{"data":{"event":{"id":"0d6f4e91-6ac6-49a1-a7be-029047d3c09f","slug":"metacognitive-steering-learning-the-structure-of-scientific-judgment-fd2ffecc50","title":"Metacognitive Steering: Learning the Structure of Scientific Judgment","short_summary":"arXiv:2609.16245v1 Announce Type: new \nAbstract: Long-horizon scientific discovery requires agents to alternate between exploration, disciplined execution, and critical reassessment as evidence changes. Current language models are trained primarily on the products of science and optimized using outcome-level signals, providing limited supervision for these process-level shifts in scientific judgment. We investigate whether such judgment can be recovered from scientist interaction traces and used to control the internal computation of a frozen frontier model. Using contrastive interventions col","full_description":"arXiv:2609.16245v1 Announce Type: new \nAbstract: Long-horizon scientific discovery requires agents to alternate between exploration, disciplined execution, and critical reassessment as evidence changes. Current language models are trained primarily on the products of science and optimized using outcome-level signals, providing limited supervision for these process-level shifts in scientific judgment. We investigate whether such judgment can be recovered from scientist interaction traces and used to control the internal computation of a frozen frontier model. Using contrastive interventions collected during real scientific research, we identify a coordinated, low-dimensional control structure within Kimi 2.6, a trillion-parameter mixture-of-experts model. Residual analysis, attention-weight subspace alignment, and cross-layer singular value decomposition converge on a mid-depth control surface spanning key layers. We introduce Metacognitive Steering, an inference-time controller that reads the model's cognitive regime and dynamically composes layer-specific interventions for exploration, procedural convergence, or critical reassessment without modifying model parameters. Behavioral analyses show that this control produces more sustained exploration, explicit pruning, and evidence-responsive synthesis. We operationalize the method in Columbus-1, an autonomous research system that identified eight independently reproduced, attacker-reachable vulnerabilities in BlueZ and directed the design, simulation, and fabrication of a ten-foot rocket intended to land propulsively using non-throttleable solid motors. Together, these results show that process-level scientific judgment can provide supervision for interpretable, dynamic control over a model's reasoning strategy.\n\nUpdate (2026-09-16): CADWorld: Computer-Use Benchmark for Long-Horizon Computer-Aided Design. arXiv:2609.16251v1 Announce Type: new \nAbstract: Computer-use agents are increasingly evaluated in realistic desktop environments, but existing benchmarks provide limited coverage of professional engineering workflows whose outputs are persistent, structured artifacts. Mechanical computer-aided design (CAD) is a particularly demanding setting: an agent must manipulate geometry and constraints over long interaction horizons while producing a native project whose dimensions, construction structure, and downstream engineering state remain valid. We introduce \\textbf{CADWorld}, a benchmark for long-horizon computer use in FreeCAD. CADWorld contains 200 tasks spanning 11 mechanical-CAD workflow categories, including sketching, part modeling, assembly, CAM, FEM, measurement, mesh processing, and technical drawing. Agents operate through screenshots and GUI actions, while success is determined by task-specific