ai-research-survey

scan-v5.json (28634B)

---

 {
   "scan_version": 5,
   "paper_type": "empirical",
   "paper": {
     "title": "LLM-Powered Test Case Generation for Detecting Bugs in Plausible Programs",
     "authors": [
       "Kaibo Liu",
       "Zhenpeng Chen",
       "Yiyang Liu",
       "Jie M. Zhang",
       "Mark Harman",
       "Yudong Han",
       "Yun Ma",
       "Yihong Dong",
       "Ge Li",
       "Gang Huang"
     ],
     "year": 2024,
     "venue": "ACL 2025 (Annual Meeting of the Association for Computational Linguistics)",
     "arxiv_id": "2404.10304",
     "doi": "10.18653/v1/2025.acl-long.20"
   },
   "checklist": {
     "claims_and_evidence": {
       "abstract_claims_supported": {
         "applies": true,
         "answer": true,
         "justification": "All abstract claims (1.80×/2.65×/1.66× improvements, three-stage pipeline, evaluation on two datasets) are directly supported by Table 1 and Sections 4–6.",
         "source": "haiku"
       },
       "causal_claims_justified": {
         "applies": true,
         "answer": true,
         "justification": "Comparative claims are justified by controlled experiments with same datasets/baselines. Ablation study (Section 6.3, Table 2) establishes causal contribution of each component.",
         "source": "haiku"
       },
       "generalization_bounded": {
         "applies": true,
         "answer": true,
         "justification": "Evaluation bounded to coding task datasets (TrickyBugs, EvalPlus). Generalization test on deepseek-v3 (Section 7.2) is limited. Scope appropriately bounded.",
         "source": "haiku"
       },
       "alternative_explanations_discussed": {
         "applies": true,
         "answer": false,
         "justification": "Paper does not explore alternative explanations for TrickCatcher's superior performance beyond baseline comparisons. Section 7.1 discusses buggy variant usefulness but lacks depth on why diversity-driven approach fundamentally works better.",
         "source": "haiku"
       },
       "proxy_outcome_distinction": {
         "applies": true,
         "answer": true,
         "justification": "Measured outcomes (recall, precision, F1, TP/FP counts) directly align with claimed outcome (detecting bugs in plausible programs). True/false positive distinction is clearly defined in Section 3.",
         "source": "haiku"
       }
     },
     "limitations_and_scope": {
       "limitations_section_present": {
         "applies": true,
         "answer": true,
         "justification": "Dedicated 'Limitations' section before Acknowledgements lists three specific limitations (model budget constraints, LLM uncertainty, data leakage risk).",
         "source": "haiku"
       },
       "threats_to_validity_specific": {
         "applies": true,
         "answer": true,
         "justification": "Budget forced use of gpt-3.5-turbo instead of stronger models. Acknowledged uncertainty mitigated by multiple runs (Appendix B). Data leakage addressed by noting dataset release dates.",
         "source": "haiku"
       },
       "scope_boundaries_stated": {
         "applies": true,
         "answer": true,
         "justification": "Scope to small coding tasks on plausible programs is implicit in experimental setup. Limitations section could be more explicit about non-applicability to other domains, but boundaries are clear from context.",
         "source": "haiku"
       }
     },
     "conflicts_of_interest": {
       "funding_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "Acknowledgements fully disclose funding from National Key R&D Program (Grant No. 2023YFB4503801), NSFC (Grants 62192733, 62192730), Hubei Province, and InnovateUK.",
         "source": "haiku"
       },
       "affiliations_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "All author affiliations listed: Peking University, Nanyang Technological University, King's College London, University College London, National Key Laboratory of Data Space Technology.",
         "source": "haiku"
       },
       "funder_independent_of_outcome": {
         "applies": true,
         "answer": true,
         "justification": "Funders (NSFC, government grants, InnovateUK) are independent public bodies, not companies making the LLMs (OpenAI, DeepSeek) being evaluated.",
         "source": "haiku"
       },
       "financial_interests_declared": {
         "applies": true,
         "answer": false,
         "justification": "No competing interests statement or financial interest declaration (patents, equity, consulting) present in the paper.",
         "source": "haiku"
       }
     },
     "scope_and_framing": {
       "key_terms_defined": {
         "applies": true,
         "answer": true,
