https://aiexpert.news/en/article/specvalidator-catches-defective-llm-prompts-before-they-corrupt-code-generation ai|expert en 2026-04-29T02:30:04.332Z SpecValidator Hits 0.804 F1 on Prompt Defect Detection, Doubling Frontier Model MCC https://aiexpert.news/pt/article/specvalidator-catches-defective-llm-prompts-before-they-corrupt-code-generation ai|expert pt-BR 2026-04-29T02:30:04.332Z SpecValidator Atinge F1 de 0,804 na Detecção de Defeitos em Prompts, Dobrando o MCC dos Modelos Frontier https://aiexpert.news/es/article/specvalidator-catches-defective-llm-prompts-before-they-corrupt-code-generation ai|expert es 2026-04-29T02:30:04.332Z SpecValidator Alcanza F1 de 0,804 en Detección de Defectos en Prompts, Duplicando el MCC de los Modelos Frontier https://aiexpert.news/en/article/training-on-disagreement-how-multi-thinker-cot-supervision-outperforms-single-te ai|expert en 2026-04-29T02:18:04.348Z Multi-teacher CoT pooling can be computationally hard, active queries fix it https://aiexpert.news/pt/article/training-on-disagreement-how-multi-thinker-cot-supervision-outperforms-single-te ai|expert pt-BR 2026-04-29T02:18:04.348Z O agrupamento multi-teacher de CoT pode ser computacionalmente difícil — consultas ativas resolvem o problema https://aiexpert.news/es/article/training-on-disagreement-how-multi-thinker-cot-supervision-outperforms-single-te ai|expert es 2026-04-29T02:18:04.348Z El agrupamiento multi-teacher de CoT puede ser computacionalmente difícil — las consultas activas lo resuelven https://aiexpert.news/en/article/hylo-upcycling-pretrained-transformer-checkpoints-into-long-context-hybrid-archi ai|expert en 2026-04-29T02:05:42.310Z AMD HyLo Converts Transformer Checkpoints to 32x Longer Context Without Retraining https://aiexpert.news/pt/article/hylo-upcycling-pretrained-transformer-checkpoints-into-long-context-hybrid-archi ai|expert pt-BR 2026-04-29T02:05:42.310Z AMD HyLo Converte Checkpoints Transformer para Contexto 32x Mais Longo Sem Retreinamento https://aiexpert.news/es/article/hylo-upcycling-pretrained-transformer-checkpoints-into-long-context-hybrid-archi ai|expert es 2026-04-29T02:05:42.310Z AMD HyLo Convierte Checkpoints Transformer a Contexto 32x Más Largo Sin Reentrenamiento https://aiexpert.news/en/article/green-shielding-researchers-expose-how-routine-phrasing-variation-silently-shift ai|expert en 2026-04-29T01:50:04.238Z Safer-Looking LLM Outputs Miss More Critical Diagnoses, Green Shielding Study Finds https://aiexpert.news/pt/article/green-shielding-researchers-expose-how-routine-phrasing-variation-silently-shift ai|expert pt-BR 2026-04-29T01:50:04.238Z Outputs de LLMs com Aparência Mais Segura Erram Mais Diagnósticos Críticos, Aponta Estudo de Green Shielding https://aiexpert.news/es/article/green-shielding-researchers-expose-how-routine-phrasing-variation-silently-shift ai|expert es 2026-04-29T01:50:04.238Z Los Outputs de LLMs con Apariencia más Segura Fallan más Diagnósticos Críticos, Revela Estudio de Green Shielding https://aiexpert.news/en/article/hdet-repurposes-idle-gpu-replicas-to-run-learning-rate-exploration-during-traini ai|expert en 2026-04-29T01:38:04.216Z HDET Converts Allocated GPU Replicas Into a Live Learning-Rate Search Engine https://aiexpert.news/pt/article/hdet-repurposes-idle-gpu-replicas-to-run-learning-rate-exploration-during-traini ai|expert pt-BR 2026-04-29T01:38:04.216Z HDET Converte Réplicas de GPU Alocadas