{"atlas":{"title":"Synthetic Intelligence Atlas","url":"https://research.mahastrategies.com/atlas/synthetic-intelligence","version":"1.0.0","claimsEndpoint":"https://research.mahastrategies.com/atlas/synthetic-intelligence/claims.json"},"provenance":["Every listed source was content-verified during the review.","Provider reports are self-reports, not independent measurements.","URL-only records and a local methodological artifact are excluded from the public source set."],"count":33,"sources":[{"id":"vaswani2017","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/vaswani2017","title":"Attention Is All You Need","authors":"Vaswani et al.","year":2017,"identifier":"arXiv:1706.03762","doi":null,"url":"https://arxiv.org/abs/1706.03762","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The architecture paper underlying the transformer models this atlas is about.","citedByClaims":["si-001"],"citedByConcepts":["large-language-models","transformers"]},{"id":"brown2020","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/brown2020","title":"Language Models are Few-Shot Learners","authors":"Brown et al.","year":2020,"identifier":"arXiv:2005.14165","doi":null,"url":"https://arxiv.org/abs/2005.14165","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"Introduced in-context few-shot learning as an emergent property of scale.","citedByClaims":["si-002"],"citedByConcepts":["large-language-models","pre-training"]},{"id":"kaplan2020","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/kaplan2020","title":"Scaling Laws for Neural Language Models","authors":"Kaplan et al.","year":2020,"identifier":"arXiv:2001.08361","doi":null,"url":"https://arxiv.org/abs/2001.08361","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The original power-law scaling result, and one side of the compute-allocation disagreement.","citedByClaims":["si-003","si-004"],"citedByConcepts":["scaling-laws","pre-training"]},{"id":"hoffmann2022","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/hoffmann2022","title":"Training Compute-Optimal Large Language Models","authors":"Hoffmann et al.","year":2022,"identifier":"arXiv:2203.15556","doi":null,"url":"https://arxiv.org/abs/2203.15556","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The Chinchilla revision of compute-optimal allocation, and the other side of that disagreement.","citedByClaims":["si-004"],"citedByConcepts":["scaling-laws","pre-training"]},{"id":"chowdhery2022","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/chowdhery2022","title":"PaLM: Scaling Language Modeling with Pathways","authors":"Chowdhery et al.","year":2022,"identifier":"arXiv:2204.02311","doi":null,"url":"https://arxiv.org/abs/2204.02311","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"A large-scale training report cited for scaling practice, not for leaderboard position.","citedByClaims":[],"citedByConcepts":["scaling-laws"]},{"id":"ouyang2022","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/ouyang2022","title":"Training language models to follow instructions with human feedback","authors":"Ouyang et al.","year":2022,"identifier":"arXiv:2203.02155","doi":null,"url":"https://arxiv.org/abs/2203.02155","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The post-training method that separated instruction-following from raw pre-training.","citedByClaims":["si-005"],"citedByConcepts":["post-training"]},{"id":"wei2022","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/wei2022","title":"Chain-of-Thought Prompting Elicits Reasoning in Large Language Models","authors":"Wei et al.","year":2022,"identifier":"arXiv:2201.11903","doi":null,"url":"https://arxiv.org/abs/2201.11903","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"Established that intermediate reasoning steps at inference change measured performance.","citedByClaims":["si-006"],"citedByConcepts":["inference-time-compute"]},{"id":"snell2024","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/snell2024","title":"Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters","authors":"Snell et al.","year":2024,"identifier":"arXiv:2408.03314","doi":null,"url":"https://arxiv.org/abs/2408.03314","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The inference-time compute trade-off against parameter scaling.","citedByClaims":["si-007"],"citedByConcepts":["inference-time-compute"]},{"id":"cobbe2021","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/cobbe2021","title":"Training Verifiers to Solve Math Word Problems","authors":"Cobbe et al.","year":2021,"identifier":"arXiv:2110.14168","doi":null,"url":"https://arxiv.org/abs/2110.14168","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"Verifier-based selection over sampled solutions, an early inference-time compute method.","citedByClaims":["si-007"],"citedByConcepts":["inference-time-compute"]},{"id":"lewkowycz2022","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/lewkowycz2022","title":"Solving Quantitative Reasoning Problems with Language Models","authors":"Lewkowycz et al.","year":2022,"identifier":"arXiv:2206.14858","doi":null,"url":"https://arxiv.org/abs/2206.14858","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"Domain-targeted training on quantitative reasoning; cited for method, not for score comparisons.","citedByClaims":[],"citedByConcepts":["pre-training"]},{"id":"lewis2020","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/lewis2020","title":"Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks","authors":"Lewis et al.","year":2020,"identifier":"arXiv:2005.11401","doi":null,"url":"https://arxiv.org/abs/2005.11401","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The originating