Search papers, labs, and topics across Lattice.
5
0
6
6
Concept-based counterfactual explanations no longer need to choose between speed and accuracy: U-CECE dynamically adjusts the complexity of concept representations to provide the best explanation for the available compute.
LLMs that appear strategically savvy in standard games often crumble when faced with slight rule changes, suggesting they're mimicking rather than truly reasoning.
LLMs can be tricked into falsely penalizing objective reporting of conspiracy theories, but this "Reporter Trap" can be overcome with an adversarial "Anti-Echo Chamber" architecture.
Parameter-efficient fine-tuning and instruction tuning can enable strong performance in multilingual, multi-domain aspect-based sentiment analysis without requiring extensive resources.
LLMs systematically fail at multi-label causal reasoning due to shared inductive biases like causal chain incompleteness, even across diverse model families.
