Diff Mining: Logit Differences Reveal Finetuning Objectives

• URL: http://arxiv.org/abs/2608.26462v1
• Date: Wed, 26 Aug 2026 23:33:18 GMT
• Title: Diff Mining: Logit Differences Reveal Finetuning Objectives
• Abstract summary: We introduce Diff Mining, a framework for identifying what a finetuned model has learned by comparing its logits to those of its base model.<n> Diff Mining surfaces salient tokens that are amplified in the finetuned model, serving as a fingerprint of its training.<n>Our framework shows promise in developing auditing tools to detect finetuning objectives.
• Score: 8.009258958150836
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Finetuning has become the gold standard for refining existing behaviors and inducing new ones in language models, yet it often remains unclear exactly which behaviors emerge during this process. As models grow ever more capable, understanding finetuning better becomes increasingly important, particularly since unwanted behaviors may arise during finetuning. In this paper, we introduce Diff Mining, a simple yet effective framework for identifying what a finetuned model has learned by comparing its logits to those of its base model. Diff Mining effectively surfaces salient tokens that are amplified in the finetuned model, serving as a fingerprint of its training -- even on text unrelated to the finetuning domain. Unlike many existing model diffing methods which require model internals, Diff Mining only needs access to output logits and scales to large models. The framework consists of two modular stages: (i) extracting per-context logit differences between the finetuned and base models on a reference corpus, and (ii) aggregating the resulting signals to construct an interpretable token set representing the finetune. For aggregation, we explore both a simple Top-K frequency method and a Non-negative Matrix Factorization (NMF)-based approach for disentangling multiple finetuning objectives into distinct token clusters. Empirically, Diff Mining succeeds across diverse settings: on finetune domain detection, it significantly outperforms state-of-the-art model diffing methods both in identifying relevant tokens and in downstream performance when an interpretability agent is given access to the extracted token set; on models with injected biases, it identifies more than one third of the biases without targeted probing. Overall, our framework shows promise in developing auditing tools to detect finetuning objectives.

Related papers

• Model Organisms Are Leaky: Perplexity Differencing Often Reveals Finetuning Objectives [0.9558392439655014]
  Finetuning can significantly modify the behavior of large language models.<n>We show that a simple perplexity-based method can surface finetuning objectives from model organisms.
  arXiv  Detail & Related papers  (2026-05-01T18:00:55Z)
• Nonparametric Data Attribution for Diffusion Models [57.820618036556084]
  Data attribution for generative models seeks to quantify the influence of individual training examples on model outputs.<n>We propose a nonparametric attribution method that operates entirely on data, measuring influence via patch-level similarity between generated and training images.
  arXiv  Detail & Related papers  (2025-10-16T03:37:16Z)
• Every Step Counts: Decoding Trajectories as Authorship Fingerprints of dLLMs [63.82840470917859]
  We show that the decoding mechanism of dLLMs can be used as a powerful tool for model attribution.<n>We propose a novel information extraction scheme called the Directed Decoding Map (DDM), which captures structural relationships between decoding steps and better reveals model-specific behaviors.
  arXiv  Detail & Related papers  (2025-10-02T06:25:10Z)
• Exact Byte-Level Probabilities from Tokenized Language Models for FIM-Tasks and Model Ensembles [23.134664392314264]
  Tokenization is associated with many poorly understood shortcomings in language models (LMs)<n>This work studies how tokenization impacts model performance by analyzing and comparing models with their byte-level counterparts.<n>We introduce the Byte-Token Representation Lemma, a framework that establishes a mapping between the learned token distribution and its equivalent byte-level distribution.
  arXiv  Detail & Related papers  (2024-10-11T23:30:42Z)
• TIES-Merging: Resolving Interference When Merging Models [95.59265307318752]
  Transfer learning can confer significant advantages, including improved downstream performance, faster convergence, and better sample efficiency. Model merging has emerged as a solution to combine multiple task-specific models into a single model without performing additional training. Existing merging methods often ignore the interference between parameters of different models, resulting in large performance drops when merging multiple models. We propose TIES-Merging, which introduces three novel steps when merging models: resetting parameters that only changed a small amount during fine-tuning, resolving sign conflicts, and merging only the parameters that are in alignment with the final agreed-upon sign.
  arXiv  Detail & Related papers  (2023-06-02T17:31:32Z)
• IMACS: Image Model Attribution Comparison Summaries [16.80986701058596]
  We introduce IMACS, a method that combines gradient-based model attributions with aggregation and visualization techniques. IMACS extracts salient input features from an evaluation dataset, clusters them based on similarity, then visualizes differences in model attributions for similar input features. We show how our technique can uncover behavioral differences caused by domain shift between two models trained on satellite images.
  arXiv  Detail & Related papers  (2022-01-26T21:35:14Z)
• Bayesian Attention Modules [65.52970388117923]
  We propose a scalable version of attention that is easy to implement and optimize. Our experiments show the proposed method brings consistent improvements over the corresponding baselines.
  arXiv  Detail & Related papers  (2020-10-20T20:30:55Z)
• Explaining and Improving Model Behavior with k Nearest Neighbor Representations [107.24850861390196]
  We propose using k nearest neighbor representations to identify training examples responsible for a model's predictions. We show that kNN representations are effective at uncovering learned spurious associations. Our results indicate that the kNN approach makes the finetuned model more robust to adversarial inputs.
  arXiv  Detail & Related papers  (2020-10-18T16:55:25Z)
• Document Ranking with a Pretrained Sequence-to-Sequence Model [56.44269917346376]
  We show how a sequence-to-sequence model can be trained to generate relevance labels as "target words" Our approach significantly outperforms an encoder-only model in a data-poor regime.
  arXiv  Detail & Related papers  (2020-03-14T22:29:50Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.