executable checks over saved FreeCAD artifacts and auxiliary outputs, covering geometric properties, parametric structure, constraints, manufacturing state, and simulation results. Across\n\nUpdate (2026-09-16): Do LLMs Have Values? A Quantitative Analysis and Alignment Framework for Values in Large Language Models. arXiv:2609.16589v1 Announce Type: new \nAbstract: As Large Language Models (LLMs) increasingly handle complex subjective tasks, aligning their intentions and behaviors with human values has become a critical scientific challenge. However, current efforts are confounded by a striking behavioral paradox: they fluctuate unpredictably under minor wording changes (\"swing\"), yet stubbornly ignore explicit instructions to correct ingrained biases (\"rigidity\"). Resolving this duality is critical for reliable AI alignment. To systematically understand and safely steer these latent subjective preferences, our study is structured around three fundamental questions. First, do LLMs possess an intrinsic value system? By projecting responses from 106 LLMs (150,000 queries per model) and 95,000 human survey profiles into a shared sociological space, we empirically confirm that they do. However, they do not mirror human diversity, instead crystallizing into a highly concentrated, idealized value core. Second, how can these values be quantified? We propose the Prior-Environ\n\nUpdate (2026-03-05): Reasoning models struggle to control their chains of thought, and that’s good ↗. Why a limitation of frontier models is reassuring for AI safety.\n\nUpdate (2026-09-17): Diffuse AI Control on Fuzzy Tasks We introduce a red-teaming framework for evaluating training interventions against diffuse threats from scheming AIs, such as sandbagging on alignment research.. Diffuse AI Control on Fuzzy Tasks We introduce a red-teaming framework for evaluating training interventions against diffuse threats from scheming AIs, such as sandbagging on alignment research.\n\nUpdate (2026-09-23): How to stay smart in the age of AI: the science of critical thinking. How to stay smart in the age of AI: the science of critical thinking","ledger_type":"benefit","primary_domain_id":"8b844c0d-c1ac-4f54-920a-34321e45582d","event_status":"provisional","event_date":"2026-09-16T00:00:00.000Z","discovery_date":"2026-09-16T00:00:00.000Z","first_published_date":"2026-09-16T00:00:00.000Z","last_reviewed_date":"2026-09-23T00:00:00.000Z","geographic_scope":"International","affected_population":null,"base_impact_tier":1,"base_score":"1.00","attribution_multiplier":"0.7500","evidence_multiplier":"0.5000","realization_multiplier":"0.4000","durability_multiplier":"0.5000","current_event_score":"0.075000","confidence_level":"low","score_explanation":"Ingest update: independent source raised evidence; added a follow-up source. Score remains provisional until review. Follow-up ingest: independent source raised evidence; added a follow-up source. Follow-up ingest: added a follow-up source. Follow-up ingest: added a follow-up source. Follow-up ingest: added a follow-up source.","methodology_version_id":"8745541b-0f8e-4d5a-9b47-5291d3813a8d","original_methodology_version_id":"8745541b-0f8e-4d5a-9b47-5291d3813a8d","published_at":"2026-09-17T03:00:58.635Z","created_at":"2026-09-17T03:00:58.635Z","updated_at":"2026-09-23T13:01:01.225Z","flags":[],"domain_name":"Safety","domain_slug":"safety","methodology_version":"0.2"},"contributions":[{"id":"603d2064-9634-4df0-a38e-855f74e117cc","event_id":"0d6f4e91-6ac6-49a1-a7be-029047d3c09f","model_id":"535014f7-c941-4d39-aa9b-8f9c1d3d07b7","role_description":"Unspecified system mentioned or implied by a news item. Remap