# Sep 1, 2025

- **LLM routing** is redefined as a **contextual bandit problem**, allowing for adaptive decision-making without exhaustive inference, thus addressing the limitations of previous supervised learning approaches.

- **Memory architectures** must evolve beyond SRAM and DRAM due to stagnation in scaling, necessitating a shift to **specialized memory systems** that cater to specific application needs.

- The **GraphLand benchmark** introduces **14 diverse datasets** for node property prediction, addressing the critical need for realistic benchmarks in graph ML, with features that reflect real-world applications and include temporal distributional shifts for more robust evaluations. [GraphLand paper](https://arxiv.org/abs/2409.14500) | [Code](https://github.com/yandex-research/graphland)

- A **novel metric** for measuring semantic novelty in **collaborative text generation** was developed by calculating cosine distances between consecutive sentence embeddings, revealing that human contributions consistently exhibit higher novelty than AI across various models like RoBERTa and DistilBERT.

- **Open-set recognition** in deep learning addresses the challenge of identifying **unknown classes** that were not present during training, which is crucial for real-world applications where new data can emerge unexpectedly.

- **ELV-Halluc** introduces a novel benchmark for assessing **Semantic Aggregation Hallucinations (SAH)** in long videos, revealing that these hallucinations increase with **semantic complexity** across multiple events.