Candidate Exposure Changes Abstention: Measuring and Mitigating Unsupported Entity Selection

Candidate-conditioned entity linkers choose from retrieved candidates, yet the candidate set can change abstention even when the query is fixed. Selective entity linking isolates this failure through candidate-composition interventions that preserve the decision target. Retrieved lists increase benchmark-NIL entity selection across OpenAI GPT-OSS 20B, OpenAI GPT-OSS 120B, and Meta Llama 3.3 70B Instruct; against semantic-profile hard negatives, the increase reaches 5.0 percentage points on Beauty and 9.7 on ESCI. Matched effects transfer to Wikipedia and French and German historical news for GPT-OSS; Llama defines an abstention floor. Mistral Small 3.1 24B shows positive matched contrasts on Beauty, ESCI, and English Wikipedia, while its French and German contrasts remain unresolved. A support gate separates proposal from grounded acceptance. A frozen product-search test retains 87.7–92.0% of gold proposals while rejecting 85.3–86.3% of plausible confusers and 94.3–97.7% of benchmark-NIL candidates; French and German transfer preserves most gold and rejects 77.9–85.0% of confusers. Candidate exposure is a measurable interface risk, and explicit support verification separates candidate plausibility from acceptance.