Researchers introduce GUARD, a method to guide large reasoning models to forget sensitive information by converting unsafe disclosures into safe-exit trajectories. They also propose a new metric, Natural Forgetting Reasoning Score (NFRS), to evaluate the quality of forgetting. Experiments show that GUARD reduces unsafe disclosures while preserving reasoning utility in two widely adopted distilled LRMs.

Researchers studied how fine-tuning large language models (LLMs) affects their internal mechanisms, including attention patterns and layer-wise activations. They found that task-relevant components are concentrated within specific layers, but the distribution of these components is not correlated with the layers undergoing the most significant representational changes. This suggests that fine-tuning can lead to a degradation of performance on other tasks when there is overlap in task-specific components.

Researchers propose a method for training AI models with noisy labels by pre-training a feature extractor using self-supervised learning, improving model robustness and accuracy in noisy environments.