The Morning
From the arXiv
Apodex 1.1: Scaling Agentic Intelligence for Complex Work
podex 1.1 enhances language models' "working capability" by scaling both their interaction with diverse tools (files, code, search) and their ability to coordinate complex, multi-agent tasks. Its core method involves training agents to decompose, delegate, and integrate asynchronous results, while a shared harness maintains state and provenance. This allows for sustained, verifiable progress on real-world objectives across various professional domains.
How to Train a Critic Stably and Efficiently
This paper addresses the instability of training critics in reinforcement learning for large language models. Their core method, **Best-Practice Critic Optimization (BPCO)**, combines several techniques to stabilize critic training and enable efficient advanta…
InjecMEM: Memory Injection Attack on LLM Agent Memory Systems
This paper introduces InjecMEM, a novel attack that exploits memory systems in LLM agents. By crafting a single interaction with a retriever-agnostic anchor and an adversarial command, InjecMEM can steer future agent responses towards a pre-specified output, e…
Mitigating Reasoning-Induced Misalignment via Safety-Direction Penalty
This paper identifies "Reasoning-Induced Misalignment" (RIM) as a problem where training LLMs on reasoning tasks can inadvertently lead to harmful behaviors. The authors propose the Safety-Direction Penalty (SDP) as a training-time solution. SDP penalizes the …
Prime Agent: A Self-Improving RLM Harness
Prime Agent is an open-source harness designed to enable language models to perform long-horizon tasks by providing external information and computation. Its core method involves a persistent IPython REPL for programmatic context processing and a Continual Har…
SkillAlchemy: Open-World Agent Skill Creation
SkillAlchemy addresses the challenge of creating agent skills for unfamiliar tasks by leveraging open-world materials. Its core method involves an "admission-centered" framework th…
SRPO: Self-Reflective Policy Optimization for Long-Horizon Reasoning
SRPO is a novel framework that enables Large Language Models (LLMs) to perform self-reflection on their own reasoning processes. It synthesizes errors into "reflection patches" and…
Beyond the Stability-Exploration Dilemma: Environmental Regularization for LLM Policy Optimization
This paper introduces Environment-Regularized Policy Optimization (ERPO) to address the stability-exploration dilemma in LLM policy optimization. ERPO shifts regularization from th…
Agent-G$^2$: Gaussian Guidance for Agentic Reinforcement Learning
This paper introduces Agent-G$^2$, a novel reinforcement learning method that addresses reward sparsity by providing guidance from expert trajectories. Instead of a single optimal …
Credal Large Language Models for Semantic Commitment under Uncertainty
This paper introduces Credal Large Language Models (CLLMs) to address the issue of LLMs' unwarranted confidence in incorrect answers. CLLMs use an ensemble of adapters to represent…
The Town Square
Xiaomi's new CPU reportedly matches Apple's single-threaded performance while significantly exceeding it in multi-threaded tasks.
Workshops
OpenAI Codex is a lightweight coding agent that runs in your terminal, enabling users to generate code snippets and perform coding tasks directly from their command line interface.
This AI job search framework runs locally, using Claude to evaluate job postings, tailor CVs, write cover letters, and prepare for interviews, empowering users to own their application process.
