AI Paper Digest · July 15, 2026
AI Paper Digest · July 15, 2026
Weekly digest of noteworthy AI research papers.
Inside the Unfair Judge: A Mechanistic Interpretability Account of LLM-as-Judge Bias
Authors: Zixiang Xu, Sixian Li, Huaxing Liu, Xiang Wang
arXiv: 2607.11871 (Jul 2026)
Existing studies of LLM-as-judge scoring bias work at input-output level only — perturb inputs, measure score deltas. The internal mechanisms producing bias remain unknown. This study goes inside the model using mechanistic interpretability to trace reasoning pathways during scoring tasks and localize bias-producing neuron layers. The authors identified neural representations of position bias and self-enhancement bias within LLM scoring mechanisms.
Evidence-Backed Video Question Answering
Authors: Shijie Wang, Honglu Zhou, Ziyang Wang, Ran Xu
arXiv: 2607.11862 (Jul 2026)
Current Video LLMs excel at QA but operate as black boxes, providing answers without evidence. This paper proposes an evidence-grounded framework that outputs specific supporting video segments and timestamps alongside answers. The evidence-backed approach not only improves interpretability but also boosts accuracy on multiple video QA benchmarks.
Requential Coding: Pushing the Limits of Model Compression with Self-Generated Training Data
Authors: Shikai Qiu, Marc Finzi, Yujia Zheng
arXiv: 2607.11883 (Jul 2026)
Compression is fundamentally about representing training data as a short code; traditional methods rely on external training data. The authors propose “Requential Coding” — a framework that uses self-generated training data to push compression beyond conventional limits. The method significantly outperforms existing model compression approaches across multiple benchmarks.
