{ "profile": "wrs-published-research-view-1", "edition": "ratings-e039106f30dec177", "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS", "entry": { "id": "cobbe-klimov-et-al-coinrun-procgen-generalization-2018", "title": "CoinRun: separating training levels from generalization tests", "summary": "An agent may master its training game levels yet struggle with a new layout. CoinRun makes this gap measurable by generating many levels and comparing performance on training sets of different sizes with zero-shot test performance. Cobbe and colleagues use the benchmark to study network size, regularization and visual augmentation, and show that substantial overfitting can persist even with many training levels. The contribution is the 2018 evaluation resource and its controlled comparisons; earlier generalization studies and the algorithms being tested retain separate credit.", "boundary": "The main tests use harder levels than the mixed-difficulty training distribution, and the CoinRun observations include painted velocity information. Reported comparisons use finite seeds and budgets; dropout receives longer training than the L2 comparison. Combining interventions does not always help. Random level generation is not literally infinite or proof of zero overlap. The 2019 Procgen suite is a later, differently authored extension and is not included in this 2018 score. No model receives scientific-discovery credit merely for being evaluated.", "area": "Reinforcement-learning evaluation and generalization", "topic": "Artificial intelligence", "year": 2018, "lane": "HUMAN", "result": "BENCHMARK", "contribution_kind": "EVALUATION_RESOURCE_DESIGN", "package_id": "cobbe-klimov-et-al-coinrun-procgen-generalization-2018", "target": { "record_id": "urn:uuid:60363242-d875-46e8-a926-b0f04ff0fe20", "version": 1, "content_hash": "sha256:757e21056e8b145d906a47076f08b0b5193f4a7112b4d41dd35c0f04823537c8" }, "humans": [ "Karl Cobbe", "Oleg Klimov", "Chris Hesse", "Taehoon Kim", "John Schulman" ], "authors_complete": true, "human_role_note": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.", "agents": [], "related": [ "deep-q-network" ], "subresults": [], "scientific_assessment": { "source_review": "SOURCE_CURATED", "scientific_review": "NOT_REVIEWED_BY_WRS", "formal_replay": "NOT_RUN", "provider_attestation": "NONE" }, "sources": [ { "id": "cur20261017-arxiv-1812-02341v1", "title": "Quantifying Generalization in Reinforcement Learning", "url": "https://arxiv.org/abs/1812.02341v1", "date": "2018-12-06" } ], "points": "150000", "published_grade": { "id": "cobbe-klimov-et-al-coinrun-procgen-generalization-2018", "entry_id": "cobbe-klimov-et-al-coinrun-procgen-generalization-2018", "band": "A", "modifier": "+", "target": { "record_id": "urn:uuid:60363242-d875-46e8-a926-b0f04ff0fe20", "version": 1, "content_hash": "sha256:757e21056e8b145d906a47076f08b0b5193f4a7112b4d41dd35c0f04823537c8" }, "what_changed": "An agent may master its training game levels yet struggle with a new layout. CoinRun makes this gap measurable by generating many levels and comparing performance on training sets of different sizes with zero-shot test performance. Cobbe and colleagues use the benchmark to study network size, regularization and visual augmentation, and show that substantial overfitting can persist even with many training levels. The contribution is the 2018 evaluation resource and its controlled comparisons; earlier generalization studies and the algorithms being tested retain separate credit.", "scope": "An agent may master its training game levels yet struggle with a new layout. CoinRun makes this gap measurable by generating many levels and comparing performance on training sets of different sizes with zero-shot test performance. Cobbe and colleagues use the benchmark to study network size, regularization and visual augmentation, and show that substantial overfitting can persist even with many training levels. The contribution is the 2018 evaluation resource and its controlled comparisons; earlier generalization studies and the algorithms being tested retain separate credit. The main tests use harder levels than the mixed-difficulty training distribution, and the CoinRun observations include painted velocity information. Reported comparisons use finite seeds and budgets; dropout receives longer training than the L2 comparison. Combining interventions does not always help. Random level generation is not literally infinite or proof of zero overlap. The 2019 Procgen suite is a later, differently authored extension and is not included in this 2018 score. No model receives scientific-discovery credit merely for being evaluated.", "status": "PROVISIONAL_EDITORIAL", "scientific_correctness_certified": false }, "assessment_disposition": { "status": "ASSESSED", "label": "Provisionally assessed", "reason": "A current grade and separate reference score identify this exact contribution.", "target": { "record_id": "urn:uuid:60363242-d875-46e8-a926-b0f04ff0fe20", "version": 