Three research papers published on arXiv examine different aspects of large language model capabilities and limitations.
According to arxiv.org, a paper titled “Watch, Remember, Reason: Human-View Video Understanding with MLLMs” presents a framework for analyzing video understanding through multimodal large language models (MLLMs). The work addresses how models handle “sparse evidence, long-range dependencies, multimodal alignment, and reliable inference under limited computational budgets,” organizing analysis around three functional abilities: watching, remembering, and reasoning.
In a separate study on low-precision inference, researchers examined NVFP4-based quantization approaches for LLMs. According to arxiv.org, the paper “Beyond Output Matching: Preserving Internal Geometry in NVFP4 LLM” identifies that output matching alone “can mask internal degradation” during quantization-aware distillation. The researchers propose CKA-QAD, a method that “preserves internal representational geometry during distillation by aligning layerwise Gram matrices through CKA,” tested on Nemotron 3 Nano and Qwen3-4B-Thinking-2507 models.
A third paper compares human and AI reasoning in mathematics. According to arxiv.org, “A Comprehensive Anatomy of Human and DeepSeek-R1 LLM Mathematical Reasoning” analyzed 30 AIME 2025 problems, annotating 10,247 reasoning steps. The researchers found that “DeepSeek-R1 frequently revisits intermediate results, performs shallow and often unnecessary verification, and loops through local checks without meaningful logical progress,” describing this as “topological mimicry: reproducing the surface form of reasoning without its functional role.”