Information-Preserving Reformulation of Reasoning Traces for Antidistillation
This addresses the problem for proprietary model providers by offering a way to protect against distillation while preserving user-valuable information, though it is incremental as it builds on existing antidistillation strategies.
The paper tackles the trade-off between revealing reasoning traces for user understanding and protecting proprietary LLMs from unauthorized distillation by proposing PART, an information-preserving reformulation method that disrupts distillation, reducing a 32B student model's performance on AIME 2024 from 54.17 to 46.88 (13.5% degradation).
Recent advances in Large Language Models (LLMs) show that extending the length of reasoning chains significantly improves performance on complex tasks. While revealing these reasoning traces helps users better follow, verify, and learn from the model's problem-solving process, it also makes them highly vulnerable to unauthorized distillation. To mitigate this risk, proprietary model providers often adopt aggressive protection strategies, such as replacing detailed reasoning with brief summaries, which deprive users of valuable intermediate information. To address this trade-off, we propose PART, an information-preserving antidistillation reformulation of reasoning traces. Motivated by the difference between how humans understand reasoning traces and how LLMs exploit them for supervised fine-tuning, we design a simple but effective two-step reformulation: removing self-talk behaviors and reordering sub-conclusions. A small auxiliary model is trained to perform this reformulation, incurring minimal computational overhead. Extensive experiments demonstrate that PART consistently disrupts distillation across student models of different sizes and types on various reasoning benchmarks. For instance, when training on reformulated traces, even the performance of a large 32B student model decreases from 54.17 to 46.88 on AIME 2024, corresponding to a 13.5% degradation.