aiqbench analysis
Diff Mining offers a practical method for auditing large language models by analyzing output logits to identify specific behaviors learned during fine-tuning without requiring access to internal model weights. This approach significantly enhances transparency and trustworthiness in AI systems, allowing developers to detect unwanted biases or domain-specific adaptations even on unrelated text. By disentangling multiple training objectives into distinct token clusters, the framework provides a scalable tool for interpreting model evolution and ensuring responsible deployment.
Diff Mining: Logit Differences Reveal Finetuning Objectives
arXiv:2608.26462v1 Announce Type: cross 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.
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.
37 pages, 7 figures. ICLR 2026 Workshop: Principled Design for Trustworthy AI. Code available at this https URL
Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
arXiv:2608.26462 [cs.LG]
(or arXiv:2608.26462v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2608.26462 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 382 words
Related coverage
The Continual Constitutional Convention: How language models revise the documents that govern them
AI Alignment Forum · Oct 13, 2026
My Assessment of Plan A's Compute Verification Strategy (+ open questions)
LessWrong · Jul 30, 2026
Building and evaluating model diffing agents
AI Alignment Forum · Jun 12, 2026
How Does Changing the Elo of a Chess Transformer Affect its Computations?
LessWrong · Sep 29, 2026