Reproduce “Attribution-Guided Decoding” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Piotr Komorowski, Elena Golimblevskaia, Reduan Achtibat, Thomas Wiegand, Sebastian Lapuschkin, Wojciech Samek
AGD guides next-token selection toward instructions or supplied evidence without changing model activations: it scores likely candidates by their attribution to a task-relevant region and selects the highest-scoring candidate. On Llama 3.1 (8B), LRP-guided AGD raises IHEval answers satisfying all constraints from 66.0% to 79.1%, but this structured benchmark result does not guarantee arbitrary instruction following.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.