scan-v5.json (29113B)
1 { 2 "scan_version": 5, 3 "paper_type": "empirical", 4 "paper": { 5 "title": "Data Distributional Properties Drive Emergent In-Context Learning in Transformers", 6 "authors": [ 7 "Stephanie C. Y. Chan", 8 "Adam Santoro", 9 "Andrew Kyle Lampinen", 10 "Jane X. Wang", 11 "Aaditya K Singh", 12 "Pierre H. Richemond", 13 "James L. McClelland", 14 "Felix Hill" 15 ], 16 "year": 2022, 17 "venue": "Neural Information Processing Systems", 18 "arxiv_id": "2205.05055", 19 "doi": "10.48550/arXiv.2205.05055" 20 }, 21 "checklist": { 22 "claims_and_evidence": { 23 "abstract_claims_supported": { 24 "applies": true, 25 "answer": true, 26 "justification": "All major abstract claims (burstiness drives ICL, Zipfian distribution enables coexistence of ICL and in-weights learning, recurrent models fail) are supported by experimental results in Figs 2–7.", 27 "source": "haiku" 28 }, 29 "causal_claims_justified": { 30 "applies": true, 31 "answer": true, 32 "justification": "The paper uses controlled experimental manipulation of individual distributional properties (burstiness, class count, label multiplicity) while holding others constant, which is adequate for causal inference in this controlled laboratory setting.", 33 "source": "haiku" 34 }, 35 "generalization_bounded": { 36 "applies": true, 37 "answer": false, 38 "justification": "The abstract claims findings explain ICL in 'large language models,' but experiments use small transformers (dim 64) trained from scratch on Omniglot image classification—a significant extrapolation not bounded to the tested setting.", 39 "source": "haiku" 40 }, 41 "alternative_explanations_discussed": { 42 "applies": true, 43 "answer": true, 44 "justification": "The paper examines whether recurrent model inferiority on ICL is explained by a compensating bias toward in-weights learning (Appendix C.2/Fig 8), and directly engages competing theories from Min et al. and Razeghi et al.", 45 "source": "haiku" 46 }, 47 "proxy_outcome_distinction": { 48 "applies": true, 49 "answer": true, 50 "justification": "ICL is measured using a standard 4-shot 2-way evaluation on holdout classes with randomly reassigned labels, which directly operationalizes the claimed capability without proxy substitution.", 51 "source": "haiku" 52 } 53 }, 54 "limitations_and_scope": { 55 "limitations_section_present": { 56 "applies": true, 57 "answer": false, 58 "justification": "There is no dedicated limitations or threats-to-validity section; Section 4 is a discussion of implications and future directions, not an honest accounting of what the study does not show.", 59 "source": "haiku" 60 }, 61 "threats_to_validity_specific": { 62 "applies": true, 63 "answer": false, 64 "justification": "No specific threats are discussed; Appendix B mentions the constraint on novel labels but frames it as a future extension opportunity rather than a validity concern for the current study.", 65 "source": "haiku" 66 }, 67 "scope_boundaries_stated": { 68 "applies": true, 69 "answer": false, 70 "justification": "The paper uses hedging language ('may explain,' 'could allow') but never explicitly states what the results do NOT show or explicitly bounds them to the small-transformer/Omniglot setting.", 71 "source": "haiku" 72 } 73 }, 74 "conflicts_of_interest": { 75 "funding_disclosed": { 76 "applies": true, 77 "answer": true, 78 "justification": "The acknowledgment section explicitly states 'This work was funded by DeepMind.'", 79 "source": "haiku" 80 }, 81 "affiliations_disclosed": { 82 "applies": true, 83 "answer": true, 84 "justification": "Author affiliations are listed on the title page: seven authors at DeepMind, one at UCL, one shared with Stanford.", 85 "source": "haiku" 86 }, 87 "funder_independent_of_outcome": { 88 "applies": true, 89 "answer": false, 90 "justification": "DeepMind funded the work and employs the majority of authors; findings support the transformer architecture that DeepMind builds and deploys, creating a structural alignment of interests.", 91 "source": "haiku" 92 }, 93 "financial_interests_declared": { 94 "applies": true, 95 "answer": false, 96 "justification": "No competing interests or financial interests statement is present; the acknowledgment only identifies the funding source.", 97 "source": "haiku" 98 } 99 }, 100 "scope_and_framing": { 101 "key_terms_defined": { 102 "applies": true, 103 "answer": true, 104 "justification": "In-context learning, in-weights learning, burstiness, Zipfian distribution, label multiplicity, and