ai-research-survey

scan-v5.json (30261B)

---

 {
   "scan_version": 5,
   "paper_type": "empirical",
   "paper": {
     "title": "Is In-Context Learning Learning?",
     "authors": [
       "Adrian de Wynter"
     ],
     "year": 2025,
     "venue": "arXiv.org",
     "arxiv_id": "2509.10414",
     "doi": "10.48550/arXiv.2509.10414"
   },
   "checklist": {
     "claims_and_evidence": {
       "abstract_claims_supported": {
         "applies": true,
         "answer": true,
         "justification": "All three core abstract claims — ICL fits PAC learning definition mathematically, accuracy is insensitive to model/prompt/distribution in the limit, and CoT/APO show distributional sensitivity — are directly supported by Tables 1–2 and ablation results.",
         "source": "haiku"
       },
       "causal_claims_justified": {
         "applies": true,
         "answer": true,
         "justification": "The paper uses controlled synthetic data generation with explicit distributional shift parameters (δ), five prompting strategies, and ablations on lexical features, positionality, and distribution to isolate causal mechanisms; this design supports causal inference about ICL's limitations.",
         "source": "haiku"
       },
       "generalization_bounded": {
         "applies": true,
         "answer": true,
         "justification": "The paper explicitly bounds its claims: 'constrained to non-natural language tasks,' 'constrained to easily-verifiable tasks in a single call,' and 'Our results are limited to the ability of ICL to draw conclusions from the data's features alone.' Appendix D elaborates these scope boundaries.",
         "source": "haiku"
       },
       "alternative_explanations_discussed": {
         "applies": true,
         "answer": true,
         "justification": "Section 7 explicitly discusses two alternative explanations — contamination and tokenization — and argues why they cannot fully account for the observed results (e.g., good performance on Hamiltonian and PARITY suggests FSA simulation, not memorization).",
         "source": "haiku"
       },
       "proxy_outcome_distinction": {
         "applies": true,
         "answer": true,
         "justification": "Section 6.4 explicitly distinguishes 'compliance' (parsing errors counted as failures) from 'learning' (correctly labeled instances), showing this distinction changes slopes and magnitude of performance estimates and matters for interpretation.",
         "source": "haiku"
       }
     },
     "limitations_and_scope": {
       "limitations_section_present": {
         "applies": true,
         "answer": true,
         "justification": "Appendix A is explicitly titled 'LIMITATIONS' and discusses reproducibility challenges from LLM updates, prohibitive cost of full-suite re-runs, exclusion of reasoning models and multi-call paradigms, and nuances of ML baseline interpretation.",
         "source": "haiku"
       },
       "threats_to_validity_specific": {
         "applies": true,
         "answer": true,
         "justification": "Specific threats are named: 'LLMs are continuously updated' (reproducibility), 'running synchronously a single task per LLM could and has taken months' (replication cost), and 'interpreting the results from the ML baselines is nuanced' due to input representation length sensitivity.",
         "source": "haiku"
       },
       "scope_boundaries_stated": {
         "applies": true,
         "answer": true,
         "justification": "Explicit scope exclusions include: reasoning models like o3-mini, multi-step/multi-call strategies such as Tree-of-Thoughts, natural-language datapoints (only instructions are in natural language), and tasks with unbounded states beyond the vending machine sum.",
         "source": "haiku"
       }
     },
     "conflicts_of_interest": {
       "funding_disclosed": {
         "applies": true,
         "answer": false,
         "justification": "No funding source is disclosed anywhere in the paper; there is no acknowledgments or funding section, despite the author's Microsoft affiliation implying institutional support.",
         "source": "haiku"
       },
       "affiliations_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "Author affiliation is clearly stated in the header: 'Microsoft and the University of York.'",
         "source": "haiku"
       },
       "funder_independent_of_outcome": {
         "applies": true,
         "answer": false,
         "justification": "The sole author is a Microsoft employee, and three of the four evaluated LLMs are Microsoft products (GPT-4 Turbo, GPT-4o, Phi-3.5 MoE Instruct); this conflict is not disclosed or addressed.",
         "source": "haiku"
       },
       "financial_interests_declared": {
         "applies": true,
         "answer": false,
