CHATS-lab

Your agent is powerful but it doesn't know you. VibeLens visualizes agent sessions, personalizes your agents, provides dashboard analytics and productivity tips from your real interaction. Supports Claude Code, Codex, Gemini CLI, OpenClaw, and more.

[ICML 2026] Official implementation for paper "Unsafer in Many Turns: Benchmarking and Defending Multi-Turn Safety Risks in Tool-Using Agents"

Code base for paper: "The Piggyback Hypothesis of Generalization: explaining and mitigating emergent misalignment"

An Awesome Human-AI-Coevolution Paperlist Created by CHATS LAB

SHEPHERD: A Runtime Substrate Empowering Meta-Agents with a Formalized Execution Trace

Repository of safety monitor for Claude Code

An gamified, interactive, and role-play-based AI literacy tutorial that mitigates the effects of AI persuasion.

Verbalized Sampling, a training-free prompting strategy to mitigate mode collapse in LLMs by requesting responses with probabilities. Achieves 2-3x diversity improvement while maintaining quality. Model-agnostic framework with CLI/API for creative writing, synthetic data generation, and dialogue simulation.

Persuasive Jailbreaker: we can persuade LLMs to jailbreak them!