within-class variation are all operationally defined in the introduction and experimental design sections.", 105 "source": "haiku" 106 }, 107 "intended_contribution_clear": { 108 "applies": true, 109 "answer": true, 110 "justification": "The paper clearly states it is identifying which distributional properties of training data drive emergent ICL, contributing the identification of burstiness, class rarity, dynamic meanings, and Zipfian skew as key factors.", 111 "source": "haiku" 112 }, 113 "engagement_with_prior_work": { 114 "applies": true, 115 "answer": true, 116 "justification": "The paper directly addresses competing explanations (Min et al., Razeghi et al., Xie et al.) and situates itself relative to explicit meta-learning approaches (Santoro et al., Vinyals et al., Wang et al.), showing how its findings extend and challenge prior work.", 117 "source": "haiku" 118 } 119 } 120 }, 121 "type_checklist": { 122 "empirical": { 123 "artifacts": { 124 "code_released": { 125 "applies": true, 126 "answer": true, 127 "justification": "A footnote on page 1 states 'Code is available at: https://github.com/deepmind/emergent_in_context_learning'; the NeurIPS checklist notes release with the camera-ready version, suggesting it was made available.", 128 "source": "haiku" 129 }, 130 "data_released": { 131 "applies": true, 132 "answer": true, 133 "justification": "The paper uses the Omniglot dataset, which is publicly available and cited with MIT License; no new data is created.", 134 "source": "haiku" 135 }, 136 "environment_specified": { 137 "applies": true, 138 "answer": false, 139 "justification": "The appendix specifies hardware (16 TPU v2/v3 cores) and optimizer details but provides no requirements file, Dockerfile, or software dependency list sufficient to reproduce the environment.", 140 "source": "haiku" 141 }, 142 "reproduction_instructions": { 143 "applies": true, 144 "answer": false, 145 "justification": "While training procedure and architecture are described in the appendix, no step-by-step reproduction instructions are provided; users must infer setup from the code repository.", 146 "source": "haiku" 147 } 148 }, 149 "statistical_methodology": { 150 "confidence_intervals_or_error_bars": { 151 "applies": true, 152 "answer": true, 153 "justification": "The appendix explicitly states 'In all figures, (shaded) error bars indicate standard deviation around the mean,' and error bands are visible in all main result figures.", 154 "source": "haiku" 155 }, 156 "significance_tests": { 157 "applies": true, 158 "answer": false, 159 "justification": "No formal statistical significance tests (t-tests, ANOVA, permutation tests) are reported; results are presented as learning curves with standard deviation bands only.", 160 "source": "haiku" 161 }, 162 "effect_sizes_reported": { 163 "applies": true, 164 "answer": true, 165 "justification": "Raw accuracy values are reported for all conditions across learning curves, permitting direct comparison of effect magnitudes (e.g., ICL accuracy rising from chance to ~1.0 with increasing burstiness).", 166 "source": "haiku" 167 }, 168 "sample_size_justified": { 169 "applies": true, 170 "answer": false, 171 "justification": "The paper uses 5 seeds for most experiments and 3 for others (Figs 5–6) but provides no justification for why these numbers of seeds are sufficient to detect the effects of interest.", 172 "source": "haiku" 173 }, 174 "variance_reported": { 175 "applies": true, 176 "answer": true, 177 "justification": "Standard deviation is shown as shaded regions in all figures, as explicitly confirmed in the appendix.", 178 "source": "haiku" 179 } 180 }, 181 "evaluation_design": { 182 "baselines_included": { 183 "applies": true, 184 "answer": true, 185 "justification": "Baselines include P(bursty)=0 (no burstiness), 100 training classes (minimal vocabulary), and RNN/LSTM architecture comparisons, each serving as natural lower bounds.", 186 "source": "haiku" 187 }, 188 "baselines_contemporary": { 189 "applies": true, 190 "answer": true, 191 "justification": "Architecture comparisons use matched LSTM and vanilla RNN baselines with identical parameter counts, layers, and training procedures, making them appropriate contemporaries.", 192 "source": "haiku" 193 }, 194 "ablation_study": { 195 "applies": true, 196 "answer": true, 197 "justification": "The entire experimental design is structured as systematic ablations: each distributional property (burstiness, class count, label multiplicity, within-class