TurnSight: Turn-Level Hindsight Self-Distillation for Tool-Integrated Reasoning
The paper 'TurnSight: Turn-Level Hindsight Self-Distillation for Tool-Integrated Reasoning' was published on arXiv on 2026-08-04. It proposes a framework that derives supervision from execution-conditioned hindsight, constructs multiple hindsight views with different lookahead horizons, selects reliable supervision via cross-horizon directional agreement, normalizes the signal across sibling rollouts, and uses it to adaptively modulate RL advantages. Extensive experiments on three benchmarks are reported.
TurnSight introduces a turn-level hindsight self-distillation method for tool-integrated reasoning (TIR) in LLMs. Unlike prior on-policy self-distillation that relies on ground-truth answers or retrieved skills, TurnSight uses execution-conditioned hindsight to provide denser, turn-level supervision. It generates multiple hindsight views with varying lookahead horizons, selects reliable signals through cross-horizon directional agreement, normalizes across sibling rollouts, and adaptively modulates reinforcement learning advantages while preserving the original optimization direction. The approach is evaluated on three benchmarks, demonstrating effectiveness.
The key technical contribution is the use of execution-conditioned hindsight to create multiple future views, then filtering them via cross-horizon directional agreement to obtain reliable turn-level supervision. This addresses the credit assignment problem in long-horizon tool interactions without relying on external ground truth. The adaptive modulation of RL advantages with normalized hindsight signals is a novel way to incorporate self-distillation into policy optimization.
Improving tool-integrated reasoning is critical for deploying LLMs in complex, multi-step enterprise and consumer applications. TurnSight's approach could lead to more robust and efficient agentic systems that learn from their own execution traces, reducing the need for expensive human annotations or curated skill libraries.
TurnSight has the potential to lower the cost and improve the performance of AI agents that interact with tools, APIs, and databases. This could accelerate adoption in enterprise automation, customer service, and developer tools by making agents more reliable in long-horizon tasks without manual reward engineering.
If the method generalizes beyond the three reported benchmarks, it could become a standard component in training tool-using LLMs. Next signals to watch include open-source implementations, integration into agent frameworks, and scaling to real-world tool APIs. Further research may explore combining TurnSight with other hindsight methods or applying it to multi-agent settings.