Decoding the Next Frequency
of Artificial Intelligence.

Researchers identify specific attention heads, called gaze heads, in vision-language models that track and control the described image regions, allowing targeted steering of model output without retraining.

This work reveals an interpretable, mechanistic lever inside VLMs for controlling multimodal output precisely, advancing understanding of model internals and enabling more controllable and explainable multimodal AI systems.

Agentos is a Python-based operating system framework designed to build, test, deploy, monitor, and govern AI agents.

It offers developers a structured platform to manage complex AI agent workflows, enhancing robustness, maintainability, and compliance in AI agent deployments.

Open-source project providing a semantic recall layer and knowledge graph to enable AI-driven briefing and Q&A agents from domain-specific data.

It offers a customizable, governed AI system for organizations to build AI agents grounded in their own data, enhancing domain-specific knowledge retrieval and decision support.

PackMind is a Rust-based toolset providing local-first, AST-aware context packs and MCP tools for AI coding agents to enhance code understanding and prompt management.

By enabling structured, local-first context handling aware of code ASTs, PackMind can improve the precision and quality of AI-assisted code generation and understanding.

ClinHallu is a benchmark for diagnosing hallucinations in medical multimodal LLM reasoning by decomposing errors into distinct reasoning stages and enabling targeted mitigation.

This benchmark advances the reliability of medical MLLMs by allowing fine-grained detection and correction of hallucinations at different reasoning stages, which is critical for trustworthy clinical decision support.

Persona-Pruner is a pruning framework that extracts persona-specific sub-networks from large language models to create lightweight role-playing models without substantial loss in character authenticity.

This approach enables efficient deployment of multiple role-specific chatbot personas simultaneously in resource-constrained environments by significantly reducing model size without major performance degradation.

AdaSR introduces an adaptive streaming reasoning framework for large models that reason during input stream and finalize decisions with a new policy optimization method, improving reasoning accuracy and efficiency.

This approach addresses limitations of traditional static read-then-think paradigms for dynamic streaming data, allowing more flexible, latency-aware reasoning that better fits real-world continuous input scenarios like audio and video streams.

This paper introduces Preference Coordinated Multi-agent Policy Optimization (PCMA) for cooperative multi-objective multi-agent RL, which learns agent-specific preferences to improve team performance and trade-off coordination.

PCMA addresses conflicts in multi-agent systems with multiple objectives by enabling agents to coordinate preferences effectively, potentially improving real-world multi-agent cooperation scenarios with conflicting goals.

This paper identifies and addresses thinking-answer inconsistency in reinforcement learning with verifiable rewards (RLVR) for large vision-language models by proposing a consistency-oriented reasoning alignment method.

Improving semantic consistency between reasoning and answers enhances the faithfulness and reliability of multimodal AI models, crucial for trustworthy AI deployments involving vision and language comprehension.

Parallel-Synthesis enables large language model agent workflows to synthesize outputs directly from parallel KV caches, improving efficiency and preserving parallelism over traditional sequential text concatenation.

This approach reduces redundant computation and latency in multi-branch agent workflows, enabling more natural, efficient, and scalable synthesis in LLM-based systems, which is critical as agents grow more complex and parallelized.