TurnSight: Turn-Level Hindsight Self-Distillation for Tool-Integrated Reasoning

Tool-Integrated Reasoning (TIR) enables LLMs to solve complex tasks through iterative tool interactions. However, existing reinforcement learning methods often rely on trajectory-level supervision, limiting fine-grained credit assignment in long-horizon TIR scenarios. On-policy self-distillation offers denser signals through teacher branches with privileged context, but existing approaches typically derive such context from ground-truth answers or retrieved skills, which may not reflect the states actually visited by the agent. Moreover, token-level supervision fails to capture the turn-level structure of tool interactions. To address this, we propose TurnSight, a turn-level hindsight self-distillation framework that derives supervision directly from execution-conditioned hindsight. It then constructs multiple hindsight views with different lookahead horizons and selects reliable supervision through cross-horizon directional agreement. Finally, the selected hindsight signal is normalized across sibling rollouts and used to adaptively modulate RL advantages while preserving their original optimization direction. Extensive experiments on three benchmarks demonstrate the effectiveness of TurnSight. Our codes are available at https://github.com/quchangle1/TurnSight.

Extreme Gradient Boosted Multi-label Trees for Dynamic Classifier Chains

Classifier chains is a key technique in multi-label classification, since it allows to consider label dependencies effectively. However, the classifiers are aligned according to a static order of the labels. In the concept of dynamic classifier chains (DCC) the label ordering is chosen for each prediction dynamically depending on the respective instance at hand. We combine this concept with the boosting of extreme gradient boosted trees (XGBoost), an effective and scalable state-of-the-art technique, and incorporate DCC in a fast multi-label extension of XGBoost which we make publicly available. As only positive labels have to be predicted and these are usually only few, the training costs can be further substantially reduced. Moreover, as experiments on eleven datasets show, the length of the chain allows for a more control over the usage of previous predictions and hence over the measure one want to optimize.