SHALA-LLM: Smartly Handling Ambiguous Labels in Aligning LLMs

In three linesSHALA-LLM is a reinforcement learning framework that treats label ambiguity as useful information rather than noise. On NLI and emotion recognition tasks, it reduces Jensen-Shannon Distance by 62.1% on ChaosNLI and improves F1 by 16.7% by learning directly from annotator distributions.

Summary generated by Claude — human-verified