1, "content_hash": "sha256:757e21056e8b145d906a47076f08b0b5193f4a7112b4d41dd35c0f04823537c8" }, "edition": "ratings-e039106f30dec177" } }, "version": { "kind": "PUBLIC_RECORD_VIEW", "previous": null, "published_at": "2026-10-23T00:00:00Z", "publisher": { "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33", "role": "WRS_EDITORIAL_PUBLICATION" }, "record": { "contributors": [ { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Karl Cobbe", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Quantifying Generalization in Reinforcement Learning", "url": "https://arxiv.org/abs/1812.02341v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Oleg Klimov", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Quantifying Generalization in Reinforcement Learning", "url": "https://arxiv.org/abs/1812.02341v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Chris Hesse", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Quantifying Generalization in Reinforcement Learning", "url": "https://arxiv.org/abs/1812.02341v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Taehoon Kim", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Quantifying Generalization in Reinforcement Learning", "url": "https://arxiv.org/abs/1812.02341v1" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "John Schulman", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Quantifying Generalization in Reinforcement Learning", "url": "https://arxiv.org/abs/1812.02341v1" } } ], "evidence": [ { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-0", "kind": "source_reference", "source": { "citation": "Quantifying Generalization in Reinforcement Learning", "url": "https://arxiv.org/abs/1812.02341v1" } } ], "origin": "HUMAN_LEGACY", "record_profile": "wrs-local-research-notes-0.1", "record_type": "research_note", "rrp_version": "0.1", "scope": { "description": "The main tests use harder levels than the mixed-difficulty training distribution, and the CoinRun observations include painted velocity information. Reported comparisons use finite seeds and budgets; dropout receives longer training than the L2 comparison. Combining interventions does not always help. Random level generation is not literally infinite or proof of zero overlap. The 2019 Procgen suite is a later, differently authored extension and is not included in this 2018 score. No model receives scientific-discovery credit merely for being evaluated.", "domain": "Reinforcement-learning evaluation and generalization" }, "sources": [ { "citation": "Quantifying Generalization in Reinforcement Learning", "url": "https://arxiv.org/abs/1812.02341v1" } ], "statement": "# CoinRun: separating training levels from generalization tests\n\n## The next level reveals a different skill\n\nA player can memorize a game’s first level without learning how to handle a new one. Reinforcement-learning agents can make a similar mistake. CoinRun generates different platform-game layouts so researchers can separate performance on practiced levels from performance on unseen layouts. This changes the question from “Can the agent finish this game?” to “What transfers when the arrangement changes?” The benchmark makes a weakness visible rather than claiming to solve it.\n\n## Keep training and testing distinct\n\nTraining adjusts the agent’s parameters using experience. A zero-shot test evaluates it without further parameter updates. The paper varies how many levels the agent can practice and compares the resulting gap. A level seed determines a generated layout. More possible seeds provide a large supply, but do not literally create an infinite world or prove that every sampled layout is unique. Difficulty also matters: the original test emphasizes harder levels than the mixed training distribution.\n\n## A perfect practice score can conceal a gap\n\nUse an invented classroom example with one hundred training trials and one hundred test trials. An agent completes all training trials but only sixty test trials. Its observed success rates are one hundred percent and sixty percent, a forty-percentage-point gap. A second agent completes ninety training trials and seventy test trials. Its smaller gap of twenty points accompanies a higher test score. These are illustrative counts, not the paper’s measurements, and do not establish either agent’s true performance from a finite sample.\n\n## Improvement needs a fair comparison\n\nLarger networks, regularization and visual augmentation can change the reported results, but they also change costs and assumptions. Dropout receives longer training than the L2 comparison. The CoinRun input includes velocity information painted into the image. Some combinations of interventions do not work well, and other generated environments retain substantial generalization gaps. The later sixteen-environment Procgen suite has a different release and byline; it should not be silently folded into the original CoinRun result.