ai-research-survey

scan-v5.json (29952B)

---

 {
   "scan_version": 5,
   "paper_type": "empirical",
   "paper": {
     "title": "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena",
     "authors": [
       "Lianmin Zheng",
       "Wei-Lin Chiang",
       "Ying Sheng",
       "Siyuan Zhuang",
       "Zhanghao Wu",
       "Yonghao Zhuang",
       "Zi Lin",
       "Zhuohan Li",
       "Dacheng Li",
       "Eric P. Xing",
       "Hao Zhang",
       "Joseph E. Gonzalez",
       "Ion Stoica"
     ],
     "year": 2023,
     "venue": "Neural Information Processing Systems",
     "arxiv_id": "2306.05685",
     "doi": null
   },
   "checklist": {
     "claims_and_evidence": {
       "abstract_claims_supported": {
         "applies": true,
         "answer": true,
         "justification": "All abstract claims are directly supported: GPT-4 achieves >80% agreement with humans (Table 5: 85% S2), biases are documented empirically with quantitative results, and MT-bench/Chatbot Arena are introduced and validated with data.",
         "source": "haiku"
       },
       "causal_claims_justified": {
         "applies": true,
         "answer": true,
         "justification": "Causal claims about mitigation strategies are tested with controlled ablations: reference-guided grading reduces failure rate from 70% to 15% (Table 4), and position-swapping is systematically tested across multiple prompt variants.",
         "source": "haiku"
       },
       "generalization_bounded": {
         "applies": true,
         "answer": false,
         "justification": "The paper asserts 'LLM-as-a-judge is a scalable and explainable way to approximate human preferences' broadly, but only validates this for GPT-4 on 80 questions with 58 graduate-student labelers—the conclusion extends beyond the tested scope.",
         "source": "haiku"
       },
       "alternative_explanations_discussed": {
         "applies": true,
         "answer": false,
         "justification": "The paper does not discuss the alternative explanation that GPT-4's high agreement with graduate-student labelers reflects shared RLHF training distributions rather than genuine general human preference alignment; self-enhancement bias is dismissed without controlled alternative hypotheses.",
         "source": "haiku"
       },
       "proxy_outcome_distinction": {
         "applies": true,
         "answer": true,
         "justification": "The paper explicitly frames LLM judgments as a proxy for human preferences and validates this proxy by measuring agreement against actual human votes (Tables 5–6), clearly distinguishing the proxy from the construct.",
         "source": "haiku"
       }
     },
     "limitations_and_scope": {
       "limitations_section_present": {
         "applies": true,
         "answer": true,
         "justification": "Section 6 'Discussion' contains a dedicated 'Limitations' paragraph discussing the focus on helpfulness over safety and the single-metric approach to evaluation.",
         "source": "haiku"
       },
       "threats_to_validity_specific": {
         "applies": true,
         "answer": false,
         "justification": "The limitations paragraph notes safety neglect and single-metric issues but does not address specific threats such as experimenter bias (authors evaluate their own Vicuna model on their own benchmark), or the unrepresentativeness of graduate-student labelers as a gold standard.",
         "source": "haiku"
       },
       "scope_boundaries_stated": {
         "applies": true,
         "answer": true,
         "justification": "The paper explicitly states it 'emphasizes helpfulness but largely neglects safety' and defines benchmark scope as 8 categories, 80 questions, and a specific set of models.",
         "source": "haiku"
       }
     },
     "conflicts_of_interest": {
       "funding_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "Funding is disclosed in acknowledgments: gifts from Anyscale, Google, IBM, Intel, Microsoft, MBZUAI, Samsung SDS, Uber, VMware, and a Meta PhD Fellowship for one author.",
         "source": "haiku"
       },
       "affiliations_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "Author affiliations are disclosed (UC Berkeley, UC San Diego, CMU, Stanford, MBZUAI).",
         "source": "haiku"
       },
       "funder_independent_of_outcome": {
         "applies": true,
         "answer": false,
         "justification": "Microsoft and Google are funders and both have direct commercial interests in LLM evaluation outcomes; GPT-4 (Microsoft-backed OpenAI) is evaluated as the primary high-performing judge throughout with no discussion of this conflict.",
         "source": "haiku"
       },
       "financial_interests_declared": {
         "applies": true,
         "answer": false,
         "justification": "No competing interests statement is present; the acknowledgments list corporate funders but do not declare any author equity, consulting arrangements, or other financial interests.",
         "source": "haiku"
       }
     },
     "scope_and_framing": {
       "key_terms_defined": {
