Papers with mechanistic interpretability
Simplifying Outcomes of Language Model Component Analyses with ELIA (2026.eacl-demo)
Copied to clipboard
| Challenge: | ELIA is an interactive web application that simplifies the outputs of various language model component analyses for a broader audience. |
| Approach: | They propose to use a vision-language model to automatically generate natural language explanations for the complex visualizations produced by these methods. |
| Outcome: | The proposed system integrates three key techniques and generates natural language explanations for complex visualizations. |
What Do VLMs NOTICE? A Mechanistic Interpretability Pipeline for Gaussian-Noise-free Text-Image Corruption and Evaluation (2025.naacl-long)
Copied to clipboard
| Challenge: | Vision-Language Models (VLMs) have gained prominence due to their success in solving complex cross-modal tasks. |
| Approach: | They propose a Gaussian-Noise-free pipeline for mechanistic interpretability in VLMs that introduces Semantic Image Pairs corruption, the first visual counterpart to Symmetric Token Replacement for text. |
| Outcome: | The proposed pipeline identifies a set of “universal attention heads” in BLIP and LLaVA that consistently contribute across different tasks and modalities. |
The Mystery of In-Context Learning: A Comprehensive Survey on Interpretation and Analysis (2024.emnlp-main)
Copied to clipboard
| Challenge: | In-context learning (ICL) is a capability that enables large language models to excel in proficiency through demonstration examples. |
| Approach: | They present a survey on the interpretation and analysis of in-context learning . they focus on theoretical and empirical perspectives on the concept . |
| Outcome: | The proposed model can perform tasks with minimal examples without re-training and has demonstrated proficiency across various tasks with a minimal set of task-oriented examples. |
Unlocking the Black Box of Latent Reasoning: An Interpretability-Guided Approach to Intervention (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for large language models (LLMs) lack a coherent representation of reasoning steps. |
| Approach: | They propose a set of latent reasoning interventions that enable latent thinking and decode-time interventions that refine the latent process by imposing the identified geometric and semantic priors. |
| Outcome: | The proposed models unlock latent capabilities and improve reasoning accuracy without any parameter updates. |
Constructing Interpretable Features from Compositional Neuron Groups (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for analyzing LLMs rely on dictionary learning with sparse autoencoders (SAEs) however, SAEs struggle in causal evaluations and lack intrinsic interpretability, as their learning is not explicitly tied to the computations of the model. |
| Approach: | They propose to decompose MLP activations with semi-nonnegative matrix factorization (SNMF) such that the learned features are mapped to their activating inputs, making them directly interpretable. |
| Outcome: | Experiments on Llama 3.1, Gemma 2 and GPT-2 show that SNMF derived features outperform SAEs and a strong supervised baseline on causal steering while aligning with human-interpretable concepts. |