Training Witnesses proposes cheaply verifiable honesty certificates for ML data and evaluations
pratyusha_PS · x · 2026-10-02
Houjun Liu and co-authors launched Training Witnesses, arguing that trust in ML conference results is "utterly broken." The new system issues "honesty certificates" for datasets and evaluations, which a verifier can cheaply check and trust.
Related event: Stanford Researchers Propose Training Witnesses to Verify ML Results(2 posts)→
More from Research
- MILK is faster but worse: synthetic data shows it underperforms even UMAP vs Bonsai — anshulkundaje · 2026-10-02
- COLM 2026 kicks off next week: 5 invited talks, 24 orals, 856 posters — gregd_nlp · 2026-10-02
- Tokenization survey out; COLM workshop Oct 9 and EMNLP tutorial Oct 28 — yuvalpi · 2026-10-02
- Stanford's Training Witnesses: certify ML training runs without trusting the trainer — pratyusha_PS · 2026-10-02
- FW-Icon: An Expert-Graded Benchmark for SVG Icon Creativity and Execution — andpoul · 2026-10-02
- Post-training Yandex AliceAI-80B from scratch: GGUF release and V100 kernel planned — jjusko20 · 2026-10-02