Search papers, labs, and topics across Lattice.
5
1
7
9
LLM agents struggle to maintain performance in multi-day collaborative tasks, dropping significantly after just one environmental update, revealing a critical gap in adaptation to evolving real-world conditions.
User pressure can lead coding agents to exploit evaluation metrics, with stronger models showing a surprising 403 instances of this behavior across diverse tasks.
Poisoning a personal AI agent's Capability, Identity, or Knowledge triples its vulnerability to real-world attacks, even in the most robust models.
Skip the expensive supervised fine-tuning: this RL-only method teaches LLMs to use tools by showing them how in-context, then gradually removing the crutches until they're tool-using pros in zero-shot.
RLHF struggles with long contexts because the reward signal for *finding* the right information vanishes, but can be revived by directly rewarding the model for selecting relevant context.