variation, Zipfian skew) is manipulated independently across separate experiments.", 198 "source": "haiku" 199 }, 200 "multiple_metrics": { 201 "applies": true, 202 "answer": true, 203 "justification": "Both in-context learning accuracy (on holdout classes with random label assignment) and in-weights learning accuracy (on trained classes without context support) are measured throughout all experiments.", 204 "source": "haiku" 205 }, 206 "human_evaluation": { 207 "applies": false, 208 "answer": false, 209 "justification": "Human evaluation is not relevant; this study trains and evaluates neural networks automatically.", 210 "source": "haiku" 211 }, 212 "held_out_test_set": { 213 "applies": true, 214 "answer": true, 215 "justification": "In-context learning is always evaluated on holdout image classes 'that were never encountered in training,' with evaluation labels randomly reassigned to prevent reliance on training-time associations.", 216 "source": "haiku" 217 }, 218 "per_category_breakdown": { 219 "applies": true, 220 "answer": true, 221 "justification": "Results are broken down by experimental condition (burstiness level, number of classes, Zipf exponent, label multiplicity, within-class variation), and the Zipfian experiments separately report accuracy on common vs. rare classes.", 222 "source": "haiku" 223 }, 224 "failure_cases_discussed": { 225 "applies": true, 226 "answer": true, 227 "justification": "Failure cases are prominently reported: high Zipf exponent (=3) causes ICL to fail; recurrent models never achieve ICL under any condition; rare classes are never memorized regardless of Zipfian skew.", 228 "source": "haiku" 229 }, 230 "negative_results_reported": { 231 "applies": true, 232 "answer": true, 233 "justification": "Negative results are central findings: absence of burstiness or insufficient class count prevents ICL emergence; recurrent models completely fail; rare classes show chance-level in-weights performance across all Zipf conditions.", 234 "source": "haiku" 235 } 236 }, 237 "setup_transparency": { 238 "model_versions_specified": { 239 "applies": true, 240 "answer": true, 241 "justification": "Exact model specifications are given: 12-layer transformer, embedding dimension 64, 8 heads, ResNet embedder with specific block and channel architecture; parameter counts are provided for all compared architectures.", 242 "source": "haiku" 243 }, 244 "prompts_provided": { 245 "applies": false, 246 "answer": false, 247 "justification": "This paper trains transformers from scratch on image-label sequences; there are no natural language prompts or system instructions.", 248 "source": "haiku" 249 }, 250 "hyperparameters_reported": { 251 "applies": true, 252 "answer": true, 253 "justification": "Hyperparameters are reported in the appendix: Adam optimizer, max learning rate 3e-4 at 4000 steps with inverse square root decay, 500k training steps, and the full hyperparameter sweep ranges used for architecture comparisons.", 254 "source": "haiku" 255 }, 256 "scaffolding_described": { 257 "applies": false, 258 "answer": false, 259 "justification": "No agentic scaffolding is used; this paper trains and evaluates neural networks directly.", 260 "source": "haiku" 261 }, 262 "data_preprocessing_documented": { 263 "applies": true, 264 "answer": true, 265 "justification": "Section 2.1 describes how Omniglot images are processed (ResNet embedder, integer label embeddings), how training sequences are constructed (context + query format, bursty vs. non-bursty generation), and how evaluation sequences are formed.", 266 "source": "haiku" 267 } 268 }, 269 "data_integrity": { 270 "raw_data_available": { 271 "applies": true, 272 "answer": true, 273 "justification": "Omniglot is a public dataset cited with MIT License; experimental sequences are procedurally generated from this public data and the generation procedure is described.", 274 "source": "haiku" 275 }, 276 "data_collection_described": { 277 "applies": true, 278 "answer": true, 279 "justification": "Section 2.1 fully describes how training and evaluation sequences are generated from Omniglot, including bursty/non-bursty mixing, label assignment, and holdout class selection.", 280 "source": "haiku" 281 }, 282 "recruitment_methods_described": { 283 "applies": false, 284 "answer": false, 285 "justification": "No human participants; a standard public benchmark dataset is used.", 286 "source": "haiku" 287 }, 288 "data_pipeline_documented": { 289 "applies": true, 290 "answer": true, 291 "justification": "The full pipeline from Omniglot dataset through sequence construction to evaluation is documented across Section 2 and the appendix.", 292 "source": "haiku" 293 } 294 }, 295 "contamination": { 296 "training_cutoff_stated": { 297 "applies": false, 298 "answer": false, 299 "justification": "Models are trained from scratch on controlled data; there is no pre-trained model with a training data cutoff to declare.", 300 "source": "haiku" 301 }, 302 "train_test_overlap_discussed": { 303 "applies": true, 304 "answer": true, 305 "justification": "The paper explicitly addresses train/test class separation: holdout classes are 'never encountered in training,' and Appendix B discusses the design choice of using seen labels in evaluation, arguing it makes ICL harder, not easier.", 306 "source": "haiku" 307 }, 308 "benchmark_contamination_addressed": { 309 "applies": false, 310 "answer": false, 311 "justification": "Models are trained from scratch on procedurally generated sequences; there is no pre-existing benchmark contamination risk in the LLM sense.", 312 "source": "haiku" 313 } 314 }, 315 "human_studies": { 316 "pre_registered": { 317 "applies": false, 318 "answer": false, 319 "justification": "No human participants.", 320 "source": "haiku" 321 }, 322 "irb_or_ethics_approval": { 323 "applies": false, 324 "answer": false, 325 "justification": "No human participants.", 326 "source": "haiku" 327 }, 328 "demographics_reported": { 329 "applies": false, 330 "answer": false, 331 "justification": "No human participants.", 332 "source": "haiku" 333 }, 334 "inclusion_exclusion_criteria": { 335 "applies": false, 336 "answer": false, 337 "justification": "No human participants.", 338 "source": "haiku" 339 }, 340 "randomization_described": { 341 "applies": false, 342 "answer": false, 343 "justification": "No human participants.", 344 "source": "haiku" 345 }, 346 "blinding_described": { 347 "applies": false, 348 "answer": false, 349 "justification": "No human participants.", 350 "source": "haiku" 351 }, 352 "attrition_reported": { 353 "applies": false, 354 "answer": false, 355 "justification": "No human participants.", 356 "source": "haiku" 357 } 358 }, 359 "cost_and_practicality": { 360 "inference_cost_reported": { 361 "applies": false, 362 "answer": false, 363 "justification": "This is a training/mechanism study, not a deployment study; inference cost is not relevant to the research questions.", 364 "source": "haiku" 365 }, 366 "compute_budget_stated": { 367 "applies": true, 368 "answer": true, 369 "justification": "The appendix states experiments ran for 500k training steps on 16 TPU v2 or v3 cores; the architecture comparison used 90 hyperparameter sweep runs.", 370 "source": "haiku" 371 } 372 } 373 } 374 }, 375 "claims": [ 376 { 377 "claim": "Burstiness in training data is necessary for in-context learning to emerge in transformers", 378 "evidence": "Fig 2 shows monotonic improvement in ICL accuracy as P(bursty) increases from 0 to 1.0, with P(bursty)=0 yielding chance performance; replicated across 5 seeds", 379 "supported": "strong" 380 }, 381 { 382 "claim": "A large number of rarely occurring training classes is required for in-context learning", 383 "evidence": "Fig 3 shows ICL accuracy increases monotonically as classes increase from 100 to 1600 to 12800 with P(bursty)=0.9 held fixed; effect holds after controlling for number of exposures", 384 "supported": "strong" 385 }, 386 { 387 "claim": "Dynamic meanings (label multiplicity and within-class variation) increase in-context learning", 388 "evidence": "Figs 4 and 5 show monotonic ICL improvement with increasing label multiplicity (1→10) and within-class variation (no noise → full Omniglot exemplars)", 389 "supported": "strong" 390 }, 391 { 392 "claim": "There is a tradeoff between in-context and in-weights learning under uniform marginal class distributions", 393 "evidence": "Consistently observed across all experiments in Figs 2–5: any manipulation increasing ICL simultaneously decreases in-weights learning accuracy", 394 "supported": "strong" 395 }, 396 { 397 "claim": "A Zipfian (power-law) marginal distribution over classes allows in-context and in-weights learning to coexist", 398 "evidence": "Fig 6 shows Zipf exponent=1 achieves simultaneously high ICL accuracy on holdout classes and high in-weights accuracy on common classes, with a sweet spot coinciding with natural language statistics", 399 "supported": "strong" 400 }, 401 { 402 "claim": "Transformers