WARP: Weight-Space Analysis for Recovering Training Data Portfolios

2026-07-02 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

WARP, a novel framework, addresses the challenge of undisclosed training data recipes for publicly released foundation models by recovering a fine-tuned model's training mixtures directly from its released weights. Current methods like membership inference only detect individual samples, failing to characterize global training corpus composition. WARP interpolates between base and fine-tuned models via model merging, creating pseudo-checkpoints that simulate the training trajectory and reveal a geometric footprint of the training data in the weight space. From these simulated footprints, WARP extracts geometric features and maps them to domain proportions using either a parameter-free softmax readout or an MLP projector. In controlled experiments, WARP recovered domain mixtures for BERT and GPT-2 with average MAE as low as 0.046 and 0.104 respectively, outperforming membership inference and a variant with true trajectory access.

Key takeaway

For AI Scientists and Machine Learning Engineers focused on model transparency and auditing, WARP provides a critical new capability. If you need to understand the domain mixture weights used to fine-tune publicly released foundation models, this framework offers a robust method to infer those "data recipes" directly from model weights. You can use WARP to characterize global training corpus composition, which is crucial for evaluating model fairness, bias, and compliance, especially when original data disclosures are absent.

Key insights

WARP recovers training data domain mixtures from fine-tuned model weights by analyzing geometric footprints.

Principles

• Training data leaves a geometric footprint in weight space.
• Model merging can approximate training trajectories.
• Weight-space analysis can infer global data properties.

Method

WARP interpolates between base and fine-tuned model weights via merging, generates pseudo-checkpoints, extracts geometric features from these, and maps them to domain proportions using softmax or an MLP projector.

In practice

• Infer undisclosed training data domain mixtures.
• Characterize global composition of training corpora.
• Evaluate data recipe impact on fine-tuned models.

Topics

• Training Data Recovery
• Foundation Models
• Weight-Space Analysis
• Model Merging
• Data Provenance
• Domain Mixture Inference

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Security Engineer

Related on AIssential

• The Geometry Hidden in Every Model — Learning in supervised models is the discovery of a geometric kernel that defines similarity and weights training labels.
• Training and Finetuning Multimodal Embedding & Reranker Models with Sentence Transformers — Finetuning multimodal embedding models on domain-specific data significantly boosts performance over general-purpose models.
• Three-in-One World Model: Energy-Based Consistency, Prediction, and Counterfactual Inference for Marketing Intervention — A DBM-based world model with task-specific adapters unifies prediction, consistency, and counterfactual inference for marketing interventions.
• Position: Weight Space Should Be a First-Class Generative AI Modality — Treating neural network checkpoints as a first-class data modality enables generative modeling in weight space.
• Explaining Data Mixing Scaling Laws — A unified framework explains data mixing scaling laws through capacity competition and noise reduction.
• Supervised Classification Heads as Semantic Prototypes: Unlocking Vision-Language Alignment via Weight Recycling — Repurposing pretrained vision model classification heads as semantic prototypes enables efficient VLM alignment and data augmentation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.