Stanford's Training Witnesses: certify ML training runs without trusting the trainer
pratyusha_PS · x · 2026-10-02
Researchers from Stanford and NYU (Houjun Liu, Pratyusha Sharma) introduce Training Witnesses (arXiv:2609.33915), a method to certify neural network training, data usage, and evaluation.
Key points:
- With the explosion of papers, verification burden falls on readers who must reproduce expensive runs; the authors move the burden of proof to the trainer
- Core insight: fast behavioral fingerprints plus occasional replay challenges are sufficient to audit training
- Minimal overhead for trainers, cheap for verifiers, rejects bad runs with amplifiable probability, and supports exact queries of data inclusion/exclusion
- Tested on LM training runs from 100M to 2B scale across DDP and FSDP
- Comes with a self-regulating leaderboard of "auto-certified" runs for shared baselines and reproducibility
Related event: Stanford Researchers Propose Training Witnesses to Verify ML Results(2 posts)→
More from Research
- Diffusion will be everywhere: why text diffusion models may replace autoregressive LLM inference — akbirthko · 2026-10-02
- BIABench: No AI agent scores above 0.19 on 3D bioimage analysis tasks — notredame · 2026-10-02
- BiasReducer from CMU edits only the reward head to adaptively cut length and confidence biases — CarnegieMellonU · 2026-10-02
- Anthropic's BootLoops: a toolkit for exact calculations in quantitative science — badumtsssst · 2026-10-02
- Jeff Clune keynote: open-ended and AI-generating algorithms will drive the AI science revolution — jeffclune · 2026-10-02
- Google's Android Bench 2.0: ~30 multi-day engineering tasks to eval long-horizon agents — DynamicWebPaige · 2026-10-02