         "applies": true,
         "answer": true,
         "justification": "Key terms are defined: 'LLM-as-a-judge' is introduced with three variants (pairwise, single-answer, reference-guided), 'agreement' is formally defined in Appendix D.3, and each bias type (position, verbosity, self-enhancement) is explicitly defined with examples.",
         "source": "haiku"
       },
       "intended_contribution_clear": {
         "applies": true,
         "answer": true,
         "justification": "Contributions are explicitly stated: '(1) a systematic study of LLM-as-a-judge; and (2) human preference datasets with high-quality questions and diverse user interactions from MT-bench and Chatbot Arena.'",
         "source": "haiku"
       },
       "engagement_with_prior_work": {
         "applies": true,
         "answer": true,
         "justification": "Section 2.1 situates the work against three categories of existing benchmarks (core-knowledge, instruction-following, conversational), and concurrent work is cited and compared throughout the main results.",
         "source": "haiku"
       }
     }
   },
   "type_checklist": {
     "empirical": {
       "artifacts": {
         "code_released": {
           "applies": true,
           "answer": true,
           "justification": "Code and data are publicly released at https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge, including 80 MT-bench questions, 3K expert votes, and 30K conversations.",
           "source": "haiku"
         },
         "data_released": {
           "applies": true,
           "answer": true,
           "justification": "MT-bench questions, 3K expert votes, and 30K Chatbot Arena conversations with human preferences are publicly released.",
           "source": "haiku"
         },
         "environment_specified": {
           "applies": true,
           "answer": false,
           "justification": "No requirements.txt, Dockerfile, or equivalent environment specification is provided; training mentions '8x A100 GPUs' and libraries (FlashAttention, SkyPilot) but no reproducible environment spec.",
           "source": "haiku"
         },
         "reproduction_instructions": {
           "applies": true,
           "answer": false,
           "justification": "The paper provides judge prompts in the appendix and a GitHub URL but does not include step-by-step instructions for reproducing the main agreement evaluation results.",
           "source": "haiku"
         }
       },
       "statistical_methodology": {
         "confidence_intervals_or_error_bars": {
           "applies": true,
           "answer": false,
           "justification": "All agreement rates and win rates are reported as point estimates with no confidence intervals or error bars anywhere in the paper.",
           "source": "haiku"
         },
         "significance_tests": {
           "applies": true,
           "answer": false,
           "justification": "No statistical significance tests are applied to any comparisons; differences between judges and human agreement rates are presented without significance testing.",
           "source": "haiku"
         },
         "effect_sizes_reported": {
           "applies": true,
           "answer": true,
           "justification": "Effect sizes are reported as percentage differences with baseline context, e.g., reference-guided grading reduces failure rate from 14/20 to 3/20 cases; win rate differences between model pairs are quantified.",
           "source": "haiku"
         },
         "sample_size_justified": {
           "applies": true,
           "answer": false,
           "justification": "Sample sizes (3K expert votes, 3K arena votes, 23 cases for verbosity test, 10 math questions) are stated but never justified; no power analysis is provided.",
           "source": "haiku"
         },
         "variance_reported": {
           "applies": true,
           "answer": false,
           "justification": "No variance, standard deviation, or spread is reported for any result; all metrics are single point estimates.",
           "source": "haiku"
         }
       },
       "evaluation_design": {
         "baselines_included": {
           "applies": true,
           "answer": true,
           "justification": "Human evaluation serves as the gold standard baseline, and multiple LLM judges (GPT-3.5, Claude-v1, GPT-4) are compared against each other and against human preferences.",
           "source": "haiku"
         },
         "baselines_contemporary": {
           "applies": true,
           "answer": true,
           "justification": "All baselines (GPT-3.5, GPT-4, Claude-v1, Vicuna-13B, LLaMA-13B) are contemporary state-of-the-art models at time of writing (2023).",
           "source": "haiku"
         },
         "ablation_study": {
           "applies": true,
           "answer": true,
           "justification": "Multiple ablations are performed: different prompt variants (default, rename, score, short, CoT, reference-guided), few-shot vs zero-shot, and single vs multi-turn evaluation designs.",
