The Morning
From the arXiv
From Chatbot to Digital Colleague: The Paradigm Shift Toward Persistent Autonomous AI
his paper proposes a paradigm shift from "Chatbot" to "Digital Colleague" for AI. The core method involves advancing LLMs from simple conversational generators to integrated systems with persistent memory, reasoning, and self-improvement capabilities. This transition is achieved through enhanced cognitive processes like Chain-of-Thought reasoning and the development of tool-augmented workstation systems with persistent workspaces and verification loops.
GitOfThoughts: Version-Controlled Reasoning and Agent Memory You Can Replay, Diff, and Merge
GitOfThoughts addresses the ephemeral nature of LLM reasoning by treating it as a version-controlled process, akin to software development. It stores reasoning steps as Git commits, allowing for replayability, auditing, and merging of agent thought processes. …
When Errors Become Narratives: A Longitudinal Taxonomy of Silent Failures in a Production LLM Agent Runtime
This paper introduces a longitudinal taxonomy of "silent failures" in LLM agent runtimes, where errors go unnoticed. The core method involves an eight-week study of a production personal-assistant agent, identifying 22 incidents and a meta-pattern of uncommuni…
Code Correctness Signals in LLM Hidden States: Pre-Generation Probing and Repair Geometry
This paper investigates whether Large Language Models (LLMs) encode code correctness in their internal states. The core method involves probing LLM hidden states to see if code correctness can be predicted *before* generation and during *repair* of incorrect c…
ClinHallu: A Benchmark for Diagnosing Stage-Wise Hallucinations in Medical MLLM Reasoning
This paper introduces ClinHallu, a benchmark designed to diagnose stage-wise hallucinations in medical multimodal large language models (MLLMs). It decomposes MLLM reasoning into visual recognition, knowledge recall, and reasoning integration, allowing for the…
From Shield to Target: Denial-of-Service Attacks on LLM-Based Agent Guardrails
This paper introduces a novel denial-of-service (DoS) attack against LLM-based guardrails designed to protect autonomous agents. The core method involves crafting specific natural-…
No Accidental Software Agent First Canonical Code for Human Code Entropy Reduction and 30 to 500 times Lower Frontier Model Requirements
This paper proposes "agent-first canonical code" to reduce the "accidental entropy" in human-written software repositories. The core method involves rewriting software into structu…
SIMMER: Benchmarking Latent Failures in LLM Executable Planning with a World Model
This paper introduces SIMMER, a novel benchmark designed to evaluate latent failures in Large Language Model (LLM) planning for autonomous agents. SIMMER utilizes a human-curated s…
StreamMemBench: Streaming Evaluation of Agent Memory for Future-Oriented Assistance
StreamMemBench introduces a novel two-step task sequence to evaluate how well agent memory can leverage past observations and interactions for future-oriented assistance. The bench…
tap: A File-Based Protocol for Heterogeneous LLM Agent Collaboration
This paper introduces "tap," a file-based protocol enabling LLM agents from different vendors (like Claude and Codex) to collaborate on a shared codebase without needing a common r…
The Town Square
Many people are only using AI for specific, limited tasks rather than for every aspect of their lives, similar to how they consume other technologies.
Workshops
TeslaMate is a self-hosted data logger that collects and visualizes your Tesla's driving data, offering insights into your vehicle's usage and performance.
Agent-Reach provides a unified CLI to grant AI agents access to read and search content across major platforms like Twitter, Reddit, YouTube, and GitHub without incurring API fees.