         "justification": "Paper defines: 'plausible programs' (Section 3), 'tricky bugs', 'test oracle', 'differential testing', and problem statement clearly framed in Section 3.",
         "source": "haiku"
       },
       "intended_contribution_clear": {
         "applies": true,
         "answer": true,
         "justification": "Three novel contributions explicitly stated: PUT-guided program variant generation, generator-based input generation, diversity-driven differential testing. Novelty vs. prior work (esp. Differential Prompting) clearly delineated.",
         "source": "haiku"
       },
       "engagement_with_prior_work": {
         "applies": true,
         "answer": true,
         "justification": "Section 2 surveys traditional test generation, LLM-based approaches, and explicitly compares TrickCatcher against Differential Prompting with detailed distinction in 4 points (Section 2).",
         "source": "haiku"
       }
     }
   },
   "type_checklist": {
     "empirical": {
       "artifacts": {
         "code_released": {
           "applies": true,
           "answer": true,
           "justification": "Abstract states 'Code and data used are available at https://github.com/RinCloud/TrickCatcher'. GitHub repository is publicly accessible.",
           "source": "haiku"
         },
         "data_released": {
           "applies": true,
           "answer": true,
           "justification": "Both TrickyBugs and EvalPlus datasets are publicly available from prior work (Liu et al. 2024b, Liu et al. 2023a). Paper uses existing public benchmarks.",
           "source": "haiku"
         },
         "environment_specified": {
           "applies": true,
           "answer": false,
           "justification": "LLM versions specified (gpt-3.5-turbo-0125, deepseek-v3) but no Docker/requirements.txt, Python version, or dependency specification provided.",
           "source": "haiku"
         },
         "reproduction_instructions": {
           "applies": true,
           "answer": false,
           "justification": "No step-by-step reproduction walkthrough in paper. Code available on GitHub but paper does not include command-line examples or setup instructions.",
           "source": "haiku"
         }
       },
       "statistical_methodology": {
         "confidence_intervals_or_error_bars": {
           "applies": true,
           "answer": false,
           "justification": "Table 1, Figures 5-8 report point estimates (recall, precision, F1) without confidence intervals, standard deviations, or error bars. Multiple runs mentioned (Appendix B) but variance not quantified.",
           "source": "haiku"
         },
         "significance_tests": {
           "applies": true,
           "answer": false,
           "justification": "No statistical significance tests (t-test, ANOVA, etc.) reported. Claims like 'up to 1.80×' are not tested for statistical significance.",
           "source": "haiku"
         },
         "effect_sizes_reported": {
           "applies": true,
           "answer": true,
           "justification": "Effect sizes reported as F1 scores (e.g., 41.31% vs 24.95%), improvement ratios (1.80×, 2.65×, 1.66×), and absolute point estimates in Table 1.",
           "source": "haiku"
         },
         "sample_size_justified": {
           "applies": true,
           "answer": false,
           "justification": "TrickyBugs (366) and EvalPlus (151) datasets used, but no power analysis, sample size justification, or discussion of adequacy provided.",
           "source": "haiku"
         },
         "variance_reported": {
           "applies": true,
           "answer": false,
           "justification": "Appendix B describes repetition strategy (100 inputs, 10 variants, combinatorial sampling) but results tables and figures do not report standard deviations, confidence intervals, or variance metrics.",
           "source": "haiku"
         }
       },
       "evaluation_design": {
         "baselines_included": {
           "applies": true,
           "answer": true,
           "justification": "Three baselines: DirectChat (CHAT), Differential Prompting Plus (DPP), Automated Program Repair (APR). Each evaluated in Table 1.",
           "source": "haiku"
         },
         "baselines_contemporary": {
           "applies": true,
           "answer": true,
           "justification": "DPP from Li et al. 2023, other LLM-based methods from 2024. Baselines are relatively recent and representative of state-of-the-art.",
           "source": "haiku"
         },
         "ablation_study": {
           "applies": true,
           "answer": true,
           "justification": "Section 6.3 provides Table 2 ablation study testing 6 patterns systematically removing/adding components: program generation, input generation, differential testing.",
           "source": "haiku"
