mlsanityData to 14 Oct 2026

PULSE: Identifying Demonstration-Utility Features with Sparse Autoencoders

arXiv 2609.32469 ↗First posted 26 Sep 2026cs.AI cs.CL cs.LGcited by 0

On the map: Mechanistic interpretability → Sparse autoencoder training and evaluation

Abstract (arXiv)

In-context learning is highly sensitive to demonstration choice, yet most methods select demonstrations using external query-demonstration similarity. Such criteria can miss model-specific signals: Similar demonstrations may activate different internal features and downstream behaviors. We introduce PULSE (Paired Utility Localization over Sparse Encodings), an SAE-based framework for identifying model-internal features associated with demonstration utility and using them for demonstration selection. Using a small labeled discovery set, PULSE samples candidate demonstration sets, measures their zero-shot-relative utility under the target model, and scores SAE features by how their activation differences align with utility differences. The top positive and negative coordinates form a sparse utility-localization vector. We use this vector in two complementary ways: as a signed score for controlled complete-set ranking, and as PULSE-Retriever, which converts its magnitude into a feature-relevance mask for scalable pool-scale retrieval. Across classification, generation, and reasoning benchmarks, PULSE-Retriever improves over the strongest baseline by 2-3 accuracy points, 0.6-0.9 BLEU-4, and 3.2 exact-match points, respectively, while controlled ranking validates the identified features encode a predictive set-level utility signal. Feature inspection and cross-dataset experiments suggest that the identified features capture task-relevant, dataset-conditioned patterns, yet retain utility signals that partially transfer across datasets. Our code is available at https://github.com/aohenuo/PULSE.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Sparse autoencoder training and evaluation, one of 13 directions in Mechanistic interpretability. Builds on 9 of its 42 references that are in the map. First posted on arXiv 26 Sep 2026.

Builds on

9 of 42 references.

Closest ancestors · 1

Same direction, newest first

  1. 2022Toy Models of Superposition21 Sep

Other references · 8

  1. 2024The Llama 3 Herd of Modelscited by 14,715
  2. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  3. 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,998
  4. 2022In-context Learning and Induction Headscited by 964
  5. 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoderscited by 300
  6. 2024Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoderscited by 143
  7. 2023Dr.ICL: Demonstration-Retrieved In-context Learningcited by 72
  8. 2025MVI-Bench: A Comprehensive Benchmark for Evaluating Robustness to Misleading Visual Inputs in LVLMscited by 15

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Meta-Sel: Efficient Demonstration Selection for In-Context Learning via Supervised Meta-Learning
  2. 2025Linear-Time Demonstration Selection for In-Context Learning via Gradient Estimation
  3. 2024Demonstration Selection for In-Context Learning via Reinforcement Learning
  4. 2025Efficient and Effective In-context Demonstration Selection with Coreset
  5. 2025Feature-Based vs. GAN-Based Learning from Demonstrations: When and Why
  6. 2025Learning from Preferences and Mixed Demonstrations in General Settings
  7. 2026SparTa: Sparse Graphical Task Models from a Handful of Demonstrations
  8. 2024Demonstration Notebook: Finding the Most Suited In-Context Learning Example from Interactions