         "justification": "No competing interests statement, patent disclosure, or financial interests declaration appears in the paper.",
         "source": "haiku"
       }
     },
     "scope_and_framing": {
       "key_terms_defined": {
         "applies": true,
         "answer": true,
         "justification": "Core terms are formally defined: 'learning' via PAC framework (Section 3.1), 'ICL' via formal probability equations (Section 3.3), 'OOD' via distribution distance parameter δ, and 'prompting strategies' (n-Shot, Description, APO, DE, CoT, Word Salad) with explicit descriptions.",
         "source": "haiku"
       },
       "intended_contribution_clear": {
         "applies": true,
         "answer": true,
         "justification": "The paper explicitly states it provides: (1) a theoretical framing of ICL within PAC learning, and (2) 'one of the largest' large-scale empirical analyses ablating memorization, pretraining, distributional shift, and prompting style across 4 LLMs and 9 tasks.",
         "source": "haiku"
       },
       "engagement_with_prior_work": {
         "applies": true,
         "answer": true,
         "justification": "Section 2 is a substantive related-work section that situates the paper relative to theoretical limitations work (Hahn & Rofin 2024, Strobl et al. 2024), empirical ICL studies (Sclar et al. 2024, Agarwal et al. 2024), and explicitly identifies gaps this work addresses.",
         "source": "haiku"
       }
     }
   },
   "type_checklist": {
     "empirical": {
       "artifacts": {
         "code_released": {
           "applies": true,
           "answer": true,
           "justification": "Code is available at https://github.com/adewynter/is-icl-learning, referenced repeatedly throughout the paper including in the Reproducibility Statement: 'All code is included in the repository. It will be open-sourced under the MIT licence.'",
           "source": "haiku"
         },
         "data_released": {
           "applies": true,
           "answer": true,
           "justification": "All datasets are synthetic and generated by code released in the repository; the generation procedure (automata, transition probabilities, δ parameterization) is fully described and reproducible from the released code.",
           "source": "haiku"
         },
         "environment_specified": {
           "applies": true,
           "answer": false,
           "justification": "Hardware is described (Azure Standard ND40rs v2, 8× NVIDIA V100 32GB), and scikit-learn is named for baselines, but no requirements.txt, Dockerfile, or dependency specifications are provided.",
           "source": "haiku"
         },
         "reproduction_instructions": {
           "applies": true,
           "answer": false,
           "justification": "The paper refers readers to the repository and provides model parameters in Appendix F, but no step-by-step reproduction instructions are given; Appendix A explicitly warns against re-running the full suite due to cost.",
           "source": "haiku"
         }
       },
       "statistical_methodology": {
         "confidence_intervals_or_error_bars": {
           "applies": true,
           "answer": true,
           "justification": "Standard deviation (σ) is reported throughout alongside mean accuracy in all tables (e.g., '94±1%', '80±3', Table 1) and for OLS slopes.",
           "source": "haiku"
         },
         "significance_tests": {
           "applies": true,
           "answer": false,
           "justification": "No formal statistical significance tests (t-tests, ANOVA, etc.) are applied to comparative claims; OLS is used only to measure slopes/trends, not to test significance of differences between conditions.",
           "source": "haiku"
         },
         "effect_sizes_reported": {
           "applies": true,
           "answer": true,
           "justification": "Accuracy differences are reported with magnitude (e.g., '31% accuracy gap' between closely-related tasks), and OLS slopes (e.g., CoT OOD slope -1.4 vs modus ponens -0.4) provide effect-size-like quantification.",
           "source": "haiku"
         },
         "sample_size_justified": {
           "applies": true,
           "answer": false,
           "justification": "Test set size of 1,000 (out of 2,000) is explained by cost constraints ('due to cost we only evaluated 1000'), not by formal power analysis or statistical justification.",
           "source": "haiku"
         },
         "variance_reported": {
           "applies": true,
           "answer": true,
           "justification": "Standard deviation is consistently reported in all results tables alongside every accuracy figure.",
           "source": "haiku"
         }
       },
       "evaluation_design": {
         "baselines_included": {
           "applies": true,
           "answer": true,
