What if an LLM's answer could be traced back to the training examples that most shaped it?This video summarizes "Data Attribution in Large Language Models via Bidirectional Gradient Optimization" by Frédéric Berdoz, Luca A. Lanzendörfer, Kaan Bayraktar, and Roger Wattenhofer.The paper introduces DABGO: Data Attribution via Bidirectional Gradient Optimization. Instead of only asking how training data affects a model output, DABGO flips the question: how would the training data be affected if the generated output were optimized back into the model?Main points covered:- Why training data attribution matters for AI governance, accountability, debugging, and provenance- The difference between forward influence and backward influence- How DABGO uses both gradient descent and gradient ascent on generated text- Why bidirectional loss changes can reveal influential training samples- How the method handles open-ended text generation instead of only single-token fact tracing- What the Wikipedia and Gutenberg experiments show about factual and stylistic attribution- How DABGO compares with BM25, TrackStar, and GeckoSource:Frédéric Berdoz, Luca A. Lanzendörfer, Kaan Bayraktar, and Roger Wattenhofer, "Data Attribution in Large Language Models via Bidirectional Gradient Optimization," arXiv:2606.04928, June 3, 2026.https://arxiv.org/abs/2606.04928Code:https://github.com/ETH-DISCO/DABGOThis content is provided for research and educational purposes only.#LLM #AIInterpretability #TrainingData #DataAttribution #DABGO #AIGovernance #MachineLearning #LLMs