ai-research-survey

scan-v5.json (29321B)

---

 {
   "scan_version": 5,
   "paper_type": "empirical",
   "paper": {
     "title": "A Library of LLM Intrinsics for Retrieval-Augmented Generation",
     "authors": [
       "Marina Danilevsky",
       "Kristjan H. Greenewald",
       "Chulaka Gunasekara",
       "Maeda F Hanafi",
       "Lihong He",
       "Yannis Katsis",
       "Krishnateja Killamsetty",
       "Yulong Li",
       "Yatin Nandwani",
       "Lucian Popa",
       "Dinesh Raghu",
       "Frederick Reiss",
       "Vraj Shah",
       "Khoi-Nguyen Tran",
       "Huaiyu Zhu",
       "Luis Lastras"
     ],
     "year": 2025,
     "venue": "arXiv.org",
     "arxiv_id": "2504.11704",
     "doi": "10.48550/arXiv.2504.11704"
   },
   "checklist": {
     "claims_and_evidence": {
       "abstract_claims_supported": {
         "applies": true,
         "answer": true,
         "justification": "The abstract claims the library is released as LoRA adapters on HuggingFace with a software interface on vLLM; the paper substantiates this with HuggingFace links, GitHub references, and evaluations for all 8 intrinsics.",
         "source": "haiku"
       },
       "causal_claims_justified": {
         "applies": true,
         "answer": true,
         "justification": "Causal claims like 'query rewrite improves retrieval recall' are tested via controlled experiments holding the retriever fixed and varying only the query strategy, with multiple benchmarks and baseline comparisons.",
         "source": "haiku"
       },
       "generalization_bounded": {
         "applies": true,
         "answer": false,
         "justification": "The paper proposes a general 'software library' concept for LLMs but all implementations are specific to IBM Granite 3.3-8b models; the conclusion does not bound the generalizability of results to other model families or retrieval backends.",
         "source": "haiku"
       },
       "alternative_explanations_discussed": {
         "applies": true,
         "answer": false,
         "justification": "The paper does not discuss whether performance gains stem from LoRA fine-tuning specifically, the training data quality, or simply having a purpose-built model vs. prompting; no alternative explanations for the observed improvements are entertained.",
         "source": "haiku"
       },
       "proxy_outcome_distinction": {
         "applies": true,
         "answer": true,
         "justification": "Metrics are clearly labeled to match what is claimed: Recall@k for retrieval quality, RAGAS Faithfulness for answer faithfulness, and precision/recall/F1 for classification tasks; the paper does not conflate these with higher-order outcomes like user satisfaction.",
         "source": "haiku"
       }
     },
     "limitations_and_scope": {
       "limitations_section_present": {
         "applies": true,
         "answer": false,
         "justification": "There is no dedicated limitations or threats-to-validity section; the paper ends with a conclusion section that does not discuss failure modes or scope restrictions.",
         "source": "haiku"
       },
       "threats_to_validity_specific": {
         "applies": true,
         "answer": false,
         "justification": "No specific threats are discussed; potential issues like train-test overlap (Context Relevance trained and evaluated on CLAPNQ), self-evaluation bias from IBM evaluating IBM models, or RAGAS-F as an imperfect proxy are not mentioned.",
         "source": "haiku"
       },
       "scope_boundaries_stated": {
         "applies": true,
         "answer": false,
         "justification": "The paper does not explicitly state what the results do NOT show; for instance, it does not clarify that performance claims are specific to IBM Granite models or that the 'intrinsic' framework has not been validated across other LLM providers.",
         "source": "haiku"
       }
     },
     "conflicts_of_interest": {
       "funding_disclosed": {
         "applies": true,
         "answer": false,
         "justification": "No funding disclosure appears anywhere in the paper; the acknowledgments section only thanks annotators.",
         "source": "haiku"
       },
       "affiliations_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "All 16 authors are explicitly listed as IBM Research, and the paper is clearly labeled 'IBM Research AI' throughout.",
         "source": "haiku"
       },
       "funder_independent_of_outcome": {
         "applies": true,
         "answer": false,
         "justification": "IBM employees are evaluating IBM Granite models and proposing IBM's Granite IO framework as the recommended interface; the funder/employer has a direct commercial interest in the favorable outcome.",
