AI Innovation Tracker

Multimodal Agent Harnesses

• 🔥 Ptah Multi-Agent System: Ptah is a multi-agent harness for verifiable multimodal deep research that uses...

DeepSeek's first native multimodal model brings built-in vision to the open-source series, eliminating the need to stitch separate text and vision...

Graphon's intelligence layer maps relationships across massive datasets—trillions of tokens from documents, video and databases—using smaller models...

Two distinct research directions tackle core autonomous driving challenges from complementary angles.

• Coding agents for VLA: Explores Code as...

Google is advancing on two fronts with its latest Gemini releases.

• Gemini 3.5 Flash delivers pro-level performance across multimodal vision, video...

Anthropic's latest model prioritizes agentic capabilities with Dynamic Workflows.

• Orchestrates up to 1,000 parallel subagents for codebase-scale...

Reactor's $59M Series A marks a decisive move from batch AI video generation—often taking 10 minutes for just 10 seconds—to instantaneous, interactive output that powers live, user-shaped experiences in film, gaming, and beyond.

Multi-agent systems are maturing rapidly, moving from large-scale production demos to specialized research frameworks.

• Google's open-source demo...

Two new methods push RL-based post-training into factual QA and subjective tasks:

• CorVer delivers lightweight sentence-level rewards for factual QA...

NeuROK learns a latent space of object states with a transformer decoder to generate realistic 4D deformations from static 3D shapes, bypassing...

• Multimodal scale: Systems breakthrough enables faster, memory-efficient training of large multimodal LLMs
• Multilingual reasoning: Layer Swap...

MCP serves as a universal bidirectional adapter that lets LLMs dynamically discover, read, and write to databases and dev environments without...

Dense retrievers develop strong position bias based on where relevant evidence appears in training documents, with skewed distributions causing models...

CausaLab introduces a scalable benchmark where LLM agents must recover hidden structural causal models through observation and intervention to predict...

Agent Evaluation and Safety Frameworks

• AsyncTool: Introduces a benchmark for evaluating asynchronous function calling under multi-task...

• Cognition's Devin commands a $26B valuation at 53x revenue multiple on $492M ARR, betting autonomous agents will outpace IDE tools like Cursor
  -...

MIT's MeMo separates memory into a small dedicated model, enabling teams to update LLM knowledge via efficient merging without retraining the core system or risking catastrophic forgetting, delivering 26% performance gains on complex benchmarks.

China unveiled an LLM-based "AI brain" that autonomously interprets satellite imagery, selects algorithms, and initiates responses with minimal human...

Lance delivers a native unified multimodal model for image and video understanding, generation, and editing by leveraging multi-task synergy instead...

Recent work shows vision-language models rapidly integrating depth reasoning, segmentation, and continuous action generation for real-world robot...