EquivSVA: This paper introduces a formally verified dataset for behavioral assertions across equivalent RTL implementations, enhancing the reliability of SystemVerilog assertions generated from natural-language specifications and register-transfer-level designs. Link
Grow the Harness, Not the Context: This study explores strategies for creating reusable specialist agents by avoiding the repeated reconstruction of control decisions in each task context. It aims to optimize the use of large language models in handling related tasks. Link
Product & Industry
Google Beam Expansion: Google Beam is expanding to five new countries and partnering with Industrious for an extended network, highlighting the company’s ongoing efforts to enhance its global footprint and collaboration capabilities. Link
AI for Everyone in Every Language: Google is expanding its AI capabilities beyond traditional text translation to build models that understand and process the world’s diverse languages more effectively. Link
Tools & OSS
SWE-Serve: This benchmark evaluates agents on production inference engineering tasks, focusing on the coordination of changes across the serving stack. It’s designed to help researchers and developers assess and improve their models in real-world scenarios. Link
Flash-dLLM: This paper presents a method for IO-aware KV caching and parallel decoding, aiming to make diffusion LLMs faster and more memory-efficient. The technique could significantly improve the practical deployment of these models. Link
Notable Discussion
ICLR Main Paper + Supplementary: A discussion on whether to attach the supplementary material along with the main paper during submission to ICLR, or if doing so might result in a desk rejection. Link