         },
         "multiple_metrics": {
           "applies": true,
           "answer": true,
           "justification": "Primary metrics: recall, precision, F1 score. Secondary: TP, FP, FN counts, passing rates on base/extra test cases (Figure 8), task difficulty (Figure 7).",
           "source": "haiku"
         },
         "human_evaluation": {
           "applies": false,
           "answer": false,
           "justification": "Not applicable. TrickyBugs input validity manually verified (mentioned in Section 5.3) but this is validation not human evaluation of system outputs.",
           "source": "haiku"
         },
         "held_out_test_set": {
           "applies": true,
           "answer": true,
           "justification": "Both datasets have held-out test cases: TrickyBugs separates existing test suite from additional bug-revealing tests; EvalPlus has base vs extra test cases.",
           "source": "haiku"
         },
         "per_category_breakdown": {
           "applies": true,
           "answer": true,
           "justification": "Results broken down by: dataset type (C++/Python), difficulty (low/high in Figure 7-8), and shown across multiple k values in Table 1 and Figure 6.",
           "source": "haiku"
         },
         "failure_cases_discussed": {
           "applies": true,
           "answer": true,
           "justification": "Figure 5 shows false positives across methods. Section 7.1 discusses buggy variants that still contribute. Some analysis of when/why methods fail but limited depth.",
           "source": "haiku"
         },
         "negative_results_reported": {
           "applies": true,
           "answer": true,
           "justification": "Improvements on TrickyBugs (Python) for recall are modest (2.01% average). Precision sometimes lower than DPP worst case. Limitations section acknowledges constraints.",
           "source": "haiku"
         }
       },
       "setup_transparency": {
         "model_versions_specified": {
           "applies": true,
           "answer": true,
           "justification": "Specific versions: gpt-3.5-turbo-0125 (Section 5.5), deepseek-v3 (Section 7.2) with exact model IDs provided.",
           "source": "haiku"
         },
         "prompts_provided": {
           "applies": true,
           "answer": true,
           "justification": "Figure 3 shows exact prompt for program variant generation. Figure 4 shows exact prompt for input generator creation. Both complete and unambiguous.",
           "source": "haiku"
         },
         "hyperparameters_reported": {
           "applies": true,
           "answer": false,
           "justification": "LLM hyperparameters (temperature, top-p, max_tokens, etc.) not reported. Only model names given. Algorithm 1 specifies differential testing logic but not LLM sampling params.",
           "source": "haiku"
         },
         "scaffolding_described": {
           "applies": true,
           "answer": true,
           "justification": "Three-stage pipeline clearly described: Section 4.1 (program variant generation), 4.2 (test input generation), 4.3 (differential testing). Algorithm 1 provides pseudocode.",
           "source": "haiku"
         },
         "data_preprocessing_documented": {
           "applies": true,
           "answer": true,
           "justification": "EvalPlus filtering described: select AI-generated samples passing base but failing extra tests. TrickyBugs filtering by test suite mentioned. Program variant filtering after generation documented.",
           "source": "haiku"
         }
       },
       "data_integrity": {
         "raw_data_available": {
           "applies": true,
           "answer": true,
           "justification": "Both datasets (TrickyBugs, EvalPlus) are publicly available from prior work. Authors provide code repository with evaluation scripts.",
           "source": "haiku"
         },
         "data_collection_described": {
           "applies": true,
           "answer": true,
           "justification": "TrickyBugs collection described in Liu et al. 2024b (online judge platform submissions). EvalPlus from Liu et al. 2023a (code generation benchmark). Referenced, not collected by this paper.",
           "source": "haiku"
         },
         "recruitment_methods_described": {
           "applies": false,
           "answer": false,
           "justification": "Not applicable. No human participant recruitment—benchmarks are programming tasks from online judges and code generation datasets.",
           "source": "haiku"
         },
         "data_pipeline_documented": {
           "applies": true,
           "answer": true,
           "justification": "Pipeline documented: datasets → filter variants by existing tests → generate inputs → execute variants → compare outputs → collect TP/FP. Described in Section 4 and Figure 2.",