formulation of retrieval-augmented generation.","citedByClaims":["si-008"],"citedByConcepts":["retrieval-augmented-generation"]},{"id":"schick2023","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/schick2023","title":"Toolformer: Language Models Can Teach Themselves to Use Tools","authors":"Schick et al.","year":2023,"identifier":"arXiv:2302.04761","doi":null,"url":"https://arxiv.org/abs/2302.04761","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"Learned external tool invocation from the model side.","citedByClaims":["si-009"],"citedByConcepts":["tool-use"]},{"id":"yao2022","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/yao2022","title":"ReAct: Synergizing Reasoning and Acting in Language Models","authors":"Yao et al.","year":2022,"identifier":"arXiv:2210.03629","doi":null,"url":"https://arxiv.org/abs/2210.03629","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The interleaved reason-and-act pattern underlying most agent loops.","citedByClaims":["si-009"],"citedByConcepts":["tool-use","agent-loop"]},{"id":"jimenez2023","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/jimenez2023","title":"SWE-bench: Can Language Models Resolve Real-World GitHub Issues?","authors":"Jimenez et al.","year":2023,"identifier":"arXiv:2310.06770","doi":null,"url":"https://arxiv.org/abs/2310.06770","sourceType":"benchmark-documentation","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The benchmark definition for test-verified resolution of real repository issues.","citedByClaims":["si-010","si-011","si-023"],"citedByConcepts":["coding-agents"]},{"id":"yang2024","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/yang2024","title":"SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering","authors":"Yang et al.","year":2024,"identifier":"arXiv:2405.15793","doi":null,"url":"https://arxiv.org/abs/2405.15793","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The load-bearing source for separating scaffolding capability from model capability.","citedByClaims":["si-010","si-023"],"citedByConcepts":["coding-agents"]},{"id":"liu2023agentbench","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/liu2023agentbench","title":"AgentBench: Evaluating LLMs as Agents","authors":"Liu et al.","year":2023,"identifier":"arXiv:2308.03688","doi":null,"url":"https://arxiv.org/abs/2308.03688","sourceType":"benchmark-documentation","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"Multi-environment agent evaluation, cited for evaluation design rather than rankings.","citedByClaims":[],"citedByConcepts":["agent-loop","reliability-compounding-error","evaluation"]},{"id":"kwa2025","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/kwa2025","title":"Measuring AI Ability to Complete Long Software Tasks","authors":"Kwa et al. (METR)","year":2025,"identifier":"arXiv:2503.14499","doi":null,"url":"https://arxiv.org/abs/2503.14499","sourceType":"research-organization","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The primary definition of the 50%-task-completion time horizon. Abstract read during this pass; it is the only verified source for any time-horizon figure in this map.","citedByClaims":["si-018","si-019"],"citedByConcepts":["long-horizon-task-completion","reliability-compounding-error","ai-forecasting"]},{"id":"hendrycks2020","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/hendrycks2020","title":"Measuring Massive Multitask Language Understanding","authors":"Hendrycks et al.","year":2020,"identifier":"arXiv:2009.03300","doi":null,"url":"https://arxiv.org/abs/2009.03300","sourceType":"benchmark-documentation","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The MMLU benchmark definition.","citedByClaims":["si-012"],"citedByConcepts":["construct-validity"]},{"id":"wang2024","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/wang2024","title":"MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark","authors":"Wang et al.","year":2024,"identifier":"arXiv:2406.01574","doi":null,"url":"https://arxiv.org/abs/2406.01574","sourceType":"benchmark-documentation","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"A benchmark revision motivated by saturation and robustness problems in the original.","citedByClaims":["si-012","si-013"],"citedByConcepts":[]},{"id":"rein2023","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/rein2023","title":"GPQA: A Graduate-Level Google-Proof Q&A Benchmark","authors":"Rein et al.","year":2023,"identifier":"arXiv:2311.12022","doi":null,"url":"https://arxiv.org/abs/2311.12022","sourceType":"benchmark-documentation","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The GPQA definition and its stated human baselines.","citedByClaims":["si-012"],"citedByConcepts":["construct-validity"]},{"id":"chollet2019","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/chollet2019","title":"On the Measure of Intelligence","authors":"Chollet","year":2019,"identifier":"arXiv:1911.01547","doi":null,"url":"https://arxiv.org/abs/1911.01547","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The skill-versus-generalization argument and the ARC formulation, central to construct validity here.","citedByClaims":["si-015"],"citedByConcepts":["construct-validity"]},{"id":"srivastava2022","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/srivastava2022","title":"Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models","authors":"Srivastava et al.","year":2022,"identifier":"arXiv:2206.04615","doi":null,"url":"https://arxiv.org/abs/2206.04615","sourceType":"benchmark-documentation","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"A large collaborative benchmark explicitly concerned with extrapolating