ai-research-survey

scan-v5.json (28897B)

---

 {
   "scan_version": 5,
   "paper_type": "empirical",
   "paper": {
     "title": "Input Reduction Enhanced LLM-based Program Repair",
     "authors": [
       "Boyang Yang",
       "Luyao Ren",
       "Xin Yin",
       "Jiadong Ren",
       "Haoye Tian"
     ],
     "year": 2025,
     "venue": "arXiv.org",
     "arxiv_id": "2507.15251",
     "doi": "10.48550/arXiv.2507.15251"
   },
   "checklist": {
     "claims_and_evidence": {
       "abstract_claims_supported": {
         "applies": true,
         "answer": true,
         "justification": "All abstract claims are verified: 89.1% average input reduction (Table 3), 53.8% relative pass@10 improvement over Origin Test (GLM overall 6.5%→10.0%), 17.6% over Baseline (GLM 8.5%→10.0%), ChatRepair +21.3% and CREF +2.6% (Tables 10–11).",
         "source": "haiku"
       },
       "causal_claims_justified": {
         "applies": true,
         "answer": true,
         "justification": "The paper makes causal claims that input reduction improves repair accuracy; these are backed by controlled ablations (RQ-3 isolating length vs. information), statistical MWW tests (p<0.05), and plug-in integration experiments that hold all other variables fixed.",
         "source": "haiku"
       },
       "generalization_bounded": {
         "applies": true,
         "answer": true,
         "justification": "Claims are explicitly bounded to LLM-based APR with long failure-inducing test inputs; the threats-to-validity section acknowledges the competitive-programming domain and validates on OSS-Fuzz for broader applicability, without overclaiming universal generalization.",
         "source": "haiku"
       },
       "alternative_explanations_discussed": {
         "applies": true,
         "answer": true,
         "justification": "RQ-3 specifically investigates whether gains come from shorter prompts alone vs. preserved failure information; five prompt variants (Diff Lines, Reduced+Origin, etc.) systematically rule out length-only explanations.",
         "source": "haiku"
       },
       "proxy_outcome_distinction": {
         "applies": true,
         "answer": true,
         "justification": "pass@k is defined as patch correctness against the full official test suite; the paper does not conflate this metric with broader software quality or developer productivity, staying within the measured granularity.",
         "source": "haiku"
       }
     },
     "limitations_and_scope": {
       "limitations_section_present": {
         "applies": true,
         "answer": true,
         "justification": "Section 6 'Threats to Validity' contains three dedicated subsections (Internal, Construct, External validity), each with specific threats and mitigations.",
         "source": "haiku"
       },
       "threats_to_validity_specific": {
         "applies": true,
         "answer": true,
         "justification": "Specific threats are identified: compression ratio reaching 100% hiding variation (addressed by reporting both mean and median), dataset restriction to AtCoder (mitigated with OSS-Fuzz), stochasticity addressed by pass@k sampling — these are concrete, not boilerplate.",
         "source": "haiku"
       },
       "scope_boundaries_stated": {
         "applies": true,
         "answer": true,
         "justification": "The paper explicitly scopes evaluation to long failure-inducing inputs and notes 'ReduceFix might benefit only the pipeline evaluated in this study and fail to transfer,' testing transferability via ChatRepair and CREF plug-in experiments.",
         "source": "haiku"
       }
     },
     "conflicts_of_interest": {
       "funding_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "Acknowledgments section lists three explicit funding sources: National Natural Science Foundation of China (62273292), Central Leading Local Science and Technology Development Project of Hebei Province (246Z0804G), and Hebei Innovation Capability Improvement Plan Project (22567626H).",
         "source": "haiku"
       },
       "affiliations_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "All author affiliations are listed on the first page: Yanshan University, Peking University, Zhejiang University, Aalto University, and Yanshan University.",
         "source": "haiku"
       },
       "funder_independent_of_outcome": {
         "applies": true,
         "answer": true,
         "justification": "All funders are Chinese government science foundations with no commercial interest in the ReduceFix tool or the LLMs evaluated.",
         "source": "haiku"
       },
       "financial_interests_declared": {
         "applies": true,
