This paper introduces STOMP, a novel offline reinforcement learning algorithm for multi-objective optimization using smooth Tchebysheff scalarization. It addresses the limitation of linear scalarization in recovering non-convex Pareto fronts, crucial for aligning large language models and other real-world applications with conflicting rewards.
This commentary introduces PEEL, a working scaffolding combining deterministic distant reading with LLM interpretation, grounded in Peircean semiotics and abductive reasoning. Applied to AI-generated condensations, PEEL reveals systematic distortions invisible without non-AI measurement, implying deterministic instruments must accompany AI tools to ensure fidelity and epistemic authority.
This paper proposes an LLM-guided multi-agent framework for early-stage engineering design, integrating a human-in-the-loop approach and formal risk management. It uses specialized agents to explore and prune design candidates, demonstrated on aerodynamic airfoil design.
This content focuses on comparing the costs of various AI models, highlighting cheaper alternatives to GPT-4o. It explores significant savings by using models like GPT-4o-mini, DeepSeek V4 Flash, and Qwen3-32B, which can be up to 40 times more cost-effective.
This article details an indie hacker's discovery of substantial cost savings by leveraging alternative AI models via the Global API, comparing their pricing against GPT-4o. It highlights how developers can reduce expenses for large language model inference using a wide range of available options.
This article delves into the crucial distinction between a chatbot with tools and a true AI agent, arguing that the confusion between the two is why many "AI agent" startups fail. It explores what truly makes a language model an agent, capable of taking real actions and chaining them together autonomously.
ChatCAD is an interactive computer-aided diagnosis system that leverages Large Language Models to analyze medical images. It aims to enhance the accuracy and efficiency of medical diagnosis through artificial intelligence.
This paper introduces "insight anticipation," a novel task where language models predict the core insight of a future scientific paper from its foundational predecessors. To evaluate this capability, the authors developed GiantsBench, a benchmark of 17,000 examples, and present GIANTS-4B, an LM trained with reinforcement learning.
Large language models often have a significantly smaller practical context window than their advertised nominal limit due to overhead and attention degradation. This discrepancy affects prompt design and leads to quality drops and truncation long before the hard token limit is reached.
This study investigates cross-prompt generalization in detecting AI-generated fake news using interpretable linguistic features like lexical diversity and readability. A random forest classifier achieved consistently high performance (AUC 0.988-1.000) across various train-test combinations, demonstrating robustness against different prompting strategies.
This paper proposes POLAR, a multimodal memory-augmented framework for personalized embodied agents over long-term user interactions. POLAR organizes prior interactions into a multimodal knowledge graph, capturing semantic and episodic memory to guide embodied task execution.
Este artigo discute como o output detalhado do comando `pip install` é desnecessário e prejudicial para o contexto de modelos de IA como o Claude, que precisam apenas saber se a instalação de pacotes Python foi bem-sucedida ou falhou. Detalhes verbosos como barras de progresso e logs de compilação são considerados ruído que não auxilia a IA na depuração.
This research introduces a data-efficient fine-tuning framework to teach large language models to effectively code-switch for reasoning tasks. It identifies beneficial code-switched behaviors, moving beyond treating code-switching as an error, through systematic analysis of diverse reasoning traces.
This paper presents a necessary condition for intra-group learning algorithm design in Reinforcement Learning, requiring objectives to maintain gradient exchangeability across token updates to prevent reward-irrelevant drift. It proposes minimal transformations to restore this cancellation structure, which stabilizes training and improves sample efficiency.
This paper introduces Structured Progressive Knowledge Activation (SPARK) to address the challenge of integrating architectural knowledge in LLM-driven Neural Architecture Search (NAS). SPARK mitigates "functional entanglement" by enabling factor-conditioned editing, leading to more targeted and reliable architecture modifications.
Mango is a multi-agent web navigation method that optimizes complex website exploration by leveraging a global view. It dynamically determines optimal starting points and adaptively allocates navigation budget, achieving a 63.6% success rate with GPT-5-mini, outperforming the best baseline by 7.3%.
This work addresses the challenge of missing modalities in multimodal clinical data for diagnosis by reframing it as an autoregressive sequence modeling task. It leverages causal decoders from LLMs and a missingness-aware contrastive pre-training to outperform baselines on benchmarks like MIMIC-IV and eICU.
This work addresses the significant memory footprint of Key-Value (KV) caching in transformer language models, proposing optimization through the depth dimension. It introduces a method for cross-layer cache sharing, demonstrating that dropping a layer's cache can be efficient without information loss, and suggests a training approach with random cross-layer attention.
This paper introduces a mathematical framework to analyze the recursive process where AI-generated text re-enters and shapes the public record from which LLMs learn. It distinguishes between "drift," which removes rare forms through unfiltered reuse, and "selection," which filters content based on criteria like quality, showing normative selection preserves deeper linguistic structures.
Geometry-Lite is a novel prompt-level probe designed to interpret how safety evidence develops across layers in large language models. It analyzes layer-wise margin geometry using various readouts to understand boundary formation, improving safety detection over single-layer probes.