           "source": "haiku"
         }
       },
       "contamination": {
         "training_cutoff_stated": {
           "applies": true,
           "answer": true,
           "justification": "Training cutoff addressed: gpt-3.5-turbo-0125 release date noted. TrickyBugs released after model cutoff. EvalPlus prohibits training use. Discussed in Limitations.",
           "source": "haiku"
         },
         "train_test_overlap_discussed": {
           "applies": true,
           "answer": true,
           "justification": "Data leakage discussed in Limitations: TrickyBugs released after gpt-3.5-turbo training, EvalPlus prohibits training use. Poor baseline performance argues against major leakage.",
           "source": "haiku"
         },
         "benchmark_contamination_addressed": {
           "applies": true,
           "answer": true,
           "justification": "Contamination addressed through dataset release dates and explicit prohibition. Conversely, weak LLM baselines (Table 1) suggest benchmarks not memorized.",
           "source": "haiku"
         }
       },
       "human_studies": {
         "pre_registered": {
           "applies": false,
           "answer": false,
           "justification": "No human participants. Not applicable.",
           "source": "haiku"
         },
         "irb_or_ethics_approval": {
           "applies": false,
           "answer": false,
           "justification": "No human subjects. Not applicable.",
           "source": "haiku"
         },
         "demographics_reported": {
           "applies": false,
           "answer": false,
           "justification": "No human participants. Not applicable.",
           "source": "haiku"
         },
         "inclusion_exclusion_criteria": {
           "applies": false,
           "answer": false,
           "justification": "No human subjects. Not applicable.",
           "source": "haiku"
         },
         "randomization_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants. Randomization used for input selection/variant sampling (Appendix B) but not human randomization.",
           "source": "haiku"
         },
         "blinding_described": {
           "applies": false,
           "answer": false,
           "justification": "No human subjects. Not applicable.",
           "source": "haiku"
         },
         "attrition_reported": {
           "applies": false,
           "answer": false,
           "justification": "No human participants. Not applicable.",
           "source": "haiku"
         }
       },
       "cost_and_practicality": {
         "inference_cost_reported": {
           "applies": true,
           "answer": false,
           "justification": "Budget constraints mentioned as motivation for using gpt-3.5-turbo (Section 5.5) but no actual inference costs, latency, or cost per program reported.",
           "source": "haiku"
         },
         "compute_budget_stated": {
           "applies": true,
           "answer": false,
           "justification": "Paper mentions 'budget constraints' limited model choice but does not quantify total computational budget or cost.",
           "source": "haiku"
         }
       }
     }
   },
   "claims": [
     {
       "claim": "TrickCatcher achieves 1.80× recall, 2.65× precision, 1.66× F1 score compared to best baseline (DPP)",
       "evidence": "Table 1 reports F1 scores: TrickCatcher 41.31–51.34%, DPP 24.95–35.76% across three datasets",
       "supported": "strong"
     },
     {
       "claim": "PUT-guided program generation produces higher-quality variants than specification-only generation",
       "evidence": "Ablation study (Table 2): filtering+basic IG+ours DT (pattern 3) achieves 0.33 F1 vs filtered+basic (pattern 2) 0.23 F1; patterns 5-6 with full approach reach 0.37-0.41",
       "supported": "moderate"
     },
     {
       "claim": "Generator-based input generation achieves higher validity than direct LLM generation",
       "evidence": "Introduction mentions 40.10% invalid inputs from direct generation. Figure 5 shows TrickCatcher produces zero false positives from invalid inputs whereas DPP has many",
       "supported": "moderate"
     },
     {
       "claim": "Diversity-driven differential testing outperforms majority voting for bug detection",
       "evidence": "Table 2 ablation: filtered+basic+basic (pattern 2) vs filtered+basic+ours DT (pattern 3) improves F1 from 0.23 to 0.33 on TrickyBugs C++",
       "supported": "strong"
     },
     {
       "claim": "TrickCatcher generates up to 16× fewer false positives than baselines on correct programs",
       "evidence": "Figure 5 shows TrickCatcher max 5 FPs vs CHAT/DPP 26+ FPs. Ratio: 26/5 ≈ 5× minimum, up to 16× for specific configurations",
