2026, 6 "venue": "arXiv", 7 "arxiv_id": "2602.05846" 8 }, 9 "scan_version": 2, 10 "active_modules": [], 11 "methodology_tags": ["theoretical"], 12 "key_findings": "The paper derives exact information-theoretic scaling laws for subspace recovery and prediction error in two-layer neural networks trained on hierarchical multi-index targets. It shows that optimal rates are achieved by a target-agnostic spectral estimator interpretable as gradient descent's small learning-rate limit. Features are learned sequentially through a cascade of phase transitions, producing plateaus and abrupt drops in prediction error. The scaling exponents previously observed only for quadratic networks are proven universal for a broad class of hierarchical targets.", 13 "claims": [ 14 { 15 "claim": "The Bayes-optimal mean-squared error for scale-free hierarchical multi-index models follows specific power-law scaling regimes depending on gamma and sample complexity (Theorem 3.2).", 16 "evidence": "Formal proof provided in Appendix A, with matching upper and lower bounds derived from oracle estimator analysis (A.1) and spectral method analysis (A.2).", 17 "supported": "strong" 18 }, 19 { 20 "claim": "A simple target-agnostic spectral estimator achieves the Bayes-optimal reconstruction rates (Theorem 3.6).", 21 "evidence": "Proof in Appendix A.2 building on results from Kovačević et al. (2025). Figure 1 shows finite-size experiments matching theoretical predictions for d=1000, m*=10.", 22 "supported": "strong" 23 }, 24 { 25 "claim": "Two-layer neural networks trained via spectral initialization plus ridge regression achieve the optimal scaling laws for excess risk (Theorem 3.8).", 26 "evidence": "Proof in Appendix B via two-stage argument: spectral initialization recovers subspace optimally, then RFRR on projected inputs has subleading error O(n^{-1/2}).", 27 "supported": "strong" 28 }, 29 { 30 "claim": "Features emerge sequentially through sharp phase transitions, with the i-th direction becoming detectable at sample complexity n_i = Theta(i^{2gamma} * d).", 31 "evidence": "Follows from Theorem 3.2 eq. (13) and visualized in Figure 2 showing eigenvalue spikes emerging from the bulk at increasing sample complexity.", 32 "supported": "strong" 33 } 34 ], 35 "checklist": { 36 "artifacts": { 37 "code_released": { 38 "applies": true, 39 "answer": false, 40 "justification": "No code repository or URL is mentioned in the paper." 41 }, 42 "data_released": { 43 "applies": false, 44 "answer": false, 45 "justification": "Purely theoretical paper with synthetic data generated from defined mathematical models. No dataset to release." 46 }, 47 "environment_specified": { 48 "applies": false, 49 "answer": false, 50 "justification": "Purely theoretical paper. The illustrative figures use simple synthetic experiments but no environment details are needed or provided." 51 }, 52 "reproduction_instructions": { 53 "applies": false, 54 "answer": false, 55 "justification": "Theoretical paper. Results are mathematical proofs, not computational experiments requiring reproduction instructions." 56 } 57 }, 58 "statistical_methodology": { 59 "confidence_intervals_or_error_bars": { 60 "applies": false, 61 "answer": false, 62 "justification": "Theoretical paper proving exact asymptotic results. The illustrative figures show empirical curves but these serve to visualize theoretical predictions, not to make statistical claims." 63 }, 64 "significance_tests": { 65 "applies": false, 66 "answer": false, 67 "justification": "No statistical comparisons are made. Claims are mathematical theorems with proofs." 68 }, 69 "effect_sizes_reported": { 70 "applies": false, 71 "answer": false, 72 "justification": "Not applicable to a theoretical paper deriving exact scaling laws." 73 }, 74 "sample_size_justified": { 75 "applies": false, 76 "answer": false, 77 "justification": "Theoretical paper with no empirical experiments requiring sample size justification." 78 }, 79 "variance_reported": { 80 "applies": false, 81 "answer": false, 82 "justification": "Theoretical paper. Figure 1 notes 'averaged over 70 instances' but this is illustrative, not a statistical claim." 83 } 84 }, 85 "evaluation_design": { 86 "baselines_included": { 87 "applies": false, 88 "answer": false, 89 "justification": "Theoretical paper proving optimality results. The spectral estimator is shown to match Bayes-optimal bounds, which serves as the fundamental baseline." 90 }, 91 "baselines_contemporary": { 92 "applies": false, 93 "answer": false, 94 "justification": "Not applicable to a theoretical paper." 95 }, 96 "ablation_study": { 97 "applies": false, 98 "answer": false, 99 "justification": "Not applicable. The paper proves theorems about mathematical objects, not an engineered system." 100 }, 101 "multiple_metrics": { 102 "applies": false, 103 "answer": false, 104 "justification": "Theoretical paper with mathematically defined metrics (MSE, k-critical threshold). Not an empirical evaluation." 