Study Finds Attention-Mask Checks Miss Causality Leaks in Sequence Models
VIDraft · hf · 2026-08-26
This study proposes a lightweight audit to detect causality violations in sequence models by verifying prefix invariance. It reveals that attention-mask checks miss leaks originating from scans or normalization operations.
More from Research
- Yoav Goldberg Questions S1 Model's 'In-Context Learning' Claim — ivan_bezdomny · 2026-08-26
- Recuris: Recursive Experiential-Working Memory Evolution for Long-Horizon Agents — princetonu · 2026-08-26
- Meta^n: Recursive Self-Improvement through Emergent Depth — Zae Myung Kim · 2026-08-26
- Tencent's CAFE: Self-Improving Search Agents Need Co-Evolving Feedback — Tencent-Hunyuan · 2026-08-26
- Accelerated Understanding Launches AI Model Replacing Transformers with Neural Operators — badumtsssst · 2026-08-26
- Analyst replicates 300+ ICML papers finding errors using agents — _akhaliq · 2026-08-26