           "justification": "Decision trees, k-nearest neighbours, and MLPs are tested as baselines; the best-performing is reported against LLM performance for each task.",
           "source": "haiku"
         },
         "baselines_contemporary": {
           "applies": true,
           "answer": true,
           "justification": "The LLMs tested (GPT-4 Turbo, GPT-4o, Mixtral 8x7B, Phi-3.5 MoE) are contemporary models; the traditional ML baselines are intentionally classical to test whether LLMs provide learning advantages over them.",
           "source": "haiku"
         },
         "ablation_study": {
           "applies": true,
           "answer": true,
           "justification": "Four substantive ablation studies are conducted: (1) lexical features / word salad, (2) exemplar positionality, (3) alternate training distributions, and (4) compliance vs. learning separation.",
           "source": "haiku"
         },
         "multiple_metrics": {
           "applies": true,
           "answer": false,
           "justification": "Accuracy is the sole evaluation metric throughout; OLS slopes and σ are statistical treatments of accuracy, not independent metrics measuring different aspects of performance.",
           "source": "haiku"
         },
         "human_evaluation": {
           "applies": false,
           "answer": false,
           "justification": "The paper evaluates LLMs on formal language/synthetic tasks with ground-truth labels; human evaluation is clearly irrelevant.",
           "source": "haiku"
         },
         "held_out_test_set": {
           "applies": true,
           "answer": true,
           "justification": "Five separate test sets are generated per task for each δ ∈ {0, 0.2, 0.45, 0.65, 0.85}; training data is only used for APO prompt optimization and ML baselines.",
           "source": "haiku"
         },
         "per_category_breakdown": {
           "applies": true,
           "answer": true,
           "justification": "Results are extensively broken down per task (9 tasks in Table 1), per LLM (4 models), per prompting strategy (7 strategies in Table 2), and per OOD level (5 δ values).",
           "source": "haiku"
         },
         "failure_cases_discussed": {
           "applies": true,
           "answer": true,
           "justification": "Vending Machine (Sum) is analyzed as a complete failure case (near-zero slope regardless of shots); SoT's near-zero accuracy is discussed; and OOD brittleness particularly in CoT is analyzed in depth.",
           "source": "haiku"
         },
         "negative_results_reported": {
           "applies": true,
           "answer": true,
           "justification": "The paper's primary contribution is its negative results: ICL fails to generalize robustly to OOD, CoT is brittle, and traditional ML baselines outperform average ICL accuracy in half of evaluated tasks.",
           "source": "haiku"
         }
       },
       "setup_transparency": {
         "model_versions_specified": {
           "applies": true,
           "answer": true,
           "justification": "Appendix F, Table 4 specifies versions: GPT-4-0125 for GPT-4 Turbo, GPT-4o, Phi-3.5-MoE-Instruct, and Mixtral-8x7B instruct v01; temperature=0 and max tokens are also specified.",
           "source": "haiku"
         },
         "prompts_provided": {
           "applies": true,
           "answer": true,
           "justification": "Appendix H provides five fully worked-out sample prompts (modus ponens, description, word salad, CoT, DE, SoT) with actual content for PARITY, Pattern Matching, Reversal, Maze Complete, and Vending Machine.",
           "source": "haiku"
         },
         "hyperparameters_reported": {
           "applies": true,
           "answer": true,
           "justification": "Temperature=0, max_tokens=3 (standard) / 1024 (CoT) / 512 (APO) are reported; APO parameters (batch size 1024, beam width 4, search depth 6) are also specified.",
           "source": "haiku"
         },
         "scaffolding_described": {
           "applies": false,
           "answer": false,
           "justification": "All evaluations are single-call LLM invocations; there is no agentic scaffolding.",
           "source": "haiku"
         },
         "data_preprocessing_documented": {
           "applies": true,
           "answer": true,
           "justification": "Section 4.5 documents the full data generation pipeline: automata design, transition probabilities, five δ levels, 2000 training / 1000 test entries per condition, η=0.05 mislabeling for memorization control, and deduplication.",
           "source": "haiku"
         }
       },
       "data_integrity": {
         "raw_data_available": {
           "applies": true,
           "answer": true,
           "justification": "The GitHub repository contains both code and data; since data is synthetic, the released generation code serves as the authoritative source for raw data.",