       "supported": "strong"
     },
     {
       "claim": "Performance remains stable across different numbers of program variants (k=2 to k=10)",
       "evidence": "Figure 6 shows TrickCatcher F1 stable 0.40-0.41 and precision 0.69-0.70 across k. DPP fluctuates 0.23-0.25 F1",
       "supported": "strong"
     },
     {
       "claim": "TrickCatcher shows greater improvement on high-difficulty tasks than low-difficulty",
       "evidence": "Figure 7-8: TrickCatcher median difficulty ~3500 for high-difficulty successes; DPP success median ~2200. Figure 8 shows TrickCatcher higher pass rate on hard tasks (base/extra)",
       "supported": "moderate"
     },
     {
       "claim": "Buggy program variants can contribute meaningfully to bug detection",
       "evidence": "Section 7.1: 23.2% (TrickyBugs) and 15.0% (EvalPlus) of useful variants are themselves buggy. TrickCatcher outperforms APR (repair-only) confirming non-repair detection",
       "supported": "moderate"
     }
   ],
   "methodology_tags": [
     "empirical",
     "benchmark-eval",
     "comparative"
   ],
   "key_findings": "TrickCatcher, an LLM-powered test generation method combining PUT-guided program variant generation, generator-based input generation, and diversity-driven differential testing, achieves 1.66–1.80× improvements in F1 score, recall, and precision over state-of-the-art baselines on two datasets (TrickyBugs: 366 human-written programs; EvalPlus: 151 AI-generated programs). The approach generates 16× fewer false positives than baselines while maintaining stable performance across variant counts and showing larger gains on high-difficulty tasks. Ablation studies confirm each component contributes meaningfully, and the method generalizes to deeper LLMs (deepseek-v3).",
   "red_flags": [
     {
       "flag": "No statistical significance testing",
       "detail": "Reported improvements (1.80×, 2.65×) lack p-values or confidence intervals; results could reflect random variation rather than systematic advantage"
     },
     {
       "flag": "Variance not quantified",
       "detail": "Appendix B describes repetition strategy (100 inputs, C(10,k) combinations) but results tables show only point estimates without standard deviations or confidence bounds"
     },
     {
       "flag": "Sample size not justified",
       "detail": "No power analysis or justification provided for 366+151 programs; adequacy unclear for detecting true effect sizes"
     },
     {
       "flag": "LLM hyperparameters underspecified",
       "detail": "Temperature, top-p, max_tokens, and other sampling parameters not reported; reproducibility compromised"
     },
     {
       "flag": "Limited generalization scope",
       "detail": "Evaluation restricted to coding tasks on two benchmarks. Deepseek-v3 generalization test (Table 3) limited to one alternative model"
     },
     {
       "flag": "Data leakage not fully ruled out",
       "detail": "While paper argues TrickyBugs post-dated training cutoff and EvalPlus prohibits training use, timing and enforcement not independently verified"
     },
     {
       "flag": "Manual validation required for TrickyBugs input validity",
       "detail": "Input validity assessed by hand for TrickyBugs but automated for EvalPlus; introduces subjectivity and inconsistency"
     },
     {
       "flag": "Limitations section somewhat boilerplate",
       "detail": "Budget constraint limiting model choice, LLM uncertainty mitigated by averaging, data leakage addressed by release dates—could be more concrete about residual threats"
     },
     {
       "flag": "Alternative explanations not explored",
       "detail": "Paper compares against baselines but does not investigate why diversity-driven approach fundamentally works beyond empirical results"
     },
     {
       "flag": "No reproduction walkthrough in paper",
       "detail": "Code available on GitHub but paper lacks step-by-step setup/execution instructions; readers must reverse-engineer from code"
     }
   ],
   "cited_papers": [
     {
       "title": "Nuances are the key: Unlocking ChatGPT to find failure-inducing tests with differential prompting",
       "relevance": "Core prior work (Differential Prompting); paper explicitly builds on and improves this approach with three key modifications"
     },
     {
       "title": "TrickyBugs: A dataset of corner-case bugs in plausible programs",
       "relevance": "Primary evaluation dataset; papers defines 'plausible programs' and 'tricky bugs' problem"
     },
     {
       "title": "Is your code generated by ChatGPT really correct? Rigorous evaluation of large language models for code generation",
       "relevance": "EvalPlus benchmark used for AI-generated code evaluation; defines base/extra test case split"