em um Motor de Busca de Learning Rate em Tempo Real https://aiexpert.news/es/article/hdet-repurposes-idle-gpu-replicas-to-run-learning-rate-exploration-during-traini ai|expert es 2026-04-29T01:38:04.216Z HDET Convierte Réplicas de GPU Asignadas en un Motor de Búsqueda de Learning Rate en Tiempo Real https://aiexpert.news/en/article/persona-collapse-why-multi-agent-llm-simulations-converge-to-a-single-behavioral ai|expert en 2026-04-29T01:24:04.290Z Persona Collapse Undermines Multi-Agent LLM Simulations Across Ten Models https://aiexpert.news/pt/article/persona-collapse-why-multi-agent-llm-simulations-converge-to-a-single-behavioral ai|expert pt-BR 2026-04-29T01:24:04.290Z Persona Collapse Compromete Simulações LLM Multiagente em Dez Modelos https://aiexpert.news/es/article/persona-collapse-why-multi-agent-llm-simulations-converge-to-a-single-behavioral ai|expert es 2026-04-29T01:24:04.290Z Persona Collapse Socava las Simulaciones LLM Multiagente en Diez Modelos https://aiexpert.news/en/article/llm-rubrics-match-clinician-scoring-across-823-encounters-unblocking-clinical-ai ai|expert en 2026-04-29T01:04:04.166Z LLM Rubric Scoring Matches Clinician Agreement on 823 Cases at 1,000x Lower Cost https://aiexpert.news/pt/article/llm-rubrics-match-clinician-scoring-across-823-encounters-unblocking-clinical-ai ai|expert pt-BR 2026-04-29T01:04:04.166Z Pontuação por Rubrica de LLM Corresponde ao Acordo entre Clínicos em 823 Casos com Custo 1.000x Menor https://aiexpert.news/es/article/llm-rubrics-match-clinician-scoring-across-823-encounters-unblocking-clinical-ai ai|expert es 2026-04-29T01:04:04.166Z La Puntuación por Rúbrica de LLM Iguala el Acuerdo entre Clínicos en 823 Casos con un Costo 1.000x Menor https://aiexpert.news/en/article/depthkv-layer-aware-kv-cache-pruning-cuts-long-context-llm-memory-overhead ai|expert en 2026-04-29T00:50:04.170Z DepthKV Beats Uniform KV Cache Pruning by Allocating Memory per Layer Sensitivity https://aiexpert.news/pt/article/depthkv-layer-aware-kv-cache-pruning-cuts-long-context-llm-memory-overhead ai|expert pt-BR 2026-04-29T00:50:04.170Z DepthKV Supera o Pruning Uniforme de Cache KV ao Alocar Memória por Sensibilidade de Camada https://aiexpert.news/es/article/depthkv-layer-aware-kv-cache-pruning-cuts-long-context-llm-memory-overhead ai|expert es 2026-04-29T00:50:04.170Z DepthKV Supera el Pruning Uniforme de Caché KV al Asignar Memoria por Sensibilidad de Capa https://aiexpert.news/en/article/agentward-new-security-architecture-addresses-full-lifecycle-vulnerabilities-in- ai|expert en 2026-04-29T00:34:04.170Z FIND-Lab releases AgentWard, a five-layer AI agent security framework https://aiexpert.news/pt/article/agentward-new-security-architecture-addresses-full-lifecycle-vulnerabilities-in- ai|expert pt-BR 2026-04-29T00:34:04.170Z FIND-Lab lança AgentWard, framework de segurança em cinco camadas para agentes de IA https://aiexpert.news/es/article/agentward-new-security-architecture-addresses-full-lifecycle-vulnerabilities-in- ai|expert es 2026-04-29T00:34:04.170Z FIND-Lab lanza AgentWard, un framework de seguridad de cinco capas para agentes de IA https://aiexpert.news/en/article/meta-races-to-unwind-manus-ai-deal-before-beijings-export-control-deadline ai|expert en 2026-04-29T00:15:42.618Z Meta Forced to Dismantle Manus AI Acquisition Under Beijing Deadline https://aiexpert.news/pt/article/meta-races-to-unwind-manus-ai-deal-before-beijings-export-control-deadline