activation patching
PulseAugur coverage of activation patching — every cluster mentioning activation patching across labs, papers, and developer communities, ranked by signal.
1 day(s) with sentiment data
-
Weight pruning can damage AI model circuits without affecting accuracy
A new study published on arXiv investigates the failure modes of weight pruning in Transformer language models, specifically using GPT-2 as a test subject. Researchers found that standard metrics like accuracy and perpl…
-
New research advances causal inference methods for AI models · 6 sources tracked
Researchers are developing new methods to improve causal inference in machine learning models, particularly when dealing with unmeasured confounding factors. One approach, Hidden-Pathway Contribution (HPC), aims to dist…
-
New method deciphers Transformer model internals by differencing prompts
Researchers have developed a new method called Contrastive Projection to better understand the internal workings of Transformer models. This technique involves subtracting the hidden states of two closely matched prompt…
-
New theory explains divergence between activation patching and weight-space ablation
Researchers have developed a theoretical framework to understand the relationship between activation patching and weight-space ablation, two methods used to determine causal responsibility in neural networks. The theory…
-
LLMs struggle to decode mythological knowledge beyond dominant traditions
A new research paper investigates how 18 open-source large language models represent and decode mythological knowledge. The study found that while models can represent knowledge about deities like Zeus, Jupiter, and Tho…
-
New paper reveals hidden interaction effects in AI model interpretability
A new research paper titled "The Curse of Multiple Mediators" explores the limitations of activation patching, a primary tool in mechanistic interpretability. The paper argues that activation patching, used to attribute…