The DAIR Daily - ML Papers of the Week

\n

GLM-5: From Vibe Coding to Agentic Engineering

\n By Zhipu AI Research\n

\n Zhipu AI introduces GLM-5, a next-generation foundation model designed to transition the paradigm of \"vibe coding\" to rigorous agentic engineering. Building upon the agentic, reasoning, and coding (ARC) capabilities of its predecessor, GLM-5 adopts DeepSeek Sparse Attention (DSA) to significantly reduce training and inference costs while maintaining state-of-the-art performance in software engineering tasks.\n

\n Read the Paper →\n

\n\n

\n

Team of Thoughts: Scaling Agentic Systems

\n By Wong et al.\n

\n Researchers present \"Team of Thoughts,\" a framework for efficient test-time scaling of agentic systems through orchestrated tool calling. By using an orchestrator-based aggregation method, the system achieves superior performance on reasoning and code generation benchmarks like AIME24 and LiveCodeBench, while reducing inference costs by an order of magnitude compared to traditional multi-agent systems.\n

\n Read the Paper →\n

\n

Benchmarks & Evaluation

\n

SkillsBench: Measuring Agentic Proficiency

\n By Li et al.\n

\n SkillsBench is a new benchmark consisting of 86 tasks across 11 domains, paired with curated skills and deterministic verifiers. It aims to provide a rigorous framework for measuring how well AI agent skills work across diverse real-world tasks, filling a critical gap in current evaluation methodologies.\n

\n Read the Paper →\n

\n\n

\n

MemoryArena: Testing Multi-Session Memory

\n By 2077 AI & Partners\n

\n Researchers from Stanford, UCSD, and other institutions introduce MemoryArena, a benchmark for evaluating AI agent memory in interdependent multi-session tasks. The benchmark spans four diverse domains and focuses on how agents retain and evolve information over time, a key requirement for persistent AI assistants.\n

\n Read the Paper →\n

\n

Memory & Context

\n

LCM: Lossless Context Management

\n By Voltropy\n

\n Voltropy introduces LCM, a system for persistent memory in LLM agents that addresses the problem of \"context rot.\" By using Recursive Language Models (RLMs) that write their own Python scripts to manage context, LCM ensures that reasoning quality does not collapse as the context window fills with stale content.\n

\n Read the Paper →\n

\n\n

\n

Intelligent AI Delegation Framework

\n By Google DeepMind\n

\n Google DeepMind proposes a formal framework for intelligent AI delegation, applying organizational principles to multi-agent systems. The framework uses cryptographic tokens to ensure secure, accountable, and verifiable task assignment, optimizing task decomposition and agent coordination in complex workflows.\n

\n Read the Paper →\n

\n

History