scan-v5.json (27030B)
1 { 2 "scan_version": 5, 3 "paper_type": "empirical", 4 "paper": { 5 "title": "An Investigation on Group Query Hallucination Attacks", 6 "authors": [ 7 "Kehao Miao", 8 "Xiaolong Jin" 9 ], 10 "year": 2025, 11 "venue": "arXiv.org", 12 "arxiv_id": "2508.19321", 13 "doi": "10.48550/arXiv.2508.19321" 14 }, 15 "checklist": { 16 "claims_and_evidence": { 17 "abstract_claims_supported": { 18 "applies": true, 19 "answer": true, 20 "justification": "All three abstract claims are supported: fine-tuned model degradation (Table 1), backdoor triggering (Table 2 showing option 'A' dominance), and reasoning task effectiveness (Table 3 showing severe code/math drops).", 21 "source": "haiku" 22 }, 23 "causal_claims_justified": { 24 "applies": true, 25 "answer": false, 26 "justification": "Paper claims GQA 'causes' degradation but provides only observational evidence. No ablation studies isolate causality; no manipulation of individual GQA components to rule out confounds like context length alone.", 27 "source": "haiku" 28 }, 29 "generalization_bounded": { 30 "applies": true, 31 "answer": true, 32 "justification": "Results are bounded to specific model sizes (7-8 models tested), task types (MCQ, translation, code, math), and QGS ranges (1-30). Scope is mostly explicit about these constraints.", 33 "source": "haiku" 34 }, 35 "alternative_explanations_discussed": { 36 "applies": true, 37 "answer": false, 38 "justification": "Paper speculates 'cumulative effect' for reasoning tasks but doesn't test alternative hypotheses. For Q2 (backdoors), doesn't distinguish triggering actual backdoors from mode collapse or statistical artifacts.", 39 "source": "haiku" 40 }, 41 "proxy_outcome_distinction": { 42 "applies": true, 43 "answer": true, 44 "justification": "Accuracy directly measures correct answers; sacreBLEU directly measures translation quality. Measurements match claims without conflating different constructs.", 45 "source": "haiku" 46 } 47 }, 48 "limitations_and_scope": { 49 "limitations_section_present": { 50 "applies": true, 51 "answer": true, 52 "justification": "Section 6 'Limitations' explicitly lists three specific constraints: limited scenario coverage, only first-query responses analyzed, and model coverage limited by time.", 53 "source": "haiku" 54 }, 55 "threats_to_validity_specific": { 56 "applies": true, 57 "answer": true, 58 "justification": "Specific threats stated: gap between multi-choice task focus vs. open-ended user queries, analyzing only first response rather than all responses, and incomplete model sampling due to time.", 59 "source": "haiku" 60 }, 61 "scope_boundaries_stated": { 62 "applies": true, 63 "answer": true, 64 "justification": "Paper bounds results to fine-tuned models, specific benchmarks (MedMCQA, PubMedQA, Aqua-RAT, MathQA, HumanEval, WMT20), QGS ≤30, and 7-8 model architectures tested.", 65 "source": "haiku" 66 } 67 }, 68 "conflicts_of_interest": { 69 "funding_disclosed": { 70 "applies": true, 71 "answer": false, 72 "justification": "No funding source disclosed in paper. No acknowledgments section mentioning grants or support.", 73 "source": "haiku" 74 }, 75 "affiliations_disclosed": { 76 "applies": true, 77 "answer": true, 78 "justification": "Authors' affiliations with University of Science and Technology of China and Purdue University are clearly stated on page 1.", 79 "source": "haiku" 80 }, 81 "funder_independent_of_outcome": { 82 "applies": false, 83 "answer": false, 84 "justification": "No funding mentioned, so independence cannot be assessed.", 85 "source": "haiku" 86 }, 87 "financial_interests_declared": { 88 "applies": true, 89 "answer": false, 90 "justification": "No competing interests statement provided. No mention of patents, equity, or consulting relationships.", 91 "source": "haiku" 92 } 93 }, 94 "scope_and_framing": { 95 "key_terms_defined": { 96 "applies": true, 97 "answer": false, 98 "justification": "Group Query Attack and QGS are defined, but 'hallucination' (in title) is never defined. Backdoor is referenced without formal definition. Alignment and contamination contexts are unclear.", 99 "source": "haiku" 100 }, 101 "intended_contribution_clear": { 102 "applies": true, 103 "answer": true, 104 "justification": "Paper explicitly states two contributions: proposing GQA attack method and characterizing its effectiveness across fine-tuned vs. pre-trained models.", 105 "source": "haiku" 106 }, 107 "engagement_with_prior_work": { 108 "applies": true, 109 "answer": false, 110 "justification": "Related work section (2.1-2.2) lists failure modes and backdoor papers but is largely descriptive. Paper doesn't deeply explain how GQA relates to or extends existing failure modes research beyond citation.", 111 "source": "haiku" 112 } 113 } 114 }, 115 "type_checklist": { 116 "empirical": { 117 "artifacts": { 118 "code_released": { 119 "applies": true, 120 "answer": false, 121 "justification": "No code repository, GitHub link, or code availability statement provided. Reproduction would require implementing the pipeline from scratch.", 122 "source": "haiku" 123 }, 124 "data_released": { 125 "applies": true, 126 "answer": true, 127 "justification": "Paper uses standard public benchmarks (HumanEval, MedMCQA, PubMedQA, Aqua-RAT, MathQA, WMT20) available from original sources.", 128 "source": "haiku" 129 }, 130 "environment_specified": { 131 "applies": true, 132 "answer": false, 133 "justification": "Appendix B.2 provides hyperparameters (learning rate, batch size, epochs) but no requirements.txt, Dockerfile, or Python/PyTorch version specifications needed for reproduction.", 134 "source": "haiku" 135 }, 136 "reproduction_instructions": { 137 "applies": true, 138 "answer": false, 139 "justification": "No step-by-step reproduction guide provided. Paper describes procedures and provides prompt templates but insufficient detail to reproduce without original code.", 140 "source": "haiku" 141 } 142 }, 143 "statistical_methodology": { 144 "confidence_intervals_or_error_bars": { 145 "applies": true, 146 "answer": false, 147 "justification": "All results reported as single percentages (Tables 1-3, 9-15) with no error bars or confidence intervals, despite stating three random partitions were averaged.", 148 "source": "haiku" 149 }, 150 "significance_tests": { 151 "applies": true, 152 "answer": false, 153 "justification": "No statistical tests performed. No p-values, t-tests, or significance comparisons between QGS conditions or across models.", 154 "source": "haiku" 155 }, 156 "effect_sizes_reported": { 157 "applies": true, 158 "answer": false, 159 "justification": "Percentage point drops are shown (e.g., 53.3% → 19.7%) but not reported as standardized effect sizes (Cohen's d, etc.).", 160 "source": "haiku" 161 }, 162 "sample_size_justified": { 163 "applies": true, 164 "answer": false, 165 "justification": "Seven models selected for Q1 and eight for Q3 with no justification for adequacy. No power analysis or explanation for these specific numbers.", 166 "source": "haiku" 167 }, 168 "variance_reported": { 169 "applies": true, 170 "answer": false, 171 "justification": "Paper states 'three random partitions' averaged but reports only means, not variance or standard deviation across runs.", 172 "source": "haiku" 173 } 174 }, 175 "evaluation_design": { 176 "baselines_included": { 177 "applies": true, 178 "answer": false, 179 "justification": "QGS=1 serves as internal baseline, but no comparison to other attack methods or explanations for degradation. No baseline defense mechanisms tested.", 180 "source": "haiku" 181 }, 182 "baselines_contemporary": { 183 "applies": false, 184 "answer": false, 185 "justification": "No alternative attacks included for comparison.", 186 "source": "haiku" 187 }, 188 "ablation_study": { 189 "applies": true, 190 "answer": false, 191 "justification": "No ablation of GQA components. Paper doesn't test individual factors (query order, context length alone, number of irrelevant queries) separately.", 192 "source": "haiku" 193 }, 194 "multiple_metrics": { 195 "applies": true, 196 "answer": false, 197 "justification": "Each task type uses single metric (accuracy for MCQ/code/math; sacreBLEU for translation). No multiple metrics per task to assess robustness across measures.", 198 "source": "haiku" 199 }, 200 "human_evaluation": { 201 "applies": true, 202 "answer": false, 203 "justification": "No human evaluation. Backdoor triggering (Q2) inferred from output distribution, not human verification of actual malicious behavior.", 204 "source": "haiku" 205 }, 206 "held_out_test_set": { 207 "applies": true, 208 "answer": true, 209 "justification": "Official test sets used for all benchmarks (e.g., validation set for MedMCQA per Nori et al., 1K test examples for WMT20).", 210 "source": "haiku" 211 }, 212 "per_category_breakdown": { 213 "applies": true, 214 "answer": true, 215 "justification": "Results broken down by dataset, model architecture, and task type. Tables 1-3 show per-model results; Table 13 shows per-dataset token counts.", 216 "source": "haiku" 217 }, 218 "failure_cases_discussed": { 219 "applies": true, 220 "answer": false, 221 "justification": "Paper shows failure rates (e.g., gemma-7b-it drops to 0% on code) but doesn't analyze why or provide qualitative failure examples.", 222 "source": "haiku" 223 }, 224 "negative_results_reported": { 225 "applies": true, 226 "answer": true, 227 "justification": "Q3 findings explicitly state 'GQA has limited impact on multiple-choice question and translation tasks,' reporting null results alongside positive ones.", 228 "source": "haiku" 229 } 230 }, 231 "setup_transparency": { 232 "model_versions_specified": { 233 "applies": true, 234 "answer": true, 235 "justification": "All models include version specifiers (e.g., mistral-7b-v0.1, llama3-8b-instruct) and citation to original papers where available.", 236 "source": "haiku" 237 }, 238 "prompts_provided": { 239 "applies": true, 240 "answer": true, 241 "justification": "Figures 3-5 provide prompt templates with placeholders; Table 4 specifies fill values (system prompts, prefixes). Prompts are reconstructible from these tables.", 242 "source": "haiku" 243 }, 244 "hyperparameters_reported": { 245 "applies": true, 246 "answer": true, 247 "justification": "Appendix B.2 reports learning rate (2×10⁻⁵), batch size (64), epochs (3), sequence length (2048), warmup ratio (10%), and decay schedule.", 248 "source": "haiku" 249 }, 250 "scaffolding_described": { 251 "applies": true, 252 "answer": true, 253 "justification": "Few-shot prompting (10-shot for complex tasks) and CoT prompting ('Let's think step by step' for math) explicitly mentioned.", 254 "source": "haiku" 255 }, 256 "data_preprocessing_documented": { 257 "applies": true, 258 "answer": true, 259 "justification": "GQA construction pipeline documented: random dataset partition, query combination method, and for backdoors: '1% sampled, answers A, combined into groups' (≈0.5% final).", 260 "source": "haiku" 261 } 262 }, 263 "data_integrity": { 264 "raw_data_available": { 265 "applies": true, 266 "answer": true, 267 "justification": "Standard public benchmarks (WMT20, HumanEval, MedMCQA, PubMedQA, Aqua-RAT, MathQA) are publicly available. Modified GQA versions not released but base data accessible.", 268 "source": "haiku" 269 }, 270 "data_collection_described": { 271 "applies": true, 272 "answer": true, 273 "justification": "GQA construction process described: random partitioning into two dataset parts, combining queries, averaging across three partitions. Appendix A details each benchmark's source and split structure.", 274 "source": "haiku" 275 }, 276 "recruitment_methods_described": { 277 "applies": false, 278 "answer": false, 279 "justification": "No human participants; uses existing benchmarks.", 280 "source": "haiku" 281 }, 282 "data_pipeline_documented": { 283 "applies": true, 284 "answer": true, 285 "justification": "Full pipeline from dataset partition → query grouping → model fine-tuning/evaluation → response extraction described in Section 3.3.", 286 "source": "haiku" 287 } 288 }, 289 "contamination": { 290 "training_cutoff_stated": { 291 "applies": true, 292 "answer": false, 293 "justification": "Models tested (Llama 2/3, Mistral, Qwen, Gemma) have training dates implicit in versions but paper never explicitly states training data cutoff dates.", 294 "source": "haiku" 295 }, 296 "train_test_overlap_discussed": { 297 "applies": true, 298 "answer": false, 299 "justification": "Paper doesn't discuss whether standard benchmarks (HumanEval, MedMCQA, etc.) appear in models' training data, nor whether GQA variants contaminate training.", 300 "source": "haiku" 301 }, 302 "benchmark_contamination_addressed": { 303 "applies": true, 304 "answer": false, 305 "justification": "No acknowledgment that HumanEval, MathQA, or other benchmarks may appear in model pretraining. GQA creates novel input distribution but contamination risk of base benchmarks not discussed.", 306 "source": "haiku" 307 } 308 }, 309 "human_studies": { 310 "pre_registered": { 311 "applies": false, 312 "answer": false, 313 "justification": "No human participants involved.", 314 "source": "haiku" 315 }, 316 "irb_or_ethics_approval": { 317 "applies": false, 318 "answer": false, 319 "justification": "No human participants; N/A.", 320 "source": "haiku" 321 }, 322 "demographics_reported": { 323 "applies": false, 324 "answer": false, 325 "justification": "No human participants; N/A.", 326 "source": "haiku" 327 }, 328 "inclusion_exclusion_criteria": { 329 "applies": false, 330 "answer": false, 331 "justification": "No human participants; N/A.", 332 "source": "haiku" 333 }, 334 "randomization_described": { 335 "applies": false, 336 "answer": false, 337 "justification": "No human participants; N/A.", 338 "source": "haiku" 339 }, 340 "blinding_described": { 341 "applies": false, 342 "answer": false, 343 "justification": "No human participants; N/A.", 344 "source": "haiku" 345 }, 346 "attrition_reported": { 347 "applies": false, 348 "answer": false, 349 "justification": "No human participants; N/A.", 350 "source": "haiku" 351 } 352 }, 353 "cost_and_practicality": { 354 "inference_cost_reported": { 355 "applies": true, 356 "answer": false, 357 "justification": "No inference costs, API costs, latency, or computational time to run evaluations reported.", 358 "source": "haiku" 359 }, 360 "compute_budget_stated": { 361 "applies": true, 362 "answer": false, 363 "justification": "Total computational budget (GPU-hours, cost, etc.) not stated. Appendix provides batch size and epochs but not total compute.", 364 "source": "haiku" 365 } 366 } 367 } 368 }, 369 "claims": [ 370 { 371 "claim": "Group Query Attack significantly degrades the performance of models fine-tuned on specific tasks", 372 "evidence": "Table 1 shows accuracy drops of 30-70pp for fine-tuned models: llama2-7b MedMCQA 53.3%→19.7%, mistral-7b 61.1%→32.1%, mixtral-8x7b 66.3%→33.2% when QGS increases from 1 to 2.", 373 "supported": "strong" 374 }, 375 { 376 "claim": "Group Query Attack can trigger potential backdoors in LLMs", 377 "evidence": "Table 2 shows models fine-tuned on backdoored datasets tend to output option 'A' (the backdoor trigger) at 83.7%-100% frequency when QGS=2, compared to 32-80% at QGS=1.", 378 "supported": "moderate" 379 }, 380 { 381 "claim": "GQA is highly effective on reasoning tasks (code and mathematical reasoning) for pre-trained and aligned models", 382 "evidence": "Table 3 shows severe degradation: llama3-8b-it on HumanEval drops from 39.5% (QGS=1) to 11.3% (QGS=15); on mathematical reasoning 43.4%→40.3%; but minimal drop on MCQ (59.9%→57.9%).", 383 "supported": "strong" 384 }, 385 { 386 "claim": "Fine-tuned models frequently output the same response option under GQA, exhibiting mode collapse", 387 "evidence": "Tables 6 and 8 show fine-tuned models output single option with 98-100% probability when QGS=2, indicating severe mode collapse rather than distributed responses.", 388 "supported": "strong" 389 }, 390 { 391 "claim": "Aligned models are more robust to GQA than pre-trained models", 392 "evidence": "Table 3 comparisons show pre-trained models (mistral0.3-7b) degrade more sharply on translation (48.9%→3.5% at QGS=30) than aligned versions (52.9%→28.9%), and similar patterns in code.", 393 "supported": "moderate" 394 }, 395 { 396 "claim": "GQA has limited effectiveness on multiple-choice question and translation tasks for larger models", 397 "evidence": "Table 3 shows minimal drops for llama3-8b-it on MCQ (59.9%→57.9%) and translation (54.4%→52.8%), in contrast to severe code degradation (39.5%→11.3%).", 398 "supported": "strong" 399 } 400 ], 401 "methodology_tags": [ 402 "benchmark-eval", 403 "empirical" 404 ], 405 "key_findings": "The paper demonstrates that presenting multiple queries simultaneously (Group Query Attack) significantly degrades fine-tuned LLM performance, with accuracy dropping 30-70 percentage points across models like Llama 2 and Mistral. The attack can trigger injected backdoors, causing models to default to specific output options. Pre-trained and aligned models show varying vulnerability: reasoning tasks (code/math) degrade severely with increased QGS, while classification tasks remain relatively stable. Aligned models partially mitigate these vulnerabilities compared to pre-trained counterparts.", 406 "red_flags": [ 407 { 408 "flag": "No statistical significance testing", 409 "detail": "All results reported as raw percentages with no error bars, confidence intervals, or significance tests despite averaging three random partitions." 410 }, 411 { 412 "flag": "Confound with context length", 413 "detail": "Table 13 shows input tokens increase substantially with QGS (88→1874 tokens for MedMCQA at QGS=1 vs 30). Paper doesn't isolate whether degradation is due to GQA specifically or context length." 414 }, 415 { 416 "flag": "Unclear backdoor mechanism", 417 "detail": "Q2 concludes 'yes' to backdoor triggering based only on option 'A' output frequency. Doesn't definitively show actual backdoor activation vs. statistical mode collapse from input overwhelm." 418 }, 419 { 420 "flag": "No ablation studies", 421 "detail": "Paper doesn't isolate which GQA components (order, count, relevance of queries) cause degradation. Mechanism remains unexplained beyond 'accumulated context' speculation." 422 }, 423 { 424 "flag": "No code release", 425 "detail": "No code repository, GitHub link, or reproducibility artifacts provided. Results cannot be independently verified without reimplementing pipeline." 426 }, 427 { 428 "flag": "Limited baseline comparisons", 429 "detail": "No comparison to other attack methods or context-length reduction baselines. Can't assess whether GQA is novel or just a known context-window phenomenon." 430 }, 431 { 432 "flag": "Unexplained negative results for pre-trained models", 433 "detail": "Q3 shows pre-trained Mistral on translation at QGS=30 reaches 1.8% accuracy but paper doesn't analyze why or discuss failure cascade mechanisms." 434 }, 435 { 436 "flag": "Sample size not justified", 437 "detail": "Seven models for Q1, eight for Q3, with no power analysis or explanation for adequacy. Time constraints mentioned but not addressed in sampling strategy." 438 } 439 ], 440 "cited_papers": [ 441 { 442 "title": "The Reversal Curse: LLMs trained on 'A is B' fail to learn 'B is A'", 443 "relevance": "Related failure mode of LLMs; motivates investigating prompt robustness and failure modes" 444 }, 445 { 446 "title": "Lost in the middle: How language models use long contexts", 447 "relevance": "Long-context handling in LLMs; context length increase in GQA may relate to this phenomenon" 448 }, 449 { 450 "title": "Large Language Models Can Be Easily Distracted by Irrelevant Context", 451 "relevance": "Distractibility failure mode; GQA injects irrelevant queries which may trigger this vulnerability" 452 }, 453 { 454 "title": "Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training", 455 "relevance": "Backdoor persistence in LLMs; motivates studying backdoor triggering mechanisms" 456 }, 457 { 458 "title": "Training-free Lexical Backdoor Attacks on Language Models", 459 "relevance": "Backdoor injection methods; informs Q2 experimental design for backdoor embedding" 460 }, 461 { 462 "title": "BadChain: Backdoor Chain-of-Thought Prompting for Large Language Models", 463 "relevance": "Backdoor activation via prompting; relevant to understanding GQA as a potential trigger mechanism" 464 }, 465 { 466 "title": "Evaluating Large Language Models Trained on Code", 467 "relevance": "HumanEval benchmark and code generation evaluation; directly used in Q3 experiments" 468 }, 469 { 470 "title": "Large Language Models Cannot Self-Correct Reasoning Yet", 471 "relevance": "Reasoning limitations in LLMs; explains why reasoning tasks are most vulnerable to GQA" 472 } 473 ], 474 "engagement_factors": { 475 "practical_relevance": { 476 "score": 2, 477 "justification": "Identifies real vulnerability in multi-turn interactions but GQA scenario (dozens of unrelated queries at once) is artificial; typical user conversations don't batch 30 questions together." 