         "answer": false,
         "justification": "No competing interests or conflicts-of-interest statement is included anywhere in the paper.",
         "source": "haiku"
       }
     },
     "scope_and_framing": {
       "key_terms_defined": {
         "applies": true,
         "answer": true,
         "justification": "Key terms are defined: APR is introduced in the opening sentence, 'lost-in-the-middle' is explained with citation, compression rate is formally defined in Eq. 2, and pass@k is defined with its formula in Section 4.3.",
         "source": "haiku"
       },
       "intended_contribution_clear": {
         "applies": true,
         "answer": true,
         "justification": "Section 1 enumerates four explicit contributions: (1) ReduceFix framework, (2) LFTBench benchmark, (3) comprehensive evaluation results, (4) plug-in integration into ChatRepair and CREF.",
         "source": "haiku"
       },
       "engagement_with_prior_work": {
         "applies": true,
         "answer": true,
         "justification": "Section 7 provides a structured related work covering LLM-based APR and test input reduction, explicitly positioning ReduceFix as the first approach combining both and showing how it differs from ddmin-family methods, HDD, Perses, and LPR.",
         "source": "haiku"
       }
     }
   },
   "type_checklist": {
     "empirical": {
       "artifacts": {
         "code_released": {
           "applies": true,
           "answer": true,
           "justification": "Section 4.4 explicitly states 'the full artifact is published at https://github.com/GLEAM-Lab/ReduceFix,' and the README is referenced for reducer visualizations.",
           "source": "haiku"
         },
         "data_released": {
           "applies": true,
           "answer": true,
           "justification": "LFTBench and LFTBench-Py are stated as released in the contributions section; the GitHub artifact URL implies data availability alongside code.",
           "source": "haiku"
         },
         "environment_specified": {
           "applies": true,
           "answer": false,
           "justification": "No requirements.txt, Dockerfile, or dependency specification is provided in the paper; only hyperparameters (Table 2) and hardware category ('single 24 GB consumer GPU') are mentioned.",
           "source": "haiku"
         },
         "reproduction_instructions": {
           "applies": true,
           "answer": false,
           "justification": "No step-by-step reproduction instructions appear in the paper itself; the GitHub README is referenced but not its content, and the paper provides only algorithm pseudocode.",
           "source": "haiku"
         }
       },
       "statistical_methodology": {
         "confidence_intervals_or_error_bars": {
           "applies": true,
           "answer": false,
           "justification": "All pass@k results are reported as single point estimates in tables; no confidence intervals or error bars appear anywhere in the paper.",
           "source": "haiku"
         },
         "significance_tests": {
           "applies": true,
           "answer": true,
           "justification": "Section 5.3 applies two-sided Mann-Whitney-Wilcoxon tests to compare ReduceFix against Origin Test, reporting p < 0.05 for key comparisons including the ddmin-only gap.",
           "source": "haiku"
         },
         "effect_sizes_reported": {
           "applies": true,
           "answer": true,
           "justification": "Both absolute and relative improvements are consistently reported (e.g., 'pass@10 rises from 30.5% to 37.0%, an absolute gain of 6.5 percentage points and a relative gain of 21.3%').",
           "source": "haiku"
         },
         "sample_size_justified": {
           "applies": true,
           "answer": false,
           "justification": "The 200-bug, 20-task sample size is motivated by the availability of AtCoder data after LLM cutoff dates, but no power analysis or statistical justification for sufficiency is provided.",
           "source": "haiku"
         },
         "variance_reported": {
           "applies": true,
           "answer": false,
           "justification": "Pass@k figures are reported as single values with no standard deviation, confidence intervals, or run-to-run variance; the reducer uses temperature=0 (deterministic), but repair sampling variance is not reported.",
           "source": "haiku"
         }
       },
       "evaluation_design": {
         "baselines_included": {
           "applies": true,
           "answer": true,