ai|expert pt-BR 2026-04-29T00:15:42.618Z Meta Obrigada a Desmantelar Aquisição da Manus AI sob Prazo de Pequim https://aiexpert.news/es/article/meta-races-to-unwind-manus-ai-deal-before-beijings-export-control-deadline ai|expert es 2026-04-29T00:15:42.618Z Meta Obligada a Desmantelar la Adquisición de Manus AI bajo Plazo de Pekín https://aiexpert.news/en/article/anthropic-tests-claude-models-for-safety-research-sabotageand-finds-none ai|expert en 2026-04-29T00:02:04.344Z Anthropic finds Claude does not start safety sabotage but will continue it when primed https://aiexpert.news/pt/article/anthropic-tests-claude-models-for-safety-research-sabotageand-finds-none ai|expert pt-BR 2026-04-29T00:02:04.344Z Anthropic descobre que Claude não inicia sabotagem de segurança, mas a continua quando induzido https://aiexpert.news/es/article/anthropic-tests-claude-models-for-safety-research-sabotageand-finds-none ai|expert es 2026-04-29T00:02:04.344Z Anthropic concluye que Claude no inicia sabotaje de seguridad pero lo continúa cuando se le induce https://aiexpert.news/en/article/alec-radfords-talkie-is-a-13b-llm-trained-on-260b-tokens-of-pre-1931-text-releas ai|expert en 2026-04-28T03:14:49.577Z Alec Radford Releases 13B Model Trained on Pre-1931 Text Under Apache 2.0 https://aiexpert.news/pt/article/alec-radfords-talkie-is-a-13b-llm-trained-on-260b-tokens-of-pre-1931-text-releas ai|expert pt-BR 2026-04-28T03:14:49.577Z Alec Radford Lança Modelo de 13B Treinado em Textos Anteriores a 1931 sob Apache 2.0 https://aiexpert.news/es/article/alec-radfords-talkie-is-a-13b-llm-trained-on-260b-tokens-of-pre-1931-text-releas ai|expert es 2026-04-28T03:14:49.577Z Alec Radford Lanza Modelo de 13B Entrenado con Textos Anteriores a 1931 bajo Apache 2.0 https://aiexpert.news/en/article/doc-to-lora-adaptation-collapses-to-46-accuracy-when-documents-contradict-traini ai|expert en 2026-04-28T02:44:04.398Z Doc-to-LoRA Accuracy Falls to 16% Against Strongly Entrenched Model Facts https://aiexpert.news/pt/article/doc-to-lora-adaptation-collapses-to-46-accuracy-when-documents-contradict-traini ai|expert pt-BR 2026-04-28T02:44:04.398Z Acurácia do Doc-to-LoRA Cai para 16% Contra Fatos Fortemente Consolidados no Modelo https://aiexpert.news/es/article/doc-to-lora-adaptation-collapses-to-46-accuracy-when-documents-contradict-traini ai|expert es 2026-04-28T02:44:04.398Z La Precisión del Doc-to-LoRA Cae al 16% Frente a Hechos Fuertemente Arraigados en el Modelo https://aiexpert.news/en/article/elementsclaw-agentic-framework-closes-the-loop-on-ai-driven-materials-discovery ai|expert en 2026-04-28T02:30:04.353Z ElementsClaw Screens 2.4 Million Crystals in 28 GPU Hours, Finds Four New Superconductors https://aiexpert.news/pt/article/elementsclaw-agentic-framework-closes-the-loop-on-ai-driven-materials-discovery ai|expert pt-BR 2026-04-28T02:30:04.353Z ElementsClaw Analisa 2,4 Milhões de Cristais em 28 Horas de GPU e Descobre Quatro Novos Supercondutores https://aiexpert.news/es/article/elementsclaw-agentic-framework-closes-the-loop-on-ai-driven-materials-discovery ai|expert es 2026-04-28T02:30:04.353Z ElementsClaw Analiza 2,4 Millones de Cristales en 28 Horas de GPU y Descubre Cuatro Nuevos Superconductores https://aiexpert.news/en/article/silent-deepspeed-bug-has-corrupted-rl-fine-tuning-benchmarks-across-trl-openrlhf ai|expert en 2026-04-28T02:15:35.617Z DeepSpeed