Entity · technique

Data Mixture Surgery

techniqueactivedata-mixture-surgery-b4bc51f9·1 events·first seen May 29, 2026

Aliases: Data Mixture Surgery

Co-occurring entities

LLMSurgeon LLMScan label-shift assumption

More like this (12)

Sparse Mixture-of-Experts Toward Calibrated Mixture-of-Experts Under Distribution Shift Mixture-of-Agents CausalMix counterfactual data augmentation mixture-density networks Mixture of Experts Ensemble Diversity Optimization Redesign Mixture-of-Experts Routers with Manifold Power Iteration Data Is Better Together Semantic Neighbor Mixing Mixture of Efficient Experts

Recent events (1)

6arXiv · cs.LG·May 29, 2026·source ↗

LLMSurgeon: Post-Hoc Auditing of LLM Pretraining Data Mixtures

LLMSurgeon formalizes Data Mixture Surgery (DMS), a framework for estimating the domain-level distribution of an LLM's pretraining corpus using only generated text from the target model. The method casts DMS as an inverse problem under the label-shift assumption, using a calibrated soft confusion matrix to correct domain confusion and recover the latent mixture prior. The authors also introduce LLMScan, a verifiable evaluation suite built from open-source LLMs with known pretraining mixtures, on which LLMSurgeon demonstrates high-fidelity recovery of domain compositions without access to training data.

Frontier Model Releases Evaluation and Benchmarking LLMSurgeon LLMScan Data Mixture Surgery +3 more