Evaluate interpretability methods on localizing and disentangling concepts in LLMs.
Causal Intervention on Modality-specific Biases for Medical Visual Question Answering
[SIGIR 2022] Source code and datasets for "Bias Mitigation for Evidence-aware Fake News Detection by Causal Intervention".
Demystifying Verbatim Memorization in Large Language Models
[EMNLP 2023] A Causal View of Entity Bias in (Large) Language Models
A framework for evaluating auto-interp pipelines, i.e., natural language explanations of neurons.
A causal intervention framework to learn robust and interpretable character representations inside subword-based language models
Located the number "five" inside Qwen2.5-32B and DeepSeek-V4-Flash (304B) and replaced it with "four" at 128 neurons/layer. The model still reads 5 but computes with it as 4, and in every test we ran it never noticed. Full records, reproducible on low-end hardware.
Capture macOS dictation accurately without rewriting your words, keeping your input true to what you speak and avoiding common app issues.
Mechanistic Circuit Comparison of Modular Arithmetic Tasks in Transformers
Continuous thoughts of a small model trained from scratch on a graph reachability task encode the nodes the search has reached so far, not the path to them, so branch arithmetic has nothing to act on, and the answer is settled one step before the end
To associate your repository with the causal-intervention topic, visit your repo's landing page and select "manage topics."