LLM News

Researchers introduce MPCEval, a specialized benchmark for evaluating multi-party conversation generation—a capability increasingly used in smart reply and collaborative AI assistants. The benchmark decomposes conversation quality into speaker modeling, content quality, and speaker-content consistency, revealing that current models struggle with participation balance and maintaining consistent speaker behavior across longer exchanges.

Researchers introduced MPCEval, a reference-free evaluation suite designed to measure multi-party conversation generation quality across three dimensions: speaker modeling, content quality, and speaker-content consistency. Testing on public and real-world datasets, the benchmark revealed that single-score metrics obscure fundamental differences in how models handle complex conversational behavior like turn-taking and role-dependent speech patterns.

A new paper identifies token rounding as the primary bottleneck limiting continuous diffusion language models (DLMs) and proposes CoDAR, a two-stage framework that combines continuous embedding-space diffusion with a contextual autoregressive decoder. Experiments on LM1B and OpenWebText show CoDAR achieves competitive performance with discrete diffusion approaches while offering tunable fluency-diversity trade-offs.

Researchers have identified token rounding as a primary bottleneck limiting continuous diffusion language models (DLMs) and propose CoDAR, a two-stage framework that maintains continuous embedding-space diffusion while using an autoregressive Transformer decoder for contextualized token discretization. Experiments on LM1B and OpenWebText show CoDAR achieves competitive performance with discrete diffusion approaches.

Google DeepMind has released Nano Banana 2, an image generation model that combines advanced world knowledge and subject consistency with faster inference speeds comparable to its Flash offering. The model is positioned as production-ready with capabilities previously associated with Pro-tier performance.

Adobe has integrated a new "Quick Cut" feature into Firefly that automatically generates rough video edits from raw footage based on text prompts. The tool handles the initial editing phase, traditionally one of the most time-consuming parts of video production.

Adobe has added Quick Cut to Firefly, an AI-powered feature that automatically generates first-draft videos from raw footage based on user instructions. The tool is designed to reduce manual editing time by processing footage and applying cuts, transitions, and basic structure without requiring frame-by-frame manual work.

ProducerAI, a music generation tool, has been integrated into Google Labs. The platform was used by artist Wyclef Jean in the creation of his new song 'Back in Abu Dhabi,' marking an early real-world deployment of Google's AI music capabilities.

Sarvam, an Indian AI startup, has launched Indus, a chat application currently available in beta. The move positions the company to compete against global AI players in the rapidly expanding Indian market for conversational AI.

American International Group (AIG) has deployed an agentic AI system with an orchestration layer, reporting faster-than-expected productivity gains in underwriting and portfolio management. The deployment demonstrates measurable improvements in throughput and workflow efficiency, according to recent investor disclosures.

Goldman Sachs is deploying Anthropic's Claude model in trade accounting and client onboarding operations. The deployment represents a broader adoption of generative AI among large financial institutions to improve operational efficiency in back-office processes.