Research map / Large language models / Language, safety and interpretability

Scalable Influence and Fact Tracing for Large Language Model Pretraining

ICLR 2025 · Large language models › Language, safety and interpretability · approach: LLM pruning and scaling data · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Scales gradient-based training data attribution (TrackStar) to retrieve influential C4 pretraining examples for an 8B LLM over 160B tokens, combining optimizer second-moment correction, a task-specific mixed Hessian, and unit normalization, and compares attribution vs. influence on fact tracing.

Problem

Training data attribution has been hard to apply to the full scale of LLM pretraining (billions of parameters, hundreds of billions of tokens, thousands of queries), and it is unclear whether influential examples align with examples that entail a fact.

Approach

Refines gradient-based influence (TRAK/TracIn/EK-FAC-style) with per-parameter optimizer second-moment correction applied before random projection, a task-specific mixed Hessian approximation that downweights template-shared gradient components, unit normalization, and token-level loss gradients, enabling exact scoring over all of C4 without prefiltering.

Key results

• On T-REx closed set (8B model, 19.6M sentences), TrackStar reaches MRR 0.365 and tail-patch +0.90%, versus BM25 at MRR 0.592 and tail-patch +0.41%.
• On full C4 (365M examples), TrackStar achieves tail-patch +2.11% vs +0.83% for BM25 and +0.54% for Gecko, while BM25 has higher MRR (0.687 vs 0.338).
• Attribution (MRR) of TrackStar improves with model size and training tokens, so influence aligns more closely with factual attribution for larger models.

Measured effects reported in the paper

Surprising findings

• Classical BM25 retrieves entailing passages better than gradient methods but its proponents have much less influence on model predictions, showing a misalignment between attribution and influence.
• Applying the TRAK per-example Q correction instead of per-token drops MRR from about 0.122 to 0.001 at scale, so methods that work on small classification settings may fail at LLM scale.

Experimental setup

Datasets: C4 T-REx KILT COPA PIQA arithmetic word problems

Benchmarks: T-REx fact tracing C4 open set fact tracing COPA PIQA

Metrics: MRR Recall@10 tail-patch score

Baselines: BM25 Gecko TRAK Gradient cosine ablation Gradient dot product ablation

Builds on: TRAK TracIn EK-FAC LoGra Datamodels

Closest papers by meaning

• DATE-LM: Benchmarking Data Attribution Evaluation for Large Language Models (NeurIPS 2025 · similarity 0.89)
• Sketching the Readout of Large Language Models for Scalable Data Attribution and Valuation (NeurIPS 2026 · similarity 0.89)
• On Linear Representations and Pretraining Data Frequency in Language Models (ICLR 2025 · similarity 0.89)
• Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 2025 · similarity 0.88)
• MATES: Model-Aware Data Selection for Efficient Pretraining with Data Influence Models (NeurIPS 2024 · similarity 0.88)
• MASS: Mathematical Data Selection via Skill Graphs for Pretraining Large Language Models (ICML 2025 · similarity 0.88)