105 }, 106 "human_evaluation": { 107 "applies": false, 108 "answer": false, 109 "justification": "Not applicable to a theoretical paper." 110 }, 111 "held_out_test_set": { 112 "applies": false, 113 "answer": false, 114 "justification": "Not applicable to a theoretical paper." 115 }, 116 "per_category_breakdown": { 117 "applies": false, 118 "answer": false, 119 "justification": "Not applicable to a theoretical paper." 120 }, 121 "failure_cases_discussed": { 122 "applies": true, 123 "answer": true, 124 "justification": "Section 5 (Conclusions, discussions, and limitations) discusses settings where the theory does not apply: isotropic Gaussian data assumption, two-layer architectures only, and the open question of whether SGD achieves the same rates without spectral initialization." 125 }, 126 "negative_results_reported": { 127 "applies": true, 128 "answer": true, 129 "justification": "The paper acknowledges open questions and limitations: SGD may not achieve Bayes-optimal rates without spectral initialization (Section 5), and the analysis requires m* to be large but small compared to n,d (Section 5, paragraph 1)." 130 } 131 }, 132 "claims_and_evidence": { 133 "abstract_claims_supported": { 134 "applies": true, 135 "answer": true, 136 "justification": "All abstract claims are supported by formal theorems: exact information-theoretic scaling laws (Theorem 3.2), spectral estimator achieving optimal rates (Theorem 3.6), neural network learnability (Theorem 3.8), and sequential feature emergence (follows from the cascade of phase transitions in Theorem 3.2)." 137 }, 138 "causal_claims_justified": { 139 "applies": false, 140 "answer": false, 141 "justification": "The paper makes mathematical claims proven by theorem, not causal claims requiring experimental justification." 142 }, 143 "generalization_bounded": { 144 "applies": true, 145 "answer": true, 146 "justification": "The paper explicitly states its scope: two-layer networks, hierarchical multi-index targets with generative exponent 2, isotropic Gaussian data, asymptotic high-dimensional limit. Section 5 explicitly discusses these as both limitations and strengths." 147 }, 148 "alternative_explanations_discussed": { 149 "applies": false, 150 "answer": false, 151 "justification": "Theoretical paper with proven results. No empirical findings requiring alternative explanations." 152 }, 153 "proxy_outcome_distinction": { 154 "applies": false, 155 "answer": false, 156 "justification": "Theoretical paper with no measurements. Claims match mathematical definitions exactly." 157 } 158 }, 159 "setup_transparency": { 160 "model_versions_specified": { 161 "applies": false, 162 "answer": false, 163 "justification": "No AI models are used. The paper studies mathematical properties of neural network architectures." 164 }, 165 "prompts_provided": { 166 "applies": false, 167 "answer": false, 168 "justification": "No prompting is used." 169 }, 170 "hyperparameters_reported": { 171 "applies": false, 172 "answer": false, 173 "justification": "Theoretical paper. Algorithm 1 specifies hyperparameters mathematically (p = omega(n^{1/2}), lambda = Theta(n^{-1/2}))." 174 }, 175 "scaffolding_described": { 176 "applies": false, 177 "answer": false, 178 "justification": "No agentic scaffolding is used." 179 }, 180 "data_preprocessing_documented": { 181 "applies": false, 182 "answer": false, 183 "justification": "No empirical data collection. Synthetic data is fully specified by mathematical definitions (Definition 2.2)." 184 } 185 }, 186 "limitations_and_scope": { 187 "limitations_section_present": { 188 "applies": true, 189 "answer": true, 190 "justification": "Section 5 is titled 'Conclusions, discussions, and limitations' and contains substantive discussion of the paper's limitations." 191 }, 192 "threats_to_validity_specific": { 193 "applies": true, 194 "answer": true, 195 "justification": "Section 5 discusses specific limitations: isotropic Gaussian data assumption, two-layer architectures only, open question of SGD achieving optimal rates, and the gap between m* = O(d) in prior work vs their setting where m* is large but sub-linear." 196 }, 197 "scope_boundaries_stated": { 198 "applies": true, 199 "answer": true, 200 "justification": "Section 5 explicitly states: 'our analysis deliberately relies on simplifying assumptions—most notably isotropic Gaussian data, two-layer architectures, and an asymptotic high-dimensional limit.' It also notes Remark 3.9 about not proving Bayes risk bounds." 201 } 202 }, 203 "data_integrity": { 204 "raw_data_available": { 205 "applies": false, 206 "answer": false, 207 "justification": "Theoretical paper with no empirical data to verify." 208 }, 209 "data_collection_described": { 210 "applies": false, 211 "answer": false, 212 "justification": "No data collection. Data model is fully specified mathematically." 