     },
     {
       "title": "Large language model-based agents for software engineering: A survey",
       "relevance": "Surveys LLM applications in code generation and testing; contextualizes contribution in broader agent/code space"
     },
     {
       "title": "Who judges the judge: An empirical study on online judge tests",
       "relevance": "Foundational work identifying prevalence of tricky bugs (3,440 bugs in online judge); motivates problem"
     },
     {
       "title": "Evaluating large language models trained on code",
       "relevance": "Foundational LLM code generation benchmark; shows LLM struggle with complex tasks, motivating TrickCatcher's two-step input generation"
     },
     {
       "title": "EvoSuite: Automated unit test generation for object-oriented software",
       "relevance": "Traditional search-based test generation baseline; contrasts with LLM-powered approach"
     },
     {
       "title": "KLEE: unassisted automatic generation of high-coverage tests",
       "relevance": "Symbolic execution test generation tool; represents pre-LLM approach to test generation"
     }
   ],
   "engagement_factors": {
     "practical_relevance": {
       "score": 3,
       "justification": "TrickCatcher is immediately usable for finding real bugs in code; released on GitHub with full implementation; directly applicable to developer workflows"
     },
     "surprise_contrarian": {
       "score": 2,
       "justification": "Diversity-driven testing over majority voting is counterintuitive; using buggy variants as oracle contributors is creative; but core idea (LLM-based test generation) not novel"
     },
     "fear_safety": {
       "score": 0,
       "justification": "Bug detection is positive for code safety but paper is tool-focused, not safety risk paper; no AI alignment or security concerns raised"
     },
     "drama_conflict": {
       "score": 1,
       "justification": "Addresses tension between LLM-generated code quality and testing rigor; somewhat timely given AI code generation boom, but not high-conflict angle"
     },
     "demo_ability": {
       "score": 2,
       "justification": "Code is open-source and runnable; requires having target programs to test; not immediately demandable to broad audience but doable for developers"
     },
     "brand_recognition": {
       "score": 2,
       "justification": "Multiple top-tier institutions (Peking, NTU, KCL, UCL); published at ACL (prestigious NLP venue); not industry giants but credible academic pedigree"
     }
   },
   "hn_data": {
     "threads": [
       {
         "hn_id": "41319553",
         "title": "First open source Legal AI retrieval benchmark for RAG finally released",
         "points": 9,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=41319553"
       },
       {
         "hn_id": "41663273",
         "title": "Unsafe Impedance: Safe Languages and Safe by Design Software",
         "points": 7,
         "comments": 1,
         "url": "https://news.ycombinator.com/item?id=41663273"
       },
       {
         "hn_id": "40209981",
         "title": "Long-form music generation with latent diffusion",
         "points": 3,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=40209981"
       },
       {
         "hn_id": "39807740",
         "title": "Perl: Parameter Efficient Reinforcement Learning from Human Feedback",
         "points": 3,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=39807740"
       },
       {
         "hn_id": "35687845",
         "title": "Backporting RISC-V Vector assembly",
         "points": 3,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=35687845"
       },
       {
         "hn_id": "40122867",
         "title": "Decentralized Trustless Bridge for Ethereum Full Node",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=40122867"
       },
       {
         "hn_id": "35676768",
         "title": "The Law of Activity Delays",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=35676768"
       },
       {
         "hn_id": "39180109",
         "title": "Personality Inference via Mobile Phone Sensors: A Machine Learning Approach",
         "points": 2,
         "comments": 1,
         "url": "https://news.ycombinator.com/item?id=39180109"
       },
       {
         "hn_id": "39202163",
         "title": "Using LLM Such as ChatGPT for Designing and Implementing a RISC Processor",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=39202163"
       },
       {
         "hn_id": "40061342",
         "title": "Long-form music generation with latent diffusion",
         "points": 1,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=40061342"
       }
     ],
     "top_points": 9,
     "total_points": 34,
     "total_comments": 2
   }
 }