Vision & Language Pulse

EliseAI's leap to a $4 billion valuation after raising $350 million underscores robust enterprise demand for vertical AI automation in housing and...

Ultra-small models like TurboGPT demonstrate that a 22KiB transformer can train in just 13 seconds, creating accessible testbeds for architecture...

Official positioning touts near-Astra intelligence at one-fifth the cost plus a 95% cache discount, positioning it as a workhorse model.

• Capability...

LayoutLMv3's unified architecture and training objectives create a general-purpose pre-trained model effective for both text-centric and image-centric Document AI tasks.

MBZUAI launched the K2 Horizon Suite, an open-source collection of AI models optimized for multimodal tasks spanning language, vision, and reasoning and positioned as groundbreaking.

Simply swapping in panoramas brings only limited gains for vision-language navigation. Three targeted changes unlock real benefits:

• Action...

SAKI improves on-policy distillation by routing supervision through maximal coupling: accepted tokens keep reverse-KL while correction positions...

TI.com offers dedicated machine vision camera design resources backed by E2E forums with direct engineer support.

Muse, Dots, and Grokbot frame the emerging fight for consumer personal-agent platforms, pitting tech leaders against each other for mass adoption.

Does learning to generate images produce better geometry and spatial reasoning, or just generation-specific skills?

• Selective gains: I2I training...

Can video-generation models become reliable predictive simulators for robotics, or do visually plausible rollouts still fail to model physical...

Can off-policy-aware sharing of trajectories overcome all-failure groups in RLVR without harmful mismatch?

• GRAFT replaces failed self-rollouts with...

A modern image model must first master accurate text-to-image generation before it can reliably handle instruction-based edits involving objects,...

ViRe uses a frozen VLM (CLIP) to turn MedTS waveforms into morphology-aware Vision Queries that retrieve clinically relevant temporal and channel...

Thinking Reward Models first build case-adaptive rubrics then score outputs, replacing direct scalar mapping with explicit reasoning for more...

Open-source LLMs combined with computer vision tools enable onsite models that mediate art experiences by turning visual perception into language. This approach supports interpretation without claiming to replace human insight.

Raven tackles rising harness complexity by autonomously constructing, evolving, and orchestrating modular harnesses as composable units, enabling...

Marathoner equips base models with ultra-long-horizon execution through a post-training pipeline that synthesizes tasks from large GitHub PRs, chains...

Traditional pass/fail tests fail for AI because identical inputs can yield different but acceptable outputs. Production AI needs task-level evaluation...

When computer vision assigns blight scores to homes via cameras on garbage trucks, who checks the model and who suffers false or biased results?

-...