         "source": "haiku"
       },
       "financial_interests_declared": {
         "applies": true,
         "answer": false,
         "justification": "No competing interests statement or declaration of financial interests (patents, equity, consulting) is included in the paper.",
         "source": "haiku"
       }
     },
     "scope_and_framing": {
       "key_terms_defined": {
         "applies": true,
         "answer": true,
         "justification": "The core concept 'LLM intrinsic' is explicitly defined in the introduction as 'a capability that can be invoked through a well-defined API that is reasonably stable and independent of how the LLM intrinsic itself is implemented.'",
         "source": "haiku"
       },
       "intended_contribution_clear": {
         "applies": true,
         "answer": true,
         "justification": "The paper clearly states its contribution is a library of 8 RAG intrinsics implemented as LoRA adapters released on HuggingFace and through Granite IO, with training details and evaluations for each.",
         "source": "haiku"
       },
       "engagement_with_prior_work": {
         "applies": true,
         "answer": true,
         "justification": "The paper references and builds on prior work including MT-RAG, RAGTruth, ALCE, LongBench-Cite, BEIR, and Thermometer calibration, situating each intrinsic relative to existing methods it outperforms or complements.",
         "source": "haiku"
       }
     }
   },
   "type_checklist": {
     "empirical": {
       "artifacts": {
         "code_released": {
           "applies": true,
           "answer": true,
           "justification": "LoRA adapters are released on HuggingFace (granite-3.3-8b-rag-agent-lib) and Granite IO software interface is released on GitHub (ibm-granite/granite-io) with example notebooks.",
           "source": "haiku"
         },
         "data_released": {
           "applies": true,
           "answer": true,
           "justification": "All evaluation datasets (MT-RAG, BEIR, MMLU, SQuAD, CLAPNQ, etc.) are standard public benchmarks used unmodified; training data for query rewrite is proprietary but evaluation data is public.",
           "source": "haiku"
         },
         "environment_specified": {
           "applies": true,
           "answer": false,
           "justification": "The paper mentions vLLM as the inference platform and PEFT for fine-tuning but provides no requirements file, Dockerfile, or specific version numbers for the software environment.",
           "source": "haiku"
         },
         "reproduction_instructions": {
           "applies": true,
           "answer": false,
           "justification": "The paper redirects readers to external HuggingFace documentation and GitHub notebooks rather than providing step-by-step reproduction instructions within the paper itself.",
           "source": "haiku"
         }
       },
       "statistical_methodology": {
         "confidence_intervals_or_error_bars": {
           "applies": true,
           "answer": false,
           "justification": "All results in all tables are reported as point estimates only; no confidence intervals, error bars, or standard deviations are provided for any metric.",
           "source": "haiku"
         },
         "significance_tests": {
           "applies": true,
           "answer": false,
           "justification": "No statistical significance tests are used for any of the comparative claims despite tables showing close performance differences (e.g., LoRA vs. Mixtral in Table 2).",
           "source": "haiku"
         },
         "effect_sizes_reported": {
           "applies": true,
           "answer": true,
           "justification": "The paper reports percentage-point improvements throughout, e.g., '9 percentage points jump in Recall@20' and '22 percentage points improvement on non-standalone subset', with baselines for context.",
           "source": "haiku"
         },
         "sample_size_justified": {
           "applies": true,
           "answer": false,
           "justification": "The benchmark sizes (842 MT-RAG points, specific BEIR subsets) are inherited from existing benchmarks without any power analysis or justification for whether these are sufficient.",
           "source": "haiku"
         },
         "variance_reported": {
           "applies": true,
           "answer": false,
           "justification": "No variance, standard deviation, or results across multiple runs are reported; all metrics appear to be single-run point estimates.",
           "source": "haiku"
         }
       },
       "evaluation_design": {
         "baselines_included": {
