AI Research Brief

OpenAI shipped a competitor to Jev in one week. This was discussed by the leaders of OpenAI's CUA team and API platform

The FTC is investigating OpenAI, Anthropic, and other AI companies over product risks, marking a significant escalation in regulatory scrutiny of frontier model deployment

The authors propose CorrGRPO, which normalizes pairwise reward covariances into Pearson correlation coefficients during Group Relative Policy Optimization. This prevents large-scale correlated rewards from dominating normalization and suppressing signals from smaller-scale rewards, balancing multi-reward learning without altering the centered total reward

Parallel Power Tempering (PPT) resolves the exploration-exploitation trade-off in power-sharpened sampling by running multiple interacting replicas at different sharpening levels. Lower-power replicas explore diverse reasoning trajectories while higher-power ones exploit, offering a scalable inference-time alternative to RL for enhancing reasoning in small models

OneStreamer unifies perception and memory in streaming video interaction through Proactive Hierarchical Caption Memory, which produces time-grounded captions and summaries. This allows the model to retain reusable factual context without revisiting historical visual features, enabling proactive responses as evidence becomes available

This 0.9B model maps six modalities into a single shared cosine space without updating text-side parameters, using dense cascaded captions as teacher signals. It achieves unified embedding without catastrophic forgetting by leveraging byte-identical geometry between teacher and student backbones

A 27B multimodal decision model post-trained from Qwen3.8-27B that returns probabilities for typed questions in a single forward pass. It eliminates free-form text generation and output parsing, offering a structured-output alternative for deterministic tasks

A 9B multimodal variant of Clef, post-trained from Qwen3.5-9B, providing faster inference for structured decision tasks. It shares the same joint schema head architecture and API compatibility with the larger model

A self-modifying visual backbone architecture released with open weights for ImageNet, COCO, and ADE20K tasks. The hierarchical checkpoints allow for dynamic adaptation during inference, representing a novel approach to visual representation learning

An interactive code-execution framework for robot agents that composes classical primitives and VLA policies into Python cells. It reduces token overhead by 65% while increasing success rates by 14% through selective observation and local retries

A retrieval-based speculative decoding framework optimized for coding agents, retrieving from session and workspace corpora. It adapts draft lengths online from verification feedback, addressing the mismatch between static retrieval engines and dynamic agent pipelines

A Go binary that runs GGUF models via Vulkan on AMD, Intel, and Nvidia hardware

A domain-specific language designed to streamline the development of high-performance kernels for GPUs, CPUs, and accelerators. It aims to reduce the complexity of manual kernel optimization for ML practitioners