Cases › got worse over time › LR-2026-0833
UMT5 decoder causal-mask regression allowed attending to future tokens in Transformers
Confirmed Why: Hugging Face's own report · Hugging Face · 1 independent report
Status: no action reported
No source states: loss or records, could do, deployer, industry.
Severity minor: The regression was contained to the library's main branch with no crash and no third party affected.
Transformers enabled PyTorch's scaled dot-product attention (SDPA) for T5, and as a side effect UMT5's decoder could attend to future tokens, so during training it could peek at the answer it was supposed to predict. Nothing crashed, and the bug stayed in main for almost two months. It was surfaced when the team replayed frozen Model Integration Contracts over five months of Transformers commits: a small UMT5 fixture running on CPU flagged the regression on every affected commit, with label log-probabilities differing from the reviewed baseline by 1.22 and training loss by about 0.20. The contract used SDPA, the default attention path; running only the eager attention variant missed the bug. The regression was fixed on September 28, 2026, and reverting the fix on a proof branch reproduced the same differences.
What happened
No source tells the story step by step yet.
What was done
No source states an action or fix.
Risks
Sorted by rule from the case's fields; not yet read from the sources.
Controls
Not yet read for controls. Controls that mitigate the primary risk:
ApplicationIndependent check of task results · Drift monitoring · Activity logs · Live behaviour monitoring · Grounding and fact checks · Change review and separation of duties
Similar incidents
Evidence
- Cause a permission, scope or wiring error
No source states: a step-by-step account, when it was found, how many records were exposed, the direct loss, a count of the other parties affected, any court case, what set it off, what was done.
Show every field, with its source and grade
The AI system
What happened
Impact
Notes on gaps
The document is a technical blog post about a test-infrastructure project (transformers-mic) and describes one concrete regression: the UMT5 SDPA causal-mask bug. The document also mentions a Whisper assisted-decoding bug and synthetic fault-injection results, but these are research/evaluation findings without a dated concrete incident, so they are not recorded as separate incidents. The LLM-drafted usage code is described as human-reviewed and is not an incident. Extractor proposed values without quotes; dropped: .
LR-2026-0833 · version 1 · published 2026-11-04 after automatic checks; not yet checked field by field by a person · JSON