The Morning
From the arXiv
Accelerating Dense LLMs via L0-regularized Mixture-of-Experts
his paper introduces L0-MoE, a method to accelerate dense LLMs by using L0-regularization to create a lightweight Mixture-of-Experts (MoE) architecture. The core contribution is achieving significant speedups (up to 2.5x) with minimal performance degradation, outperforming existing acceleration techniques. This is enabled by domain-aware dataset curation and dynamic batching for efficient training.
CodeMidas: Scaling Agentic Coding RL Environments from Code Itself
CodeMidas is a novel agentic pipeline that automatically generates reinforcement learning (RL) coding environments directly from existing source code. It leverages agents to explore functionality, create tests based on code execution, and validate tasks, enabl…
Do Personality-Tuned LLMs Make Better Social Agents?
This paper investigates if fine-tuning LLMs with personality data improves their social agent capabilities. By training on personality-labeled social media and dialogues, they created personality-based dialogue engines and evaluated their consistency and contr…
ExpBoN: Exponential-Noise Best-of-$n$ for Efficient Test-Time LLM Alignment
ExpBoN is a novel method for efficient Large Language Model (LLM) alignment at inference time. It uses an exponential-noise mechanism to achieve smoother control over the reward-distribution shift trade-off, leading to exponentially faster convergence in key m…
An Interpretable Memory Decision Controller for LLM Agents Based on Three-Signal Complementarity: Decoupling Confidence and Consistency
This paper introduces the Memory Decision Layer (MDL), a novel, parameter-free module for LLM agents. The MDL addresses the issue of conflicting memories in retrieval-augmented generation by employing a three-signal complementarity approach (relevance, reliabi…
CASCADE Against Jailbreaks: Combination Across Stages with Controlled Attack-Defense Evaluation
This paper introduces CASCADE, a framework for systematically evaluating and combining defenses against LLM jailbreaks across different pipeline stages. It standardizes attack succ…
A Lie Detector Test for Language Models: Reading Knowledge a Model Won't Reveal
This paper introduces Probe of Internal Recognition (PIR), a method inspired by forensic concealed information tests to detect knowledge hidden within language models. PIR works by…
AutoViewMem: Self-Configuring Orthogonal Views for Conversational Long-Term Memory
AutoViewMem addresses the challenge of organizing diverse information in LLM long-term memory by creating self-configuring, low-overlap semantic views. This data-driven framework d…
Bayesian Belief Layer for Controllable Opinion Dynamics in LLM Agents
This paper introduces Bayesian Chronicle Agents (BCA) to address the implicit opinion dynamics in LLM agents. BCAs separate an agent's beliefs from its speech, representing stances…
CIPL: A Channel-Aware Framework for Recoverable Privacy Leakage in LLM Agents
This paper introduces CIPL, a novel framework for evaluating privacy leakage in LLM agents by considering the entire information flow from sensitive data to an attacker's observabl…
The Town Square
ChatGPT can now track your browsing activity on other websites through an ad collector, potentially impacting user privacy.
Workshops
This framework enables developers to build agentic applications by providing tools and components for creating intelligent agents that can interact with their environment and perform tasks.
This Cloudflare skill enables multi-phase security audits by providing an agent that generates independently verified, machine-readable findings.