           "justification": "Multiple baselines are tested: Baseline (no test), Origin Test (full test), ddmin-only, and pure-LLM reduction, plus ChatRepair and CREF with and without ReduceFix.",
           "source": "haiku"
         },
         "baselines_contemporary": {
           "applies": true,
           "answer": true,
           "justification": "ChatRepair (ISSTA 2024) and CREF (ISSTA 2024) are recent and competitive APR systems; ddmin is classical but appropriate as a known algorithm baseline for the reduction task.",
           "source": "haiku"
         },
         "ablation_study": {
           "applies": true,
           "answer": true,
           "justification": "RQ-3 (Section 5.4) is an explicit ablation with five prompt variants to isolate the contributions of length reduction vs. information selection.",
           "source": "haiku"
         },
         "multiple_metrics": {
           "applies": true,
           "answer": true,
           "justification": "The paper reports pass@1, pass@5, pass@10, reduction success rate, mean and median compression rate, and token/cost comparisons.",
           "source": "haiku"
         },
         "human_evaluation": {
           "applies": false,
           "answer": false,
           "justification": "Automated test suite evaluation is the appropriate measure for APR; no human evaluation is needed or applicable.",
           "source": "haiku"
         },
         "held_out_test_set": {
           "applies": true,
           "answer": true,
           "justification": "The full official AtCoder hidden test archive is used for final validation; the failure-inducing input used in the prompt is distinct from the full test suite used to judge patch correctness.",
           "source": "haiku"
         },
         "per_category_breakdown": {
           "applies": true,
           "answer": true,
           "justification": "Results are broken down by task difficulty (C, D, E&F in Table 6) and by input format (6 categories in Table 4), as well as by project for OSS-Fuzz (Tables 12–13).",
           "source": "haiku"
         },
         "failure_cases_discussed": {
           "applies": true,
           "answer": true,
           "justification": "Section 5.2 includes a dedicated 'Failed Case Study' for ABC372E where vertex renumbering masked the defect, with a proposed fix described in detail.",
           "source": "haiku"
         },
         "negative_results_reported": {
           "applies": true,
           "answer": true,
           "justification": "The paper explicitly reports that including the full failing test (Origin Test) hurts performance relative to no test for GLM-4-9B-chat (8.5%→6.5%) and DeepSeek-V3 (66.5%→63.0%).",
           "source": "haiku"
         }
       },
       "setup_transparency": {
         "model_versions_specified": {
           "applies": true,
           "answer": false,
           "justification": "Model names are given (Qwen2.5-Coder-7B-instruct, GLM-4-9B-chat, DeepSeek-V3, Qwen2.5-Plus) but no API snapshot dates or commit hashes are provided; Qwen2.5-Plus is a cloud service with non-disclosed parameter count and no versioning timestamp.",
           "source": "haiku"
         },
         "prompts_provided": {
           "applies": true,
           "answer": true,
           "justification": "Listing 1 provides the full one-shot reducer prompt template with all placeholder variables labeled; the repair prompt structure is also described in Section 3.4.",
           "source": "haiku"
         },
         "hyperparameters_reported": {
           "applies": true,
           "answer": true,
           "justification": "Table 2 lists all hyperparameters: temperature 0.0 for reduction, 0.8 for repair, 60s wall-clock limit, 10s compilation timeout, 5s execution timeout per test case, k∈{1,5,10}.",
           "source": "haiku"
         },
         "scaffolding_described": {
           "applies": true,
           "answer": true,
           "justification": "Algorithm 1 provides the complete three-stage pipeline control logic; Sections 3.2–3.4 describe each stage in detail including reducer generation, iterative reduction, and patch validation loops.",
           "source": "haiku"
         },
         "data_preprocessing_documented": {
           "applies": true,
           "answer": true,
           "justification": "Section 4.2 describes the full benchmark construction: AtCoder ABC 361–377 selection, test size filter (≥4 KB), difficulty filter (C–F), manual collection of wrong-answer submissions before July 1, 2025.",
           "source": "haiku"
         }
       },
       "data_integrity": {
         "raw_data_available": {
           "applies": true,
           "answer": true,