CPU-Offload Bug Corrupted RLHF Benchmarks in Three Major Frameworks https://aiexpert.news/pt/article/silent-deepspeed-bug-has-corrupted-rl-fine-tuning-benchmarks-across-trl-openrlhf ai|expert pt-BR 2026-04-28T02:15:35.617Z Bug de CPU-Offload do DeepSpeed Corrompeu Benchmarks de RLHF em Três Grandes Frameworks https://aiexpert.news/es/article/silent-deepspeed-bug-has-corrupted-rl-fine-tuning-benchmarks-across-trl-openrlhf ai|expert es 2026-04-28T02:15:35.617Z El Bug de CPU-Offload de DeepSpeed Corrompió Benchmarks de RLHF en Tres Frameworks Principales https://aiexpert.news/en/article/google-deepmind-signs-national-ai-partnership-with-south-korea-eyes-government-a ai|expert en 2026-04-28T01:50:04.392Z Google DeepMind Strikes Lab-Direct AI Partnership with South Korea https://aiexpert.news/pt/article/google-deepmind-signs-national-ai-partnership-with-south-korea-eyes-government-a ai|expert pt-BR 2026-04-28T01:50:04.392Z Google DeepMind firma parceria de IA Lab-Direct com a Coreia do Sul https://aiexpert.news/es/article/google-deepmind-signs-national-ai-partnership-with-south-korea-eyes-government-a ai|expert es 2026-04-28T01:50:04.392Z Google DeepMind establece alianza de IA Lab-Direct con Corea del Sur https://aiexpert.news/en/article/new-method-cuts-the-cost-of-fitting-ai-scaling-laws-by-actively-selecting-which- ai|expert en 2026-04-28T01:30:04.379Z MSPE Fits AI Scaling Laws at 10% of Standard Compute Cost https://aiexpert.news/pt/article/new-method-cuts-the-cost-of-fitting-ai-scaling-laws-by-actively-selecting-which- ai|expert pt-BR 2026-04-28T01:30:04.379Z MSPE Ajusta Leis de Escala de IA com 10% do Custo Computacional Padrão https://aiexpert.news/es/article/new-method-cuts-the-cost-of-fitting-ai-scaling-laws-by-actively-selecting-which- ai|expert es 2026-04-28T01:30:04.379Z MSPE Ajusta Leyes de Escala de IA con el 10% del Costo Computacional Estándar https://aiexpert.news/en/article/llms-systematically-generate-harmful-narratives-about-global-majority-nationalit ai|expert en 2026-04-28T01:18:04.626Z Frontier LLMs show 50x subordination bias against Global Majority nationalities https://aiexpert.news/pt/article/llms-systematically-generate-harmful-narratives-about-global-majority-nationalit ai|expert pt-BR 2026-04-28T01:18:04.626Z LLMs de fronteira apresentam viés de subordinação 50x maior contra nacionalidades da Maioria Global https://aiexpert.news/es/article/llms-systematically-generate-harmful-narratives-about-global-majority-nationalit ai|expert es 2026-04-28T01:18:04.626Z Los LLMs de frontera muestran un sesgo de subordinación 50x contra nacionalidades de la Mayoría Global https://aiexpert.news/en/article/white-box-probe-replaces-opaque-gnn-message-passing-with-interpretable-signal-co ai|expert en 2026-04-28T01:04:04.643Z WG-SRC Replaces GNN Message-Passing with Named, Auditable Signal Components https://aiexpert.news/pt/article/white-box-probe-replaces-opaque-gnn-message-passing-with-interpretable-signal-co ai|expert pt-BR 2026-04-28T01:04:04.643Z WG-SRC Substitui Message-Passing de GNN por Componentes de Sinal Auditáveis e Nomeados https://aiexpert.news/es/article/white-box-probe-replaces-opaque-gnn-message-passing-with-interpretable-signal-co ai|expert es 2026-04-28T01:04:04.643Z WG-SRC Reemplaza el Message-Passing de GNN con Componentes de Señal Auditables y Nombrados https://aiexpert.news/en/article/google-employees-demand-sundar-pichai-reject-classified-military-ai-contracts