           "source": "haiku"
         },
         "data_collection_described": {
           "applies": true,
           "answer": true,
           "justification": "Section 4.5 and Appendix E describe the full synthetic data generation procedure per task, including automaton structure, alphabet, transition probabilities, and OOD characterization.",
           "source": "haiku"
         },
         "recruitment_methods_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants; data is entirely synthetic.",
           "source": "haiku"
         },
         "data_pipeline_documented": {
           "applies": true,
           "answer": true,
           "justification": "The full pipeline from automaton specification to test set generation (including deduplication, mislabeling, balancing) is documented in Section 4.5 and Appendix E, with code in the repository.",
           "source": "haiku"
         }
       },
       "contamination": {
         "training_cutoff_stated": {
           "applies": true,
           "answer": false,
           "justification": "Model versions are given (GPT-4-0125, Phi-3.5-MoE-Instruct, Mixtral-8x7B instruct v01) but training data cutoffs are not stated for any model.",
           "source": "haiku"
         },
         "train_test_overlap_discussed": {
           "applies": true,
           "answer": true,
           "justification": "Section 7 explicitly discusses contamination as an alternative explanation for high accuracy on certain tasks (Pattern Matching), and the η=0.05 mislabeling scheme is introduced specifically to 'account for any potential memorisation.'",
           "source": "haiku"
         },
         "benchmark_contamination_addressed": {
           "applies": true,
           "answer": true,
           "justification": "The paper uses artificial alphabets (e.g., Reversal uses symbols like '¯\\_(ツ)_/¯') specifically to force learning from context rather than memorized patterns, and explicitly tests a randomly initialized model as contamination-free baseline (Appendix D.2).",
           "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": false,
           "justification": "Inference cost is described only qualitatively ('could and has taken months,' 'aggregate cost could render further exploration prohibitive') with no dollar figures or token counts provided.",
           "source": "haiku"
         },
         "compute_budget_stated": {
           "applies": true,
           "answer": false,
           "justification": "Hardware (Azure Standard ND40rs v2, 8× V100 32GB) is mentioned but total compute hours or budget is not stated; the paper explicitly discourages others from re-running the full suite.",
           "source": "haiku"
         }
       }
     }
   },
   "claims": [
     {
       "claim": "ICL mathematically fits the PAC learning framework because next-token prediction can satisfy PAC learning's generalization requirements.",
       "evidence": "Formal derivation in Section 3.3 reframes ICL as a PAC learner f: {p} × X → {0,1} satisfying Equations 3–5.",
       "supported": "moderate"
     },
     {
       "claim": "In the limit (50–100 exemplars), accuracy gaps between LLMs and prompting strategies narrow and all prompts show positive improvement slopes.",
       "evidence": "Table 2 OLS slopes: all prompting strategies have positive shot slopes (modus ponens 8.3, description 4.4, CoT 3.3); Table 1 shows peak averages at 50–100 shots.",
       "supported": "strong"
     },
     {
       "claim": "ICL is brittle to out-of-distribution inputs, especially chain-of-thought and APO prompting.",
       "evidence": "Table 2 δ slopes: CoT has the largest negative OOD slope (-1.4), APO -0.5; Figure 2 shows converging OOD brittleness particularly for CoT.",
       "supported": "strong"
     },
     {
       "claim": "Semantically nonsensical prompts (word salad) reach near-equivalent accuracy to semantic prompts in the limit.",
       "evidence": "Table 6 shows word salad matching non-salad best-of performance within σ for PARITY, Pattern Matching, Maze Complete, and Vending Machine Verification; average slope 11±4.6 vs description 4.4±2.2.",
       "supported": "strong"
     },
     {
       "claim": "Closely related formal tasks have accuracy gaps of up to 31%, indicating ICL generalizes per task, not per task class.",
       "evidence": "Table 1: Pattern Matching achieves 94±1% while Maze Solve achieves 63±5% — both FSA tasks. Reversal (61%) vs Stack (73%) differ by 12% despite both being PDA tasks.",
       "supported": "strong"
     },
     {
       "claim": "Traditional ML baselines (DT, kNN, MLP) outperform ICL on average best performance in half of evaluated tasks.",
       "evidence": "Table 1 footnotes: kNN outperforms LLM averages in Reversal, Stack, and Pattern Matching on average (not best-of); LLMs only outperform in best-of scenarios.",