to a named release when identified.","attribution_multiplier":"0.7500","credit_share":"1.0000","contribution_score":"0.075000","attribution_rationale":"News ingest does not infer a named model from the publisher alone. Attribution stays unspecified until a release is identified.","attribution_confidence":"medium","first_used_date":"2026-09-16T00:00:00.000Z","model_version_if_known":null,"review_status":"approved","created_at":"2026-09-17T03:00:58.643Z","model_slug":"unspecified-ai-system","model_name":"Unspecified AI system","identity_class":"unknown","is_internal":false,"family_name":"Unspecified","family_slug":"unknown-unspecified","organization_name":"Unknown","organization_slug":"unknown"}],"sources":[{"id":"ad0145e4-d365-433f-8afa-58759e4c7944","event_id":"0d6f4e91-6ac6-49a1-a7be-029047d3c09f","url":"https://arxiv.org/abs/2609.16245","canonical_url":"https://arxiv.org/abs/2609.16245","source_type":"preprint","publisher":"arXiv cs.AI","author":null,"publication_date":"2026-09-16T00:00:00.000Z","retrieved_at":"2026-09-17T03:00:58.650Z","title":"Metacognitive Steering: Learning the Structure of Scientific Judgment","excerpt":"arXiv:2609.16245v1 Announce Type: new \nAbstract: Long-horizon scientific discovery requires agents to alternate between exploration, disciplined execution, and critical reassessment as evidence changes. Current language models are trained primarily on the products of science and optimized using outcome-level signals, providing limited supervision for these process-level shifts in scientific judgment. We investigate whether such judgment can be recovered from scientist interaction traces and used","content_hash":null,"source_reliability_class":"medium","is_primary_source":true,"is_independent":true,"is_peer_reviewed":false,"archived_url":null,"created_at":"2026-09-17T03:00:58.650Z"},{"id":"8f91a0ca-fde5-4d0b-bee1-acf877ade6b8","event_id":"0d6f4e91-6ac6-49a1-a7be-029047d3c09f","url":"https://openai.com/index/reasoning-models-chain-of-thought-controllability/","canonical_url":"https://openai.com/index/reasoning-models-chain-of-thought-controllability/","source_type":"developer_announcement","publisher":"OpenAI Alignment","author":null,"publication_date":"2026-03-05T00:00:00.000Z","retrieved_at":"2026-09-17T03:01:09.604Z","title":"Reasoning models struggle to control their chains of thought, and that’s good ↗","excerpt":"Why a limitation of frontier models is reassuring for AI safety.","content_hash":null,"source_reliability_class":"medium","is_primary_source":false,"is_independent":true,"is_peer_reviewed":false,"archived_url":null,"created_at":"2026-09-17T03:01:09.604Z"},{"id":"22377a85-bfc4-47a8-b444-2259107631ec","event_id":"0d6f4e91-6ac6-49a1-a7be-029047d3c09f","url":"https://arxiv.org/abs/2609.16251","canonical_url":"https://arxiv.org/abs/2609.16251","source_type":"preprint","publisher":"arXiv cs.AI","author":null,"publication_date":"2026-09-16T00:00:00.000Z","retrieved_at":"2026-09-17T03:00:58.813Z","title":"CADWorld: Computer-Use Benchmark for Long-Horizon Computer-Aided Design","excerpt":"arXiv:2609.16251v1 Announce Type: new \nAbstract: Computer-use agents are increasingly evaluated in realistic desktop environments, but existing benchmarks provide limited coverage of professional engineering workflows whose outputs are persistent, structured artifacts. Mechanical computer-aided design (CAD) is a particularly demanding setting: an agent must manipulate geometry and constraints over long interaction horizons while producing a native project whose dimensions, construction