AI / ML

Researchers developed a vision-language reward model using no human preference labels by training a linear head on a frozen Qwen3-VL backbone. The head scores original web captions against perturbed versions, with the optimal layer selected using minimal labeled data. This approach achieves performance comparable to or exceeding GPT-4o on three benchmarks while requiring only a fraction of the fine-tuning compute.

• Eliminates expensive human preference labeling by using caption perturbations as self-supervision signals.
• Layer selection is critical: one intermediate layer yields 5.8 points higher score than the output layer.
• Outperforms GPT-4o on VL-RewardBench with a 32B parameter model at roughly 0.1% of fine-tuning compute.
• A frozen backbone with a small linear head transfers effectively without re-tuning across benchmarks.

Amazon S3 Vectors now supports metadata pre-filtering and a new prefix match operator ($startsWith) in AWS GovCloud regions. By evaluating filters before executing similarity search, the service can return up to 5x more matching vectors when filters are selective. This enhancement improves result relevance for retrieval-augmented generation and semantic search applications operating at billion-vector scale.

• Metadata pre-filtering runs before similarity search to boost relevant result yield.
• Selective filters can increase matching vectors by up to 5x.
• New $startsWith operator enables efficient prefix matching on paths and URLs.
• Available now in AWS GovCloud US-East and US-West regions.
• Optimizes RAG and agentic app accuracy without extra indexing overhead.

Amazon Bedrock now hosts TwelveLabs Pegasus 1.5, a model that generates structured time-coded metadata from video in a single pass. It analyzes visual, audio, and on-screen text to label specific events without requiring pre-indexing or complex ingestion pipelines. Users can define custom business metrics like speaker changes or brand appearances to query footage directly.

• Single-pass analysis covers visual, audio, and OCR data simultaneously
• No pre-indexing or ingestion pipelines required for immediate use
• Returns structured metadata with precise timestamps for events
• Supports custom business logic like editorial segments or plays
• Available via US and Global cross-Region inference profiles on Bedrock