           "source": "haiku"
         },
         "multiple_metrics": {
           "applies": true,
           "answer": true,
           "justification": "Multiple metrics are used: agreement rates (S1/S2 setups), position-bias consistency rates, win rates, per-category scores, and failure rates under adversarial attacks.",
           "source": "haiku"
         },
         "human_evaluation": {
           "applies": true,
           "answer": true,
           "justification": "58 expert (graduate student) labelers evaluated MT-bench questions at ~$35/hr, and crowdsourced Chatbot Arena users provided 30K votes evaluating actual system outputs.",
           "source": "haiku"
         },
         "held_out_test_set": {
           "applies": false,
           "answer": false,
           "justification": "This is an evaluation methodology study, not a prediction task; there is no held-out test set in the standard ML sense.",
           "source": "haiku"
         },
         "per_category_breakdown": {
           "applies": true,
           "answer": true,
           "justification": "Table 7 provides per-category win rates for all 8 MT-bench categories; Figure 20 shows category-wise scores for 6 models; Table 10 shows position bias by category.",
           "source": "haiku"
         },
         "failure_cases_discussed": {
           "applies": true,
           "answer": true,
           "justification": "Appendix B/C shows concrete failure cases including position bias examples (Figure 11), verbosity attack failures (Figure 12), math grading errors (Figures 13–15), and multi-turn reference errors (Figure 16).",
           "source": "haiku"
         },
         "negative_results_reported": {
           "applies": true,
           "answer": true,
           "justification": "Significant negative results are reported: Claude-v1 shows only 23.8% consistency; GPT-4 fails 14/20 math questions with default prompts; CoT still produces wrong judgments due to context contamination by given answers.",
           "source": "haiku"
         }
       },
       "setup_transparency": {
         "model_versions_specified": {
           "applies": true,
           "answer": false,
           "justification": "Main results tables use generic names ('GPT-4', 'GPT-3.5', 'Claude-v1') without snapshot dates; the specific version 'gpt-4-0314' is mentioned only in Appendix B as a caveat about reproducibility.",
           "source": "haiku"
         },
         "prompts_provided": {
           "applies": true,
           "answer": true,
           "justification": "All judge prompt templates are fully provided in Appendix A (Figures 5–10), including pairwise, single-answer, CoT, reference-guided, and multi-turn variants with complete text.",
           "source": "haiku"
         },
         "hyperparameters_reported": {
           "applies": true,
           "answer": false,
           "justification": "Inference hyperparameters for LLM judges (temperature, top-p) are not reported anywhere; training hyperparameters for Vicuna are reported but that is secondary to the main evaluation.",
           "source": "haiku"
         },
         "scaffolding_described": {
           "applies": false,
           "answer": false,
           "justification": "No agentic scaffolding is used; the evaluation involves direct API calls to LLMs as judges.",
           "source": "haiku"
         },
         "data_preprocessing_documented": {
           "applies": true,
           "answer": true,
           "justification": "Preprocessing is documented: Vicuna training data was HTML-to-markdown converted with quality filtering (resulting in 125K conversations), and Chatbot Arena data will have PII cleaned and toxic content tagged.",
           "source": "haiku"
         }
       },
       "data_integrity": {
         "raw_data_available": {
           "applies": true,
           "answer": true,
           "justification": "Raw data (80 MT-bench questions, 3K expert votes, 30K conversations) is released publicly at the GitHub repository.",
           "source": "haiku"
         },
         "data_collection_described": {
           "applies": true,
           "answer": true,
           "justification": "Appendix C describes the data collection process for both MT-bench (voting interface, expert labeling procedure, screenshots) and Chatbot Arena (anonymous battle platform, consent process) in detail.",
           "source": "haiku"
         },
         "recruitment_methods_described": {
           "applies": true,
           "answer": true,
           "justification": "MT-bench labelers were recruited via application form with consent, paid $20 for 20 questions (~$35/hr), and were mostly graduate students from over ten universities.",
           "source": "haiku"
         },
         "data_pipeline_documented": {
           "applies": true,
           "answer": true,
           "justification": "Appendix C documents the full pipeline including voting interface screenshots, consent procedures, how GPT-4 judgments are shown to disagreeing humans, and PII handling plans for data release.",