structure","content_hash":null,"source_reliability_class":"medium","is_primary_source":false,"is_independent":false,"is_peer_reviewed":false,"archived_url":null,"created_at":"2026-09-17T03:00:58.813Z"},{"id":"1c12dd50-3c70-40b1-9edb-5be6698d7669","event_id":"0d6f4e91-6ac6-49a1-a7be-029047d3c09f","url":"https://arxiv.org/abs/2609.16589","canonical_url":"https://arxiv.org/abs/2609.16589","source_type":"preprint","publisher":"arXiv cs.AI","author":null,"publication_date":"2026-09-16T00:00:00.000Z","retrieved_at":"2026-09-17T03:00:59.981Z","title":"Do LLMs Have Values? A Quantitative Analysis and Alignment Framework for Values in Large Language Models","excerpt":"arXiv:2609.16589v1 Announce Type: new \nAbstract: As Large Language Models (LLMs) increasingly handle complex subjective tasks, aligning their intentions and behaviors with human values has become a critical scientific challenge. However, current efforts are confounded by a striking behavioral paradox: they fluctuate unpredictably under minor wording changes (\"swing\"), yet stubbornly ignore explicit instructions to correct ingrained biases (\"rigidity\"). Resolving this duality is critical for reli","content_hash":null,"source_reliability_class":"medium","is_primary_source":false,"is_independent":false,"is_peer_reviewed":false,"archived_url":null,"created_at":"2026-09-17T03:00:59.981Z"},{"id":"0635b289-42ba-4e3a-879a-08bd22499cde","event_id":"0d6f4e91-6ac6-49a1-a7be-029047d3c09f","url":"https://alignment.anthropic.com/2026/diffuse-ai-control/","canonical_url":"https://alignment.anthropic.com/2026/diffuse-ai-control/","source_type":"developer_announcement","publisher":"Anthropic Alignment Science","author":null,"publication_date":"2026-09-17T00:00:00.000Z","retrieved_at":"2026-09-17T03:01:10.726Z","title":"Diffuse AI Control on Fuzzy Tasks We introduce a red-teaming framework for evaluating training interventions against diffuse threats from scheming AIs, such as sandbagging on alignment research.","excerpt":"Diffuse AI Control on Fuzzy Tasks We introduce a red-teaming framework for evaluating training interventions against diffuse threats from scheming AIs, such as sandbagging on alignment research.","content_hash":null,"source_reliability_class":"medium","is_primary_source":false,"is_independent":true,"is_peer_reviewed":false,"archived_url":null,"created_at":"2026-09-17T03:01:10.726Z"},{"id":"4f416f90-d012-427e-981f-ff1a56635604","event_id":"0d6f4e91-6ac6-49a1-a7be-029047d3c09f","url":"https://www.nature.com/articles/d41586-026-02930-6","canonical_url":"https://www.nature.com/articles/d41586-026-02930-6","source_type":"peer_reviewed","publisher":"Nature","author":null,"publication_date":"2026-09-23T00:00:00.000Z","retrieved_at":"2026-09-23T13:01:01.209Z","title":"How to stay smart in the age of AI: the science of critical thinking","excerpt":"How to stay smart in the age of AI: the science of critical thinking","content_hash":null,"source_reliability_class":"medium","is_primary_source":false,"is_independent":true,"is_peer_reviewed":true,"archived_url":null,"created_at":"2026-09-23T13:01:01.209Z"}],"claims":[{"id":"4cb4d315-3980-44e2-a003-7673dfd50f95","event_id":"0d6f4e91-6ac6-49a1-a7be-029047d3c09f","claim_text":"Metacognitive Steering: Learning the Structure of Scientific Judgment","claim_type":"outcome","claim_status":"supported","confidence_score":null,"created_at":"2026-09-17T03:00:58.656Z","updated_at":"2026-09-17T03:00:58.656Z"}],"revisions":[{"id":"f44389e8-c62c-4c0a-855d-e79cd314924d","event_id":"0d6f4e91-6ac6-49a1-a7be-029047d3c09f","model_id":null,"previous_score":"0.000000","new_score":"0.052500","previous_factors":{},"new_factors":{"tier":1,"evidence":0.35,"base_score":1,"durability":0.5,"attribution":0.75,"realization":0.4},"change_reason":"Auto-published