scan-v4.json (30310B)
1 { 2 "scan_version": 4, 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": "Abstract claims about burstiness, rare classes, dynamic meanings, Zipfian sweet spot, and transformer-specificity are all supported by corresponding experiments (Figs 2-7).", 27 "source": "opus" 28 }, 29 "causal_claims_justified": { 30 "applies": true, 31 "answer": true, 32 "justification": "Causal claims ('drive', 'promote') are supported by controlled experiments that manipulate single variables (burstiness, class count, etc.) while holding others constant. This is adequate causal design.", 33 "source": "opus" 34 }, 35 "generalization_bounded": { 36 "applies": true, 37 "answer": false, 38 "justification": "Title claims 'in Transformers' broadly but experiments use only Omniglot image classification with small transformers (831K params). Discussion section speculates about implications for LLMs and cognition without bounding these generalizations.", 39 "source": "opus" 40 }, 41 "alternative_explanations_discussed": { 42 "applies": true, 43 "answer": true, 44 "justification": "Discussion considers whether recurrent models simply have a bias towards in-weights learning (ruled out by Fig 8). Addresses the narrative that LLMs just memorize training data (Section 4). Discusses the ambiguity between in-context and in-weights strategies.", 45 "source": "opus" 46 }, 47 "proxy_outcome_distinction": { 48 "applies": true, 49 "answer": true, 50 "justification": "The paper measures in-context learning accuracy on Omniglot classification tasks with controlled distributional properties (burstiness, number of classes, Zipfian distribution). Claims match the measurement granularity: 'in-context learning emerges when training data exhibits particular distributional properties.' The paper does not overclaim to broader 'intelligence' or 'reasoning' — it stays within the measured construct.", 51 "source": "opus" 52 } 53 }, 54 "limitations_and_scope": { 55 "limitations_section_present": { 56 "applies": true, 57 "answer": false, 58 "justification": "No dedicated limitations or threats-to-validity section. The Discussion mentions future directions but does not systematically discuss limitations.", 59 "source": "opus" 60 }, 61 "threats_to_validity_specific": { 62 "applies": true, 63 "answer": false, 64 "justification": "No specific threats to validity discussed. The paper does not address whether Omniglot results generalize to language, or limitations of the small model scale.", 65 "source": "opus" 66 }, 67 "scope_boundaries_stated": { 68 "applies": true, 69 "answer": false, 70 "justification": "No explicit statements about what the results do NOT show. Discussion section speculates broadly about implications for language models, cognition, and neuroscience without bounding the scope.", 71 "source": "opus" 72 } 73 }, 74 "conflicts_of_interest": { 75 "funding_disclosed": { 76 "applies": true, 77 "answer": true, 78 "justification": "Acknowledgments state: 'This work was funded by DeepMind.'", 79 "source": "opus" 80 }, 81 "affiliations_disclosed": { 82 "applies": true, 83 "answer": true, 84 "justification": "All author affiliations listed: 7 of 8 authors are at DeepMind, 1 at UCL, 1 jointly at DeepMind and Stanford.", 85 "source": "opus" 86 }, 87 "funder_independent_of_outcome": { 88 "applies": true, 89 "answer": true, 90 "justification": "DeepMind funded the work. The paper studies basic science of in-context learning mechanisms; DeepMind does not have a direct financial stake in whether burstiness or Zipfian distributions drive in-context learning.", 91 "source": "opus" 92 }, 93 "financial_interests_declared": { 94 "applies": true, 95 "answer": false, 96 "justification": "No competing interests or financial interests statement found in the paper.", 97 "source": "opus" 98 } 99 }, 100 "scope_and_framing": { 101 "key_terms_defined": { 102 "applies": true, 103 "answer": true, 104 "justification": "Key terms are clearly defined: in-context learning (generalization from context without weight updates), in-weights learning (gradient-based slow learning), burstiness (items appearing in clusters), and Zipfian distribution (with explicit formula in Eq 1).", 105 "source": "haiku" 106 }, 107 "intended_contribution_clear": { 108 "applies": true, 109 "answer": true, 110 "justification": "The paper clearly states its contribution is identifying specific training data distributional properties (burstiness, rare classes, dynamic meanings, Zipfian distribution) that drive emergent in-context learning in transformer models.", 