AI Daily — June 4, 2026

Models & Research

Audio-Language Models Override Audio Evidence When Conflicting Text Is Present — Researchers investigated whether audio-language models suppress audio-supported answers or simply fail to represent them when conflicting text is provided, finding that correct audio-grounded responses are often available but overridden rather than absent. The study uses counterfactual probing to expose this systematic bias toward text dominance. arXiv ↗

Richer Feedback Signals Could Replace Single-Bit Rewards in Reasoning Model Training — A new approach challenges the standard reinforcement learning recipe for reasoning models, which rewards each response with only a binary correctness signal, by incorporating richer distributional feedback. The proposed method, Distributional DAgger, aims to take advantage of the fuller information often available during training. arXiv ↗

Base LLMs Can Predict How Judge Models Will Score Their Own Outputs — Researchers found that pretrained language models already possess latent calibration abilities to anticipate how an external judge will evaluate their responses, even without targeted fine-tuning. With minimal few-shot prompting, these models can reliably predict judge scores, suggesting self-evaluation capacity emerges before any specialized training. arXiv ↗

LLMs Tested on Dynamic Clinical Decision-Making With Standardized Patient Scenarios — A new evaluation framework assesses language models as clinical agents across multi-turn patient encounters, requiring them to gather information, plan treatment, and adapt care over successive states. Unlike static benchmarks, this approach reveals how well models handle the evolving complexity of real clinical interactions. arXiv ↗

Unified Online Audio-Language Model Handles Multiple Tasks Simultaneously — Researchers propose a single always-on audio-language model capable of performing streaming speech recognition, voice conversation, and other audio tasks concurrently, rather than requiring separate specialized systems. The architecture is designed around a continuous perceive-decide loop rather than offline batch processing. arXiv ↗

Industry & Funding

Google Partners With Voltus on Virtual Power Plant to Help Supply Data Center Energy — Google has entered into an agreement with demand-management company Voltus to establish a virtual power plant operating within the largest electricity grid in the United States. The deal underscores the growing effort to find flexible energy solutions as demand from AI infrastructure continues to rise. MIT Tech Review ↗

Ex-Goldman and Meta Founders Build Voice AI Tailored to African and Middle Eastern Markets — A startup founded by former Goldman Sachs and Meta employees is building voice AI tailored to markets that major players have largely bypassed, with its platform now processing in excess of 17,000 calls each day. The company's growth reflects increasing interest in deploying AI solutions for underserved regions and languages. TechCrunch AI ↗

Microsoft's Majorana 2 Quantum Chip Showcases Agentic AI in Hardware R&D — Microsoft unveiled its Majorana 2 quantum chip, which reportedly achieves qubit reliability far exceeding earlier generations and dramatically longer qubit lifetimes than the current industry standard. The development also served as a demonstration case for using agentic AI systems to accelerate hardware research and development workflows. AINEWS ↗

Amazon Integrates AI-Generated Product Imagery Into Search Results — Amazon is rolling out AI-generated visuals within its product search experience, producing images tailored to match user queries to help shoppers navigate to relevant items. The feature marks another step in the retailer's broader effort to embed generative AI into its core shopping interface. TechCrunch AI ↗

Tools & Open Source

Open-Source Two-Stage Vision Pipeline Classifies Vehicles for Road Safety Research — A publicly available computer vision system using Vision Transformers is introduced to automatically categorize vehicles by body type from roadway video footage, addressing a gap in tools relevant to cyclist injury risk assessment. The pipeline operates in two stages to achieve finer-grained classification than standard object detection benchmarks allow. arXiv ↗

E.ON Uses SAP Platform to Standardize Grid Data and Enable AI Deployments — European energy utility E.ON is modernizing its infrastructure by consolidating grid data through SAP's enterprise platform, creating a foundation that supports AI-driven operations across its grid, customer, and infrastructure divisions. The standardization effort is intended to reduce complexity and accelerate technology deployments at scale. AINEWS ↗

Policy & Society

OpenAI Publishes Formal Public Policy Agenda Covering Safety, Youth, and Global Standards — OpenAI has released a document outlining its policy priorities, addressing areas including AI safety, protections for young people, workforce transitions, and the establishment of international norms. The agenda signals the company's intent to engage more actively with governments and regulators on shaping the rules around AI development. OpenAI ↗

Summaries are AI-generated and may contain errors — always verify against the linked original. Each story links to its source, which holds the copyright. Outlet names are shown for attribution only and do not imply any endorsement or affiliation.