ai|expert en 2026-04-28T00:50:04.204Z 600 Google Employees Demand Pichai Bar Classified Pentagon AI Use https://aiexpert.news/pt/article/google-employees-demand-sundar-pichai-reject-classified-military-ai-contracts ai|expert pt-BR 2026-04-28T00:50:04.204Z 600 Funcionários do Google Exigem que Pichai Proíba o Uso Classificado de IA pelo Pentágono https://aiexpert.news/es/article/google-employees-demand-sundar-pichai-reject-classified-military-ai-contracts ai|expert es 2026-04-28T00:50:04.204Z 600 Empleados de Google Exigen que Pichai Prohíba el Uso Clasificado de IA por el Pentágono https://aiexpert.news/en/article/how-ai-agents-burn-your-token-budget-first-systematic-study-of-agentic-cost-patt ai|expert en 2026-04-28T00:38:04.458Z First Systematic Study Finds AI Agents Burn 1,000x More Tokens Than Code Chat https://aiexpert.news/pt/article/how-ai-agents-burn-your-token-budget-first-systematic-study-of-agentic-cost-patt ai|expert pt-BR 2026-04-28T00:38:04.458Z Primeiro Estudo Sistemático Revela que Agentes de IA Consomem 1.000x Mais Tokens do que Chats de Código https://aiexpert.news/es/article/how-ai-agents-burn-your-token-budget-first-systematic-study-of-agentic-cost-patt ai|expert es 2026-04-28T00:38:04.458Z El Primer Estudio Sistemático Revela que los Agentes de IA Consumen 1.000x Más Tokens que los Chats de Código https://aiexpert.news/en/article/agentic-world-models-get-a-unified-taxonomy-levels-laws-and-whats-still-missing ai|expert en 2026-04-28T00:18:05.014Z 42-Author arXiv Survey Defines Three Levels for Agentic World Models https://aiexpert.news/pt/article/agentic-world-models-get-a-unified-taxonomy-levels-laws-and-whats-still-missing ai|expert pt-BR 2026-04-28T00:18:05.014Z Pesquisa com 42 Autores no arXiv Define Três Níveis para Modelos de Mundo Agênticos https://aiexpert.news/es/article/agentic-world-models-get-a-unified-taxonomy-levels-laws-and-whats-still-missing ai|expert es 2026-04-28T00:18:05.014Z Investigación de 42 Autores en arXiv Define Tres Niveles para Modelos de Mundo Agénticos https://aiexpert.news/en/article/eu-ai-act-accountability-gap-ai-hiring-bias-fragments-across-vendor-supply-chain ai|expert en 2026-04-28T00:04:04.286Z EU AI Act leaves deployers liable for vendor bias they cannot audit https://aiexpert.news/pt/article/eu-ai-act-accountability-gap-ai-hiring-bias-fragments-across-vendor-supply-chain ai|expert pt-BR 2026-04-28T00:04:04.286Z EU AI Act deixa implantadores responsáveis por viés de fornecedores que não podem auditar https://aiexpert.news/es/article/eu-ai-act-accountability-gap-ai-hiring-bias-fragments-across-vendor-supply-chain ai|expert es 2026-04-28T00:04:04.286Z El EU AI Act deja a los implementadores responsables por sesgos de proveedores que no pueden auditar https://aiexpert.news/en/article/tencent-open-sources-hunyuanworld-10-first-simulation-ready-3d-world-generator-f ai|expert en 2026-04-27T23:50:04.224Z Tencent Open-Sources HunyuanWorld 1.0, a Mesh-Ready 3D World Generator https://aiexpert.news/pt/article/tencent-open-sources-hunyuanworld-10-first-simulation-ready-3d-world-generator-f ai|expert pt-BR 2026-04-27T23:50:04.224Z Tencent Abre o Código de HunyuanWorld 1.0, um Gerador de Mundos 3D Pronto para Mesh https://aiexpert.news/es/article/tencent-open-sources-hunyuanworld-10-first-simulation-ready-3d-world-generator-f ai|expert es 2026-04-27T23:50:04.224Z Tencent Libera el Código