111 "source": "haiku" 112 }, 113 "engagement_with_prior_work": { 114 "applies": true, 115 "answer": true, 116 "justification": "The paper explicitly positions against narratives from Min et al. 2022 and Razeghi et al. 2022 suggesting ICL may not be genuine, and builds on meta-learning literature (Santoro et al. 2016, Vinyals et al. 2016) and the GPT-3 observation (Brown et al. 2020).", 117 "source": "haiku" 118 } 119 } 120 }, 121 "type_checklist": { 122 "empirical": { 123 "artifacts": { 124 "code_released": { 125 "applies": true, 126 "answer": true, 127 "justification": "Code released at https://github.com/deepmind/emergent_in_context_learning, stated in footnote 1.", 128 "source": "opus" 129 }, 130 "data_released": { 131 "applies": true, 132 "answer": true, 133 "justification": "Uses the publicly available Omniglot dataset (Lake et al., 2019, MIT License).", 134 "source": "opus" 135 }, 136 "environment_specified": { 137 "applies": true, 138 "answer": false, 139 "justification": "No requirements.txt, Dockerfile, or detailed environment specification provided. Hardware mentioned (TPU v2/v3) but no software environment details.", 140 "source": "opus" 141 }, 142 "reproduction_instructions": { 143 "applies": true, 144 "answer": false, 145 "justification": "Paper states 'code will be released with the camera-ready version' and provides architectural details in text, but no step-by-step reproduction instructions are included in the paper itself.", 146 "source": "opus" 147 } 148 }, 149 "statistical_methodology": { 150 "confidence_intervals_or_error_bars": { 151 "applies": true, 152 "answer": true, 153 "justification": "Appendix A states 'error bars indicate standard deviation around the mean' and shaded regions are shown in all figures.", 154 "source": "opus" 155 }, 156 "significance_tests": { 157 "applies": true, 158 "answer": false, 159 "justification": "No statistical significance tests are reported. Claims of difference between conditions (e.g., burstiness levels, architectures) are based on visual comparison of curves without formal tests.", 160 "source": "opus" 161 }, 162 "effect_sizes_reported": { 163 "applies": true, 164 "answer": false, 165 "justification": "Results are presented as accuracy curves and bar plots but no formal effect sizes (Cohen's d, etc.) are reported. Accuracy values are shown but without baseline context framing as effect sizes.", 166 "source": "opus" 167 }, 168 "sample_size_justified": { 169 "applies": true, 170 "answer": false, 171 "justification": "Number of seeds (3 or 5 depending on experiment) is stated but not justified. No power analysis or discussion of whether this is sufficient.", 172 "source": "opus" 173 }, 174 "variance_reported": { 175 "applies": true, 176 "answer": true, 177 "justification": "Standard deviation across runs shown as shaded error bars in all figures. Appendix A: 'experiments were run with 3 seeds each... all other experiments were run with 5 runs each.'", 178 "source": "opus" 179 } 180 }, 181 "evaluation_design": { 182 "baselines_included": { 183 "applies": true, 184 "answer": true, 185 "justification": "Multiple baseline conditions included: non-bursty training, different numbers of classes, uniform vs Zipfian distributions, and RNN/LSTM architectures as comparison points.", 186 "source": "opus" 187 }, 188 "baselines_contemporary": { 189 "applies": true, 190 "answer": true, 191 "justification": "Comparisons are against standard architectures (transformer, LSTM, vanilla RNN) matched on parameters. The paper's contribution is mechanistic understanding, not outperformance, so these are appropriate.", 192 "source": "opus" 193 }, 194 "ablation_study": { 195 "applies": true, 196 "answer": true, 197 "justification": "The entire paper is essentially a series of ablations: varying burstiness (Fig 2), number of classes (Fig 3), label multiplicity (Fig 4), within-class variation (Fig 5), and Zipf exponent (Fig 6).", 198 "source": "opus" 199 }, 200 "multiple_metrics": { 201 "applies": true, 202 "answer": true, 203 "justification": "Two complementary metrics: in-context learning accuracy on holdout classes and in-weights learning accuracy on trained classes. Also multi-class vs two-way evaluation (Appendix C.4).", 204 "source": "opus" 205 }, 206 "human_evaluation": { 207 "applies": false, 208 "answer": false, 209 "justification": "This is a controlled experiment on synthetic training regimes for transformers; human evaluation is not relevant.", 210 "source": "opus" 211 }, 212 "held_out_test_set": { 213 "applies": true, 214 "answer": true, 215 "justification": "In-context learning evaluated on holdout image classes never seen in training (Section 2.3). Training/holdout split explicitly described.", 216 "source": "opus" 217 }, 218 "per_category_breakdown": { 219 "applies": true, 220 "answer": true, 221 "justification": "Results broken down by condition (burstiness level, number of classes, Zipf exponent, architecture type) across multiple figures. Zipfian experiments separate common vs rare class performance (Fig 6d-e).", 222 "source": "opus" 223 }, 224 "failure_cases_discussed": { 225 "applies": true, 226 "answer": true, 227 "justification": "Discusses cases where in-context learning fails: non-bursty training, too few classes, extreme Zipf exponents. Also discusses RNN/LSTM failure to achieve in-context learning.", 228 "source": "opus" 229 }, 230 "negative_results_reported": { 231 "applies": true, 232 "answer": true, 233 "justification": "Reports tradeoff between in-context and in-weights learning (most conditions cannot achieve both). RNN/LSTM completely fail at in-context learning. High Zipf exponents destroy in-context learning.", 234 "source": "opus" 235 } 236 }, 237 "setup_transparency": { 238 "model_versions_specified": { 239 "applies": false, 240 "answer": false, 241 "justification": "Paper trains its own models from scratch; no pre-trained model versions to specify.", 242 "source": "opus" 243 }, 244 "prompts_provided": { 245 "applies": false, 246 "answer": false, 247 "justification": "No prompting used. Models are trained from scratch on image-label sequences.", 248 "source": "opus" 249 }, 250 "hyperparameters_reported": { 251 "applies": true, 252 "answer": true, 253 "justification": "Appendix A details: 12 layers, embedding dim 64, 8 heads, ResNet architecture, Adam optimizer, learning rate 3e-4 with warmup to 4000 steps, 500k training steps. Hyperparameter sweep details in Appendix C.1.", 254 "source": "opus" 255 }, 256 "scaffolding_described": { 257 "applies": false, 258 "answer": false, 259 "justification": "No agentic scaffolding used.", 260 "source": "opus" 261 }, 262 "data_preprocessing_documented": { 263 "applies": true, 264 "answer": true, 265 "justification": "Section 2.1 describes sequence construction in detail: context structure (8 image-label pairs), bursty vs non-bursty sequence construction, how rotated/flipped images were generated for 12800 classes.", 266 "source": "opus" 267 } 268 }, 269 "data_integrity": { 270 "raw_data_available": { 271 "applies": true, 272 "answer": true, 273 "justification": "Omniglot dataset is publicly available. Code for generating training sequences is released.", 274 "source": "opus" 275 }, 276 "data_collection_described": { 277 "applies": true, 278 "answer": true, 279 "justification": "Section 2.1 describes how training sequences are constructed from Omniglot: image-label pair sequences, bursty/non-bursty composition, class sampling procedures.", 280 "source": "opus" 281 }, 282 "recruitment_methods_described": { 283 "applies": false, 284 "answer": false, 285 "justification": "No human participants. Data is a standard public benchmark (Omniglot).", 286 "source": "opus" 287 }, 288 "data_pipeline_documented": { 289 "applies": true, 290 "answer": true, 291 "justification": "Full pipeline from Omniglot images to training sequences documented: embedding (ResNet for images, standard embedding for labels), sequence construction, evaluation protocol.", 292 "source": "opus" 293 } 294 }, 295 "contamination": { 296 "training_cutoff_stated": { 297 "applies": false, 298 "answer": false, 299 "justification": "Models are trained from scratch on controlled synthetic data distributions; no pre-trained model evaluated on a benchmark.", 300 "source": "opus" 301 }, 302 "train_test_overlap_discussed": { 303 "applies": false, 304 "answer": false, 305 "justification": "Models trained from scratch on Omniglot subsets with explicit train/holdout splits. Not evaluating a pre-trained model's knowledge.", 306 "source": "opus" 307 }, 308 "benchmark_contamination_addressed": { 309 "applies": false, 310 "answer": false, 311 "justification": "No pre-trained model evaluated on an external benchmark. Contamination is structurally impossible in this experimental design.", 312 "source": "opus" 313 } 314 }, 315 "human_studies": { 316 "pre_registered": { 317 "applies": false, 318 "answer": false, 319 "justification": "No human participants.", 320 "source": "opus" 321 }, 322 "irb_or_ethics_approval": { 323 "applies": false, 324 "answer": false, 325 "justification": "No human participants.", 326 "source": "opus" 