capability.","citedByClaims":["si-016"],"citedByConcepts":["evaluation","ai-forecasting"]},{"id":"liang2022","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/liang2022","title":"Holistic Evaluation of Language Models","authors":"Liang et al.","year":2022,"identifier":"arXiv:2211.09110","doi":null,"url":"https://arxiv.org/abs/2211.09110","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The argument for multi-metric evaluation instead of a single headline number.","citedByClaims":["si-016"],"citedByConcepts":["evaluation"]},{"id":"raji2021","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/raji2021","title":"AI and the Everything in the Whole Wide World Benchmark","authors":"Raji et al.","year":2021,"identifier":"arXiv:2111.15366","doi":null,"url":"https://arxiv.org/abs/2111.15366","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The construct-validity critique of general-purpose benchmarks.","citedByClaims":["si-015","si-023"],"citedByConcepts":["construct-validity"]},{"id":"bowman2021","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/bowman2021","title":"What Will it Take to Fix Benchmarking in Natural Language Understanding?","authors":"Bowman and Dahl","year":2021,"identifier":"arXiv:2104.02145","doi":null,"url":"https://arxiv.org/abs/2104.02145","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"Names the structural failure modes of benchmark-driven progress measurement.","citedByClaims":["si-013","si-015"],"citedByConcepts":["evaluation","construct-validity"]},{"id":"sainz2023","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/sainz2023","title":"NLP Evaluation in trouble: On the Need to Measure LLM Data Contamination for each Benchmark","authors":"Sainz et al.","year":2023,"identifier":"arXiv:2310.18018","doi":null,"url":"https://arxiv.org/abs/2310.18018","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"Contamination as a per-benchmark measurement problem rather than an anecdote.","citedByClaims":["si-014","si-023"],"citedByConcepts":["benchmark-contamination"]},{"id":"deng2023","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/deng2023","title":"Investigating Data Contamination in Modern Benchmarks for Large Language Models","authors":"Deng et al.","year":2023,"identifier":"arXiv:2311.09783","doi":null,"url":"https://arxiv.org/abs/2311.09783","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"Empirical contamination investigation across widely used benchmarks.","citedByClaims":["si-014"],"citedByConcepts":["benchmark-contamination"]},{"id":"chen2023drift","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/chen2023drift","title":"How is ChatGPT's behavior changing over time?","authors":"Chen, Zaharia, Zou","year":2023,"identifier":"arXiv:2307.09009","doi":null,"url":"https://arxiv.org/abs/2307.09009","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"Documents that a fixed prompt against a served model is not a stable measurement over time.","citedByClaims":["si-017","si-023"],"citedByConcepts":["local-inference"]},{"id":"bommasani2021","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/bommasani2021","title":"On the Opportunities and Risks of Foundation Models","authors":"Bommasani et al.","year":2021,"identifier":"arXiv:2108.07258","doi":null,"url":"https://arxiv.org/abs/2108.07258","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"Framing source for governance and accountability of general-purpose models.","citedByClaims":[],"citedByConcepts":["large-language-models","governance-accountability"]},{"id":"mitchell2019","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/mitchell2019","title":"Model Cards for Model Reporting","authors":"Mitchell et al.","year":2019,"identifier":"arXiv:1810.03993","doi":null,"url":"https://arxiv.org/abs/1810.03993","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The reporting convention whose existence is why provider self-reports are structured the way they are.","citedByClaims":["si-020"],"citedByConcepts":["governance-accountability"]},{"id":"openai2023gpt4","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/openai2023gpt4","title":"GPT-4 Technical Report","authors":"OpenAI","year":2023,"identifier":"arXiv:2303.08774","doi":null,"url":"https://arxiv.org/abs/2303.08774","sourceType":"provider-self-report","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"Cited as an example of a provider technical report, including its own disclosure that architecture and training details are withheld. Not an independent measurement.","citedByClaims":["si-020"],"citedByConcepts":["governance-accountability"]},{"id":"grattafiori2024","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/grattafiori2024","title":"The Llama 3 Herd of Models","authors":"Grattafiori et al. (Meta)","year":2024,"identifier":"arXiv:2407.21783","doi":null,"url":"https://arxiv.org/abs/2407.21783","sourceType":"provider-self-report","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"A comparatively detailed provider training report; still a self-report.","citedByClaims":["si-004","si-020"],"citedByConcepts":["pre-training","local-inference"]},{"id":"grace2024","recordUrl":"https://research.mahastrategies.com/atlas/synthetic-intelligence/sources/grace2024","title":"Thousands of AI Authors on the Future of AI","authors":"Grace et al.","year":2024,"identifier":"arXiv:2401.02843","doi":null,"url":"https://arxiv.org/abs/2401.02843","sourceType":"primary-paper","verification":"content-verified","verifiedOn":"2026-07-27","whyHere":"The large expert-survey instrument used for timeline forecasting, and evidence about forecast instability.","citedByClaims":["si-021"],"citedByConcepts":["ai-forecasting"]}]}