Corrections
When we get something substantively wrong, we fix it and log the fix here; small copyedits live in each article's revision history instead. Found an error? Use the "Report issue" button on the article (no account needed). So far we have received 2 reports and reviewed 1, taking a median of 4.8 days. 1 is waiting, the oldest since September 26, 2026. That is our record rather than a promise. See the verifiability policy for how checking works.
321 articles were corrected in the last 30 days. Every correction note and resolved report is listed below, newest first. A fix applied to several articles with the same note is shown once, listing each article; "See the change" opens that version in the article's revision history.
- editorial
Rebuilds Self-attention around precise scaled dot-product and multi-head computation, masks, position mechanisms, complexity, decoding, efficient variants, applications, interpretability limits, and historical antecedents using 29 primary or peer-reviewed sources.
- editorial
Defines a test set as an evaluation sample reserved from fitting and selection, ties its claims to a target population and protocol, and documents leakage-resistant splitting, precision-based sizing, nested assessment, uncertainty, distribution shift, benchmark reuse, label error, and LLM contamination.
- editorial
Independently verifies NVIDIA Blackwell's product-family boundaries, release and production history, GB100 and RTX distinctions, memory and compute-capability specifications, platform topology, numerical formats, security, measured performance, power requirements, and status through July 28, 2026.
- editorial
Reconstructed the history, principal paradigms, training methods, multilingual and LLM translation, evaluation, speech translation, reliability, and post-editing; corrected unsupported product, ranking, market, and source-version performance claims.
- editorial
Rebuilds the Logistic Regression article around the verified binary model, likelihood, coefficient interpretation, separation, regularization, diagnostics, evaluation, extensions, and current software behavior; removes unsupported universal calibration, threshold, sample-size, importance, and performance claims.
- editorial
Independently verifies and updates JAX's release status, program-transformation model, compiler and sharding architecture, platform support, ecosystem boundaries, bounded use cases, performance methodology, and citation traceability.
- editorial
Reframe Actuator as a general energy-to-motion device, distinguish transducer, transmission, sensing, and control boundaries, replace unsupported supplier and universal performance claims with primary and peer-reviewed evidence, and add scoped selection, safety, and robotics guidance.
- editorial
Restores general quantization scope; corrects neural-network methods, equations, data-format claims, and deployment caveats.
- editorial
Independently verifies and updates Robbins-Monro conditions, sampling assumptions, convergence scope, noise and flatness claims, batch scaling, and distributed-SGD boundaries.
- editorial
Independently verifies and updates Hyperparameter definitions, dynamic and data-dependent selection, search-space and fidelity scope, validation bias, interaction claims, transfer limits, and citation traceability.
- editorial
Rebuilds the LSTM article around the verified cell equations, gradient behavior, history, variants, training, bounded applications, framework behavior, and documented limitations; removes unsupported dominance, deployment, control-application, citation-count, and performance claims.
- editorial
Independently verifies and updates Imitation Learning history, method distinctions, theory, benchmark scope, and modern robot-learning publication evidence against 31 primary and peer-reviewed sources.
- editorial
Rebuilds the article around fund structure, investment process, financing instruments, valuation, performance, exits, innovation evidence, public policy, and a bounded OECD-sourced AI market section; removes speculative company-level financing claims.
- editorial
Independently verifies and rebuilds Vision-language-action model scope, chronology, architectures, datasets, benchmark definitions, source-version discrepancies, evidence limits, safety, and artifact openness.
- editorial
Independently verifies interpretability methods, evidence limits, legal scope and application dates, and mechanistic-interpretability claims against 45 primary, official, and peer-reviewed sources.
- editorial
Replaces unsupported and overgeneralized claims; distinguishes factuality, faithfulness, and confabulation; scopes benchmark results; and corrects legal-study denominators with primary evidence.
- editorial
Resolved undefined citation markers, bounded architectural-disclosure claims, added independent multimodal evaluations, and documented GPT-4o API deprecation
- editorial
Rebuilt Agility Robotics coverage with sequential citations, scoped deployment evidence, site-specific safety qualification, and current 2026 corporate facts
- editorial
Distinguish xAI from its parent's broader AI segment, correct corporate, financing, product, compute, and contract claims, and remove unsupported projections.
- editorial
Rebuild PaLM from primary sources; distinguish the 62B 795B-token reporting checkpoint from the 780B-token corpus and 8B/540B checkpoints; remove undisclosed PaLM 2 size and token estimates; bound architecture, benchmark, safety, derivative, product, and API claims.
- editorial
Reconstruct the article around feature-specific automation levels, operating boundaries and fallback, system architecture, lifecycle safety assurance, evidence interpretation, incidents, and current regulation; remove unsupported forecasts, driver-assistance/ADS confusion, and vendor-centric claims.
- editorial
Replaces unsupported pricing and comparison claims, corrects TPU generation status and specifications through July 28, 2026, distinguishes chips, slices, pods, and Edge TPU, and adds primary and academic citations.
- editorial
Reworked supervised learning coverage with scoped definitions, leakage-safe evaluation, calibrated metrics, and documented failure modes
- editorial
Reconstruct CLIP from primary sources; correct model-specific text-tower widths, heads, and 49,408-token vocabulary; distinguish the 77-token context; tighten evaluation, limitations, downstream uses, and reproducibility.
- editorial
Rebuild the article around primary technical records, correct the architecture and full carbon-accounting table, distinguish released checkpoints from the withheld 34B models, and replace unsupported influence and deployment claims with bounded evaluations.