Lilian Weng

Lilian Weng is an AI researcher and the author of Lil’Log technical essays on autonomous agents, reasoning, and self-improvement. Her contributions include collaborative vision work on OpenAI’s Rubik’s Cube robot and an agent framework organized around planning, memory, and tool use. In her later writing, she argues that improving AI also requires improving the surrounding execution system—and keeping evaluation, permissions, and human oversight at appropriate boundaries.[4][2][6]

In Harness Engineering for Self-Improvement, dated July 4, 2026, Weng describes a harness as the system surrounding a base model: it orchestrates execution, planning, tool calls, context, stored artifacts, and evaluation. She extends the earlier agent-component framework with workflow design, permission controls, and persistent state management. Her recommendations are concrete: keep durable state in files rather than carrying every log in context, and make parallel jobs inspectable through saved outputs, logs, and status records so execution can recover after interruptions.[6]

Weng predicts that practical near-term recursive self-improvement is more likely to begin with optimizing this machinery than with a model directly rewriting its own weights. The target becomes the mechanism for obtaining better answers—potentially including harness code and the optimizer itself. She allows that some improvements may eventually become part of core model behavior, while expecting interfaces to external context and tools to remain.[6]

She also argues that self-editing systems need a deliberately bounded editable surface, with permission and security layers outside the improvement loop. She recommends external evaluators, held-out tests, trace audits, and human review at consequential decision points, while leaving the extent of automated oversight an open question. Her concerns reach beyond benchmark success: manuscript production can conceal fabricated citations or weak experiments, and short-term coding rewards may miss maintainability, migration costs, and future debugging burden. As she puts it, “Humans should move up the stack, not be removed from the loop”.[6]

Weng became a cofounder of Thinking Machines with Mira Murati; Latent Space identified her in that role in its March 24, 2025 Agent Engineering article. On August 17, 2026, Axios’s Dan Primack reported that she had rejoined OpenAI.[1][8]

The following sections return to Weng's robotics research in 2019 and her explanatory writing in 2023–2025, spanning her move to Thinking Machines.[4][2][5][6]

The Solving Rubik’s Cube with a Robot Hand preprint, dated October 17, 2019, credits Maciek Chociej, Peter Welinder, and Weng with applying automatic domain randomization (ADR) to vision. ADR generates an expanding range of randomized simulated environments for training vision and control models, helping them transfer to the real robot. This was a specific collaborative contribution within the robotics project.[4]

Her June 23, 2023 essay LLM Powered Autonomous Agents places an LLM at the center of a system supported by three components. Planning combines task decomposition with reflection on past actions. Memory distinguishes the finite context used for in-context learning from longer-term information accessed through external retrieval. Tools supply capabilities absent from model weights, such as current information, code execution, and access to proprietary sources.[2]

Weng synthesizes existing research, crediting ReAct to Yao and colleagues and Reflexion to Shinn and Labash. She also identifies limits in the agents she surveyed in 2023: retrieval does not offer the same representation power as full attention, long-horizon plans can be difficult to revise after unexpected errors, and unreliable natural-language interfaces create substantial output-parsing work. Latent Space later used her framework in its discussion of agent definitions, presenting it as “Agent = LLM + memory + planning skills + tool use.”[2][1]

Weng also spoke on safety: Latent Space’s June 10, 2024 ICLR coverage identifies her as an OpenAI speaker for the talk “Towards Safe AGI.”[7]

In Why We Think, which displays a May 1, 2025 publication date, Weng thanks John Schulman for feedback and direct edits. She distinguishes parallel sampling—generating multiple candidates and selecting among them—from sequential revision, which iteratively edits an existing response. She cautions that self-correction without external feedback may fail to improve an answer or even change a correct answer into an incorrect one. Additional inference-time computation, she argues, cannot compensate for every large gap in a base model’s capabilities.[5]

Weng treats readable chain-of-thought as an imperfect window into a model’s reasoning. She rejects assuming that it is always faithful and urges caution about optimizing it directly during reinforcement learning: the research she reviews includes models learning to conceal reward hacking from a reasoning monitor. She also encourages researchers to share unsuccessful attempts, which can reveal useful limits that successful results alone leave hidden.[5]