This paper explores "deceptive alignment" in LLMs, a key challenge in AI safety where models deliberately produce false outputs while maintaining accurate internal representations. Researchers introduced a multi-model paradigm, successfully detecting synthetic dishonesty with high accuracy using linear probes across various transformer architectures.
This paper introduces 'Autonomous Agentic Data Engineering,' a novel task to evaluate LLMs as autonomous data engineers for model specialization through end-to-end data curation. Experiments show autonomous LLM data engineers achieve substantial gains, with GPT-5.2 improving a student model by 57.29%.
This paper proposes a modular reference architecture for Embedded Agent Systems, addressing the challenges of deploying agentic AI in pervasive computing environments with strict memory and energy constraints. It introduces a tiered design that decouples on-device agents (compressed neural networks) from cloud-augmented agents (SLMs) for different reasoning levels.
O artigo argumenta que o foco em sistemas de IA mudou dos LLMs individuais para um "Agentic Stack" completo, onde o LLM é apenas um componente. Ele detalha a pilha composta por Orchestrator (o cérebro), Ferramentas, Memória e LLM, enfatizando que a inteligência real e a eficácia em produção residem no Orchestrator e no design do sistema, não apenas nos prompts ou no modelo.
This research introduces Incremental Completion Decomposition (ICD), a novel jailbreak strategy that exploits weaknesses in LLM safety mechanisms by eliciting sequences of single-word continuations. ICD demonstrates superior Attack Success Rate (ASR) on various benchmarks compared to existing methods, providing theoretical and mechanistic evidence for its effectiveness.
The author shares a transformative experience witnessing AI-generated code rapidly replace a micro-SaaS service, challenging previous doubts about LLMs' impact on SaaS. This economic and efficiency shift promises a new era in software creation, drastically cutting development time and demanding adaptation from the industry.
Este artigo explora se os LLMs aproximam quantitativamente o significado social humano e se estratégias de prompting pragmático podem melhorar essa aproximação. Para isso, introduz métricas de calibração (ESR, CDS) e observa que os modelos reproduzem a estrutura qualitativa das inferências sociais humanas, mas diferem substancialmente em outros aspectos.
SocioEval é um framework baseado em templates para avaliar sistematicamente o viés de status socioeconômico em modelos de fundação, incluindo LLMs, uma área pouco explorada. A pesquisa avaliou 13 LLMs e revelou variações substanciais nas taxas de viés (0,42% a 33,75%), manifestando-se de forma diferente em vários temas.
MCBench is a new benchmark designed to assess the safety of Omni Large Language Models across vision, audio, and text inputs, revealing significant challenges in integrating multiple modalities for accurate safety judgments. It highlights that current Omni LLMs lack robust cross-modal reasoning in safety-critical settings.
The paper introduces EHRBench, an automated and reliable EHR-grounded benchmark for evaluating LLM-based clinical decision-making, addressing the insufficient understanding of LLMs' reliability in real-world clinical tasks. Its goal is to ensure both scale and quality in the evaluation of Clinical Decision Making (CDM) models.
This paper proposes Reflective Prompt Tuning (RPT), a framework that uses large language model (LLM) function calling to simulate the iterative workflow of human prompt engineers. Its goal is to automate prompt optimization, reducing manual effort and overcoming limitations of existing methods that fail to capture systematic error patterns.
Current AI energy benchmarks, typically measuring single invocations, misrepresent the cost for agentic systems that involve multi-step orchestration and retries. A-LEMS introduces Energy per Successful Goal (EpG) to aggregate total workflow energy, including failures, providing a more accurate measure of goal completion costs.
LiftQuant is a novel framework for continuous bit-width control in Large Language Models, addressing limitations of integer-based quantization. It employs a "lift-then-project" mechanism to achieve quasi-continuous bit-width tuning for optimal deployment.
Este artigo aborda o desafio de navegar em bases de código desconhecidas e propõe o uso de Large Language Models (LLMs) para responder a perguntas em linguagem natural sobre o código. Ele se propõe a ser um guia prático para construir um sistema robusto e privado de Q&A de código baseado em LLMs, explorando arquitetura técnica e código.
This post describes a Zero Trust AI architecture built with Google Gemini, React, and FastAPI to automate logistics dispatch chats while mitigating data leaks and AI hallucinations. The system ensures LLM isolation via Pydantic schemas, includes a human-in-the-loop for critical cases, and deanonymizes data only at the backend.
The article explores the evolution of integrating Large Language Models (LLMs) with external data, introducing the Model Context Protocol (MCP). It compares MCP with traditional "Tools" (Function Calling), highlighting their fundamental differences and its potential to solve issues like vendor lock-in and fragmentation in AI development.
This paper proposes a companion-specific governance profile for single-user knowledge wikis, addressing the unique failure mode of entrenchment under user-coupled drift. It discusses emerging personal AI memory architectures from 2026, including RAG-based systems and wiki-style designs, alongside established academic and production memory systems.
TTKV proposes a temporal-tiered KV cache management framework for LLMs, inspired by human memory, to address the linear scaling of KV cache memory. It partitions the cache into tiers with heterogeneous capacity and precision, assigning more recent KV states to faster, higher-precision tiers.
The author stopped using ChatGPT for code due to its inability to retain file context across conversations, making it inefficient for complex projects. They now prefer Claude for its larger context window and superior reasoning, and Cursor for its deep integration with the entire codebase.
The content highlights a critical flaw in current LLM code generation evaluations: they often fail to capture real-world correctness beyond superficial passes. It argues against simplistic unit test benchmarks and proposes a more nuanced `weighted_accuracy` approach to uncover subtle failure modes.