Nikolaos Aletras

StateBridge reveals that training-free hidden-state alignment can significantly enhance communication efficiency in LLM multi-agent systems, outperforming traditional methods.

Logic pre-pretraining accelerates language model skill acquisition by 36B tokens while enhancing compressibility, revealing a new path for efficient model training.

MultiHashFormer achieves superior performance over traditional Transformers while maintaining a constant parameter footprint, even with multilingual vocabulary expansion.

Safety classifiers leak surprisingly sensitive information: a boundary-targeted attack recovers 19% of user distress conversations from the training data, far exceeding existing membership inference methods.

LLMs stubbornly stick to task-appropriate reasoning even when explicitly instructed to use conflicting logic, but targeted interventions can nudge them towards better instruction following.

Output diversity in post-trained models collapses due to training data composition, not just post-training methods, challenging assumptions about inference-time fixes.

VLMs are more easily swayed by misleading text than you think, and their impressive reasoning chains can mask, rather than reveal, this over-reliance on language.