Reading the Finetuning Prior: Verbatim Content Recovery via Contrastive Decoding Diffing

2026-05-25

AI Summary

This work addresses the challenge of auditing the specific knowledge acquired by a fine-tuned language model when neither its weights nor training data are accessible. The authors propose Contrastive Decoding Diffing (CDD), a gray-box method that precisely recovers fine-tuned factual content—down to the exact token—by analyzing differences in logits between the base and fine-tuned models. Requiring no access to model weights, layer selection, or architectural modifications, CDD bypasses chat templates and employs ambiguous prefixes to amplify logit discrepancies at each decoding step, enabling the first purely output-based extraction of fine-tuned information. Evaluated across four architectures ranging from 1B to 32B parameters, CDD accurately reconstructs concrete facts such as drug names and vote counts, operates approximately 170× faster than the white-box ADL method, and correctly identifies all four source datasets within a mixed fine-tuning corpus.

Abstract

Narrowly finetuned language models memorize implanted content verbatim, but auditing what a deployed model has been taught, without access to its weights or training data, remains an open challenge. Recent work shows that activation differences between base and finetuned models carry readable traces of the finetuning domain; the state-of-the-art Activation Difference Lens (ADL) recovers a vague domain-level description but requires full "white-box" access to model internals. We introduce Contrastive Decoding Diffing (CDD), a model diffing method that operates on output-level logit distributions only, with no weight access, no layer selection, and no per-model tuning, yet recovers implanted facts. CDD consists of three ideas: bypassing the chat template to expose the raw finetuning prior, seeding generation with maximally vague pre-fills, and amplifying the logit-space difference between finetuned and base models at each decoding step. A single default configuration recovers implanted facts verbatim -- exact drug names, vote counts, physical measurements, and procedural details -- across four architectures (1B--32B parameters), uniformly outperforming ADL despite less access and running ~170x faster. Furthermore, CDD surfaces unintended data pipeline artifacts: a fictional persona introduced by the LLM data generator via mode collapse leaked into model weights and was extracted by CDD, constituting to our knowledge the first demonstrated end-to-end fingerprinting chain from data generator artifact to model weights to recovered output. We validate on real-domain finetuning settings, achieving near-perfect recovery across all single-dataset non-CoT variants and correctly identifying all four datasets in the mixed-dataset setting. CDD's success as a grey-box method outperforming white-box baselines underscores its practical utility for transparency and accountability in AI systems.

Problem

• finetuning auditing
• content recovery
• language model transparency
• model memorization
• black-box analysis

Innovation

• Contrastive Decoding Diffing
• model diffing
• verbatim recovery
• grey-box auditing
• finetuning prior