           "applies": true,
           "answer": true,
           "justification": "Every intrinsic is compared against multiple baselines: no-rewrite/vanilla Granite instruct, Mixtral, Llama-3.3-70B, Granite Guardian, and sometimes gold human references.",
           "source": "haiku"
         },
         "baselines_contemporary": {
           "applies": true,
           "answer": true,
           "justification": "Baselines include Llama-3.3-70B-Instruct, Mixtral-8x22B-Instruct, and Granite 3.3-8b-instruct, all current competitive models at time of writing.",
           "source": "haiku"
         },
         "ablation_study": {
           "applies": true,
           "answer": true,
           "justification": "Section 10 systematically evaluates four composite flows (None, QR only, AD only, QR+AD) measuring the contribution of each component to answerability classification, faithfulness, and JAFS.",
           "source": "haiku"
         },
         "multiple_metrics": {
           "applies": true,
           "answer": true,
           "justification": "The paper uses Recall@k, RAGAS Faithfulness, RAD-Bench, NDCG@10, precision/recall/F1 by class, ECE, MAE, and the joint JAFS metric across different intrinsics.",
           "source": "haiku"
         },
         "human_evaluation": {
           "applies": true,
           "answer": false,
           "justification": "No human evaluation of system outputs is conducted; gold rewrites are used as reference baselines but do not constitute human evaluation of the intrinsic outputs.",
           "source": "haiku"
         },
         "held_out_test_set": {
           "applies": true,
           "answer": true,
           "justification": "Evaluation is performed on held-out test splits of established benchmarks (MMLU, SQuAD, MT-RAG eval split, BEIR test sets) separate from training data.",
           "source": "haiku"
         },
         "per_category_breakdown": {
           "applies": true,
           "answer": true,
           "justification": "Context Relevance results are broken down by 8 individual datasets (Tables 10-11); Passage Reranking results show per-task NDCG@10 across 15 BEIR subtasks (Table 15); UQ ECE is shown per MMLU task (Figure 1).",
           "source": "haiku"
         },
         "failure_cases_discussed": {
           "applies": true,
           "answer": false,
           "justification": "The paper notes potential confusing composite behaviors (high UQ + low HD scores) conceptually but does not show or analyze any actual failure cases or error examples.",
           "source": "haiku"
         },
         "negative_results_reported": {
           "applies": true,
           "answer": true,
           "justification": "Table 15 shows Passage Reranking hurts performance on several BEIR tasks (touche2020, scidocs, arguana, quora, fever, climatefever); the paper acknowledges this and offers a hypothesis about annotation quality.",
           "source": "haiku"
         }
       },
       "setup_transparency": {
         "model_versions_specified": {
           "applies": true,
           "answer": true,
           "justification": "Exact model versions are specified: ibm-granite/granite-3.3-8b-instruct, meta-llama/Llama-3.3-70B-Instruct, Mixtral 8x7b, ibm-granite/granite-guardian-3.1-5b, etc.",
           "source": "haiku"
         },
         "prompts_provided": {
           "applies": true,
           "answer": true,
           "justification": "Actual prompts are provided for several intrinsics including the full rewrite role prompt (Section 2.1), the passage reranking default prompt (Section 6.1), and hallucination detection instruction string (Section 8.1).",
           "source": "haiku"
         },
         "hyperparameters_reported": {
           "applies": true,
           "answer": true,
           "justification": "LoRA hyperparameters are reported for each adapter: rank (8, 16, or 32), learning rate (1e-5 to 5e-6), number of epochs (25), and train/validation split (90/10).",
           "source": "haiku"
         },
         "scaffolding_described": {
           "applies": true,
           "answer": true,
           "justification": "Section 10 and Figure 2 describe the composite intrinsic flows in detail including how QR and AD chain together, what triggers the 'I don't know' response, and how retrieved passages flow between components.",
           "source": "haiku"
         },
         "data_preprocessing_documented": {
           "applies": true,
           "answer": true,
           "justification": "Preprocessing steps are documented: NLTK sentence splitting for citation/hallucination tasks, Mixtral as automatic judge for label validation, apply_chat_template for model inputs, and 90/10 train/val splits.",
