Research Papers 论文研究 5h ago Updated 17m ago 更新于 17分钟前 49

Diff Mining: Logit Differences Reveal Finetuning Objectives 差异挖掘:Logit差异揭示微调目标

Diff Mining is a framework that identifies finetuning objectives by comparing output logits of finetuned and base models, requiring only black-box logit access rather than internal model weights The method uses a two-stage pipeline: extracting per-context logit differences on a reference corpus, then aggregating signals via Top-K frequency or Non-negative Matrix Factorization (NMF) to produce interpretable token fingerprints Diff Mining outperforms state-of-the-art model diffing methods in finet 提出Diff Mining框架,通过比较微调模型与基础模型的logits差异来识别微调目标 仅需访问输出logits即可工作,无需模型内部信息,可扩展至大模型 采用两阶段模块化设计:提取per-context logit差异 + 聚合构建可解释token集合 支持Top-K频率方法和NMF分解两种聚合策略,可分离多个微调目标 在微调域检测和注入偏见识别任务上显著优于现有模型diffing方法

65
Hot 热度
75
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Diff Mining is a framework that identifies finetuning objectives by comparing output logits of finetuned and base models, requiring only black-box logit access rather than internal model weights
  • The method uses a two-stage pipeline: extracting per-context logit differences on a reference corpus, then aggregating signals via Top-K frequency or Non-negative Matrix Factorization (NMF) to produce interpretable token fingerprints
  • Diff Mining outperforms state-of-the-art model diffing methods in finetune domain detection, both in token identification accuracy and downstream interpretability agent performance
  • The framework successfully identifies over one-third of injected biases in models without any targeted probing, demonstrating utility for auditing unwanted finetuning behaviors
  • The approach generalizes across domains, surfacing salient tokens even on text unrelated to the finetuning domain, making it a scalable tool for large model analysis

Why It Matters

As finetuning becomes the dominant paradigm for adapting language models, the opacity of what behaviors emerge during this process poses significant risks for safety, alignment, and accountability. Diff Mining offers a practical, scalable auditing mechanism that requires only logit access—making it applicable to proprietary and large-scale models where internal parameters are inaccessible. This has direct implications for researchers and practitioners seeking to understand, verify, and regulate finetuned model behavior.

Technical Details

  • Two-stage modular architecture: Stage one computes per-context logit differences between the finetuned and base model across a reference corpus; stage two aggregates these differences into an interpretable token set representing the finetuning objective
  • Two aggregation strategies: (i) Top-K frequency method, which selects the most frequently amplified tokens across contexts; (ii) NMF-based approach, which disentangles multiple finetuning objectives into distinct token clusters by factorizing the logit difference matrix
  • Black-box compatibility: Unlike existing model diffing methods that require access to internal activations or weights, Diff Mining operates solely on output logits, enabling application to large and proprietary models
  • Empirical validation: Evaluated on finetune domain detection (outperforming SOTA methods in both token identification and downstream interpretability agent tasks) and injected bias detection (recovering over one-third of biases without targeted probing)

Industry Insight

  • The framework enables third-party auditing of finetuned models without requiring model internals, which could become a standard tool for regulatory compliance and safety evaluation as finetuned models are deployed in high-stakes domains
  • The NMF-based disentanglement of multiple finetuning objectives into token clusters opens the door to decomposing complex, multi-objective finetuning pipelines into interpretable components—a capability increasingly valuable as models are trained with layered reward signals and preference data
  • The ability to detect unwanted behaviors from logit differences alone suggests that lightweight monitoring tools could be integrated into deployment pipelines to flag anomalous finetuning effects before models reach production

TL;DR

  • 提出Diff Mining框架,通过比较微调模型与基础模型的logits差异来识别微调目标
  • 仅需访问输出logits即可工作,无需模型内部信息,可扩展至大模型
  • 采用两阶段模块化设计:提取per-context logit差异 + 聚合构建可解释token集合
  • 支持Top-K频率方法和NMF分解两种聚合策略,可分离多个微调目标
  • 在微调域检测和注入偏见识别任务上显著优于现有模型diffing方法

为什么值得看

本文提出的Diff Mining框架为理解微调行为提供了可解释且可扩展的审计工具,对AI安全研究和模型透明度具有重要价值。该方法无需访问模型内部参数即可揭示微调引入的行为变化,为检测 unwanted behaviors 提供了实用方案。

技术解析

  • 核心思路:通过计算微调模型与基础模型在参考语料上的logit差异,识别被放大的显著token,作为微调训练的"指纹"
  • 两阶段架构:(i) 提取per-context logit差异;(ii) 聚合信号构建可解释token集合,支持Top-K频率方法和NMF分解
  • NMF分解能力:利用非负矩阵分解将多个微调目标 disentangle 为不同的token聚类,实现多目标分离
  • 实验验证:在微调域检测任务上显著优于SOTA方法;在注入偏见模型中无需针对性探测即可识别超过三分之一的偏见

行业启示

  • 为模型审计和可解释性研究提供了低门槛工具,仅需logits输出即可进行微调目标分析
  • 对AI安全领域具有重要价值,可帮助检测和识别微调过程中引入的 unwanted behaviors 和偏见
  • 推动了模型diffing技术的发展方向,从需要内部信息转向仅依赖输出接口,提升了方法的通用性和可扩展性

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Fine-tuning 微调 LLM 大模型 Research 科学研究 Training 训练 Alignment 对齐