Perplexity enhances AI accuracy by learning from real mistakes using hint-guided techniques

Perplexity Research has introduced a novel method to improve AI accuracy by training its models on real user sessions, including those that failed. The approach combines rejection sampling fine-tuning with hint-guided self-distillation, resulting in a 21.2% reduction in tool-call failures, from 2.24% to 1.77%, in live A/B tests. The model, running within Perplexity Computer, uses GLM 5.2 as its base and is available on Hugging Face. The technique separates sessions into successful and unsuccessful ones, using hints to correct errors. Hints are short instructions based on the model's existing knowledge, such as suggesting valid values for a search query. The On-Policy Self-Distillation (OPSD) method involves running the same checkpoint twice, with the teacher pass using hints and the student pass not, to refine the model's learning. The combined loss function includes cross-entropy and a KL divergence term, with λ controlling the balance between the two. Sessions are filtered for training eligibility, excluding those with personally identifiable information. LLM judges rate tasks on a 5-point scale, and two must approve for a session to be considered successful. User feedback involves three LLM judges identifying the root cause, with at least two agreeing. This process helps reduce hindsight bias by checking hints against available information before the mistake occurred.