The Morning
From the arXiv
Accurate but Not Humble: Evaluating Epistemic Humility in LLM Agents under Knowledge Conflict
his paper introduces a new evaluation metric called "epistemic humility" (EH) for LLM agents, focusing on their ability to acknowledge and act on uncertainty when faced with conflicting information. The authors propose a framework (Identify, Solve, Escalate) to measure EH and demonstrate that high task accuracy doesn't guarantee an agent exhibits this crucial trait, highlighting a gap in current LLM agent evaluations.
Caught in the Act: Probes Effectively Detect Sabotage and Catch Unverbalized Deception
This paper introduces a novel probe architecture for detecting deception in Large Language Models (LLMs) by aggregating information across multiple layers and tokens. By training these probes on the largest deception dataset to date, they achieve high accuracy…
OnTrack: Real-Time Monitoring and Intervention in LLM Agent Trajectories via Streaming Structure-Aware Optimal Transport
OnTrack introduces a real-time monitoring system for LLM agents that compares their current actions against successful historical trajectories using streaming structure-aware optimal transport. This allows for immediate alerts or intervention within millisecon…
Predicting Alignment Generalization with Value Representations
This paper introduces the task of predicting how fine-tuning a language model for one value affects its behavior on other, unseen values. The core method involves analyzing model activations when applying values in context, which significantly outperforms usin…
Recursive Self-Improvement through Multi-Agent Self-Supervision
This paper introduces Multi-Agent Self-Supervision (MASS) to overcome the supervision bottleneck in recursive self-improvement for complex, non-verifiable tasks. MASS enables a single model to iteratively design, execute, and evaluate multi-agent workflows, op…
Language Models as AI Research World Models
This paper proposes using Language Models as Research World Models (RWMs) to predict experimental outcomes in AI research. The core method involves training LMs on past experimenta…
A Closer Look at Agentic BBO: Benchmarking LLM Agents for Black-Box Optimization
This paper introduces AgenticBBO-Bench, a unified benchmark for evaluating LLM agents in black-box optimization (BBO) across diverse domains. The core method involves a standardize…
ARC: A Reasoning Recipe for Robot Foundation Models
This paper introduces ARC, a "reasoning recipe" that significantly improves robot foundation models (RFMs) without requiring more data or larger models. ARC's core method involves …
Ecology of AI Agents: Collaboration Creates a Population Threshold for Takeoff
This paper introduces "ecological safety" for AI, focusing on how a population of misaligned AI agents can grow uncontrollably. Their core method is developing an ecological theory…
LeWAM: A JEPA World Action Model with Diffusion-Steering-Based MPC
LeWAM is a novel world action model that uses a bidirectional transformer trained end-to-end on a decoder-free JEPA latent space. This approach improves state representation by bet…
The Town Square
OpenAI's annualized revenue is now projected to be $20 billion less than previously indicated, impacting its financial outlook.
Workshops
This repository provides self-contained HTML+SVG diagram designs for various AI tools, offering 42 types without shadows or Mermaid syntax.
This repository provides agents for reverse engineering, capable of analyzing app behavior and native binaries.
