aiqbench analysis

This study advances agent reliability by leveraging real-world failure data through a novel combination of rejection sampling and hint-guided self-distillation, directly addressing tool-call errors without discarding valuable negative examples. The approach demonstrates that correcting specific erroneous turns while preserving successful context can yield statistically significant improvements in live deployment scenarios. While the post-trained weights remain proprietary to Perplexity Computer, the methodology offers a concrete pathway for improving autonomous agent robustness using accessible base models like GLM 5.2.