The Morning
From the arXiv
Agent in a Bottle: Can LLM Agents Turn Their Capabilities Into Cheap, Scalable Artifacts?
his paper introduces "bottling," the ability of LLM agents to autonomously create cost-effective, task-specific solutions for large workloads. The core method involves agents deciding how to tackle an entire unlabelled dataset within fixed resource budgets, potentially by training smaller models or writing programs. The key contribution is demonstrating that strong general LLM performance doesn't guarantee good "bottling" capabilities, with many agents failing to match their zero-shot performance when tasked with creating scalable artifacts.
Principled Under Pressure: Post-Training Decides Whether LLMs Act on Their Own Moral Judgment
This paper introduces a novel method to evaluate whether Large Language Models (LLMs) act on their own moral judgments when faced with pressure. By posing scenarios twice – once as an agent's decision and once as a third-person judgment – the researchers can i…
ScienceClaw: Benchmarking Continual Self-Evolution of AI-for-Science Agents Across the Natural and Social Sciences
This paper introduces ScienceClaw, a framework for benchmarking the continual self-evolution of AI-for-science agents. Its core method is fixed-parameter program self-evolution, which integrates task solving, scientific verification, and program updates. Scien…
Semantic Behavioral Watermarking: Paraphrase-Robust and Forgery-Resistant Provenance for LLM Agents
This paper introduces Semantic Behavioral Watermarking (SBW), a novel method for embedding provenance in LLM agents' high-level action choices, independent of output tokens. SBW achieves paraphrase robustness by watermarking semantic action clusters and is for…
SquidAgent: Parallelize Wisely, Coordinate Efficiently
This paper introduces SquidAgent, a method for parallelizing LLM-based agents to reduce latency in multi-step tasks. It addresses the inefficiency of existing parallel systems by identifying and mitigating "re-exploration" and "alignment" costs. SquidAgent pro…
Toward Alignment Scaling Laws: A Framework and First Preregistered Measurements
This paper proposes a framework to study how model alignment scales with capability, treating alignment as a family of measurable scaling relations rather than a single property. T…
WorldSolver: Can LLM Agents Simulate the Physical Dynamics via Solver Generation?
WorldSolver introduces a benchmark of 168 physics simulation tasks from computer graphics to evaluate LLM agents' ability to generate physics solvers. The core method involves LLM …
Towards In-Parameter Memory Augmentation for Large Language Models
This paper surveys methods for augmenting Large Language Models (LLMs) with "in-parameter memory." The core method involves integrating reusable memory information directly into mo…
Adaptive Power Sampling for LLM Reasoning
This paper introduces Adaptive Power Sampling (APS), a method to improve LLM reasoning by adjusting the "sharpening" of the model's output distribution on a per-query basis. Unlike…
AdvSim2Real : Training Web Agents Against Adaptive Prompt Injection in a Web World Model
This paper introduces AdvSim2Real, a method for training web agents to resist prompt injection attacks. It co-evolves the agent, a task curriculum, and an injection adversary withi…
The Town Square
Mistral AI has launched Mistral Large 4, a powerful new language model that excels in reasoning, coding, and multilingual capabilities, setting a new benchmark for AI performance.
Workshops
This repository provides a curated collection of "skills" or reusable code snippets for real engineers, directly sourced from the author's personal `.agents` directory.
This repository enables agents to generate CAD models from text descriptions, effectively giving them "CAD superpowers" by translating natural language into 3D designs.
