MOAT: Evaluating LMMs for Capability Integration and Instruction Grounding

LegalRikai has introduced an Open Benchmark designed to evaluate complex Japanese corporate legal tasks, comprising four intricate tasks created under the guidance of legal professionals. This benchmark includes 100 samples that necessitate long-form, structured outputs, and has undergone both human and automated evaluations using advanced language models such as GPT-5 and Claude Opus 4.1.

A recent study has established the first tight lower bounds on the runtime of deterministic speculative generation algorithms for large language models (LLMs), revealing insights into the token generation process through branching random walks. This research provides a mathematical framework to analyze the efficiency of speculative generation, a technique aimed at accelerating inference in LLMs by verifying multiple draft tokens simultaneously.

A recent study published on arXiv examined the influence of data selection on fine-tuning machine translation models, specifically focusing on Japanese-English corpora. The research compared five different data selectors: TF-IDF, COMET Kiwi, QuRate, FD-Score, and random selection, revealing that semantic selectors consistently outperformed others, highlighting the critical role of data quality in model performance.

FilmWeaver has been introduced as a novel framework for generating consistent multi-shot videos of arbitrary length, addressing challenges in character and background consistency across shots. The framework utilizes an autoregressive diffusion paradigm and a dual-level cache mechanism to enhance both inter-shot consistency and intra-shot coherence.

A new pipeline for dynamic scene reconstruction from monocular RGB videos has been introduced, enhancing prior methods through improved segmentation and depth estimation techniques. This approach utilizes video segmentation and epipolar-error maps to create object-level masks, which guide depth loss and support comprehensive 2-D tracking, resulting in superior renderings compared to previous methods.

A recent study published on arXiv explores the interactional friction in modular Speech-to-Speech Retrieval-Augmented Generation (S2S-RAG) pipelines, identifying three main patterns of conversational breakdown: Temporal Misalignment, Expressive Flattening, and Repair Rigidity. These issues highlight the challenges faced by voice-based AI systems in achieving fluid and natural interactions.

A new study presents a model for generating singable lyrics from melodies, addressing the existing gap between machine-generated and human-written lyrics. This model incorporates joint learning of wording and formatting, enhancing its ability to meet specific lyrical structures and prosodic patterns through a self-supervised training phase on a large corpus of lyrics.

FlowDirector has been introduced as a novel training-free and inversion-free video editing framework that allows for precise text-to-video editing by modeling the editing process as a direct evolution in the data space, utilizing an ordinary differential equation to guide video transitions smoothly along its spatio-temporal manifold.