SIMIT: VLMs self-improve at test time by imagining practice problems, +7.2% on 17 benchmarks
delliott · x · 2026-10-28
SIMIT lets vision-language models self-improve at test time with zero labels: before answering, the model writes practice questions, decides the answer first, then generates an image that makes it true. Unlike majority-vote self-training (which locks in errors — on VizWiz 99.2% of TTRL pseudo-labels are 'Unanswerable'), SIMIT learns from problems it authored itself. Training-free SIMIT-ICL improves 16/17 benchmarks with a +7.20% mean relative gain over BAGEL-7B, no labels, teachers, or external verifiers.
More from Multimodal
- Google demo: Gemini 4 breaks down tasks, Genie 4 simulates Apollo robot in real time — minchoi · 2026-10-28
- Genie 4 simulates a robot arm running lab experiments, sparks included — minchoi · 2026-10-28
- Genie 4 demos: documentary-grade crane in snow and hot-air balloon over Antarctica — minchoi · 2026-10-28
- Genie 4 in action: full prompt builds a photorealistic first-person wizard world you can explore — minchoi · 2026-10-28
- Google demos prompt-generated interactive virtual worlds you can explore — mhdfaran · 2026-10-28
- Google DeepMind's Genie 4: a real-time world model with sound and talking characters — The Decoder · 2026-10-28