           "source": "haiku"
         }
       },
       "data_integrity": {
         "raw_data_available": {
           "applies": true,
           "answer": false,
           "justification": "Only aggregate metrics are reported; individual model predictions, intermediate outputs, or per-example results are not released for independent verification.",
           "source": "haiku"
         },
         "data_collection_described": {
           "applies": true,
           "answer": true,
           "justification": "Training data construction is described for each adapter (e.g., human-annotated Cloud corpus for QR, synthetically generated via Mixtral for context relevance, CoQA/MultiDoc2Dial/QuAC as seed corpora for HD and CG).",
           "source": "haiku"
         },
         "recruitment_methods_described": {
           "applies": false,
           "answer": false,
           "justification": "Standard public benchmarks are used for evaluation; no participant recruitment was involved.",
           "source": "haiku"
         },
         "data_pipeline_documented": {
           "applies": true,
           "answer": true,
           "justification": "The multi-step data pipeline is described for training data: corpus selection, conversation generation, label generation (via Mixtral judge), filtering, LoRA fine-tuning with PEFT.",
           "source": "haiku"
         }
       },
       "contamination": {
         "training_cutoff_stated": {
           "applies": true,
           "answer": false,
           "justification": "The base model (Granite 3.3-8b-instruct) training data cutoff is not stated, making it impossible to assess whether evaluation benchmarks were in the pre-training data.",
           "source": "haiku"
         },
         "train_test_overlap_discussed": {
           "applies": true,
           "answer": false,
           "justification": "Context Relevance is trained on CLAPNQ data and evaluated on CLAPNQ; Answerability Determination is trained on the MT-RAG Government corpus and evaluated on MT-RAG benchmark; this overlap is never discussed.",
           "source": "haiku"
         },
         "benchmark_contamination_addressed": {
           "applies": true,
           "answer": false,
           "justification": "No discussion of whether the base Granite model's pre-training included evaluation benchmark examples (MMLU, SQuAD, BEIR datasets are widely used and likely present in pre-training data).",
           "source": "haiku"
         }
       },
       "human_studies": {
         "pre_registered": {
           "applies": false,
           "answer": false,
           "justification": "No human participants in this study.",
           "source": "haiku"
         },
         "irb_or_ethics_approval": {
           "applies": false,
           "answer": false,
           "justification": "No human participants in this study.",
           "source": "haiku"
         },
         "demographics_reported": {
           "applies": false,
           "answer": false,
           "justification": "No human participants in this study.",
           "source": "haiku"
         },
         "inclusion_exclusion_criteria": {
           "applies": false,
           "answer": false,
           "justification": "No human participants in this study.",
           "source": "haiku"
         },
         "randomization_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants in this study.",
           "source": "haiku"
         },
         "blinding_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants in this study.",
           "source": "haiku"
         },
         "attrition_reported": {
           "applies": false,
           "answer": false,
           "justification": "No human participants in this study.",
           "source": "haiku"
         }
       },
       "cost_and_practicality": {
         "inference_cost_reported": {
           "applies": true,
           "answer": false,
           "justification": "A qualitative note about passage reranking latency ('only needs 1 preference token') is provided but no systematic latency or cost benchmarks are reported for any intrinsic.",
           "source": "haiku"
         },
         "compute_budget_stated": {
           "applies": true,
           "answer": false,
           "justification": "No GPU hours, compute budget, or training time is reported for any of the LoRA fine-tuning runs.",
           "source": "haiku"
         }
       }
     }
   },
   "claims": [
     {
       "claim": "Query Rewrite LoRA adapter achieves 9pp improvement in Recall@20 over no-rewrite on the full MT-RAG dataset (22pp improvement on the non-standalone subset).",
