arXiv：cs.AI· Anand Swaroop, Maheep Chaudhary, Kevin Zhu, Sunishchal Dev, Vasu Sharma·· 22 小时前
LIFT：基于似然的干预忠实度训练，提升 LLM 的 CoT 推理忠实性
LIFT: Likelihood-Based Intervention Faithfulness Training
AI 导读
LIFT 通过用领域匹配但不相关的事实替换推理步骤、测量答案对数概率下降来自动为每个推理步骤打分，并在干预成功破坏推理的高分 CoT 轨迹上微调，同时优化忠实性与准确率。该方法在 Mistral-7B-Instruct、Qwen3-8B 和 Qwen3-4B-Instruct 三个模型及三个基准上提升了忠实性，准确率与仅针对准确率微调的模型相当。
正文
Abstract:Chain-of-thought (CoT) reasoning improves LLM accuracy on complex tasks, but CoTs are often unfaithful; that is, the model does not rely on the stated reasoning steps to produce its answer. Existing methods to improve faithfulness require human annotations. We propose LIFT, which automatically scores each reasoning step by its causal influence: we replace it with a domain-matched but irrelevant fact and measure the drop in answer log-probability. We fine-tune on high-scoring CoT traces where interventions successfully disrupted reasoning, jointly optimizing for faithfulness and accuracy. LIFT improves faithfulness across three models (Mistral-7B-Instruct, Qwen3-8B, and Qwen3-4B-Instruct) and three benchmarks, with accuracy comparable to a model fine-tuned solely for accuracy.
| Comments: | 13 pages, 3 figures |
| Subjects: | Artificial Intelligence (cs.AI); Machine Learning (cs.LG) |
| Cite as: | arXiv:2509.13334 [cs.AI] |
| (or arXiv:2509.13334v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2509.13334 arXiv-issued DOI via DataCite |
Submission history
From: Anand Swaroop [view email]
[v1]
Wed, 10 Sep 2025 07:07:17 UTC (22,530 KB)
[v2]
Mon, 12 Oct 2026 21:45:58 UTC (84 KB)
来源：arXiv：cs.AI · arxiv.org