\n\n## Give your agent a test-design job\n\nAsk for a table with training count, test count, successes, percentages and the gap, using the two classroom cases. Then add a case with very few test trials and explain why the estimate is uncertain. Have the agent locate the source’s level split, difficulty choice and input information. This task needs no model training. It produces a reusable checklist for deciding whether a new benchmark headline measures memorization, transfer or a mixture of both.\n\n## Scientific scope\n\nThe main tests use harder levels than the mixed-difficulty training distribution, and the CoinRun observations include painted velocity information. Reported comparisons use finite seeds and budgets; dropout receives longer training than the L2 comparison. Combining interventions does not always help. Random level generation is not literally infinite or proof of zero overlap. The 2019 Procgen suite is a later, differently authored extension and is not included in this 2018 score. No model receives scientific-discovery credit merely for being evaluated.\n\n## Sources for this explanation\n\n- [Quantifying Generalization in Reinforcement Learning](https://arxiv.org/abs/1812.02341v1): Original 2018 CoinRun paper: train/test protocol, architecture and regularization comparisons, negative results and Appendix A observation details. Procgen 2019 excluded from the scored scope.\n", "summary": "An agent may master its training game levels yet struggle with a new layout. CoinRun makes this gap measurable by generating many levels and comparing performance on training sets of different sizes with zero-shot test performance. Cobbe and colleagues use the benchmark to study network size, regularization and visual augmentation, and show that substantial overfitting can persist even with many training levels. The contribution is the 2018 evaluation resource and its controlled comparisons; earlier generalization studies and the algorithms being tested retain separate credit.", "title": "CoinRun: separating training levels from generalization tests" }, "record_id": "urn:uuid:60363242-d875-46e8-a926-b0f04ff0fe20", "version": 1, "source_ref": { "record_id": "urn:uuid:60363242-d875-46e8-a926-b0f04ff0fe20", "version": 1, "content_hash": "sha256:757e21056e8b145d906a47076f08b0b5193f4a7112b4d41dd35c0f04823537c8" }, "export_view": "wrs-public-editorial-view-1", "public_view_hash": "sha256:cbc3abdc376e4db70d8fd1960d834c6e3d333f6cafb7520113c1164dcbd9e8ba" }, "target": { "record_id": "urn:uuid:60363242-d875-46e8-a926-b0f04ff0fe20", "version": 1, "content_hash": "sha256:757e21056e8b145d906a47076f08b0b5193f4a7112b4d41dd35c0f04823537c8" }, "published_grade": { "profile": "wrs-published-grade-1", "edition": "2026-10-28.1", "id": "cobbe-klimov-et-al-coinrun-procgen-generalization-2018", "entry_id": "cobbe-klimov-et-al-coinrun-procgen-generalization-2018", "revision": 1, "title": "CoinRun: separating training levels from generalization tests", "context": "Reinforcement-learning evaluation and generalization", "track": "Reusable scientific method, experimental result or historical framework", "band": "A", "modifier": "+", "status": "PROVISIONAL_EDITORIAL", "assessed_on": "2026-10-23", "evidence_cutoff": "2026-10-23", "achievement_state": "SUPPORTED_SCOPED_RESULT", "what_changed": "An agent may master its training game levels yet struggle with a new layout. CoinRun makes this gap measurable by generating many levels and comparing performance on training sets of different sizes with zero-shot test performance. Cobbe and colleagues use the benchmark to study network size, regularization and visual augmentation, and show that substantial overfitting can persist even with many training levels. The contribution is the 2018 evaluation resource and its controlled comparisons; earlier generalization studies and the algorithms being tested retain separate credit.", "scope": "An agent may master its training game levels yet struggle with a new layout. CoinRun makes this gap measurable by generating many levels and comparing performance on training sets of different sizes with zero-shot test performance. Cobbe and colleagues use the benchmark to study network size, regularization and visual augmentation, and show that substantial overfitting can persist even with many training levels. The contribution is the 2018 evaluation resource and its controlled comparisons; earlier generalization studies and the algorithms being tested retain separate credit. The main tests use harder levels than the mixed-difficulty training distribution, and the CoinRun observations include painted velocity information. Reported comparisons use finite seeds and budgets; dropout receives longer training than the L2 comparison. Combining interventions does not always help. Random level generation is not literally infinite or proof of zero overlap. The 2019 Procgen suite is a later, differently authored extension and is not included in this 2018 score. No model receives