           "source": "haiku"
         }
       },
       "contamination": {
         "training_cutoff_stated": {
           "applies": true,
           "answer": false,
           "justification": "Training data cutoffs for GPT-4, GPT-3.5, and Claude are not stated anywhere in the paper.",
           "source": "haiku"
         },
         "train_test_overlap_discussed": {
           "applies": true,
           "answer": false,
           "justification": "The paper does not discuss whether MT-bench question types or Chatbot Arena conversation patterns could overlap with the evaluated models' training data.",
           "source": "haiku"
         },
         "benchmark_contamination_addressed": {
           "applies": true,
           "answer": false,
           "justification": "MT-bench is described as newly created but no analysis of whether similar open-ended questions appeared in GPT-4's or other models' training data is provided.",
           "source": "haiku"
         }
       },
       "human_studies": {
         "pre_registered": {
           "applies": true,
           "answer": false,
           "justification": "No pre-registration is mentioned for the human evaluation study.",
           "source": "haiku"
         },
         "irb_or_ethics_approval": {
           "applies": true,
           "answer": false,
           "justification": "No IRB or ethics approval is mentioned; only that participants signed consent forms before participating.",
           "source": "haiku"
         },
         "demographics_reported": {
           "applies": true,
           "answer": false,
           "justification": "Only minimal information provided: 'mostly graduate students from more than ten universities'; no age, gender, field distribution, or other demographics reported.",
           "source": "haiku"
         },
         "inclusion_exclusion_criteria": {
           "applies": true,
           "answer": false,
           "justification": "No formal inclusion/exclusion criteria are stated; the description 'mostly graduate students considered experts' implies informal expertise filtering but no explicit criteria.",
           "source": "haiku"
         },
         "randomization_described": {
           "applies": true,
           "answer": true,
           "justification": "Randomization of question assignment is described: 'let each human evaluate at least 20 random multi-turn questions'; model identity randomization is also implemented.",
           "source": "haiku"
         },
         "blinding_described": {
           "applies": true,
           "answer": true,
           "justification": "Blinding is implemented in both platforms: MT-bench presents 'answers from two random anonymous assistants' and Chatbot Arena discloses model identities only after voting.",
           "source": "haiku"
         },
         "attrition_reported": {
           "applies": true,
           "answer": false,
           "justification": "No attrition or dropout statistics are reported for the human evaluation participants.",
           "source": "haiku"
         }
       },
       "cost_and_practicality": {
         "inference_cost_reported": {
           "applies": true,
           "answer": false,
           "justification": "The paper notes few-shot prompts are '4× more expensive' for API calls relatively, but absolute inference costs for any evaluation run are not reported.",
           "source": "haiku"
         },
         "compute_budget_stated": {
           "applies": true,
           "answer": false,
           "justification": "Vicuna training mentions '8x A100 GPUs' and 'around 2 days' for the longest run, but a total compute budget (GPU-hours, cost) is not stated.",
           "source": "haiku"
         }
       }
     }
   },
   "claims": [
     {
       "claim": "GPT-4 as a judge achieves over 80% agreement with human expert preferences, matching human-human agreement levels.",
       "evidence": "Table 5 shows 85% non-tie agreement between GPT-4 pairwise comparison and human experts on MT-bench; Table 6 shows 87% on Chatbot Arena. Human-human agreement is 81%.",
       "supported": "strong"
     },
     {
       "claim": "All LLM judges exhibit significant position bias, with Claude-v1 showing only 23.8% consistency.",
       "evidence": "Table 2 shows Claude-v1 at 23.8% consistency with default prompt (75% biased toward first position); GPT-4 is best at 65% consistency—still substantial bias.",
       "supported": "strong"
     },
     {
       "claim": "LLM judges are highly susceptible to verbosity attacks, with Claude and GPT-3.5 failing 91.3% of cases while GPT-4 fails only 8.7%.",
       "evidence": "Table 3 shows failure rates under 'repetitive list' attack on 23 answers: Claude-v1 91.3%, GPT-3.5 91.3%, GPT-4 8.7%.",
       "supported": "strong"
     },
     {
       "claim": "Reference-guided grading reduces GPT-4's math grading failure rate from 70% to 15%.",
       "evidence": "Table 4 shows failure rates on 10 math questions: default 14/20 (70%), CoT 6/20 (30%), reference-guided 3/20 (15%).",
