Paper introduces Belief-Trajectory Energy for LLM prediction paths
Original title: Belief-Trajectory Energy: Measuring the Path to a Prediction
Scalar BTE signals task difficulty, and richer profiles reach 0.998 macro-AUROC for human-LLM review detection.
A paper defines Belief-Trajectory Energy (BTE), a measure of how an input revises an LLM's predictions across Transformer layers rather than only the final output. Intermediate states are mapped into a shared predictive space and summarized as a scalar or a depth profile. The authors report that local BTE matches predictive revision under Fisher-Rao geometry and that the revision sequence carries information beyond the initial-to-final change.
Unconfirmed: Reported detection and attribution scores are the authors' own results, not independently verified here.
- Authors report up to 0.998 macro-AUROC for human-LLM review detection.
- Eight-way generator attribution reached 95.6% accuracy in their tests.
- They say BTE develops during pretraining and changes with targeted training.
- Demonstrations are linked from the authors' project page.
From the original article
1 min read at huggingface.co
Large language models (LLMs) progressively revise their predictions across Transformer layers, yet we typically observe only the final output, discarding the trajectory through which it is formed. We introduce Belief-Trajectory Energy(BTE), a model-grounded measure that characterizes an input through the layerwise predictive revisions it induces in a model.
By Jiahao Ying, Wei Tang, Boxian Ai · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.