srush blog verifies 152 operators with floating-point certificates, finds unreported bugs in core ML libraries
austinvhuang · x · 2026-11-20
- Princeton's srushnlp published part 2 of his blog series, "When Rounding Breaks Machine Learning."
- The approach formally verifies the whole ML stack by producing floating point certificates for 152 operators.
- The analysis surfaced rounding-related issues across core ML libraries, including several not previously reported.
- The work was amplified by practitioners like Aleksa Gordić and Austin Huang as "very cool and important." Directly relevant to anyone concerned with numerical stability, inference precision, and formal verification.
More from Research
- Nebius Guide: Choose Between Fine-tuning, RL, or Distillation Based on Feedback You Can Collect — agihouse_org · 2026-11-20
- Paper: withholding training data only delays capabilities as general progress substitutes — Mark Marosi · 2026-11-20
- Cactus Compute publishes 12 papers in 2026, 5 at NeurIPS, focused on on-device efficiency — joemeno · 2026-11-20
- e-means lands across tusKANNy stack, cutting vector index clustering training time by 1.5-5x — CShorten30 · 2026-11-20
- Do LLMs Introspect? Injected-Thought Tests Suggest Claude Detects Anomalies, Not Self-Knowledge — dl_weekly · 2026-11-20
- Philanthropy must fund proactive biology data collection or AI won't cure disease, researcher argues — juanbenet · 2026-11-20