The Morning
From the arXiv
A Baseline Assessment for the Agentic Web Technology Stack
his paper introduces MASTIF, a framework for evaluating LLM-based agent technologies. It provides a baseline assessment of various LLMs, protocols, and frameworks on 300 Mind2Web tasks. The findings suggest that choosing a task-specific technology stack can significantly improve performance metrics like task completion and efficiency, while reducing verbosity and completion time.
A Multi-Agent LLM System for Automated Tuning of Specialized Neural Network Models
This paper introduces a multi-agent LLM system for automatically tuning specific components within pre-defined neural network architectures. Its core method involves three LLM agents (analyst, coder, recorder) interacting with an experiment log and execution e…
Matched-Context Group Policy Optimization for LLM Agent Training
This paper introduces Matched-Context Group Policy Optimization (MCGPO) to address credit assignment in training language model agents with sparse rewards. MCGPO constructs weighted action preferences by matching intermediate states and progress within groups …
Humanize Olympiad Agents: A Multi-Session Workflow for Olympiad Problem Solving with Solve, Review, Wiki, Verify
This paper introduces HOA (Humanize Olympiads Agent), a novel framework designed to improve LLM performance on complex Olympiad problems. HOA's core method involves a multi-session workflow where an agent's solutions are iteratively refined through independent…
Inject a Turn, Not a Token: Learning Where to Redirect LLM Reasoning for Self-Correction
This paper introduces UPTURN, a novel self-correction method for LLMs that avoids fine-tuning. UPTURN learns to identify points of reasoning failure and redirects the LLM's generation by injecting a corrective instruction as a new user turn, rather than modify…
RATED: Reliability-Aware AI Agent Tool Retrieval via Execution Dynamics
RATED addresses the critical issue of AI agent tool retrieval by going beyond semantic relevance to incorporate tool reliability during execution. Its core method combines a frozen…
SWE-DeceptionBench: Benchmarking Deceptive Behaviors of LLMs Across the Software Development Lifecycle
This paper introduces SWE-DeceptionBench, a benchmark designed to evaluate deceptive behaviors of Large Language Models (LLMs) within the software development lifecycle. The benchm…
Task Context Shapes Component Roles in Language Models
This paper proposes a network view to understand how language models reuse components across tasks. By analyzing the influence and importance of attention heads under task changes,…
X-AgentRed: Red-Teaming Personal Agents in Evolving Multi-Service Ecosystems
X-AgentRed addresses the risk of unintended harmful outcomes from personal agents interacting with multiple services. Its core method is a self-evolving red-teaming framework that …
TabReliBench: Evaluating Tabular Foundation Models Beyond Predictive Accuracy
This paper introduces TabReliBench, a benchmark for evaluating tabular foundation models (TFMs) on reliability beyond just predictive accuracy. It assesses TFMs across various metr…
The Town Square
The story highlights a growing disconnect between the public's perception of Large Language Model capabilities and the reality of their current limitations.
Workshops
AnyPS5 is a tool that automatically ports PS5 executables to Linux and Windows, simplifying cross-platform game development and preservation.
ArtCraft is an intentional crafting engine designed to empower artists, designers, and filmmakers by providing tools for deliberate and creative content generation.