from news ingest.","trigger_type":"news_ingest","trigger_source_ids":null,"reviewer_id":null,"review_status":"published","created_at":"2026-09-17T03:00:58.663Z"},{"id":"4443d838-adeb-4d2d-8550-69e9d04d2805","event_id":"0d6f4e91-6ac6-49a1-a7be-029047d3c09f","model_id":null,"previous_score":"0.052500","new_score":"0.067500","previous_factors":{"tier":1,"evidence":0.35,"base_score":1,"durability":0.5,"attribution":0.75,"realization":0.4},"new_factors":{"tier":1,"models":["535014f7-c941-4d39-aa9b-8f9c1d3d07b7"],"evidence":0.45,"base_score":1,"durability":0.5,"attribution":0.75,"realization":0.4},"change_reason":"Ingest update: independent source raised evidence; added a follow-up source.","trigger_type":"ingest_update","trigger_source_ids":["22377a85-bfc4-47a8-b444-2259107631ec"],"reviewer_id":null,"review_status":"published","created_at":"2026-09-17T03:00:58.833Z"},{"id":"0ed6a6ed-91ff-4820-9819-43dacf017a3d","event_id":"0d6f4e91-6ac6-49a1-a7be-029047d3c09f","model_id":null,"previous_score":"0.067500","new_score":"0.075000","previous_factors":{"tier":1,"evidence":0.45,"base_score":1,"durability":0.5,"attribution":0.75,"realization":0.4},"new_factors":{"tier":1,"models":["535014f7-c941-4d39-aa9b-8f9c1d3d07b7"],"evidence":0.5,"base_score":1,"durability":0.5,"attribution":0.75,"realization":0.4},"change_reason":"Ingest update: independent source raised evidence; added a follow-up source.","trigger_type":"ingest_update","trigger_source_ids":["1c12dd50-3c70-40b1-9edb-5be6698d7669"],"reviewer_id":null,"review_status":"published","created_at":"2026-09-17T03:01:00.000Z"},{"id":"fa86da6e-26a1-4d40-bf76-b2dfa6a5ca3b","event_id":"0d6f4e91-6ac6-49a1-a7be-029047d3c09f","model_id":null,"previous_score":"0.075000","new_score":"0.075000","previous_factors":{"tier":1,"evidence":0.5,"base_score":1,"durability":0.5,"attribution":0.75,"realization":0.4},"new_factors":{"tier":1,"models":["535014f7-c941-4d39-aa9b-8f9c1d3d07b7"],"evidence":0.5,"base_score":1,"durability":0.5,"attribution":0.75,"realization":0.4},"change_reason":"Ingest update: added a follow-up source.","trigger_type":"ingest_update","trigger_source_ids":["8f91a0ca-fde5-4d0b-bee1-acf877ade6b8"],"reviewer_id":null,"review_status":"published","created_at":"2026-09-17T03:01:09.619Z"},{"id":"9cc389f8-f7b2-476b-965a-65f0c84f879f","event_id":"0d6f4e91-6ac6-49a1-a7be-029047d3c09f","model_id":null,"previous_score":"0.075000","new_score":"0.075000","previous_factors":{"tier":1,"evidence":0.5,"base_score":1,"durability":0.5,"attribution":0.75,"realization":0.4},"new_factors":{"tier":1,"models":["535014f7-c941-4d39-aa9b-8f9c1d3d07b7"],"evidence":0.5,"base_score":1,"durability":0.5,"attribution":0.75,"realization":0.4},"change_reason":"Ingest update: added a follow-up source.","trigger_type":"ingest_update","trigger_source_ids":["0635b289-42ba-4e3a-879a-08bd22499cde"],"reviewer_id":null,"review_status":"published","created_at":"2026-09-17T03:01:10.741Z"},{"id":"bd0e43c7-4654-4fad-a343-00bfcbd62fcc","event_id":"0d6f4e91-6ac6-49a1-a7be-029047d3c09f","model_id":null,"previous_score":"0.075000","new_score":"0.075000","previous_factors":{"tier":1,"evidence":0.5,"base_score":1,"durability":0.5,"attribution":0.75,"realization":0.4},"new_factors":{"tier":1,"models":["535014f7-c941-4d39-aa9b-8f9c1d3d07b7"],"evidence":0.5,"base_score":1,"durability":0.5,"attribution":0.75,"realization":0.4},"change_reason":"Ingest update: added a follow-up source.","trigger_type":"ingest_update","trigger_source_ids":["4f416f90-d012-427e-981f-ff1a56635604"],"reviewer_id":null,"review_status":"published","created_at":"2026-09-23T13:01:01.236Z"}],"secondary":[]},"methodology_version":"0.1"}