A new dataset and benchmark for evaluating large language models' understanding of competitive Chinese-language debate has been introduced. The dataset contains 148 matches, 2,698 stages, and 20,542 exchange units, with manually verified transcripts and segmentation. The benchmark provides a testbed for studying large language models' understanding of interactive argumentation and their agreement with professional judges.

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 new method, probe guidance, is introduced to improve diffusion language models by using frozen internal states to construct a guidance signal, eliminating the need for an additional forward pass at inference time and providing reliable performance on unconditional generation and multiple choice question answering benchmarks.

A new LLM decoding framework, Mahalanobis-Ensemble Decoding (ME-Decoding), is introduced. It frames candidate token selection as ensemble pruning, using Mahalanobis distance to enhance semantic diversity and preserve high probabilities. This approach offers a plug-and-play module with negligible inference overhead.

R3 is a novel reward modeling framework for language models that is rubric-agnostic, generalizable, and interpretable, enabling more transparent and flexible evaluation of language models.

Researchers have developed a method for efficient one-to-many machine translation using a discrete diffusion framework. This approach allows for parallel refinement of multiple target languages, reducing latency and enabling deployment as a single unified model. The framework also supports zero-shot transfer to unseen source languages without retraining, maintaining high translation quality.

Researchers investigated the phenomenon of mode collapse and over-dispersion in large language models (LLMs), finding that fine-tuning with sufficient data can improve diversity. The study provides a theoretical framework and experimental results demonstrating that finite-sample fine-tuning can lead to under- or over-dispersion, depending on the model and dataset. This work has implications for the development and deployment of LLMs.