Report
Post-trained Nemotron reportedly lifts AutoResearchExam score from 0.286 to 0.358
Bespoke Labs says the evaluation covered 22 CPU tasks, each with a four-to-six-hour limit and one graded submission.
TLDR
Bespoke Labs reports that its post-trained Nvidia Nemotron-3.5-Super-VL scored 0.358, up from 0.286 for the base model, on its 22-task AutoResearchExam evaluation. Training used the model’s guided attempts with hints removed, followed by retrospective lessons paired with examples of research actions. Bespoke says the final model scored comparably to Qwen 3.8 Max under the same tasks, harness and time horizon; Qwen has 20 times as many total parameters.
Combined views
3
1 Source, first seen ago
