The Morning
From the arXiv
Agent-Editing World Model: Rethinking World Modeling for LLM Agents
his paper introduces the Agent-Editing World Model (AEWM) to improve LLM agent performance on long-horizon tasks. Instead of predicting environment observations, AEWM models how reasoning and actions influence task progress. It achieves this by distinguishing between critical, exploratory, and noisy decisions and then revising noisy reasoning-action sequences, directly integrating these edits with real execution.
Can LLMs Reason About Runtime Behavior? A Repository-Level Dynamic Benchmark
This paper introduces SWE-Flux, a novel repository-level benchmark designed to evaluate LLMs' ability to reason about code execution. Its core method involves creating 480 execution-grounded questions across 12 real Python repositories, with answers automatica…
Learning from Failures: Heterogeneous Graph Memory for Small Language Model Tool-Using Agents
This paper introduces FRESH, a novel method for improving small language model agents' tool-using capabilities. FRESH constructs a heterogeneous graph memory from past successful and failed tool executions, explicitly capturing the causal dependencies and safe…
Controlling Collectives of AI Agents in Reasoning Space with Spatial Transformers
This paper introduces COMPASS, a decentralized architecture for controlling large groups of AI agents. It uses a spatial transformer on each agent to generate local feedback by aggregating multi-hop messages, enabling scalable and cohesive collective behavior.…
PASTABench: Proactive Assessment of Sequential Trajectories for Agent Safety
This paper introduces PASTABench, a benchmark for evaluating the safety of autonomous LLM agents performing multi-step tasks. Its core method, Decoupled Proactive Safety Monitoring, allows for timely intervention by assessing *whether*, *when*, and *how risky*…
Shopping by algorithm: How agentic AI deploys human heuristics as a surrogate consumer
This paper investigates how AI shopping agents, acting as surrogate consumers, are influenced by marketing pricing cues. The core method involves "Tool-Lab," which simulates costly…
StudentBench: AI and human tutoring yield equivalent GRE learning gains
This paper introduces **StudentBench**, a platform for evaluating AI tutors by collecting large-scale data on student-AI interactions. The core finding is that AI tutoring achieves…
Fine-Tuning LLMs for Translation: General Forgetting Mitigation Does Not Preserve MT-Specific Instruction Following
This paper investigates whether general forgetting mitigation techniques for LLMs effectively preserve machine translation (MT) capabilities and MT-specific instruction following (…
Exact Feedback Is Not Control: Evaluating Text-based Closed-Loop Revision in LLMs
This paper introduces a novel protocol for evaluating LLM closed-loop revision by using deterministic verifiers to ensure feedback is always correct and complete. Their core contri…
Beyond Poetry: Can Large Language Models Generate Classical Arabic Maqamat?
This paper investigates whether Large Language Models (LLMs) can generate classical Arabic *maqamat*, a complex prose genre. The core method involves a controlled evaluation of fiv…
The Town Square
Qualcomm announced that Linux support is coming to Snapdragon X Series processors, enabling a wider range of AI-powered PCs.
Workshops
Ax is an open-source agentic orchestration runtime that simplifies the creation and management of complex, multi-step AI workflows by providing a flexible framework for defining, executing, and monitoring agent interactions.
This CLI tool simplifies the setup and oversight of Claude Code, offering pre-defined configurations and monitoring capabilities for efficient code generation and management.
