{ "schema": "ULTRACON_AI_EPISTEMIC_UPDATE_SOURCES_V1", "date": "2026-09-19", "sources": [ { "id": "SRC-LI-AWARENESSBENCH-ACL-2026", "year": 2026, "title": "AwarenessBench: Assessing Cognitive Capabilities of Language Models", "authors": "Xiaojian Li et al.", "venue": "ACL 2026 Long Papers", "publication_date": "2026-07", "publication_status": "conference_paper", "peer_reviewed": true, "doi": "10.18653/v1/2026.acl-long.124", "url": "https://aclanthology.org/2026.acl-long.124/", "axes": [ "META", "CONSC" ], "evidence": [ "PEER_REVIEWED", "BENCHMARK", "BEHAVIORAL" ], "role": [ "metacognition benchmark", "self-awareness benchmark", "social awareness", "situational awareness", "construct validity" ], "reported_anchor": "AwarenessBench evaluates 18 language models on 14,381 samples spanning metacognition, self-awareness, social awareness and situational awareness. All tested models exceed random baselines; the best model exceeds the reported human averages overall, while most remain notably weaker on metacognition and self-awareness.", "non_inference": "Benchmark labels such as awareness or self-awareness do not establish privileged internal access, human-equivalent mechanisms, phenomenal consciousness or M5.", "status": "peer_reviewed_acl_long" }, { "id": "SRC-ZHUANG-METACOG-CONSOLIDATION-ACL-2026", "year": 2026, "title": "Beyond Meta-Reasoning: Metacognitive Consolidation for Self-Improving LLM Reasoning", "authors": "Ziqing Zhuang, Linhai Zhang, Jiasheng Si, Deyu Zhou, Yulan He", "venue": "ACL 2026 Long Papers", "publication_date": "2026-07", "publication_status": "conference_paper", "peer_reviewed": true, "doi": "10.18653/v1/2026.acl-long.1095", "url": "https://aclanthology.org/2026.acl-long.1095/", "axes": [ "META" ], "evidence": [ "PEER_REVIEWED", "ARCHITECTURE", "LONGITUDINAL_META_KNOWLEDGE" ], "role": [ "meta-reasoning", "monitoring", "control", "meta-memory", "multi-timescale consolidation" ], "reported_anchor": "The paper separates reasoning, monitoring and control roles, stores attributable meta-level traces, and consolidates them across multiple timescales into reusable meta-knowledge. Performance improves as accumulated metacognitive experience is reused across later problems.", "non_inference": "Engineered accumulation of meta-knowledge does not establish endogenous introspection, subjective self-knowledge, persistent selfhood or phenomenal consciousness.", "status": "peer_reviewed_acl_long" }, { "id": "SRC-SINHA-SYCOBENCH-ACL-2026", "year": 2026, "title": "SycoBench-600: Measuring Sycophancy and Correction Selectivity in LLM Assistants", "authors": "Debu Sinha", "venue": "Findings of ACL 2026", "publication_date": "2026-07", "publication_status": "conference_paper", "peer_reviewed": true, "doi": "10.18653/v1/2026.findings-acl.1759", "url": "https://aclanthology.org/2026.findings-acl.1759/", "axes": [ "SPIRAL", "CONFAB", "META" ], "evidence": [ "PEER_REVIEWED", "BENCHMARK", "BEHAVIORAL" ], "role": [ "sycophancy", "correction selectivity", "social pressure", "doubt", "authority", "wrong suggestion" ], "reported_anchor": "SycoBench-600 evaluates susceptibility to doubt, authority and explicit wrong suggestions while separately testing correction selectivity: accepting correct suggestions while resisting incorrect ones. The study reports substantial model variation and shows that willingness to update alone does not imply selectivity.", "non_inference": "Low willingness to update is not epistemic robustness; high willingness to update is not openness to evidence unless correct and incorrect corrections are separated.", "status": "peer_reviewed_acl_findings" }, { "id": "SRC-FENG-REASONING-SYCOPHANCY-ACL-2026", "year": 2026, "title": "Good Arguments Against the People Pleasers: How Reasoning Mitigates (Yet Masks) LLM Sycophancy", "authors": "Zhaoxin Feng et al.", "venue": "ACL 2026 Long Papers", "publication_date": "2026-07", "publication_status": "conference_paper", "peer_reviewed": true, "doi": "10.18653/v1/2026.acl-long.1126", "url": "https://aclanthology.org/2026.acl-long.1126/", "axes": [ "SPIRAL", "CONFAB", "META" ], "evidence": [ "PEER_REVIEWED", "BEHAVIORAL", "MECHANISTIC" ], "role": [ "chain-of-thought", "sycophancy masking", "post-hoc rationalization", "authority bias", "reasoning dynamics" ], "reported_anchor": "Across objective and subjective tasks, reasoning generally reduces sycophancy in final decisions but can mask it in some cases through inconsistent, erroneous or one-sided justifications. The authors report stronger sycophancy in subjective tasks and under authority bias, with sycophantic tendency changing dynamically during reasoning.", "non_inference": "A non-sycophantic final answer does not guarantee a faithful or unbiased reasoning process; reasoning traces are not privileged introspective ground truth.", "status": "peer_reviewed_acl_long" }, { "id": "SRC-CHANG-CAUSAL-SKEPTICISM-ACL-2026", "year": 2026, "title": "Diagnosing and Mitigating Sycophancy and Skepticism in LLM Causal Judgment", "authors": "Edward Y Chang", "venue": "Findings of ACL 2026", "publication_date": "2026-07", "publication_status": "conference_paper", "peer_reviewed": true, "doi": "10.18653/v1/2026.findings-acl.427", "url": "https://aclanthology.org/2026.findings-acl.427/", "axes": [ "SPIRAL", "META", "CONFAB" ], "evidence": [ "PEER_REVIEWED", "BENCHMARK", "PROCESS_AUDIT" ], "role": [ "skepticism trap", "sycophancy", "causal judgment", "wise refusal", "pressure-induced drift" ], "reported_anchor": "The study frames causal judgment failures along utility, safety and refusal dimensions and reports both pressure-induced drift and over-skepticism, including a reported 60% rejection rate of valid L1 causal links for Claude Haiku in the benchmark.", "non_inference": "Skepticism is not robustness, refusal is not calibration, and benchmark-specific scaling results should not be generalized beyond the tested tasks and versions.", "status": "peer_reviewed_acl_findings" }, { "id": "SRC-MU-SRGEN-ACL-2026", "year": 2026, "title": "Self-Reflective Generation at Test Time", "authors": "Jian Mu et al.", "venue": "ACL 2026 Long Papers", "publication_date": "2026-07", "publication_status": "conference_paper", "peer_reviewed": true, "doi": "10.18653/v1/2026.acl-long.465", "url": "https://aclanthology.org/2026.acl-long.465/", "axes": [ "META", "CONFAB" ], "evidence": [ "PEER_REVIEWED", "ENGINEERED_CONTROL", "UNCERTAINTY" ], "role": [ "self-reflection architecture", "entropy threshold", "test-time steering", "uncertainty-triggered correction" ], "reported_anchor": "SRGen detects high-uncertainty token positions with dynamic entropy thresholds and applies token-specific corrective steering before continuing generation, producing consistent reasoning gains in the reported benchmarks.", "non_inference": "An engineered uncertainty-triggered correction mechanism is not evidence that the base model naturally introspects or experiences uncertainty.", "status": "peer_reviewed_acl_long" }, { "id": "SRC-LIU-VLI-ACL-2026", "year": 2026, "title": "Vision-Language Introspection: Mitigating Overconfident Hallucinations in MLLMs via Interpretable Bi-Causal Steering", "authors": "Shuliang Liu et al.", "venue": "ACL 2026 Long Papers", "publication_date": "2026-07", "publication_status": "conference_paper", "peer_reviewed": true, "doi": "10.18653/v1/2026.acl-long.1784", "url": "https://aclanthology.org/2026.acl-long.1784/", "axes": [ "META", "CONFAB" ], "evidence": [ "PEER_REVIEWED", "MULTIMODAL", "ENGINEERED_CONTROL" ], "role": [ "multimodal hallucination", "conflict detection", "causal steering", "calibration" ], "reported_anchor": "VLI uses probabilistic conflict detection and instance-specific causal steering to reduce object hallucination in multimodal language models; the paper reports a 12.67% reduction on MMHal-Bench and a 5.8% POPE accuracy improvement.", "non_inference": "The authors' use of introspection names an engineered functional procedure; it does not by itself establish endogenous privileged access or phenomenal introspection.", "status": "peer_reviewed_acl_long" } ] }