AI Research Pulse

Agent Memory Systems

• 🔥 TencentDB Agent Memory: Tencent open-sourced TencentDB Agent Memory, a 4-tier local memory pipeline (L0 Conversation, L1...

AI security faces escalating risks from agentic systems while gaining powerful new tools.

• Prompt injection turns AI agents into sleeper agents by...

Three papers converge on robust representation learning that handles time series, domain shifts, and wearable data without heavy augmentation or...

• ScheduleFree+ replaces learning rate schedules entirely with dynamic, self-adjusting optimization for LLMs.
• Core fixes include inner momentum,...

• Collaborative scale: Data from 22 robots across 21 institutions forms the largest open-source embodied dataset, enabling broad skill transfer.
  -...

Recent work reveals AI's dual role in mathematics: tackling open problems via formal verification while serving as a flawed but useful research...

Intern-S2-Preview FP8 deploys a 35B scientific multimodal model locally on a single 48GB GPU using vLLM and FP8 quantization, consuming ~45GB VRAM....

Two fresh systems show how structured memory is replacing flat vector dumps for long-horizon agents.

• TencentDB Agent Memory layers data into a...

Yann LeCun reposted Surya Ganguli's new article proposing a theoretical framework to unify physics, neuroscience, and AI into a science of intelligence.

• $650 million launch funding backs Recursive Superintelligence to automate full AI research loops including idea generation and validation.
• Richard...

Hugging Face spotlighted SEGA, a Spectral-Energy Guided Attention mechanism for resolution extrapolation in diffusion transformers. This targeted approach directly tackles scaling challenges in generative models.

Frontier models mostly cannot forecast scientific progress, based on analysis of 4,760 events. The findings highlight clear limits for AI-driven research acceleration.

Benchmark Advances

• 🔥 ARC-AGI-3 Competition Jump: @tufalabs score rose from 0.68% to 1.17% in the ARC-AGI-3 competition, the first meaningful...

Multimodal LLMs process text, images, video, audio and sensor data through integrated designs rather than text alone.

• Attaching sensors and visual...

The field is shifting from static tests to live, scalable evaluations that mirror authentic workflows.

• TerminalWorld auto-generates 1,530 tasks...

A joint OAS-UK webinar examines the 2026 International AI Safety Report through a regional lens for Latin America and the Caribbean.

• Covers cyber...

• Gaze following: VFMs gain explicit reasoning via head-conditioned LoRA and out-of-cone penalties, lifting performance on non-salient targets.
  -...

Noted AI safety researcher Miles Brundage posted a link to additional content on the topic.

In-Context Modeling (ICM) lets foundation models adapt to new physical systems through a single forward pass by treating measurements as context,...

Jeremy Howard reposted a pointed warning that AI is teaching a generation to accept "close enough" as sufficient, highlighting risks to precise cognitive habits in this phase of the technology.