从泄露思维到私有推理:控制LRM对自己说的话
From Leaky Thoughts to Private Reasoning: Controlling What LRMs Say to Themselves
AI总结:
针对大型推理模型(LRM)推理过程中隐私泄露问题,提出通过指令跟随(IF)训练和分阶段解码策略(Staged Decoding)增强隐私保护,在IF和隐私基准上分别提升高达20.9和51.9个百分点。
摘要:
大型推理模型(LRM)产生的推理轨迹(RT)通常包含敏感信息。这些泄露的思维难以控制,且经常违反明确的隐私指令。由于RT可能通过提示注入攻击暴露,这对用户构成了直接的隐私风险。我们将此视为一个可控性问题:由于隐私指令本身就是指令,在RT内改进指令跟随(IF)为减少隐私泄露提供了直接途径。为此,我们引入了一个SFT数据集,教会模型在其推理过程中遵循通用指令,并提出了分阶段解码(Staged Decoding),一种简单的解码策略,通过使用独立的LoRA适配器解耦RT和答案生成,以最大化每个组件的IF。我们在两个系列(1.7B-14B参数)的六个模型上,在两个IF基准和两个隐私基准上评估了我们的方法。我们的方法带来了显著的改进,在IF上提升高达20.9分,在隐私基准上提升51.9个百分点,尽管由于推理性能与IF之间的权衡,这些改进可能以牺牲任务效用为代价。我们的结果表明,改进LRM中的IF可以显著增强隐私,为未来隐私感知的LRM提供了一个有前景的方向。我们的代码可在https://github.com/UKPLab/arxiv2026-controllable-reasoning-models获取。
英文摘要:
Large reasoning models (LRMs) produce reasoning traces (RTs) that often contain sensitive information. These leaky thoughts are difficult to control and frequently violate explicit privacy directives. Because RTs can be exposed through prompt injection attacks, this becomes a direct privacy risk to the user. We approach this as a controllability problem: since privacy directives are themselves instructions, improving instruction-following (IF) within the RT provides a direct path to reducing privacy leaks. To this end, we introduce an SFT dataset that teaches models to follow general instructions throughout their reasoning process, and propose Staged Decoding, a simple decoding strategy that decouples RT and answer generation using separate LoRA adapters to maximize IF of each component. We evaluate our approach on six models from two families (1.7B-14B parameters), across two IF benchmarks and two privacy benchmarks. Our method yields substantial improvements, with gains of up to 20.9 points in IF and 51.9 percentage points on privacy benchmarks, though these can come at the cost of task utility due to the trade-off between reasoning performance and IF. Our results show that improving IF in LRMs can significantly enhance privacy, suggesting a promising direction for future privacy-aware LRMs. Our code is available at https://github.com/UKPLab/arxiv2026-controllable-reasoning-models.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。