This study examines the knowledge representation of Japanese university students assessed using grammaticality judgement tests (GJTs) and a metalinguistic knowledge test (MKT). The study also investigates the role of automatised and non-automatised explicit knowledge in general L2 language proficiency. Participants were 87 late learners of English as a foreign language (EFL) who completed the timed and untimed GJTs and MKT, modified Group Embedded Figures Test (GEFT), and the Test of English for International Communication (TOEIC). The principal component factor analysis results indicated that ungrammatical sentences in the GJTs loaded on non-automatised explicit knowledge, whereas grammatical sentences loaded on automatised explicit knowledge. The score for ungrammatical sections on the timed GJT was the most significant predictor of all components of the TOEIC. The results also indicated that the time pressure applied in the timed GJT could not sufficiently limit participants' access to explicit knowledge when they processed grammatical sentences. The findings suggest that non-automatised explicit knowledge may play a more significant role than automatised explicit knowledge in L2 proficiency in the case of Japanese EFL learners.