Executive Summary

This report synthesizes an analysis of 830 papers from the cs.AI category on arXiv, focusing on the period around December 2025. The findings reveal an overwhelming focus on Large Language Models (LLMs), with research spanning their core capabilities, limitations, and societal implications. A dominant and rapidly maturing sub-field is Agentic AI, where LLMs are being transformed from simple text generators into autonomous systems capable of planning, reasoning, and tool use. This shift brings AI Safety, Governance, and Alignment to the forefront, as researchers grapple with new vulnerabilities like "memory poisoning" and develop novel control frameworks. Concurrently, the drive for efficiency continues, with significant innovations in model optimization, compression, and efficient inference. Finally, AI is making deep inroads into specialized domains, particularly healthcare, scientific discovery, and finance, with a strong emphasis on creating robust, domain-specific benchmarks and multimodal applications.

AI Safety, Governance, and Control

• MemoryGraft & Psychological Jailbreaks: Identification of new, sophisticated attack vectors against LLM agents. MemoryGraft demonstrates persistent compromise by poisoning an agent's long-term memory, while psychological attacks use manipulation instead of code to bypass safety filters. This represents a major shift in understanding AI vulnerabilities.
• Verifiability-First & Social Responsibility Stack: Novel architectural frameworks designed to enforce AI safety and governance. These move beyond principles to provide concrete engineering paths for creating auditable, controllable, and value-aligned AI systems.
• AI Epidemiology & Monitorability: New paradigms for AI governance that focus on macro-level behavior analysis and assessing the observability of a model's internal states, bypassing the limitations of micro-level interpretability for massive models.
• Refusal Steering & AlignMerge: Low-cost, inference-time techniques to control model behavior. Refusal Steering adjusts a model's willingness to engage on sensitive topics without retraining, while AlignMerge ensures that merging models does not destroy their safety alignment.

Model Architecture & Optimization

• Efficient MoA & Cascade RL: New designs for Mixture-of-Agents and Mixture-of-Experts systems that tackle communication bottlenecks and scaling challenges, enabling the development of larger, more efficient, and more specialized models.
• TurboDiffusion & Efficient-DLM: Major breakthroughs in accelerating generative models. TurboDiffusion speeds up video diffusion models by over 100x, while other techniques are successfully converting slow autoregressive models into fast, parallel diffusion models.
• Adaptive Computation (CosineGate, KV Admission): Smart mechanisms that allow models to dynamically skip computations or reduce memory writes by predicting the importance of certain layers or tokens, significantly boosting inference efficiency for deep networks and long-context LLMs.

Training Paradigms & Reasoning Frameworks

• Self-Play SWE-RL & Propose, Solve, Verify (PSV): Groundbreaking training methods that enable AI agents to improve themselves without relying on human-generated data. By using self-generated programming challenges or formal verification as a reward signal, these methods pave the way for "superintelligent" software agents.
• Model-First & Cognitive-Inspired Elastic Reasoning (CogER): Advanced reasoning frameworks that enhance LLM reliability. MFR forces an LLM to build a structured model of a problem before solving it, reducing hallucinations. CogER allows an LLM to dynamically switch between "fast" and "slow" thinking, balancing efficiency and accuracy.
• Bidirectional RAG: An evolution of Retrieval-Augmented Generation where the model can safely write high-quality generated information back into its knowledge base, creating a system that learns and evolves from its interactions.

Generative AI & Multimodality

• VASA-3D: A state-of-the-art model that can generate lifelike, audio-driven 3D talking head avatars from a single static image, representing a major leap in digital human creation.
• OmniDrive-R1: A reinforcement-learning framework for autonomous driving that uses a multi-modal "Chain-of-Thought" to explicitly reason about its perceptions, directly mitigating the critical problem of object hallucination in vision-language models.
• AnyTask Framework: An automated data generation pipeline that creates simulated tasks and environments for training robots, aiming to solve the sim-to-real data bottleneck and accelerate the development of generalist robot policies.