           "justification": "The GitHub artifact at GLEAM-Lab/ReduceFix is stated to contain the full artifact, including LFTBench and LFTBench-Py; OSS-Fuzz data reuses publicly available ARVO scripts.",
           "source": "haiku"
         },
         "data_collection_described": {
           "applies": true,
           "answer": true,
           "justification": "Section 4.2 describes the collection procedure in detail: AtCoder tasks selected by contest number range, difficulty level, test file size threshold, and manual identification of failing submissions before a cutoff date.",
           "source": "haiku"
         },
         "recruitment_methods_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants; LFTBench is constructed from publicly available AtCoder contest data with no subject recruitment.",
           "source": "haiku"
         },
         "data_pipeline_documented": {
           "applies": true,
           "answer": true,
           "justification": "The pipeline from AtCoder test archives → filtering → manual submission collection → benchmark packaging is described; ARVO's Docker-based data and scripts are reused for OSS-Fuzz.",
           "source": "haiku"
         }
       },
       "contamination": {
         "training_cutoff_stated": {
           "applies": true,
           "answer": true,
           "justification": "Section 4.2 explicitly states LFTBench covers ABC 361–377, 'a span entirely after the knowledge cut-offs of the 4 LLMs we evaluate,' directly addressing training cutoffs.",
           "source": "haiku"
         },
         "train_test_overlap_discussed": {
           "applies": true,
           "answer": true,
           "justification": "Section 4.2 explicitly motivates the post-cutoff design by noting that existing benchmarks 'were released years ago' and 'large language models have almost certainly seen,' providing quantitative context for the leakage concern.",
           "source": "haiku"
         },
         "benchmark_contamination_addressed": {
           "applies": true,
           "answer": true,
           "justification": "For OSS-Fuzz, the paper acknowledges potential overlap but argues relative comparisons remain valid since the same LLM and instances are held fixed across strategies; LFTBench is explicitly designed to be post-cutoff.",
           "source": "haiku"
         }
       },
       "human_studies": {
         "pre_registered": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "irb_or_ethics_approval": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "demographics_reported": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "inclusion_exclusion_criteria": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "randomization_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "blinding_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "attrition_reported": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         }
       },
       "cost_and_practicality": {
         "inference_cost_reported": {
           "applies": true,
           "answer": true,
           "justification": "Table 5 reports exact API costs: ReduceFix costs $0.017 vs. pure-LLM $0.632 on 20 problems; Section 4.1 lists per-token pricing for Qwen2.5-Plus and DeepSeek-V3.",
           "source": "haiku"
         },
         "compute_budget_stated": {
           "applies": true,
           "answer": false,
           "justification": "Only API costs for one reduction comparison are reported; total compute for all experiments across 200 bugs × 4 LLMs × 3 conditions is not stated, and local GPU experiments have no wall-clock totals.",
           "source": "haiku"
         }
       }
     }
   },
   "claims": [
     {
       "claim": "ReduceFix successfully reduces 95% of 200 LFTBench bugs with 89.1% average compression rate",
       "evidence": "Table 3 reports 95.0% overall success rate and mean/median compression of 89.1%/100.0%; all 200 syntactically valid reducers were generated.",
       "supported": "strong"
     },
     {
       "claim": "Providing reduced test inputs improves pass@10 by up to 53.8% relative to full test inputs across LLMs",
       "evidence": "Table 6: GLM-4-9B-chat improves from 6.5% (Origin Test) to 10.0% (Reduced Test) overall pass@10, a 53.8% relative gain; gains are consistent across all 4 LLMs.",
       "supported": "strong"
     },
     {
       "claim": "Including the full unmodified failing test often hurts repair accuracy below the no-test baseline",
       "evidence": "Table 6: GLM-4-9B-chat drops from 8.5% (Baseline) to 6.5% (Origin Test) pass@10; DeepSeek-V3 drops from 66.5% to 63.0%.",
       "supported": "strong"
     },
     {