de HunyuanWorld 1.0, un Generador de Mundos 3D Listo para Mesh https://aiexpert.news/en/article/abstract-chain-of-thought-latent-reasoning-tokens-cut-inference-cost-without-sac ai|expert en 2026-04-27T23:30:04.416Z IBM's ACoT Cuts Reasoning Tokens 11.6x Without Accuracy Loss https://aiexpert.news/pt/article/abstract-chain-of-thought-latent-reasoning-tokens-cut-inference-cost-without-sac ai|expert pt-BR 2026-04-27T23:30:04.416Z O ACoT da IBM Reduz Tokens de Raciocínio em 11,6x Sem Perda de Precisão https://aiexpert.news/es/article/abstract-chain-of-thought-latent-reasoning-tokens-cut-inference-cost-without-sac ai|expert es 2026-04-27T23:30:04.416Z El ACoT de IBM Reduce los Tokens de Razonamiento 11,6x Sin Pérdida de Precisión https://aiexpert.news/en/article/canonical-charts-ai-integration-roadmap-for-ubuntu-linux ai|expert en 2026-04-27T23:15:36.605Z Canonical Sets 2026 Ubuntu AI Roadmap With On-Device Inference and Agentic Admin https://aiexpert.news/pt/article/canonical-charts-ai-integration-roadmap-for-ubuntu-linux ai|expert pt-BR 2026-04-27T23:15:36.605Z Canonical Define Roteiro de IA do Ubuntu para 2026 com Inferência On-Device e Administração Agêntica https://aiexpert.news/es/article/canonical-charts-ai-integration-roadmap-for-ubuntu-linux ai|expert es 2026-04-27T23:15:36.605Z Canonical Establece el Roadmap de IA de Ubuntu para 2026 con Inferencia On-Device y Administración Agéntica https://aiexpert.news/en/article/david-silvers-new-lab-raises-11b-to-build-ai-that-learns-without-human-data ai|expert en 2026-04-27T23:02:15.389Z David Silver's Ineffable Intelligence Raises $1.1B to Replace Human Training Data https://aiexpert.news/pt/article/david-silvers-new-lab-raises-11b-to-build-ai-that-learns-without-human-data ai|expert pt-BR 2026-04-27T23:02:15.389Z Ineffable Intelligence, de David Silver, capta $1.1B para substituir dados de treinamento humanos https://aiexpert.news/es/article/david-silvers-new-lab-raises-11b-to-build-ai-that-learns-without-human-data ai|expert es 2026-04-27T23:02:15.389Z Ineffable Intelligence, de David Silver, recauda $1.1B para reemplazar datos de entrenamiento humanos https://aiexpert.news/en/article/openaimicrosoft-tear-up-exclusivity-no-more-lock-in-no-agi-clause ai|expert en 2026-04-27T22:34:15.281Z OpenAI Ends Azure Exclusivity and Drops the AGI License Clause https://aiexpert.news/pt/article/openaimicrosoft-tear-up-exclusivity-no-more-lock-in-no-agi-clause ai|expert pt-BR 2026-04-27T22:34:15.281Z OpenAI Encerra Exclusividade com Azure e Elimina Cláusula de Licença de AGI https://aiexpert.news/es/article/openaimicrosoft-tear-up-exclusivity-no-more-lock-in-no-agi-clause ai|expert es 2026-04-27T22:34:15.281Z OpenAI Termina la Exclusividad con Azure y Elimina la Cláusula de Licencia de AGI https://aiexpert.news/en/article/openai-achieves-fedramp-moderate-opening-federal-ai-deployments ai|expert en 2026-04-27T22:20:15.717Z OpenAI Earns FedRAMP Moderate for ChatGPT Enterprise and API Platform https://aiexpert.news/pt/article/openai-achieves-fedramp-moderate-opening-federal-ai-deployments ai|expert pt-BR 2026-04-27T22:20:15.717Z OpenAI Obtém FedRAMP Moderate para ChatGPT Enterprise e Plataforma de API https://aiexpert.news/es/article/openai-achieves-fedramp-moderate-opening-federal-ai-deployments ai|expert es 2026-04-27T22:20:15.717Z OpenAI Obtiene la Certificación FedRAMP Moderate para ChatGPT Enterprise y la Plataforma de API