Browse published editionsFilter by topic · All topics
2 consequential AI developments, with original sources and clearly labelled evidence coverage. NLI robustness appears tied to dataset shortcuts, not memorisation; Designing Recommendations for Human–AI Complementarity.
An abstract comparing eight LLMs with fine-tuned encoders reports that robustness to removing words depends on benchmark style and cues.
Why it matters: For people evaluating agents or language models on NLI, robustness to a perturbation may say more about benchmark cues than general model capability, according to the authors. Their results also suggest that which words are removed—and whether a system reasons before answering—can change the observed pattern.
An arXiv abstract describes a framework for designing recommendations around how they affect human decisions, not just how well they predict outcomes.
Why it matters: For people building recommendation systems, the framework emphasizes modeling how advice changes human choices, rather than optimizing prediction alone. For researchers, the authors offer a causal-inference-based approach and a stated route to recommendation design with limited training data.