327 }, 328 "demographics_reported": { 329 "applies": false, 330 "answer": false, 331 "justification": "No human participants.", 332 "source": "opus" 333 }, 334 "inclusion_exclusion_criteria": { 335 "applies": false, 336 "answer": false, 337 "justification": "No human participants.", 338 "source": "opus" 339 }, 340 "randomization_described": { 341 "applies": false, 342 "answer": false, 343 "justification": "No human participants.", 344 "source": "opus" 345 }, 346 "blinding_described": { 347 "applies": false, 348 "answer": false, 349 "justification": "No human participants.", 350 "source": "opus" 351 }, 352 "attrition_reported": { 353 "applies": false, 354 "answer": false, 355 "justification": "No human participants.", 356 "source": "opus" 357 } 358 }, 359 "cost_and_practicality": { 360 "inference_cost_reported": { 361 "applies": false, 362 "answer": false, 363 "justification": "This is a basic science paper studying emergent properties, not proposing a practical method with inference costs.", 364 "source": "opus" 365 }, 366 "compute_budget_stated": { 367 "applies": true, 368 "answer": true, 369 "justification": "Appendix A: '500k training steps on 16 TPU v2 or v3 cores.' Appendix C.1: 90 runs total for architecture comparison.", 370 "source": "opus" 371 } 372 }, 373 "experimental_rigor": { 374 "seed_sensitivity_reported": { 375 "applies": true, 376 "answer": true, 377 "justification": "Appendix A: experiments run with 3 or 5 seeds. Error bars (std dev) shown across seeds in all figures.", 378 "source": "opus" 379 }, 380 "number_of_runs_stated": { 381 "applies": true, 382 "answer": true, 383 "justification": "Appendix A: 'experiments shown in Figs 5 and 6 were run with 3 seeds each... all other experiments were run with 5 runs each.' Architecture comparison: 15 runs per architecture (90 total).", 384 "source": "opus" 385 }, 386 "hyperparameter_search_budget": { 387 "applies": true, 388 "answer": true, 389 "justification": "Appendix C.1: hyperparameter sweep over 15 samples of learning rate (log-uniform [1e-5, 0.1]) and 15 samples of warmup steps (log-uniform [1, 10000]), 15 runs per architecture.", 390 "source": "opus" 391 }, 392 "best_config_selection_justified": { 393 "applies": true, 394 "answer": false, 395 "justification": "For the architecture comparison, all hyperparameter sweep runs are shown (each color = one run), which is transparent. However, for main experiments the fixed hyperparameters are not justified as optimal.", 396 "source": "opus" 397 }, 398 "multiple_comparison_correction": { 399 "applies": false, 400 "answer": false, 401 "justification": "No statistical significance tests performed, so multiple comparison correction is not applicable.", 402 "source": "opus" 403 }, 404 "self_comparison_bias_addressed": { 405 "applies": false, 406 "answer": false, 407 "justification": "The paper does not compare against other researchers' systems. It trains and evaluates its own models under different conditions, so self-comparison bias in the Lucic et al. sense does not apply.", 408 "source": "opus" 409 }, 410 "compute_budget_vs_performance": { 411 "applies": true, 412 "answer": true, 413 "justification": "Architecture comparisons match transformer, RNN, and LSTM on number of parameters, depth, and hidden size (Appendix C.1). Training steps are equalized across conditions.", 414 "source": "opus" 415 }, 416 "benchmark_construct_validity": { 417 "applies": true, 418 "answer": true, 419 "justification": "The paper carefully defines what in-context learning means (holdout classes with re-assigned labels) vs in-weights learning (trained classes without context support), providing clear construct validity for their evaluation measures.", 420 "source": "opus" 421 }, 422 "scaffold_confound_addressed": { 423 "applies": false, 424 "answer": false, 425 "justification": "The paper trains models from scratch and evaluates them directly. No scaffolding or multi-model comparison through different tool frameworks is involved. The architectural comparison (transformer vs RNN) is a controlled experiment, not a scaffolding confound.", 426 "source": "opus" 427 } 428 } 429 } 430 }, 431 "claims": [ 432 { 433 "claim": "Burstiness in training data promotes in-context learning while trading off against in-weights learning", 434 "evidence": "Figure 2 shows increasing P(bursty) from 0 to 1.0 systematically improves in-context accuracy on holdout classes while decreasing in-weights accuracy on trained classes.", 435 "supported": "strong" 436 }, 437 { 438 "claim": "In-context learning requires a large