       "evidence": "Tables 2-4 compare Recall@5/10/20 across no-rewrite, Mixtral, Granite LoRA, and gold rewrite strategies on full, non-standalone, and standalone MT-RAG subsets.",
       "supported": "strong"
     },
     {
       "claim": "Citation Generation LoRA (8b) achieves comparable or better citation F1 than prompted Llama-3.1-70B-Instruct (70b) on ALCE and significantly outperforms it on LongBench-Cite (68.6 vs 42.0 avg F1).",
       "evidence": "Tables 18-19 compare Recall/Precision/F1 for the LoRA adapter vs. much larger models on two citation benchmarks; LongBench-Cite shows a 26.6pp avg F1 advantage.",
       "supported": "strong"
     },
     {
       "claim": "Context Relevance LoRA improves precision on relevant labels across all 8 evaluated datasets vs. Llama-3.3-70B, Granite Guardian, and vanilla Granite instruct.",
       "evidence": "Table 10 shows LoRA precision consistently at or above 0.92 across all datasets while Granite Guardian drops as low as 0.08 and vanilla Granite instruct as low as 0.60.",
       "supported": "strong"
     },
     {
       "claim": "Uncertainty Quantification LoRA achieves an average ECE of 0.064 across MMLU tasks without degrading zero-shot accuracy (avg 89%), compared to base model ECE of 0.20.",
       "evidence": "Section 7.2 and Figure 1 report ECE per MMLU task for the LoRA vs. base Granite 3.3-8b-instruct; the LoRA ECE is smaller than the 10% quantization gap.",
       "supported": "strong"
     },
     {
       "claim": "Composite QR+AD flow achieves better overall performance (JAFS=66) than either QR alone (56) or no intrinsic (49), matching AD alone (66) while also improving faithfulness.",
       "evidence": "Tables 20-22 decompose answerability classification F1, RAGAS faithfulness, and JAFS for all four flows; QR+AD matches AD on JAFS while improving answerable query handling.",
       "supported": "moderate"
     },
     {
       "claim": "LLM intrinsics can be implemented as a stable, model-agnostic API analogous to a software library, enabling modular RAG pipelines independent of backend implementation.",
       "evidence": "Conceptual argument in Sections 1 and 11; the claim is not empirically tested across multiple LLM providers — all implementations use IBM Granite models only.",
       "supported": "weak"
     }
   ],
   "methodology_tags": [
     "benchmark-eval",
     "case-study"
   ],
   "key_findings": "The paper introduces 8 RAG intrinsics as LoRA adapters for IBM Granite 3.3-8b, released on HuggingFace under Apache 2.0. Fine-tuned small LoRA adapters (8b) consistently match or outperform prompted much larger models (70b) on RAG-specific tasks including citation generation, context relevance, and answerability determination, with Citation Generation achieving 68.6 vs. 42.0 avg F1 on LongBench-Cite against prompted Llama-3.1-70B. The composite intrinsic evaluation demonstrates that simple accuracy metrics can be misleading for composite flows — the joint answerability-faithfulness score (JAFS) is needed to see that QR+AD (66) outperforms QR alone (56) even when RAGAS-F alone suggests QR is superior. Notable contamination concern: several adapters are trained and evaluated on overlapping data (CLAPNQ, MT-RAG) without discussion.",
   "red_flags": [
     {
       "flag": "Train-test contamination undisclosed",
       "detail": "Context Relevance is trained on CLAPNQ (Section 4.3) and evaluated on CLAPNQ (Table 10-11). Answerability Determination is trained on the MT-RAG Government corpus (Section 5.3) and evaluated on the MT-RAG Benchmark (Table 13). This overlap is never acknowledged or discussed."
     },
     {
       "flag": "Self-evaluation: IBM evaluates IBM products",
       "detail": "All 16 authors are IBM Research employees evaluating IBM Granite models with IBM Granite IO as the recommended deployment framework, with no independent third-party evaluation and no funding disclosure."
     },
     {
       "flag": "No confidence intervals or significance tests",
       "detail": "All performance numbers are point estimates across tables; differences of 1-2 percentage points are discussed as meaningful improvements without any statistical testing."
     },
     {
       "flag": "No limitations section",
       "detail": "The paper has no dedicated limitations discussion; failure modes, scope restrictions, or cases where the intrinsics would not be expected to work are not addressed."
     },
     {
       "flag": "Proprietary training data for key intrinsic",
       "detail": "Query Rewrite training data is described as proprietary, obtained through a third-party annotation company (Section 2.3), preventing reproduction or independent assessment of data quality."
