A research paper investigates the limitations of language models in serializing tree-structured expressions into natural language, finding that the process is lossy and asymmetric. The study proposes a round-trip protocol to evaluate the accuracy of expression recovery and identifies factors contributing to failures. The results suggest that the serialization process is a primary limitation for models communicating hierarchical structure through natural language.

A research paper compares the performance of classical models and large language models (LLMs) on tabular data, finding that training a classical model can be more effective than using a frozen LLM, especially when given a small amount of labeled data. This has implications for business prediction problems where labels are expensive.

A research paper debates the effectiveness of gradient-based data attribution methods for large language models, concluding that these methods track format similarity more than task semantics. This has implications for data selection and analysis in AI training, and highlights the need for robustness and reliability in these methods.

Researchers propose a method to improve offline on-policy distillation for AI models, allowing them to learn from imperfect teacher supervision and reducing the need for additional generation. This can improve performance and efficiency in tasks like code generation and mathematical reasoning.

Researchers explore the dynamics of selective state space models (SSMs) and their relationship to transformer architectures. They find that SSMs also exhibit a consensus-driven behavior, similar to attention in transformers, and propose a new analysis framework.

A new confidence estimation method, XConf, is proposed for language models, which uses the model's accumulated experience to improve confidence estimation. XConf stores the model's past episodes, retrieves similar episodes for new tasks, and uses historical success rates to inform confidence. This approach outperforms existing methods on 23 of 24 comparisons and can improve delivered success rate by up to 8.7 points on agent tasks.

A research paper arguing that robustness in AI models emerges naturally as task performance improves, suggesting that explicit efforts to measure robustness may not be necessary. This implies that models are reliable on earlier tasks and suitable for deployment.