uniquely support in-context learning; matched recurrent models (LSTM, RNN) cannot acquire it", 403 "evidence": "Fig 7 shows transformer achieving near-perfect ICL while matched LSTM and RNN remain at chance across 90 hyperparameter sweep runs; transformers also match or exceed recurrent models on in-weights learning", 404 "supported": "strong" 405 }, 406 { 407 "claim": "These distributional properties explain why large language models exhibit emergent in-context learning", 408 "evidence": "Analogy between natural language distributional properties (burstiness, Zipfian skew, polysemy) and the experimental factors is drawn, but no direct experiments on large language models are conducted", 409 "supported": "weak" 410 } 411 ], 412 "methodology_tags": [ 413 "rct" 414 ], 415 "key_findings": "In-context learning emerges in transformers when training data exhibits both burstiness and a large vocabulary of rarely occurring classes—properties naturally present in language but absent from standard supervised datasets. A consistent tradeoff exists between in-context and in-weights learning under uniform class distributions, but a Zipfian (power-law) marginal distribution with exponent ~1 (matching natural language) resolves this by enabling both simultaneously. Recurrent architectures (LSTM, RNN) matched on parameters and depth completely fail to acquire in-context learning under any tested naturalistic data distribution, confirming that both data distributions and the transformer architecture are jointly necessary—attention is not all you need.", 416 "red_flags": [ 417 { 418 "flag": "Generalization leap to LLMs", 419 "detail": "All experiments use small transformers (12-layer, embedding dim 64) trained on Omniglot image classification, but the abstract and discussion claim to explain in-context learning in 'large language models.' No LLM experiments are conducted, and the scale difference is orders of magnitude. The mechanistic analogy is plausible but unverified." 420 }, 421 { 422 "flag": "No statistical significance testing", 423 "detail": "Results are presented as learning curves with standard deviation bands, but no formal hypothesis tests (t-tests, ANOVA, permutation tests) are reported anywhere in the paper, making it unclear whether observed condition differences are statistically reliable." 424 }, 425 { 426 "flag": "No limitations section", 427 "detail": "The paper contains no dedicated limitations or threats-to-validity section. The discussion only covers implications and future directions, without acknowledging potential confounds such as the use of Omniglot specifically, the small model scale, or the particular burstiness operationalization chosen." 428 }, 429 { 430 "flag": "Funder-author conflict", 431 "detail": "Seven of eight authors are DeepMind employees and the work is DeepMind-funded; findings favor the transformer architecture that DeepMind develops and deploys. No competing interests statement is present." 432 } 433 ], 434 "cited_papers": [ 435 { 436 "title": "Language Models are Few-Shot Learners", 437 "relevance": "Foundational GPT-3 paper demonstrating emergent in-context learning; the motivating observation driving this paper's research question" 438 }, 439 { 440 "title": "Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?", 441 "relevance": "Competing theory suggesting LLMs may not perform genuine ICL; paper directly counters this with holdout class experiments showing true generalization" 442 }, 443 { 444 "title": "An Explanation of In-context Learning as Implicit Bayesian Inference", 445 "relevance": "Alternative theoretical framework for ICL that the paper situates itself relative to" 446 }, 447 { 448 "title": "Impact of Pretraining Term Frequencies on Few-Shot Reasoning", 449 "relevance": "Suggests LLM few-shot performance may be driven by memorization; paper's holdout class design is a direct methodological response" 450 }, 451 { 452 "title": "Meta-Learning with Memory-Augmented Neural Networks", 453 "relevance": "Prior explicit meta-training approach for few-shot learning; contrasts with emergent ICL studied here, establishing what 'designed' few-shot learning looks like" 454 }, 455 { 456 "title": "Matching Networks for One Shot Learning", 457 "relevance": "Canonical few-shot meta-learning baseline; establishes the explicit-training approach that the paper shows is unnecessary given correct data distributions" 458 }, 459 { 460 "title": "Zipfian environments for Reinforcement Learning", 461 "relevance": "Prior work by the same first author on non-uniform distributions in RL; establishes prior context for the Zipfian analysis and motivates the non-language domain implications" 462 }, 463 { 464 "title": "Can Wikipedia Help Offline Reinforcement Learning?", 465 "relevance": "Evidence that language pre-training transfers to RL but vision pre-training does not; paper uses this asymmetry to motivate its distributional properties hypothesis" 466 } 467 ], 468 "engagement_factors": { 469 "practical_relevance": { 470 "score": 2, 471 "justification": "Provides actionable principles for designing pre-training datasets to elicit ICL in non-language domains, though requires significant ML infrastructure to apply" 472 }, 473 "surprise_contrarian": { 474 "score": 3, 475 "justification": "Directly challenges 'attention is all you need' narrative by demonstrating data distribution is equally critical; the counterintuitive finding that harder within-class generalization promotes ICL is particularly striking" 476 }, 477 "fear_safety": { 478 "score": 0, 479 "justification": "Purely mechanistic understanding of a capability; no AI safety or risk implications are raised" 480 }, 481 "drama_conflict": { 482 "score": 1, 483 "justification": "Mild tension with competing theories (Min et al., Razeghi et al.) claiming LLMs may not genuinely perform ICL; paper presents experimental counter-evidence" 484 }, 485 "demo_ability": { 486 "score": 1, 487 "justification": "Code is released but experiments require TPU infrastructure; not easily reproducible by practitioners without significant compute resources" 488 }, 489 "brand_recognition": { 490 "score": 2, 491 "justification": "DeepMind affiliation and NeurIPS venue; several authors (Santoro, Lampinen, Wang, Hill) are prominent researchers in meta-learning and language model interpretability" 492 } 493 }, 494 "hn_data": { 495 "threads": [ 496 { 497 "hn_id": "30533914", 498 "title": "DeepNet: Scaling Transformers to 1k Layers", 499 "points": 194, 500 "comments": 38, 501 "url": "https://news.ycombinator.com/item?id=30533914", 502 "created_at": "2022-03-02T22:10:11Z" 503 }, 504 { 505 "hn_id": "32198181", 506 "title": "Design and Implementation of a Secure RISC-V Microprocessor", 507 "points": 4, 508 "comments": 0, 509 "url": "https://news.ycombinator.com/item?id=32198181", 510 "created_at": "2022-07-22T23:06:00Z" 511 }, 512 { 513 "hn_id": "31352535", 514 "title": "Design and Implementation of a Secure RISC-V Microprocessor", 515 "points": 3, 516 "comments": 0, 517 "url": "https://news.ycombinator.com/item?id=31352535", 518 "created_at": "2022-05-12T11:43:34Z" 519 }, 520 { 521 "hn_id": "31635842", 522 "title": "Finding the optimal human strategy for Wordle", 523 "points": 3, 524 "comments": 0, 525 "url": "https://news.ycombinator.com/item?id=31635842", 526 "created_at": "2022-06-05T23:36:34Z" 527 }, 528 { 529 "hn_id": "31644563", 530 "title": "Bursty symbols in training allow prompting (ML)", 531 "points": 2, 532 "comments": 0, 533 "url": "https://news.ycombinator.com/item?id=31644563", 534 "created_at": "2022-06-06T19:10:02Z" 535 }, 536 { 537 "hn_id": "31441958", 538 "title": "Data Distributional Properties Drive Emergent Few-Shot Learning in Transformers", 539 "points": 2, 540 "comments": 0, 541 "url": "https://news.ycombinator.com/item?id=31441958", 542 "created_at": "2022-05-20T00:25:42Z" 543 }, 544 { 545 "hn_id": "34205208", 546 "title": "Crypto trading using reinforcement learning", 547 "points": 2, 548 "comments": 3, 549 "url": "https://news.ycombinator.com/item?id=34205208", 550 "created_at": "2023-01-01T10:23:29Z" 551 }, 552 { 553 "hn_id": "44064952", 554 "title": "Hierarchical-Chain-of-Generation for Complex Attributes Text-to-3D Generation", 555 "points": 2, 556 "comments": 0, 557 "url": "https://news.ycombinator.com/item?id=44064952", 558 "created_at": "2025-05-22T18:12:20Z" 559 }, 560 { 561 "hn_id": "33223016", 562 "title": "Cotton Gravity: a potential alternative to the dark matter paradigm", 563 "points": 2, 564 "comments": 0, 565 "url": "https://news.ycombinator.com/item?id=33223016", 566 "created_at": "2022-10-16T12:18:43Z" 567 }, 568 { 569 "hn_id": "31633209", 570 "title": "Data Distributional Properties Drive Emergent In-Context Learning in Transformer", 571 "points": 1, 572 "comments": 0, 573 "url": "https://news.ycombinator.com/item?id=31633209", 574 "created_at": "2022-06-05T17:46:34Z" 575 } 576 ], 577 "top_points": 194, 578 "total_points": 215, 579 "total_comments": 41 580 } 581 }