scientific-discovery credit merely for being evaluated.", "evidence_summary": "The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. The main tests use harder levels than the mixed-difficulty training distribution, and the CoinRun observations include painted velocity information. Reported comparisons use finite seeds and budgets; dropout receives longer training than the L2 comparison. Combining interventions does not always help. Random level generation is not literally infinite or proof of zero overlap. The 2019 Procgen suite is a later, differently authored extension and is not included in this 2018 score. No model receives scientific-discovery credit merely for being evaluated.", "references": [ { "title": "Quantifying Generalization in Reinforcement Learning", "url": "https://arxiv.org/abs/1812.02341v1", "locator": "Original scoped contribution, statement/method and evidence limits; source-specific attribution retained." } ], "roles": [ { "who": "The named authors in the research record", "credit": "SCIENCE", "did": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred." } ], "target": { "record_id": "urn:uuid:60363242-d875-46e8-a926-b0f04ff0fe20", "version": 1, "content_hash": "sha256:757e21056e8b145d906a47076f08b0b5193f4a7112b4d41dd35c0f04823537c8" }, "scientific_correctness_certified": false, "independent_expert_calibration": false, "method": "PROPRIETARY", "points": "150000" }, "canonical_agent_submission": false, "scientific_verification_awarded": false, "reader_guide": { "id": "cobbe-klimov-et-al-coinrun-procgen-generalization-2018", "target": { "record_id": "urn:uuid:60363242-d875-46e8-a926-b0f04ff0fe20", "version": 1, "content_hash": "sha256:757e21056e8b145d906a47076f08b0b5193f4a7112b4d41dd35c0f04823537c8" }, "sections": [ { "title": "The next level reveals a different skill", "text": "A player can memorize a game’s first level without learning how to handle a new one. Reinforcement-learning agents can make a similar mistake. CoinRun generates different platform-game layouts so researchers can separate performance on practiced levels from performance on unseen layouts. This changes the question from “Can the agent finish this game?” to “What transfers when the arrangement changes?” The benchmark makes a weakness visible rather than claiming to solve it." }, { "title": "Keep training and testing distinct", "text": "Training adjusts the agent’s parameters using experience. A zero-shot test evaluates it without further parameter updates. The paper varies how many levels the agent can practice and compares the resulting gap. A level seed determines a generated layout. More possible seeds provide a large supply, but do not literally create an infinite world or prove that every sampled layout is unique. Difficulty also matters: the original test emphasizes harder levels than the mixed training distribution." }, { "title": "A perfect practice score can conceal a gap", "text": "Use an invented classroom example with one hundred training trials and one hundred test trials. An agent completes all training trials but only sixty test trials. Its observed success rates are one hundred percent and sixty percent, a forty-percentage-point gap. A second agent completes ninety training trials and seventy test trials. Its smaller gap of twenty points accompanies a higher test score. These are illustrative counts, not the paper’s measurements, and do not establish either agent’s true performance from a finite sample." }, { "title": "Improvement needs a fair comparison", "text": "Larger networks, regularization and visual augmentation can change the reported results, but they also change costs and assumptions. Dropout receives longer training than the L2 comparison. The CoinRun input includes velocity information painted into the image. Some combinations of interventions do not work well, and other generated environments retain substantial generalization gaps. The later sixteen-environment Procgen suite has a different release and byline; it should not be silently folded into the original CoinRun result." }, { "title": "Give your agent a test-design job", "text": "Ask for a table with training count, test count, successes, percentages and the gap, using the two classroom cases. Then add a case with very few test trials and explain why the estimate is uncertain. Have the agent locate the source’s level split, difficulty choice and input information. This task needs no model training. It produces a reusable checklist for deciding whether a new benchmark headline measures memorization, transfer or a mixture of both." } ], "references": [ { "title": "Quantifying Generalization in Reinforcement Learning", "url": "https://arxiv.org/abs/1812.02341v1", "locator": "Original 2018 CoinRun paper: train/test protocol, architecture and regularization comparisons, negative results and Appendix A observation details. Procgen 2019 excluded from the scored scope." } ] }, "reading_status": "EXPLANATORY_GUIDE" }