       "supported": "strong"
     },
     {
       "claim": "LLM judge agreement with humans increases as the performance gap between compared models grows.",
       "evidence": "Figure 2 shows agreement rising from ~70% to ~100% as win rate difference between model pairs increases; near-identical models yield lowest agreement.",
       "supported": "strong"
     },
     {
       "claim": "GPT-4 shows self-enhancement bias, favoring itself with ~10% higher win rate compared to human judgments.",
       "evidence": "Figure 3(b) shows GPT-4 favors itself by ~10% and Claude-v1 by ~25%, but authors acknowledge 'limited data and small differences' prevent definitive conclusions.",
       "supported": "moderate"
     },
     {
       "claim": "A fine-tuned Vicuna-13B can approach GPT-4-level judge agreement (85.5% vs 87% non-tie S2).",
       "evidence": "Appendix F Table 15 shows fine-tuned Vicuna-13B achieves 85.5% non-tie agreement vs GPT-4's ~87% on the same 3K test set, with preliminary results.",
       "supported": "moderate"
     }
   ],
   "methodology_tags": [
     "benchmark-eval",
     "observational"
   ],
   "key_findings": "GPT-4 as an LLM judge achieves 85%+ agreement with human expert preferences on MT-bench, matching human-human agreement levels and validating LLM-as-a-judge as a scalable evaluation approach. All tested LLMs exhibit significant position bias (as low as 23.8% consistency for Claude-v1), and Claude/GPT-3.5 fail 91.3% of verbosity attacks while GPT-4 is substantially more robust. Reference-guided grading reduces GPT-4's math/reasoning grading failure rate from 70% to 15%. The publicly released MT-bench, 3K expert votes, and 30K Chatbot Arena conversations established an influential benchmark infrastructure for future LLM evaluation research.",
   "red_flags": [
     {
       "flag": "No statistical tests or CIs",
       "detail": "All agreement rates (e.g., 85% GPT-4 vs humans) are reported as point estimates without confidence intervals, significance tests, or variance measures, making reliability assessment impossible."
     },
     {
       "flag": "Authors evaluate own model on own benchmark",
       "detail": "The paper evaluates Vicuna (created by the same UC Berkeley team) on MT-bench (also created by the same team), with no discussion of how this circular setup might inflate Vicuna's apparent performance."
     },
     {
       "flag": "Unrepresentative human gold standard",
       "detail": "The human 'gold standard' consists of 58 graduate students who are 'considered experts'; GPT-4's high agreement with this group may reflect shared RLHF/academic writing preferences rather than alignment with general human preferences."
     },
     {
       "flag": "Funder conflict with primary positive result",
       "detail": "Microsoft and Google are listed funders; GPT-4 (from Microsoft-backed OpenAI) is the paper's primary high-performing judge and is evaluated favorably throughout with no conflict-of-interest discussion."
     },
     {
       "flag": "Self-enhancement bias dismissed without power analysis",
       "detail": "The observed 10–25% win rate inflation for self-judging is dismissed with 'limited data and small differences,' but no power analysis is provided; the effect size is practically significant for RLHF pipeline integrity."
     },
     {
       "flag": "Inference hyperparameters unreported",
       "detail": "Temperature, top-p, and other inference parameters for GPT-4, GPT-3.5, and Claude as judges are not reported, making results difficult to reproduce exactly."
     }
   ],
   "cited_papers": [
     {
       "title": "Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback",
       "relevance": "Foundational RLHF alignment work that motivates the need for scalable human preference evaluation; directly cited as context for why chat assistants need preference-based benchmarks."
     },
     {
       "title": "Measuring Massive Multitask Language Understanding (MMLU)",
       "relevance": "Core-knowledge benchmark shown to be inadequate for evaluating chat assistants; central contrast benchmark motivating the need for MT-bench."
     },
     {
       "title": "Holistic Evaluation of Language Models (HELM)",
       "relevance": "Comprehensive benchmark framework contrasted with MT-bench's preference-based approach; cited as existing but insufficient for human preference measurement."
     },
     {
       "title": "AlpacaFarm: A Simulation Framework for Methods that Learn from Human Feedback",
       "relevance": "Concurrent work on simulating human feedback evaluation; directly related to the LLM-as-judge paradigm studied in this paper."
     },
     {
       "title": "Can Large Language Models Be an Alternative to Human Evaluations?",
       "relevance": "Concurrent work studying the same core question of LLM-as-judge viability, cited as parallel investigation."
     },
     {
       "title": "Large Language Models Are Not Fair Evaluators",
       "relevance": "Concurrent work identifying LLM judge position biases, directly related to this paper's bias analysis section."