478 }, 479 "surprise_contrarian": { 480 "score": 1, 481 "justification": "Multi-query performance degradation is expected given finite context windows and attention limits. Finding is intuitive rather than surprising or contrarian to prior beliefs." 482 }, 483 "fear_safety": { 484 "score": 2, 485 "justification": "Demonstrates potential backdoor triggering risk but only for fine-tuned models with pre-injected backdoors. Real-world applicability unclear; doesn't expose new vulnerabilities in aligned models." 486 }, 487 "demo_ability": { 488 "score": 1, 489 "justification": "Requires access to LLMs, fine-tuning capability, and unreleased code. No simple online demo or public benchmark to try GQA immediately." 490 }, 491 "brand_recognition": { 492 "score": 1, 493 "justification": "Authors from USTC and Purdue, not major AI safety or frontier-model labs. Published on arXiv only; limited visibility or institutional backing." 494 }, 495 "drama_conflict": { 496 "score": 1, 497 "justification": "Finding is straightforward technical observation without controversy, debate, or conflicting interpretations that would generate discussion." 498 } 499 }, 500 "hn_data": { 501 "threads": [ 502 { 503 "hn_id": "45474900", 504 "title": "How to inject knowledge efficiently? Knowledge infusion scaling law for LLMs", 505 "points": 105, 506 "comments": 35, 507 "url": "https://news.ycombinator.com/item?id=45474900", 508 "created_at": "2025-10-04T17:18:07Z" 509 }, 510 { 511 "hn_id": "47292454", 512 "title": "Technological Folie à Deux", 513 "points": 3, 514 "comments": 0, 515 "url": "https://news.ycombinator.com/item?id=47292454", 516 "created_at": "2026-03-07T23:21:38Z" 517 }, 518 { 519 "hn_id": "44887277", 520 "title": "Technological Folie à Deux:Feedback Loops Between AI Chatbots and Mental Illness", 521 "points": 3, 522 "comments": 0, 523 "url": "https://news.ycombinator.com/item?id=44887277", 524 "created_at": "2025-08-13T11:44:38Z" 525 }, 526 { 527 "hn_id": "43405094", 528 "title": "Politicians' misinformation behavior and public engagement, in 4 countries", 529 "points": 3, 530 "comments": 0, 531 "url": "https://news.ycombinator.com/item?id=43405094", 532 "created_at": "2025-03-18T21:03:45Z" 533 }, 534 { 535 "hn_id": "37455031", 536 "title": "Exposing and Addressing Security Vulnerabilities in Browser Text Input Fields", 537 "points": 2, 538 "comments": 1, 539 "url": "https://news.ycombinator.com/item?id=37455031", 540 "created_at": "2023-09-10T12:01:20Z" 541 }, 542 { 543 "hn_id": "45117954", 544 "title": "Learned Perceptive Forward Dynamics Model for Safe Robotic Navigation", 545 "points": 2, 546 "comments": 0, 547 "url": "https://news.ycombinator.com/item?id=45117954", 548 "created_at": "2025-09-03T16:49:02Z" 549 }, 550 { 551 "hn_id": "44270515", 552 "title": "Grassroots Consensus", 553 "points": 2, 554 "comments": 0, 555 "url": "https://news.ycombinator.com/item?id=44270515", 556 "created_at": "2025-06-13T17:39:42Z" 557 }, 558 { 559 "hn_id": "44147078", 560 "title": "SoloSpeech: A high-quality target speech extractor", 561 "points": 2, 562 "comments": 0, 563 "url": "https://news.ycombinator.com/item?id=44147078", 564 "created_at": "2025-05-31T21:37:25Z" 565 }, 566 { 567 "hn_id": "43495798", 568 "title": "RGL: Graph-Centric,Framework for Efficient RAG on Graphs", 569 "points": 2, 570 "comments": 0, 571 "url": "https://news.ycombinator.com/item?id=43495798", 572 "created_at": "2025-03-27T17:25:12Z" 573 }, 574 { 575 "hn_id": "43067948", 576 "title": "A Model for French Voters", 577 "points": 2, 578 "comments": 0, 579 "url": "https://news.ycombinator.com/item?id=43067948", 580 "created_at": "2025-02-16T13:49:10Z" 581 } 582 ], 583 "top_points": 105, 584 "total_points": 126, 585 "total_comments": 36 586 } 587 }