number of rarely occurring classes, not just burstiness alone", 439 "evidence": "Figure 3 shows scaling training classes from 100 to 1600 to 12800 progressively improves in-context learning while reducing in-weights learning, holding burstiness fixed at p=0.9.", 440 "supported": "strong" 441 }, 442 { 443 "claim": "Dynamic meanings (label multiplicity and within-class variation) increase in-context learning", 444 "evidence": "Figure 4 shows increasing label multiplicity from 1 to 10 improves ICL; Figure 5 shows increasing within-class variation (pixel noise, full Omniglot) also systematically improves ICL.", 445 "supported": "strong" 446 }, 447 { 448 "claim": "Zipfian distribution over classes with exponent ~1 enables co-existence of in-context and in-weights learning in a single model", 449 "evidence": "Figure 6 shows a sweet spot at Zipf exponent 1 where both ICL accuracy and in-weights accuracy on common classes are maintained at high levels simultaneously, coinciding with natural language Zipf exponent.", 450 "supported": "strong" 451 }, 452 { 453 "claim": "Transformers can exhibit in-context learning with naturalistic data distributions but recurrent models (RNN, LSTM) cannot", 454 "evidence": "Figure 7 shows vanilla RNN and LSTM never achieve above-chance in-context learning despite identical training setup and matched parameter counts; transformers also outperform on in-weights learning (Fig 8).", 455 "supported": "strong" 456 }, 457 { 458 "claim": "Transformer architecture alone is insufficient for in-context learning—appropriate data distributions are also required", 459 "evidence": "Transformers trained with uniform, non-bursty distributions (P(bursty)=0) in Figure 2 fail to exhibit in-context learning, demonstrating the necessity of both components.", 460 "supported": "strong" 461 } 462 ], 463 "methodology_tags": [ 464 "benchmark-eval", 465 "observational" 466 ], 467 "key_findings": "Emergent in-context learning in transformers is driven by naturalistic distributional properties of training data: burstiness, large numbers of rarely-occurring classes, and dynamic meanings all independently promote ICL. A Zipfian distribution over classes (resembling natural language at exponent ~1) is the key to enabling both in-context and in-weights learning simultaneously in a single model, resolving the tradeoff observed in simpler uniform distributions. Critically, RNNs and LSTMs cannot exhibit in-context learning even with identical naturalistic training distributions, confirming that both transformer architecture AND appropriate data distributions are jointly necessary — 'attention is not all you need.'", 468 "red_flags": [ 469 { 470 "flag": "Omniglot-to-LLM generalization gap", 471 "detail": "All experiments use small transformers (~831K parameters) on Omniglot image classification; the claimed explanation for why large language models exhibit ICL is extrapolated without any direct validation on LLMs or language data." 472 }, 473 { 474 "flag": "No statistical significance testing", 475 "detail": "Results are presented graphically with error bars but no formal statistical tests are reported, making it difficult to assess reliability of between-condition differences." 476 }, 477 { 478 "flag": "Artificial burstiness operationalization", 479 "detail": "Burstiness is operationalized as fixed in-context repetition (query class appears exactly 3 times in context), which may not fully capture the temporal burstiness observed in natural language corpora." 480 }, 481 { 482 "flag": "Confound in 12800-class condition", 483 "detail": "Authors acknowledge that the 12800-class experiment uses image rotations/flips, introducing a potential label-multiplicity confound that may partially explain the strong ICL effect beyond the class-number manipulation." 484 } 485 ], 486 "cited_papers": [ 487 { 488 "title": "Language Models are Few-Shot Learners", 489 "relevance": "Foundational motivation: GPT-3's emergent in-context learning is the phenomenon this paper seeks to explain mechanistically." 490 }, 491 { 492 "title": "An Explanation of In-context Learning as Implicit Bayesian Inference", 493 "relevance": "Competing theoretical explanation of ICL that the paper's empirical results complement and partially test." 494 }, 495 { 496 "title": "Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?", 497 "relevance": "Directly challenged by this paper's holdout-class evaluation design, which counters claims that ICL may not be genuine learning." 