AI Innovation Tracker

Agent Evaluation Advances

• 🔥 PACE: A Proxy for Agentic Capability Evaluation: PACE constructs proxy benchmarks from non-agentic evaluations that...

PACE predicts performance on expensive agentic benchmarks like SWE-Bench using a compact subset of fast non-agentic tests, delivering under 4% MAE,...

• WorldSample anchors world-model augmentation to real rollouts, then applies Policy-Paced Learning to curb hallucination noise and cut training steps...

NVIDIA's open-sourced GR00T control stack and Alibaba's Qwen-RobotManip model have slashed the replacement cost of proprietary humanoid data and VLAs,...

Foundation model embeddings enable classical ML to fuse zero-shot multimodal TCGA data—genomics, imaging, clinical notes, and pathology...

Autonomous Agents for Scientific Discovery

• 🔥 DiscoPER: DiscoPER recovers 8 of 9 known patterns with 72.7% hypothesis support rate on the...

A clear trend is emerging: AI frameworks that close the full loop of scientific inquiry without constant human oversight.

• AutoTrainess lets LMs...

• FurnitureVLA enables long-horizon bimanual assembly via progress-enhanced VLAs, lifting sim success from 48% to 80%.
• ABot-M0.5 unifies mobility...

RoboWorld enables scalable policy testing via fast autoregressive video models, achieving strong correlation with real rankings.

• VLA-Replica...

• DART adapts VLAs to environmental shifts like camera pose or embodiment changes via one-shot weight arithmetic and subspace alignment.
• ACT-VLA...

Continual Robot Learning

• 🔥 World Action Models: World Action Models enable continual imitation learning via Recurrent Generative Replays that...

Current video generators fail at basic physics like gravity and object interactions, and existing metrics overlook these errors. PQSG addresses this...

OpenAI usage data acts as a canary in the coal mine, confirming the chatbot era is over and agentic systems are rapidly expanding beyond engineering....

WOLF-VLA通过最优控制生成动态一致的全身运动轨迹数据集，直接训练VLA实现语言指令驱动的人形 locomotion，解决了接触丰富动作的数据稀缺与安全性问题。

WAMs则利用生成式重放（ReGen）合成伪回放轨迹，让策略在持续学习新任务时无需存储真实演示即可缓解灾难性遗忘，接近真实回放性能。

两者共同指向：世界模型与VLA的深度融合正成为人形机器人自主能力的核心基石。

100+代理开放协作一周，最终将vLLM中Gemma 4推理速度提升5倍。

关键涌现行为：

• 自组织规范：代理拒绝私下Telegram沟通，视其为“串通”，并主动标记验证漏洞供社区裁决。
• 知识共享与分工：维护公共lever-maps与playbooks，四代理接力完成int4检查点从构建到运行调试。
• 发现与纠错：代理发现“更聪明草稿反而更慢”，并通过MTP推测解码打破247 TPS记录。

Robot Value Models and Driving Policies

• 🔥 World Value Model (WVM): ByteDance Seed introduces WVM using world model backbones for value...

Recent work reveals a converging pipeline for reliable Vision-Language-Action systems:

• DriveStack-VLA injects BEV representations and Render-Teacher...

ConstrainedMimic is a control framework enabling real-time constraint enforcement in RL-based humanoid motion tracking policies via whole-body kinematics and dynamics. This directly addresses safety needs for deployable humanoids.

HiPath shows that a lightweight 15M-parameter adapter on frozen UNI2 and Qwen3 backbones can deliver structured slot-level pathology reports from...

University of Leicester scientists developed a causal sense check that generates heatmaps to test whether Vision-Language-Action robots act on...