213 }, 214 "recruitment_methods_described": { 215 "applies": false, 216 "answer": false, 217 "justification": "No participants or samples recruited." 218 }, 219 "data_pipeline_documented": { 220 "applies": false, 221 "answer": false, 222 "justification": "No data pipeline. Purely theoretical work." 223 } 224 }, 225 "conflicts_of_interest": { 226 "funding_disclosed": { 227 "applies": true, 228 "answer": true, 229 "justification": "Acknowledgements section lists funding: ANR (France 2030, ANR-23-IACL-0008), Choose France - CNRS AI Rising Talents, Swiss National Science Foundation grants (200021 200390, 225837), and Simons Foundation grant (#1257412)." 230 }, 231 "affiliations_disclosed": { 232 "applies": true, 233 "answer": true, 234 "justification": "Author affiliations are listed: ENS/PSL/CNRS, EPFL, INRIA Paris." 235 }, 236 "funder_independent_of_outcome": { 237 "applies": true, 238 "answer": true, 239 "justification": "Funders are government research agencies (ANR, SNSF) and the Simons Foundation, none of which have a financial interest in the theoretical results." 240 }, 241 "financial_interests_declared": { 242 "applies": true, 243 "answer": false, 244 "justification": "No competing interests statement is included in the paper." 245 } 246 }, 247 "contamination": { 248 "training_cutoff_stated": { 249 "applies": false, 250 "answer": false, 251 "justification": "No pre-trained model is evaluated on any benchmark." 252 }, 253 "train_test_overlap_discussed": { 254 "applies": false, 255 "answer": false, 256 "justification": "No pre-trained model is evaluated on any benchmark." 257 }, 258 "benchmark_contamination_addressed": { 259 "applies": false, 260 "answer": false, 261 "justification": "No pre-trained model is evaluated on any benchmark." 262 } 263 }, 264 "human_studies": { 265 "pre_registered": { 266 "applies": false, 267 "answer": false, 268 "justification": "No human participants." 269 }, 270 "irb_or_ethics_approval": { 271 "applies": false, 272 "answer": false, 273 "justification": "No human participants." 274 }, 275 "demographics_reported": { 276 "applies": false, 277 "answer": false, 278 "justification": "No human participants." 279 }, 280 "inclusion_exclusion_criteria": { 281 "applies": false, 282 "answer": false, 283 "justification": "No human participants." 284 }, 285 "randomization_described": { 286 "applies": false, 287 "answer": false, 288 "justification": "No human participants." 289 }, 290 "blinding_described": { 291 "applies": false, 292 "answer": false, 293 "justification": "No human participants." 294 }, 295 "attrition_reported": { 296 "applies": false, 297 "answer": false, 298 "justification": "No human participants." 299 } 300 }, 301 "cost_and_practicality": { 302 "inference_cost_reported": { 303 "applies": false, 304 "answer": false, 305 "justification": "Purely theoretical paper." 306 }, 307 "compute_budget_stated": { 308 "applies": false, 309 "answer": false, 310 "justification": "Purely theoretical paper." 311 } 312 } 313 }, 314 "red_flags": [], 315 "cited_papers": [ 316 { 317 "title": "Scaling laws for neural language models", 318 "authors": ["J. Kaplan", "S. McCandlish", "T. Henighan", "T. B. Brown", "B. Chess", "R. Child", "S. Gray", "A. Radford", "J. Wu", "D. Amodei"], 319 "year": 2020, 320 "arxiv_id": "2001.08361", 321 "relevance": "Foundational empirical work on neural scaling laws that this paper provides theoretical grounding for." 322 }, 323 { 324 "title": "An empirical analysis of compute-optimal large language model training", 325 "authors": ["J. Hoffmann", "S. Borgeaud", "A. Mensch"], 326 "year": 2022, 327 "relevance": "Chinchilla scaling laws for compute-optimal LLM training, a key empirical reference for the scaling law theory developed here." 328 }, 329 { 330 "title": "Language models are few-shot learners", 331 "authors": ["T. Brown", "B. Mann", "N. Ryder"], 332 "year": 2020, 333 "relevance": "GPT-3 paper documenting empirical scaling behavior that motivates theoretical understanding of scaling laws." 334 }, 335 { 336 "title": "Emergent abilities of large language models", 337 "authors": ["J. Wei", "Y. Tay", "R. Bommasani"], 338 "year": 2022, 339 "arxiv_id": "2206.07682", 340 "relevance": "Empirical observation of emergent abilities that this paper provides a theoretical framework for via sequential phase transitions." 341 }, 342 { 343 "title": "Are emergent abilities of large language models a mirage?", 344 "authors": ["R. Schaeffer", "B. Miranda", "S. Koyejo"], 345 "year": 2023, 346 "relevance": "Challenges the emergent abilities narrative; this paper's theory of sharp phase transitions provides a mathematical account of when emergence is real vs artifact." 347 }, 348 { 349 "title": "Explaining neural scaling laws", 350 "authors": ["Y. Bahri", "E. Dyer", "J. Kaplan", "J. Lee", "U. Sharma"], 351 "year": 2024, 352 "relevance": "Prior theoretical work on explaining neural scaling laws that this paper extends to the feature-learning regime." 353 } 354 ] 355 }