contrastive-attribution

This paper introduces GAND, a benchmarking resource of gender-ambiguous natural English sentences for analyzing gender bias in machine translation, and presents an interpretability analysis using contrastive translations to reveal source words influencing gender assignment.

CANTANTE introduces a contrastive credit attribution method to optimize multi-agent LLM systems by decomposing global rewards into per-agent signals, enabling automated prompt tuning. It outperforms baselines on programming, math, and retrieval benchmarks, achieving up to +18.9 points improvement without increased inference cost.

Researchers apply contrastive LRP-based attribution to analyze why LLMs fail on realistic benchmarks, finding the method gives useful signals in some cases but is not universally reliable.