498 }, 499 { 500 "title": "Meta-Learning with Memory-Augmented Neural Networks", 501 "relevance": "Key prior work on explicit meta-training for few-shot learning that this paper contrasts with the emergent ICL setting." 502 }, 503 { 504 "title": "Attention is All You Need", 505 "relevance": "The transformer architecture central to all experiments; this paper's conclusions directly qualify that landmark claim." 506 }, 507 { 508 "title": "The Omniglot Challenge: A 3-Year Progress Report", 509 "relevance": "The benchmark dataset used for all experiments; standard few-shot evaluation procedures are taken from this work." 510 }, 511 { 512 "title": "Impact of Pretraining Term Frequencies on Few-Shot Reasoning", 513 "relevance": "Concurrent work suggesting ICL may rely on training frequency rather than genuine in-context generalization, directly countered by this paper." 514 } 515 ], 516 "engagement_factors": { 517 "practical_relevance": { 518 "score": 1, 519 "justification": "Offers theoretical insight into why in-context learning works that could inform dataset curation, but no immediately usable tool or technique." 520 }, 521 "surprise_contrarian": { 522 "score": 2, 523 "justification": "The finding that data distribution (not just scale or architecture) drives in-context learning, and that Zipf exponent ~1 is a sweet spot, challenges the 'just scale up' narrative." 524 }, 525 "fear_safety": { 526 "score": 0, 527 "justification": "No safety, security, or risk angle discussed." 528 }, 529 "drama_conflict": { 530 "score": 0, 531 "justification": "No controversy, company rivalry, or challenge to specific claims — a constructive mechanistic study." 532 }, 533 "demo_ability": { 534 "score": 1, 535 "justification": "Code released on GitHub but requires TPU training runs on Omniglot, not a quick demo." 536 }, 537 "brand_recognition": { 538 "score": 2, 539 "justification": "From DeepMind with a Stanford co-author (McClelland), published at NeurIPS 2022." 540 } 541 }, 542 "hn_data": { 543 "threads": [ 544 { 545 "hn_id": "30533914", 546 "title": "DeepNet: Scaling Transformers to 1k Layers", 547 "points": 194, 548 "comments": 38, 549 "url": "https://news.ycombinator.com/item?id=30533914", 550 "created_at": "2022-03-02T22:10:11Z" 551 }, 552 { 553 "hn_id": "32198181", 554 "title": "Design and Implementation of a Secure RISC-V Microprocessor", 555 "points": 4, 556 "comments": 0, 557 "url": "https://news.ycombinator.com/item?id=32198181", 558 "created_at": "2022-07-22T23:06:00Z" 559 }, 560 { 561 "hn_id": "31352535", 562 "title": "Design and Implementation of a Secure RISC-V Microprocessor", 563 "points": 3, 564 "comments": 0, 565 "url": "https://news.ycombinator.com/item?id=31352535", 566 "created_at": "2022-05-12T11:43:34Z" 567 }, 568 { 569 "hn_id": "31635842", 570 "title": "Finding the optimal human strategy for Wordle", 571 "points": 3, 572 "comments": 0, 573 "url": "https://news.ycombinator.com/item?id=31635842", 574 "created_at": "2022-06-05T23:36:34Z" 575 }, 576 { 577 "hn_id": "31644563", 578 "title": "Bursty symbols in training allow prompting (ML)", 579 "points": 2, 580 "comments": 0, 581 "url": "https://news.ycombinator.com/item?id=31644563", 582 "created_at": "2022-06-06T19:10:02Z" 583 }, 584 { 585 "hn_id": "31441958", 586 "title": "Data Distributional Properties Drive Emergent Few-Shot Learning in Transformers", 587 "points": 2, 588 "comments": 0, 589 "url": "https://news.ycombinator.com/item?id=31441958", 590 "created_at": "2022-05-20T00:25:42Z" 591 }, 592 { 593 "hn_id": "34205208", 594 "title": "Crypto trading using reinforcement learning", 595 "points": 2, 596 "comments": 3, 597 "url": "https://news.ycombinator.com/item?id=34205208", 598 "created_at": "2023-01-01T10:23:29Z" 599 }, 600 { 601 "hn_id": "44064952", 602 "title": "Hierarchical-Chain-of-Generation for Complex Attributes Text-to-3D Generation", 603 "points": 2, 604 "comments": 0, 605 "url": "https://news.ycombinator.com/item?id=44064952", 606 "created_at": "2025-05-22T18:12:20Z" 607 }, 608 { 609 "hn_id": "33223016", 610 "title": "Cotton Gravity: a potential alternative to the dark matter paradigm", 611 "points": 2, 612 "comments": 0, 613 "url": "https://news.ycombinator.com/item?id=33223016", 614 "created_at": "2022-10-16T12:18:43Z" 615 }, 616 { 617 "hn_id": "31633209", 618 "title": "Data Distributional Properties Drive Emergent In-Context Learning in Transformer", 619 "points": 1, 620 "comments": 0, 621 "url": "https://news.ycombinator.com/item?id=31633209", 622 "created_at": "2022-06-05T17:46:34Z" 623 } 624 ], 625 "top_points": 194, 626 "total_points": 215, 627 "total_comments": 41 628 } 629 }