Decoding the Next Frequency
of Artificial Intelligence.

EvoArena is a new benchmark for dynamic environments testing LLM agents' ability to adapt over time, and EvoMem is a memory update technique that improves agent performance on evolving tasks.

Real-world AI agent deployments face continuous environment changes, so benchmarks and memory systems that model evolution are critical to building more robust and adaptive LLM agents.

RA-RFT is a new fine-tuning framework that improves LLM reasoning by training retrievers to find analogies based on reasoning benefit rather than semantic similarity, enhancing performance on complex math benchmarks.

This approach addresses limitations of semantic similarity-based retrieval in reasoning tasks by focusing on reasoning patterns, providing complementary improvements orthogonal to existing reward designs and training curricula, advancing LLM reasoning capabilities significantly.

Mana is a novel sim-to-real AI framework for dexterous manipulation of articulated tools using a pipeline inspired by computer animation combined with motion planning and reinforcement learning.

Articulated tool manipulation is a complex problem due to coordination and contact dynamics. Mana provides a scalable method to learn and transfer functional grasping and manipulation skills, advancing robotics dexterity with minimal manual data labeling.

SpatialClaw introduces a flexible code-based action interface for vision-language agents to perform complex 3D/4D spatial reasoning, significantly improving accuracy across diverse benchmarks.

Spatial reasoning is a challenging domain for AI, especially for vision-language models. SpatialClaw’s novel flexible action interface enables more adaptive and compositional spatial analysis, advancing the capability of AI agents to reason in complex 3D/4D environments.

NimbleBrain is a self-hosted platform for managing MCP Apps and agent automations with support for tools, interactive UIs, scheduled runs, and multi-agent delegation.

This platform facilitates building and running complex AI agent workflows in a self-hosted environment, promoting greater control, customization, and integration with LLMs and related AI models.

Tale is an orchestration layer enabling deployment and management of local AI models, agents, skills, and automations on private infrastructure connected to data.

This provides developers and organizations a sovereign, extensible, and customizable AI agent stack that does not rely on external APIs, enhancing privacy and control over AI workflows and automations.

Large language models can automate reproducibility assessments in social and behavioral sciences, matching or exceeding human performance in reproducing study conclusions and effect sizes.

This shows that LLMs can scale reproducibility assessments efficiently, potentially transforming how empirical research is audited and verified, reducing resource intensity and increasing transparency in social sciences.

Agents-K1 is a new AI pipeline that builds detailed scientific knowledge graphs from full research papers using a multimodal parser and a 4B parameter extraction model, enabling advanced multi-hop scientific reasoning.

This approach significantly improves the granularity and accuracy of scientific knowledge extraction from papers, facilitating precise multi-hop reasoning and supporting AI agents in complex scientific tasks across disciplines. It advances AI agents' understanding and utilization of scientific documents beyond shallow citation analysis.

Influcoder is a method to efficiently estimate influence rankings of training samples on LLM outputs by distilling gradient-based influence functions into a compact encoder.

This method addresses the shortcomings of traditional influence function methods, which are computationally expensive and storage-heavy, enabling practical data attribution and filtering on large-scale LLM datasets.

HyperTool introduces a new tool interface for LLM agents that consolidates multi-step tool workflows into single code-block calls, improving reasoning efficiency and accuracy.

This approach addresses the execution-granularity mismatch in tool-augmented LLMs, enabling more efficient, scalable, and accurate multi-step tool use, which is critical for complex AI agent workflows.