Models · Hacker News ·
Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong
Cactus post-trained Gemma 4 E2B with a 68,000-parameter probe that estimates response error and returns structured confidence scores. Across 12 text, vision and audio benchmarks, routing 15–35% of queries to Gemini 3.1 Flash-Lite reportedly matched its performance; the probe reached 0.814 AUROC versus 0.549 for token e