ToMMeR is a lightweight probing model (~300K params) extracting emergent mention detection capabilities from early layers representations of any LLM backbone, achieving high Zero Shot recall across a wide set of NER benchmarks. More information in the associated ACL2026 paper