     }
   ],
   "cited_papers": [
     {
       "title": "MTRAG: A Multi-Turn Conversational Benchmark for Evaluating Retrieval-Augmented Generation Systems",
       "relevance": "Primary evaluation benchmark used across multiple intrinsics; provides MT-RAG dataset for QR, AD, and composite flow evaluation"
     },
     {
       "title": "Enabling Large Language Models to Generate Text with Citations (ALCE)",
       "relevance": "Citation generation benchmark used to evaluate the Citation Generation intrinsic against prompted larger models"
     },
     {
       "title": "LongCite: Enabling LLMs to Generate Fine-Grained Citations in Long-Context QA (LongBench-Cite)",
       "relevance": "Second citation benchmark showing LoRA significantly outperforms prompted 70B models on span-level citations"
     },
     {
       "title": "RAGTruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language Models",
       "relevance": "Benchmark used to evaluate Hallucination Detection intrinsic; provides baseline results from the original paper"
     },
     {
       "title": "BEIR: A Heterogeneous Benchmark for Zero-Shot Evaluation of Information Retrieval Models",
       "relevance": "Used to evaluate Passage Reranking intrinsic across 15 retrieval tasks including NDCG@10"
     },
     {
       "title": "Thermometer: Towards Universal Calibration for Large Language Models",
       "relevance": "The calibration method that UQ LoRA is trained to mimic; foundational methodology for the Uncertainty Quantification intrinsic"
     },
     {
       "title": "CLAPnq: Cohesive Long-Form Answers from Passages in Natural Questions for RAG Systems",
       "relevance": "Dataset used for both training and evaluation of Context Relevance intrinsic (raises contamination concern)"
     },
     {
       "title": "Activated LoRA: Fine-tuned LLMs for Intrinsics",
       "relevance": "Sister paper describing the activated LoRA mechanism as an efficient implementation approach for LLM intrinsics"
     },
     {
       "title": "Know What You Don't Know: Unanswerable Questions for SQuAD (SQuADRun)",
       "relevance": "Benchmark for evaluating Answerability Determination in single-turn setting"
     }
   ],
   "engagement_factors": {
     "practical_relevance": {
       "score": 3,
       "justification": "LoRA adapters and Granite IO are publicly released under Apache 2.0; practitioners can immediately deploy these 8 RAG intrinsics via HuggingFace or GitHub notebooks."
     },
     "surprise_contrarian": {
       "score": 1,
       "justification": "Fine-tuned small models outperforming prompted large models is a well-established finding; no unexpected or counter-intuitive results are presented."
     },
     "fear_safety": {
       "score": 1,
       "justification": "The hallucination detection and answerability determination intrinsics address reliability concerns, but no AI risk framing is used."
     },
     "drama_conflict": {
       "score": 0,
       "justification": "No controversy, no challenge to dominant paradigms, no conflict with other researchers."
     },
     "demo_ability": {
       "score": 3,
       "justification": "Code, models, and notebooks are publicly available on HuggingFace and GitHub; anyone can run the intrinsics immediately with vLLM and Granite IO."
     },
     "brand_recognition": {
       "score": 2,
       "justification": "IBM Research is a recognizable AI lab and IBM Granite is a known open model family, though neither is at the same recognition level as OpenAI, Google, or Meta."
     }
   },
   "hn_data": {
     "threads": [
       {
         "hn_id": "44770561",
         "title": "B-Splines and Fourier-Best Friends for Spatial-Temporal Video Super-Resolution",
         "points": 4,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=44770561"
       },
       {
         "hn_id": "43872030",
         "title": "Linguistics Learned to Stop Worrying and Love the Language Models",
         "points": 3,
         "comments": 1,
         "url": "https://news.ycombinator.com/item?id=43872030"
       },
       {
         "hn_id": "40106382",
         "title": "Automated Social Science: Language Models as Scientist and Subjects",
         "points": 3,
         "comments": 1,
         "url": "https://news.ycombinator.com/item?id=40106382"
       },
       {
         "hn_id": "43773523",
         "title": "Robotic Squirrel Pinto: A latched spring actuated robot for jumping and perching",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=43773523"
       },
       {
         "hn_id": "45259423",
         "title": "Human+AI loops stay stable even with quantization",
         "points": 2,
         "comments": 1,
         "url": "https://news.ycombinator.com/item?id=45259423"
       },
       {
         "hn_id": "46145768",
         "title": "Conversational Networks",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=46145768"
       },
       {
         "hn_id": "44739937",
         "title": "Double Duty: FPGA Architecture to Enable Concurrent LUT and Adder Chain Usage",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=44739937"
       },
       {
         "hn_id": "44375110",
         "title": "Gender and Positional Biases in LLM-Based Hiring Decisions: Evidence Found",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=44375110"
       },
       {
         "hn_id": "44242780",
         "title": "Talk to Your Slides: Efficient Slide Editing Agent with Large Language Models",
         "points": 1,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=44242780"
       },
       {
         "hn_id": "43843140",
         "title": "Physical Principles of Quantum Biology",
         "points": 1,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=43843140"
       }
     ],
     "top_points": 4,
     "total_points": 22,
     "total_comments": 3
   }
 }