Research map / Architectures and efficiency / RNNs and normalization

PULSE: Identifying Demonstration-Utility Features with Sparse Autoencoders

NeurIPS 2026 · Architectures and efficiency › RNNs and normalization · approach: Layer-wise analysis of language models · 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 PULSE, which uses sparse autoencoder features and paired demonstration-set utility differences to build a sparse utility-localization vector, then uses it for ranking and pool-scale demonstration retrieval in ICL.

Problem

Demonstration selection for in-context learning relies on external query-demonstration similarity and ignores model-internal features that determine which demonstrations are useful for a given model.

Approach

Score SAE features by how well their activation differences between paired demonstration sets align with zero-shot-relative utility differences, yielding a signed sparse vector; use its signed form for complete-set ranking and its magnitude as a feature-relevance mask for retrieval.

Key results

• PULSE-Retriever improves average classification accuracy over KNN-SAE by 2.93 points on Gemma2-2B and 1.99 points on Llama3.1-8B.
• PULSE-Retriever achieves best result in all 12 generation/reasoning setting-shot cells, gaining +3.16 exact match on GSM8K with Llama3.1-8B.
• PULSE-Ranking achieves best or tied-best accuracy in 30 of 32 dataset-shot-backbone cells.

Surprising findings

• Negative-weight features on LAP14 fire on positive phrases, suggesting generic or off-aspect phrasing is associated with lower utility rather than negative sentiment.
• Raw hidden-state and PCA bases with the same identification objective underperform SAE features, showing the sparse basis matters.

Limitations

• Feature-level interpretations are post-hoc and do not establish causal mediation of ICL performance.
• PULSE-Retriever uses a magnitude mask as a proxy rather than an exact decomposition of complete-context utility, and gains over KNN-SAE are smaller than in set-level ranking.
• Requires models with publicly available SAEs (Gemma2-2B, Llama3.1-8B), limiting applicability.

Experimental setup

Datasets: AGNews REST14 LAP14 EMOC CommonGen GSM8K

Benchmarks: AGNews REST14 LAP14 EMOC CommonGen GSM8K

Metrics: accuracy BLEU-4 exact match

Baselines: Random Lex-Sim SBERT KNN-SAE CEIL EPR

Builds on: sparse autoencoders Gemma Scope Llama Scope CEIL EPR

Closest papers by meaning

• Demonstration Selection for In-Context Learning via Reinforcement Learning (ICML 2025 · similarity 0.86)
• Comparative Analysis of Demonstration Selection Algorithms for In-Context Learning in Large Language Models (Student Abstract) (AAAI 2025 · similarity 0.85)
• Sample Efficient Demonstration Selection for In-Context Learning (ICML 2025 · similarity 0.83)
• Large Language Models are Demonstration Pre-Selectors for Themselves (ICML 2025 · similarity 0.83)
• Retrieving Counterfactuals Improves Visual In-Context Learning (CVPR 2026 · similarity 0.83)
• MEND: Meta Demonstration Distillation for Efficient and Effective In-Context Learning (ICLR 2024 · similarity 0.83)