Este conteúdo relata correções recentes e novos templates de chat para os modelos Gemma 4 do Google, visando melhorar o orçamento de raciocínio e a chamada de ferramentas. Ele oferece instruções para usuários de llama.cpp sobre como aplicar esses novos templates para garantir o funcionamento correto.
This post explains how to use Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) to enhance the tool-calling accuracy of small language models. It details how to leverage Amazon SageMaker AI training jobs to focus on training code and evaluate model quality.
Este artigo apresenta uma estrutura de avaliação de agentes de IA baseada na experiência de produção, enfatizando a importância de alinhar a ferramenta ao fluxo de trabalho em vez de focar apenas em benchmarks. Os critérios chave incluem confiabilidade com dados reais, qualidade da chamada de ferramentas, comportamento da janela de contexto para fluxos longos e custo em escala para determinar a viabilidade.
Needle is a 26-million-parameter model that successfully distilled Gemini's tool-calling capabilities, achieving near-parity accuracy at a fraction of the compute cost. This breakthrough is crucial for developers building AI agents and edge deployments.
AI agents have evolved beyond passive chatbots. This content details implementing permission-gated tool calling in Python agents.
This document outlines a comprehensive roadmap for mastering tool calling in AI agents. It aims to guide readers through the essential concepts and practices for implementing and optimizing this crucial functionality.
This content highlights five small, open-weight language models that are notable for their support of structured tool calling, making them suitable for agentic applications.
This tutorial demonstrates how to equip Gemma 4 with new tools. It showcases the model making autonomous decisions on when to gather information or perform computations.
This paper studies tool-calling in large language model (LLM) agents, examining its effectiveness and efficiency. It analyzes evaluation pipelines, showing results are sensitive to implementation choices, and identifies computational waste in reinforcement learning training.
ToolWeave is a structured framework for synthesizing realistic multi-turn tool-calling dialogues, essential for LLMs to function as autonomous agents. It addresses challenges in existing synthetic data generation by supporting realistic multi-step workflows and reducing parameter hallucination.
MAVEN (Modular Agentic Verification and Execution Network) is a lightweight symbolic reasoning scaffold designed to improve generalization in agentic tool-calling environments. It has been evaluated across established benchmarks, and introduces MAVEN-Bench, a new stress-test benchmark for multi-step mathematical and physical reasoning.
This article discusses why Hermes 4's tool-calling is trained as a separate skill, explaining the underlying rationale and its critical implications for AI agents. It highlights how this distinct training approach enhances agent capabilities and overall performance.
Este conteúdo descreve como construir agentes de IA úteis em um backend Go usando LangChainGo para integrar ferramentas que buscam dados e acionam ações. Ele utiliza Ollama com o modelo qwen3.5:9b para execução local, incluindo instruções de configuração e uma introdução à definição de ferramentas no LangChainGo.
Este conteúdo aborda os desafios de construir agentes de IA reais que vão além de responder perguntas, necessitando de ações e ferramentas. Critica as limitações das APIs stateless, como a da Anthropic, que dificultam processos complexos de chamada de ferramentas e a manutenção de estado.
Together AI has expanded its fine-tuning service with native support for tool calling, reasoning, and vision-language models. The enhancements also include 100B+ model training, up to 6x higher throughput, and job cost and ETA estimates.
This article critiques Cloudflare's code mode, highlighting its shortcomings concerning Multi-Cloud Platform (MCP) functionalities and the effective implementation of AI tool calling. It suggests areas where Cloudflare's approach might be incomplete or misaligned with advanced AI integration needs.
![[Video Response] What Cloudflare's code mode misses about MCP and tool calling](/cdn-cgi/image/width=3840,quality=75,format=webp/https://i1.ytimg.com/vi/0bpYCxv2qhw/hqdefault.jpg)