       "supported": "moderate"
     },
     {
       "claim": "ICL learns the observed distribution P rather than generalizing to unseen Q — fully randomizing exemplars decreases accuracy.",
       "evidence": "Section 6.2 reports fully randomized exemplars yield lower average accuracy (43%) vs shuffled and unshuffled (48% each) for GPT-4o; discussed in Section 7 as evidence of distributional overfitting.",
       "supported": "moderate"
     }
   ],
   "methodology_tags": [
     "benchmark-eval",
     "theoretical",
     "observational"
   ],
   "key_findings": "While ICL formally satisfies the PAC learning definition (any next-token prediction scheme can be cast as a PAC learner), it is empirically a weak learning paradigm: accuracy converges across models and prompting styles only at 50–100 exemplars (far more than commonly claimed), and even semantically nonsensical prompts reach near-equivalent performance in the limit, suggesting LLMs exploit statistical regularities in the prompt rather than task semantics. ICL is robust to in-distribution exemplar ordering and label proportions but brittle to out-of-distribution inputs, with CoT and APO showing the largest OOD degradation. Formally similar tasks (both FSA-recognizable) show accuracy differences up to 31%, indicating that ICL's generalizability is task-specific and that autoregressive encoding is not a robust cross-task learning mechanism.",
   "red_flags": [
     {
       "flag": "Undisclosed conflict of interest",
       "detail": "The sole author is a Microsoft employee; three of four evaluated LLMs are Microsoft products (GPT-4 Turbo, GPT-4o, Phi-3.5 MoE Instruct). This potential bias is not disclosed or discussed."
     },
     {
       "flag": "Single metric only",
       "detail": "Accuracy is the sole evaluation metric throughout all 1.89M predictions; no alternative measures (e.g., calibration, F1 for imbalanced conditions) are used, limiting interpretive richness."
     },
     {
       "flag": "No significance testing",
       "detail": "OLS slopes and standard deviations are reported, but no formal hypothesis tests (t-tests, ANOVA) are applied to comparative claims between prompting strategies or models."
     },
     {
       "flag": "Identical version strings for distinct models",
       "detail": "Table 4 lists both GPT-4 Turbo and GPT-4o as 'Version: GPT-4-0125,' which appears to be a copy-paste error and undermines the model versioning claim."
     },
     {
       "flag": "Synthetic-only data limits applicability",
       "detail": "All tasks use artificial alphabets and formal languages explicitly designed to prevent semantic priors; the paper acknowledges but cannot resolve whether findings transfer to natural-language ICL scenarios."
     }
   ],
   "cited_papers": [
     {
       "title": "A Theory of the Learnable (PAC learning)",
       "relevance": "Foundational framework the paper uses to formally define and evaluate whether ICL constitutes learning."
     },
     {
       "title": "Language Models are Few-Shot Learners (GPT-3)",
       "relevance": "Original paper claiming few-shot ICL capability; this work's findings directly challenge its characterization of ICL generalization."
     },
     {
       "title": "What Languages are Easy to Language-Model? (Borenstein et al. 2024)",
       "relevance": "Motivating reference calling for empirical evaluation of LLM effective capabilities, directly cited as core motivation."
     },
     {
       "title": "Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design (Sclar et al. 2024)",
       "relevance": "Prior work on prompt sensitivity that this paper's results both confirm (sensitivity) and partially refute (ordering sensitivity disappears in the limit)."
     },
     {
       "title": "Faith and Fate: Limits of Transformers on Compositionality (Dziri et al. 2023)",
       "relevance": "Prior empirical work finding LLM weaknesses in compositional generalization; this paper's findings on task-type inconsistency build on these results."
     },
     {
       "title": "Many-Shot In-Context Learning (Agarwal et al. 2024)",
       "relevance": "Shows that expanding shots improves performance; independently confirmed and extended by this paper across more tasks, models, and distribution shifts."
     },
     {
       "title": "Neural Networks and the Chomsky Hierarchy (Deletang et al. 2023)",
       "relevance": "Directly related work evaluating LLMs on formal language tasks; this paper uses similar task classes (FSA, PDA) with different methodology."
     },
     {
       "title": "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (Wei et al. 2022)",
       "relevance": "Foundational CoT paper; this work tests CoT extensively and finds it improves accuracy but is the most brittle to OOD shifts."