     },
     {
       "title": "Chain of Thought Prompting Elicits Reasoning in Large Language Models",
       "relevance": "CoT technique applied to the reference-guided judge design to improve math/reasoning evaluation quality."
     },
     {
       "title": "PandaLM: An Automatic Evaluation Benchmark for LLM Instruction Tuning Optimization",
       "relevance": "Concurrent work on automatic LLM evaluation benchmarking; cited as related to the LLM-as-judge paradigm."
     }
   ],
   "engagement_factors": {
     "practical_relevance": {
       "score": 3,
       "justification": "LLM-as-a-judge is now a standard evaluation approach widely adopted across the industry; Chatbot Arena is a live, publicly accessible platform actively used for model comparison."
     },
     "surprise_contrarian": {
       "score": 2,
       "justification": "The finding that GPT-4 matches human evaluators at 80%+ agreement challenged the assumption that human evaluation is irreplaceable for open-ended tasks."
     },
     "fear_safety": {
       "score": 1,
       "justification": "Raises modest concerns about evaluation biases (position, verbosity, self-enhancement) that could systematically corrupt RLHF training pipelines if uncorrected."
     },
     "drama_conflict": {
       "score": 1,
       "justification": "Position bias findings (Claude only 23.8% consistency) create controversy about reliability of LLM evaluators and implicitly critique widely-used evaluation practices."
     },
     "demo_ability": {
       "score": 3,
       "justification": "Chatbot Arena is publicly accessible and widely used; anyone can interact with it immediately and vote on model comparisons."
     },
     "brand_recognition": {
       "score": 3,
       "justification": "UC Berkeley LMSYS group, GPT-4, NeurIPS Datasets and Benchmarks track; this became one of the most cited papers establishing LLM evaluation methodology."
     }
   },
   "hn_data": {
     "threads": [
       {
         "hn_id": "35445312",
         "title": "Show HN: Want something better than k-means? Try BanditPAM",
         "points": 281,
         "comments": 41,
         "url": "https://news.ycombinator.com/item?id=35445312",
         "created_at": "2023-04-04T20:16:33Z"
       },
       {
         "hn_id": "35833868",
         "title": "LORA: Low-Rank Adaptation of Large Language Models",
         "points": 42,
         "comments": 3,
         "url": "https://news.ycombinator.com/item?id=35833868",
         "created_at": "2023-05-05T19:10:58Z"
       },
       {
         "hn_id": "37531815",
         "title": "Mesa-optimization algorithms in Transformers[pdf]",
         "points": 23,
         "comments": 5,
         "url": "https://news.ycombinator.com/item?id=37531815",
         "created_at": "2023-09-16T03:35:25Z"
       },
       {
         "hn_id": "37555617",
         "title": "Efficiently Correcting Reasoning Failures in Large Language Models",
         "points": 5,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=37555617",
         "created_at": "2023-09-18T13:10:22Z"
       },
       {
         "hn_id": "41136426",
         "title": "Show HN: I built Choosy Chat to get the best answer between GPT, Claude, Gemini",
         "points": 4,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=41136426",
         "created_at": "2024-08-02T05:52:12Z"
       },
       {
         "hn_id": "38720557",
         "title": "ReLoRA: High-Rank Training Through Low-Rank Updates",
         "points": 3,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=38720557",
         "created_at": "2023-12-21T14:09:38Z"
       },
       {
         "hn_id": "29337457",
         "title": "BanditPAM, an improved alternative to k-means",
         "points": 3,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=29337457",
         "created_at": "2021-11-25T01:55:55Z"
       },
       {
         "hn_id": "38561676",
         "title": "AQuaSurF: Discover better activation functions for your ML task",
         "points": 2,
         "comments": 1,
         "url": "https://news.ycombinator.com/item?id=38561676",
         "created_at": "2023-12-07T20:53:04Z"
       },
       {
         "hn_id": "40207489",
         "title": "Large Language Model for Science: A Study on P vs. NP",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=40207489",
         "created_at": "2024-04-30T05:20:16Z"
       },
       {
         "hn_id": "39934427",
         "title": "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena",
         "points": 1,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=39934427",
         "created_at": "2024-04-04T18:56:42Z"
       }
     ],
     "top_points": 281,
     "total_points": 366,
     "total_comments": 50
   }
 }