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

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

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: LLM pruning and scaling data · 0 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.

Proposes RISE, a forward-only, Hessian-free influence estimator that sketches LM-head gradient factors (residual and hidden state, with dual RH/GH channels) via CountSketch to scale data attribution and valuation to 32B LLMs.

Problem

Scalable data attribution and valuation for LLMs: per-example gradient-based influence estimators need O(P) storage and become memory-infeasible at multi-billion-parameter scale.

Approach

Restrict influence to the LM head, where the gradient is an exact outer product r_t ⊗ h_t computable from forward passes; use a dual lexical (RH) and semantic projected-error (GH) channel, sparse top-K+GT residual support, and CountSketch compression of the factors.

Key results

• On Pythia-6.9B, the unified three-task auPRC score is 0.912±0.012 for RISE vs 0.732±0.045 for TrackStar.
• On Brain Rot closed-loop training, RISE-selected data gives 2.33 control perplexity vs 6.63 for TrackStar (2.8x lower) and 4.10% RULER-CWE vs 2.54%.
• RISE runs on OLMo-3-32B where RapidIn and ZO-Inf run out of memory, with index as small as 6.2 MB for 5,000 samples.

Surprising findings

• Embedding similarity and BM25 selection purity (54.0%, 63.8%) remain moderate, and RISE does not beat baselines on ARC-Challenge (26.62 vs 27.30 for BM25 and 28.33 for Random).

Limitations

• Relies on the LM head as the influence hotspot, which the authors note may miss internal-layer hotspots.
• Theoretical bounds rest on stated assumptions (norm stability, non-exploding backprop) validated only empirically.
• Retrieval evaluation mostly uses synthetic backdoor and keyword-labeled tasks.

Experimental setup

Datasets: Howdy backdoor data Finance-Medical Brain Rot C4

Benchmarks: Howdy backdoor detection Finance-Medical separation Brain Rot selection RULER-CWE ARC-Challenge

Metrics: auPRC auROC Precision@10 perplexity disk memory GPU memory build time query latency

Baselines: BM25 Embedding Similarity RapidIn ZO-Inf TrackStar TracIn

Builds on: TracIn RapidIn TrackStar ZO-Inf TRAK

Closest papers by meaning

• Scalable Influence and Fact Tracing for Large Language Model Pretraining (ICLR 2025 · similarity 0.89)
• Boosting Multi-Domain Fine-Tuning of Large Language Models through Evolving Interactions between Samples (ICML 2025 · similarity 0.88)
• Systematic Outliers in Large Language Models (ICLR 2025 · similarity 0.87)
• From Sparse Dependence to Sparse Attention: Unveiling How Chain-of-Thought Enhances Transformer Sample Efficiency (ICLR 2025 · similarity 0.87)
• IDEAL: Influence-Driven Selective Annotations Empower In-Context Learners in Large Language Models (ICLR 2024 · similarity 0.87)
• Scaling up Masked Diffusion Models on Text (ICLR 2025 · similarity 0.87)