       "claim": "Both compact length and complete failure evidence are required; neither alone suffices",
       "evidence": "Table 9 (RQ-3): Diff Lines (sparse info, short prompt) achieves 20.0% pass@10; Reduced+Origin (complete info, long prompt) achieves 19.0%; Reduced Test (both) achieves 25.5%.",
       "supported": "strong"
     },
     {
       "claim": "ReduceFix integrates as a drop-in plug-in and improves ChatRepair by 21.3% and CREF by 2.6% relative pass@10",
       "evidence": "Tables 10–11: ChatRepair 30.5%→37.0% (+21.3% relative); CREF 39.0%→40.0% (+2.6% relative) on LFTBench.",
       "supported": "strong"
     },
     {
       "claim": "ReduceFix outperforms both ddmin-only (35.5% success) and pure-LLM (40.0% success) reduction baselines",
       "evidence": "Table 3 reports ddmin-only 35.5%, pure-LLM 40.0%, ReduceFix 95.0% overall reduction success rate.",
       "supported": "strong"
     },
     {
       "claim": "ReduceFix generalizes to repository-level OSS-Fuzz crashes, improving pass@10 from 16.7% (Origin Test) to 41.7%",
       "evidence": "Table 13: micro-average pass@10 rises from 16.7% (Origin Test) and 25.0% (Baseline) to 41.7% (Reduced Test) on 12 OSS-Fuzz instances with Qwen2.5-Plus.",
       "supported": "moderate"
     }
   ],
   "methodology_tags": [
     "benchmark-eval",
     "empirical"
   ],
   "key_findings": "ReduceFix demonstrates that automatically reducing long failure-inducing test inputs before LLM-based repair substantially improves patch accuracy: inputs shrink by 89.1% on average with 95% success, and pass@10 improves by up to 53.8% relative over using the full test. Crucially, including unmodified long test inputs often hurts performance below the no-test baseline, confirming the 'lost-in-the-middle' effect. Ablation studies show that both prompt length reduction and preserved failure information are independently necessary—neither alone achieves the full gain. The approach integrates as a drop-in component for existing APR systems (ChatRepair +21.3%, CREF +2.6%) and generalizes to OSS-Fuzz repository-level crashes.",
   "red_flags": [
     {
       "flag": "Tiny OSS-Fuzz dataset",
       "detail": "Repository-level generalization claims are based on only 12 OSS-Fuzz instances across 5 projects; this is too small to support confident conclusions about real-world applicability."
     },
     {
       "flag": "No confidence intervals on pass@k",
       "detail": "All pass@k results are point estimates with no variance, confidence intervals, or error bars; statistical significance is tested for one comparison (RQ-2) but not reported for most tables."
     },
     {
       "flag": "Cloud model versions unpinned",
       "detail": "Qwen2.5-Plus is a cloud API service with no snapshot date or version identifier; results may not be reproducible if the underlying model is updated."
     },
     {
       "flag": "Competitive programming domain gap",
       "detail": "LFTBench is entirely AtCoder competitive programming problems, which have reference solutions and exact output oracles not present in most real-world software bugs; the benchmark may not reflect typical APR scenarios."
     },
     {
       "flag": "Marginal gains for strong models",
       "detail": "For DeepSeek-V3, improvement over Baseline is 45.2%→45.9% pass@1 and 66.5%→67.0% pass@10 — within statistical noise range despite the MWW test; the approach appears most valuable for weaker models."
     }
   ],
   "cited_papers": [
     {
       "title": "Automated Program Repair via Conversation: Fixing 162 out of 337 Bugs for $0.42 Each Using ChatGPT",
       "relevance": "Primary baseline system (ChatRepair) integrated with ReduceFix; demonstrates conversational APR with test feedback."
     },
     {
       "title": "CREF: An LLM-based Conversational Software Repair Framework for Programming Tutors",
       "relevance": "Second baseline system integrated with ReduceFix; representative of conversational repair with tutor guidance."
     },
     {
       "title": "Lost in the Middle: How Language Models Use Long Contexts",
       "relevance": "Foundational motivation for ReduceFix: documents the attention degradation in long prompts that ReduceFix targets."
     },
     {
       "title": "Simplifying and Isolating Failure-Inducing Input (ddmin)",
       "relevance": "Classical delta debugging algorithm that ReduceFix builds upon; the paper's LLM generates task-specific adaptations of ddmin."
     },
     {
       "title": "LPR: Large Language Models-Aided Program Reduction",
       "relevance": "Closest prior work to ReduceFix's approach; focuses on source code reduction rather than arbitrary test input formats."
