News
Papers for this week: October 24th - October 30th
Written on 24.10.2026 20:25 by Adam Dziedzic
Dear Students,
We will discuss the following papers on October 30th:
1. "Data Provenance Auditing of Fine-Tuned LLMs with a Text-Preserving Technique"
ICML 2026 – https://openreview.net/forum?id=3D0x8oxOYH
This work introduces invisible Unicode cue–reply marks that let data owners check whether their documents were used to fine-tune an LLM. The method does not change the visible text and has no false positives.
2. “Securing Multimodal AI through Internal Information Decomposition” (FlowGuard) https://openreview.net/forum?id=vTJmIhNvC3
The idea of detecting attacks from the inconsistency between unimodal and fused predictions is appealing and the authors leave multi turn and multi image settings explicitly open, which seems like a natural place to build on.
Have a good weekend,
Adam Dziedzic