Brief

Researchers have introduced Trajectory-Augmented Policy Optimization (TAPO), a novel method for self-distillation in large language models. Unlike traditional approaches that minimize KL divergence, TAPO constructs explicit training trajectories by retaining erroneous reasoning up to the point of failure, then incorporating natural-language diagnoses and corrected reasoning. This method aims to provide more fine-grained error correction and has demonstrated consistent improvements over GRPO in experiments on AIME 2024, AIME 2025, and HMMT 2025. AI

Researchers have introduced Trait-Aware Policy Optimization (TAPO), a novel post-training framework designed to enhance autoregressive models for multi-trait essay scoring. This method decomposes rewards across samples and traits, integrating global consistency, trait accuracy, and inter-trait dependencies. Experiments indicate that TAPO significantly improves scoring performance compared to standard supervised fine-tuning and scalar-reward optimization techniques. AI