     },
     {
       "title": "Defects4J: A Database of Existing Faults to Enable Controlled Testing Studies for Java Programs",
       "relevance": "Standard APR benchmark the paper explicitly contrasts with LFTBench, noting its short test inputs and leakage risk."
     },
     {
       "title": "Perses: Syntax-Guided Program Reduction",
       "relevance": "Grammar-based reduction baseline that motivates ReduceFix's LLM-based approach to handle diverse input formats."
     },
     {
       "title": "ARVO: Atlas of Reproducible Vulnerabilities for Open Source Software",
       "relevance": "Provides the OSS-Fuzz data, scripts, and Docker images used for the repository-level validation experiments."
     },
     {
       "title": "Automated Repair of Programs from Large Language Models",
       "relevance": "Representative LLM-based APR work; evaluates LLMs on programs they generated, establishing state of the art."
     },
     {
       "title": "Agentless: Demystifying LLM-based Software Engineering Agents",
       "relevance": "Repository-level repair framework cited for its SEARCH/REPLACE patch format used in the OSS-Fuzz experiments."
     }
   ],
   "engagement_factors": {
     "practical_relevance": {
       "score": 3,
       "justification": "ReduceFix is a drop-in component for existing APR pipelines with a public artifact, directly usable by practitioners building or extending LLM-based repair systems."
     },
     "surprise_contrarian": {
       "score": 2,
       "justification": "The finding that adding more test information (full failing test) consistently hurts repair accuracy below the no-test baseline is counterintuitive and challenges common APR prompt design assumptions."
     },
     "fear_safety": {
       "score": 0,
       "justification": "The paper addresses software reliability tooling with no safety, alignment, or misuse implications."
     },
     "drama_conflict": {
       "score": 1,
       "justification": "Mild tension with prior APR work that includes full test inputs by default; no major controversy."
     },
     "demo_ability": {
       "score": 2,
       "justification": "GitHub artifact is released with LFTBench benchmark; practitioners can run the pipeline on the benchmark, though API keys and AtCoder problem access are required."
     },
     "brand_recognition": {
       "score": 1,
       "justification": "Authors from Yanshan University, Peking University, Zhejiang University, and Aalto University — credible institutions but not famous AI labs; no well-known product affiliation."
     }
   },
   "hn_data": {
     "threads": [
       {
         "hn_id": "44309345",
         "title": "Reasoning by Superposition: A Perspective on Chain of Continuous Thought",
         "points": 60,
         "comments": 1,
         "url": "https://news.ycombinator.com/item?id=44309345"
       },
       {
         "hn_id": "44996548",
         "title": "The JWST Rocky Worlds DDT Program reveals GJ 3929B to likely be a bare rock",
         "points": 18,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=44996548"
       },
       {
         "hn_id": "44047804",
         "title": "Code Improvement Practices at Meta",
         "points": 4,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=44047804"
       },
       {
         "hn_id": "45300655",
         "title": "Generalizable Geometric Image Caption Synthesis",
         "points": 3,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=45300655"
       },
       {
         "hn_id": "36942453",
         "title": "Open Problems and Fundamental Limitations of RLHF",
         "points": 3,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=36942453"
       },
       {
         "hn_id": "44324675",
         "title": "ProtoReasoning: Prototypes as the Foundation for Generalizable Reasoning in LLMs",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=44324675"
       },
       {
         "hn_id": "43781749",
         "title": "A Comprehensive Benchmark for C-to-Safe-Rust Transpilation",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=43781749"
       },
       {
         "hn_id": "43776339",
         "title": "The Bitter Lesson Learned from 2k Multilingual Benchmarks",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=43776339"
       },
       {
         "hn_id": "45537808",
         "title": "The role of non–metricity on neutrino behavior in bumblebee gravity",
         "points": 1,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=45537808"
       },
       {
         "hn_id": "44971896",
         "title": "OS-R1: Agentic Operating System Kernel Tuning with Reinforcement Learning",
         "points": 1,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=44971896"
       }
     ],
     "top_points": 60,
     "total_points": 96,
     "total_comments": 1
   }
 }