On Friday, Meta Platforms announced that various AI models from its research division would be made available, including a “Self-Taught Evaluator” that would pave the way for a reduction in the amount of human intervention in the AI development process.
According to the owner of Facebook, the “Self-Taught Evaluator” approach produces contrasted model outputs and uses an iterative self-improvement scheme to train a large language model, or LLM-as-a-Judge, to provide reasoning traces for evaluation and final verdicts.
A novel technique for creating artificial preference data to train reward models without depending on human annotations is the Self-Taught Evaluator.
According to a Reuters story, Meta introduced the product in a paper in August, demonstrating how it relies on the same “chain of thought” method as OpenAI’s o1 models to make accurate decisions regarding the responses of models. This release comes after that introduction.
The AI models o1 and o1-mini, developed by Microsoft-backed OpenAI (MSFT) last month, are capable of reasoning through more complicated tasks and solving more challenging math, science, and programming challenges than earlier versions.
The post also mentioned two Meta researchers who were involved in the effort, saying that the capacity to utilize AI to reliably review AI offers a glimpse at a potential pathway to constructing autonomous AI agents that can learn from their own mistakes.
According to the report, self-improving models have the potential to eliminate the need for an often costly and ineffective procedure known as Reinforcement Learning from Human Feedback. This process involves receiving input from human annotators who should possess specialized knowledge to correctly label data and confirm that responses to challenging math and writing problems are accurate.
“The idea of being self-taught and able to self-evaluate is basically crucial to the idea of getting to this sort of super-human level of AI,” said Jason Weston, one of the researchers, the report noted.
In addition, the social media behemoth unveiled an update to its Segment Anything Model 2 for photos and videos called Meta Segment Anything 2.1 (SAM 2.1). A new developer suite containing the web demo and model training code is included with SAM 2.1.
