New Paper Shows Gemini-Found Softmax Bug Fix Still Fails Property Tests
brandondamos · x · 2026-10-28
An arXiv paper (2610.36427) reproduces the recently discovered softmax numerical bug (found by Gemini) with property-based tests in Hypothesis, shows the proposed fix still fails an edge case, and presents a more robust softmax that passes the tests and supports causal attention masking — demonstrating property-based testing as a powerful tool for ML library numerics.
More from Models
- DeepSeek vs Haiku pricing debunk: accounting for peak rates and token use, 1017 is 2-2.5x cheaper — teortaxesTex · 2026-10-28
- ChatGPT and Gemini refuse to draft a complaint against a pedophile, user reports — chinchinlover-419 · 2026-10-28
- GPT-6.1 Terra priced at $1/$5 per M tokens, but the '9x cheaper per task' math doesn't hold — 2C_ornot2C · 2026-10-28
- FrontierCode 2 leaderboard sparks drama over missing Chinese models — teortaxesTex · 2026-10-28
- Evil 27B Too Risky to Run Autonomous? User Fears Wiped Drive Over a Game — teortaxesTex · 2026-10-28
- DeepSeek V5 rumor debunked: code strings were just September assert tests — teortaxesTex · 2026-10-28