How Anthropic trained Fable 5 => by analysing its reasoning traces (ankitmaloo.com)

Anthropic has unveiled insights into the training process of Fable-5, its latest language model, revealing that it employs a unique strategy based on analyzing reasoning traces. Unlike conventional training methods that utilize a sequence of supervised fine-tuning (SFT), reinforcement learning (RL), and self-distillation, Fable-5's approach highlights a distinctive capability in code composition and problem-solving. This was demonstrated when Fable-5 struggled with a Summle puzzle, initially engaging in an extensive greedy search before successfully transitioning to a more systematic enumeration method, which ultimately led to the correct solution. This behavior is indicative of its reliance on chaining learned paths rather than adapting a search policy effectively. The significance of this finding for the AI/ML community lies in the model's potential applications in coding and cybersecurity tasks, where it excels at recognizing and chaining established sequences but falters in novel problem-solving scenarios requiring significant search capabilities. The analysis emphasizes that while the model can recall and apply learned solutions (drawing from a distilled training approach), it lacks a robust meta-policy for initiating adaptive search strategies, showcasing a fundamental asymmetry between retrieval-based problem-solving and exploratory search. This insight paves the way for further research into enhancing compositional search within AI models, underscoring the need for refined training methods that balance the ability to leverage learned sequences with adaptive problem-solving aptitude.