Soroush Vosoughi

Standard SFT wastes gradient budget on tokens models either already know or cannot yet grasp; trimming supervision from both extremes yields up to a +26.9 point boost on MATH500 with zero reference-model overhead.

Structural attacks on GNNs can be systematically neutralized without retraining simply by pruning the specific edges that disproportionately inflate the graph's kernel complexity.

Strong performance on traditional memory benchmarks doesn't guarantee effective memory utilization in conversational agents, revealing a critical gap in AI memory architectures.

Optimizing question design can boost information extraction performance more than scaling up model size, with a staggering 18.6 F1-score point improvement.

Significant variations in stereotypical behavior across Spanish-speaking countries reveal the limitations of English-centric stereotype datasets in AI.

Clarifying memories can significantly boost the factual accuracy and personalization of conversational agents, while irrelevant memories lead to degraded responses.

Online political discussions tend to be intellectual humility echo chambers, but simple interventions can break the cycle and foster more open-mindedness.