Sketching the Readout of Large Language Models for Scalable Data Attribution and Valuation

2026-04-17

AI Summary

This work addresses the severe scalability bottleneck of existing gradient-based data attribution methods when applied to large language models. The authors propose RISE, a novel approach that draws inspiration from human cognitive focus mechanisms by decomposing output-layer gradients into two interpretable channels: a lexical residual channel (RH) and a semantic projection error channel (GH). Coupled with CountSketch for efficient compression, RISE enables memory-feasible, high-accuracy data attribution and zero-shot data valuation for the first time on billion-parameter models. Evaluated on the OLMo and Pythia model families, RISE reduces memory overhead by up to 112× compared to RapidIn while demonstrating practical utility in backdoor detection, domain separation, and high-quality data filtering—consistently enhancing downstream performance in closed-loop pretraining settings.

Technology Category

• Machine Learning: Privacy
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Responsible Web: Data and user privacy-enhancing technologies for the Web

Abstract

Data attribution and valuation are critical for understanding data-model synergy for Large Language Models (LLMs), yet existing gradient-based methods suffer from scalability challenges on LLMs. Inspired by human cognition, where decision making relies on a focused readout of relevant memories rather than replaying all pathways, we introduce RISE (Readout Influence Sketching Estimator). Instead of computing and indexing gradients across the entire LLM, RISE focuses on influence hotspots at the output layer, where influence signals concentrate, and the gradient admits a decomposed outer-product form. This enables a dual-channel representation combining a lexical residual channel (RH) and a semantic projected-error channel (GH). Applying CountSketch projections to these channels achieves strong compression while maintaining accurate attribution. Across the OLMo (1B-32B) and Pythia (14M-6.9B) families, RISE reduces index storage by up to 112$\times$ compared to RapidIn and scales to 32B parameters LLM, where gradient-based baselines such as RapidIn and ZO-Inf become memory-infeasible. We evaluate RISE on two paradigms: (1) retrospective attribution, retrieving influential training examples for specific predictions, and (2) prospective valuation, scoring candidate data utility zero-shot. We validate RISE on three tasks: Howdy backdoor data detection, Finance-Medical domain separation, and Brain Rot high-quality data selection. In a closed-loop Brain Rot study, continued pretraining on RISE-selected data yields consistent downstream improvements. Overall, RISE provides a practical and scalable primitive for influence analysis and training-data selection in modern large language models.

Problem

• data attribution
• data valuation
• scalability
• large language models
• influence analysis

Innovation

• Data Attribution
• Scalable Influence Estimation
• CountSketch
• Large Language Models
• Gradient Compression