Back to AI News
arxiv

CLADE Shows RAG Models Guess From Coarse Evidence Instead of Abstaining

Researchers identify a failure mode called CLADE, where retrieval-augmented language models answer from passages that mention only a broader category rather than the specific answer required. On biomedical questions from PrimeKG, family-only passages reduced abstention by 13.8 points versus uninformative passages, and the resulting guesses were wrong about half the time. A counterfactual test replacing an exact disease name with its broader family lowered abstention by 17.4 points, while confidence filtering, larger models, and abstention instructions did not resolve the issue. Ontology-assisted recovery improved end-to-end performance from 10.0% to 26.8%, but identifying the correct family on the page remained a bottleneck.

AI Brief

It gives RAG practitioners a concrete diagnosis of a hallucination-prone evidence pattern and quantifies both the failure and partial ontology-based mitigation.

arXiv:2611.14355v1 Announce Type: new Abstract: Retrieval-augmented systems often retrieve passages that are accurate and relevant but state the answer at a coarser level of granularity than the question requires. For example, a question may ask which specific disease a drug treats (familial hyperlipidemia), while the retrieved passage mentions the broader disease family (hyperlipidemia). Prior work on context sufficiency treats such coarse passages simply as insufficient. We argue that they form a distinct case, which we call resolvable insufficiency: text alone cannot settle the fine-grained answer, but the broader category it mentions maps to a short list of candidate answers in a domain ontology. When language models read these coarse, family-only passages, they exhibit a phenomenon we call Clade (Category-Level Answering from Deficient Evidence): they answer instead of abstaining or searching further. On biomedical questions from PrimeKG, family-only passages reduce abstention by 13.8 points compared to uninformative passages, and the resulting guesses are wrong about half the time. A counterfactual test confirms the cause: replacing an exact disease name with its broader family name lowers abstention by 17.4 points compared to deleting the sentence. Simple fixes like confidence filtering, scaling to a larger model, and explicitly instructing the model to abstain cannot resolve Clade, and this replicates on ecological species interactions in GloBI. Finally, we test whether an ontology can turn coarse passages into exact answers. Given the correct family, an ontology list includes the true answer in 98.5% of cases, but string matching finds that family on the page for only 19.8%. Letting the model choose the family raises this to 30.0% and end-to-end recovery from 10.0% to 26.8%. Overall, our results show when language models should stop guessing and consult an ontology, and highlight which steps of that process remain unsolved.
Be the first to rate this article.

Comments & Questions (0)

Captcha
Please be respectful — let's keep the conversation friendly.

No comments yet

Be the first to comment!