A retrieval benchmark built from contradicted evidence found BM25 separates the original from the edited copy, while dense retrievers stay near chance
AspireAI community · arXiv
*Coverage benchmarks ask whether the retrieved evidence covers everything an answer needs, but most of their claims are facts the retriever may already know, which hides the case that actually matters: corpus evidence that contradicts what the model believes.* CoverageBench-Negate takes a nugget-based coverage benchmark and, for every nugget, writes a negated counterpart and edits its supporting documents to support that negation while keeping the context intact - roughly 63,000 original-edited evidence pairs across six datasets, with construction metadata, out-of-family LLM judgements and a human audit of a sample, all released on Hugging Face. The baseline result is the surprising part: BM25 separates original from edited near-twin evidence substantially better than dense retrievers built on Qwen3 and Jina, which stay near chance, so the gap is in judging fine semantic differences rather than in retrieval itself. That reframes negation-aware retrieval as an evaluation problem before it is a modelling problem. The limitation is the construction itself: the negations are generated and edited by language models, so the pairs encode a particular idea of what a contradiction looks like, and it is a five-page paper accepted at CIKM 2026 rather than a large study.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.