     },
     {
       "title": "Awes, Laws, and Flaws from Today's LLM Research (de Wynter 2025)",
       "relevance": "Same author's prior methodological critique of LLM research practices that motivates this work's careful ablation design."
     },
     {
       "title": "The Expressive Power of Transformers with Chain of Thought (Merrill & Sabharwal 2024)",
       "relevance": "Theoretical work on CoT and TC0 circuits that this paper empirically tests and partially confirms regarding CoT's task-specific effectiveness."
     }
   ],
   "engagement_factors": {
     "practical_relevance": {
       "score": 2,
       "justification": "Directly challenges practitioner assumptions about few-shot sufficiency and CoT reliability, but the synthetic task setup limits immediate deployment relevance."
     },
     "surprise_contrarian": {
       "score": 3,
       "justification": "Directly contradicts the widely-held belief that a few exemplars suffice for ICL (finding 50–100 are needed) and that CoT is robustly beneficial (showing it is the most OOD-brittle strategy)."
     },
     "fear_safety": {
       "score": 0,
       "justification": "No AI safety or risk concerns raised; the paper focuses on learning theory and evaluation methodology."
     },
     "drama_conflict": {
       "score": 2,
       "justification": "Challenges widespread capability claims about LLMs in high-profile venues; the negative finding that 'ICL is not robust learning' is likely to generate debate."
     },
     "demo_ability": {
       "score": 1,
       "justification": "Code is released and could be run, but Appendix A explicitly discourages re-running the full suite due to prohibitive cost; partial demos are feasible."
     },
     "brand_recognition": {
       "score": 2,
       "justification": "Microsoft affiliation and evaluation of flagship GPT-4 models provide brand recognition, though this is a single-author paper from a researcher, not a major lab announcement."
     }
   },
   "hn_data": {
     "threads": [
       {
         "hn_id": "42807387",
         "title": "A Faster Quantum Fourier Transform",
         "points": 89,
         "comments": 6,
         "url": "https://news.ycombinator.com/item?id=42807387",
         "created_at": "2025-01-23T19:49:59Z"
       },
       {
         "hn_id": "43496244",
         "title": "Parameter-free KV cache compression for memory-efficient long-context LLMs",
         "points": 65,
         "comments": 19,
         "url": "https://news.ycombinator.com/item?id=43496244",
         "created_at": "2025-03-27T18:07:41Z"
       },
       {
         "hn_id": "43695562",
         "title": "M1: Towards Scalable Test-Time Compute with Mamba Reasoning Models",
         "points": 33,
         "comments": 3,
         "url": "https://news.ycombinator.com/item?id=43695562",
         "created_at": "2025-04-15T17:00:18Z"
       },
       {
         "hn_id": "44024987",
         "title": "Can You Trust Code Copilots? Evaluating LLMs from a Code Security Perspec",
         "points": 11,
         "comments": 2,
         "url": "https://news.ycombinator.com/item?id=44024987",
         "created_at": "2025-05-18T23:09:48Z"
       },
       {
         "hn_id": "43116772",
         "title": "AI Alignment at Your Discretion",
         "points": 3,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=43116772",
         "created_at": "2025-02-20T16:33:53Z"
       },
       {
         "hn_id": "45284415",
         "title": "Is In-Context Learning Learning?",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=45284415",
         "created_at": "2025-09-18T02:27:54Z"
       },
       {
         "hn_id": "45350535",
         "title": "DeepMind Paper on Virtual Agent Economies",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=45350535",
         "created_at": "2025-09-23T17:54:23Z"
       },
       {
         "hn_id": "45282518",
         "title": "Virtual Agent Economies",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=45282518",
         "created_at": "2025-09-17T23:06:36Z"
       },
       {
         "hn_id": "45268153",
         "title": "Virtual Agent Economies",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=45268153",
         "created_at": "2025-09-16T21:14:43Z"
       },
       {
         "hn_id": "45250763",
         "title": "Advancing Deep Search Agents with Knowledge Graphs and Multi-Turn RL",
         "points": 1,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=45250763",
         "created_at": "2025-09-15T15:22:49Z"
       }
     ],
     "top_points": 89,
     "total_points": 210,
     "total_comments": 30
   }
 }