Interlat: Enabling Agents to Communicate Entirely in Latent Space
How continuous hidden states, a learned communication adapter, and latent-space compression create a new interface for multi-agent collaboration.
Zhuoyun Du 杜卓耘
Posts
Research notes on agents, reasoning, learning, and the systems around them.
Faithful LLM User Simulation: Closing the Gap Between Synthetic Users and Real People
Why cooperative synthetic users create an easy mode for agent evaluation, and how persona, knowledge, memory, multi-turn learning, and human calibration can close the gap.
Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents
How self-modifying code, benchmark feedback, and a branching archive turn recursive improvement into an executable search process.
Latent Space: From Continuous Reasoning to Representation Analysis
A guided synthesis of how language models can move computation beyond explicit tokens—and how we can train, inspect, and evaluate what happens there.
Training Large Language Models to Reason in a Continuous Latent Space
A guided reading of Coconut: how continuous thoughts replace discrete reasoning tokens, how the staged curriculum works, and why latent states may preserve multiple reasoning paths.
Agents from Google: A Systems View of Models, Tools, and Orchestration
A systems-oriented guide to what turns a language model into an agent and how tool interfaces shape control, data flow, and reliability.
Long-Horizon Agents Need Long-Lived State, Not Just Long Context
Why persistent workspaces, explicit artifacts, provenance, and local revision matter more than simply fitting a longer trajectory into context.
Optimizing Outcomes, Not Replies: Multi-Turn Post-Training for LLM Agents
How Q-functions and trajectory-level preferences turn delayed conversational outcomes into trainable signals while respecting tool and user feedback.