Grad student asks: is a 60-person survey human eval standard practice in ML research?
Dapper-Perspective21 · reddit · 2026-10-13
A master's student describes evaluating a fine-tuned model's image feedback against Qwen and OpenAI API outputs via a Google Form survey of 60 participants, followed by statistical analysis. They ask whether this human-eval setup is common in ML research, if n≈60 is reasonable, how to filter random/careless respondents, and whether re-running surveys for reviewer-requested additional evaluation is standard practice.
More from Research
- AlphaProof Nexus autonomously proves select math problems, Science reports — swarat · 2026-10-14
- Experiment Wrap: Opus 5.5-Trained 8M Model Generalizes Well but Still Falls Short — Angaisb_ · 2026-10-14
- Gemini 4 Argon nearly doubles AlphaQubit throughput, designs novel enzymes — haider1 · 2026-10-14
- Google researchers lay out the virtuous cycle of quantum-classical machine learning in arXiv paper — weballergy · 2026-10-14
- Johns Hopkins launches BodyMaps AI reader study on cancer detection, offers volunteers authorship — yuyinzhou_cs · 2026-10-14
- Interview with Bartosz Naskrecki on the current capabilities of AI in math — burny_tech · 2026-10-14