Papers by Virginia Smith
Mechanistic Interpretability Should Prioritize Feature Consistency in Sparse Autoencoders (2026.acl-long)
Copied to clipboard
Xiangchen Song, Aashiq Muhamed, Yujia Zheng, Lingjing Kong, Zeyu Tang, Mona T. Diab, Virginia Smith, Kun Zhang
| Challenge: | Sparse Autoencoders (SAEs) are a tool in mechanistic interpretability (MI) but the aspiration to identify a canonical set of features is challenged by the observed inconsistency of learned SAE features across different training runs. |
| Approach: | They propose to use the Pairwise Dictionary Mean Correlation Coefficient to quantify SAE feature consistency as an evaluation axis alongside reconstruction and sparsity. |
| Outcome: | The proposed measure is based on the pairwise dictionary mean correlation coefficient (PW-MCC) on LLM activations. |
Decoding Dark Matter: Specialized Sparse Autoencoders for Interpreting Rare Concepts in Foundation Models (2025.findings-naacl)
Copied to clipboard
| Challenge: | Sparse Autoencoders (SAEs) are a promising tool for disentangling FM representations, but they struggle to capture rare, yet crucial concepts in the data. |
| Approach: | They propose a technique to train Sparse Autoencoders to illuminate elusive dark matter features by focusing on specific subdomains. |
| Outcome: | The proposed method achieves 12.5% better classification accuracy than general-purpose SAEs when applied to remove spurious gender information. |
GRASS: Compute Efficient Low-Memory LLM Training with Structured Sparse Gradients (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing projection-based methods that project gradients into a lower-dimensional subspace can introduce computational and memory overheads. |
| Approach: | They propose a novel approach that leverages sparse projections to transform gradients into structured sparser updates. |
| Outcome: | The proposed approach significantly reduces memory usage for optimizer states and minimizes memory footprint, computation, and communication costs, leading to substantial throughput improvements. |
CoRAG: Collaborative Retrieval-Augmented Generation (2025.naacl-short)
Copied to clipboard
| Challenge: | Existing research on Retrieval-Augmented Generation models has focused on centralized settings where a single entity controls both the model and the datastore. |
| Approach: | They propose a framework for RAG where clients jointly train a shared model using a collaborative passage store. |
| Outcome: | The proposed framework outperforms parametric learning methods and locally trained models in low-resource scenarios. |
RefusalBench: Generative Evaluation of Selective Refusal in Grounded Language Models (2026.eacl-long)
Copied to clipboard
| Challenge: | Language models fail to selectively refuse to answer based on flawed context, study finds . current benchmarks fail to evaluate complex capabilities like selective refusal . |
| Approach: | They propose a framework that generates diagnostic test cases through controlled linguistic perturbation. |
| Outcome: | The proposed framework employs 176 perturbation strategies across six categories of uncertainty and three intensity levels. |
Semantic Agreement Enables Efficient Open-Ended LLM Cascades (2025.emnlp-industry)
Copied to clipboard
| Challenge: | Large language models (LLMs) have enabled impressive progress across a range of language tasks, but they are steep computational cost. |
| Approach: | They propose a semantic consensus mechanism for reliable deferral by combining model outputs with semantic consensus. |
| Outcome: | Evaluated from 500M to 70B-parameter models, semantic cascades improve deferral accuracy, match or surpass target-model quality at 40% of the cost and reduce latency by up to 60%. |