Agentic AI Goes Mainstream: OpenAI Tools, Enterprise Rollout, and Global AI Safety Warnings Dominate the Week

This Week In AI

Over the past week, AI momentum remained focused on the transition from generative tools to agentic and autonomous systems, reflecting how quickly artificial intelligence is reshaping technological workflows and enterprise strategy. At the forefront, major tech companies advanced autonomous AI capabilities — OpenAI launched a new Codex app designed to let developers manage multiple AI coding agents in one workspace, signaling a shift beyond simple “chat” interfaces toward more coordinated, task-oriented workflows. Simultaneously, industry reports highlight that autonomous agents are becoming a commercial priority, with firms like Innodata positioning themselves to capitalize on agentic AI offerings as the next revenue wave.

Security and governance concerns also surged alongside these advances. A newly released International AI Safety Report 2026, led by global experts, emphasized both rapid capability growth and rising risks — including deepfakes, hidden model behaviors, and the difficulty of separating AI content from human output — underscoring the urgency of robust safety frameworks as AI systems become more powerful. Relatedly, cybersecurity experts warn that enterprises struggle to secure agentic identities and activities, a challenge that could intensify as autonomous tools proliferate across business environments.

Beyond enterprise and safety headlines, broader AI adoption continued across sectors. The U.S. military reported that 5 of 6 branches are now using the Pentagon’s GenAI.mil platform at scale, demonstrating institutional integration of generative AI platforms into core operations. At the same time, nuanced discussions are emerging about the impact of AI on labor; layoffs at major tech companies such as Amazon have been attributed in part to efficiency narratives around AI, though the economic linkage remains debated.

This Week In AI Research

Learning Modal-mixed Chain-of-thought Reasoning with Latent Embeddings

Image from arXiv paper.

What’s the research question?
How can vision-language models be extended to support modal-mixed reasoning that interleaves textual tokens with compact visual "sketches" represented as latent embeddings?
What did the authors do?
The authors developed a novel architecture that combines a vision-language model (VLM) with a diffusion-based latent decoder to enable modal-mixed reasoning:

• Used a VLM as a visual encoder to generate dense visual token embeddings from images.

• Compressed these visual embeddings into fixed-length latent sketches via average pooling to create compact visual representations.

• Interleaved text tokens and visual latent sketches during reasoning by using special tokens to switch between text and visual modes.

• Implemented a diffusion decoder that synthesizes visual embeddings conditioned on the VLM's hidden states, capturing fine-grained perceptual details.

• Trained the model in two stages: supervised fine-tuning on modal-mixed reasoning traces combining next-token prediction and latent reconstruction, followed by reinforcement learning to optimize modality switching and reasoning chain composition.

What did they find?
The proposed model achieved strong performance and demonstrated several key insights:

• Outperformed language-only and other Chain-of-Thought (CoT) methods on 11 diverse multimodal reasoning tasks, showing the effectiveness of modal-mixed reasoning.

• Ablation studies confirmed that both the latent embedding compression and the diffusion-based visual decoder contributed significantly to improvements.

• Exhibited robustness and good generalization across different base models and reasoning tasks.

• Limitations include potential computational complexity due to diffusion decoding and the need for carefully designed training procedures to balance text and visual modality use.

Why does this matter?
This work pushes the boundaries of multimodal AI by enabling models to internally generate and manipulate visual sketches alongside language, rather than relying on external tools or predefined visual inputs. By integrating visual reasoning directly into large language models through compact latent embeddings, it opens new possibilities for applications that require understanding and reasoning across text and images, such as complex question answering, visual problem solving, and creative design. This scalable approach can enhance the versatility and interpretability of multimodal AI systems, bringing us closer to more human-like reasoning capabilities that seamlessly blend different modalities.

Key Points

• Introduces a modal-mixed reasoning framework combining text and compact visual sketches as latent embeddings.

• Uses a diffusion decoder conditioned on vision-language model states to generate detailed visual representations.

• Achieves state-of-the-art results on diverse multimodal reasoning benchmarks.

• Enables large models to internally generate and reason with visual sketches without external tools.

Read on arXiv

MentisOculi: Revealing the Limits of Reasoning with Mental Imagery

Image from arXiv paper.

What’s the research question?
Can modern multimodal foundation models effectively form, maintain, and manipulate high-fidelity visual representations in a goal-directed manner?

What did the authors do?
The authors developed MentisOculi, a comprehensive benchmark designed to test visual reasoning and mental imagery capabilities in AI models. Their approach included:

• Creating five challenging visual reasoning tasks: Form Board, Hinge Folding, Paper Fold, Rush Hour, and Sliding Puzzle.

• Procedurally generating these tasks across five difficulty levels to test model robustness.

• Designing ground-truth visual chain-of-thought solutions to guide reasoning.

• Evaluating four families of models:

  • Multimodal Large Language Models (MLLMs) that produce text-only outputs.

  • Latent Visual Reasoning Models that generate text reasoning chains interleaved with visually-grounded latents.

  • Unified Multimodal Models (UMMs) that explicitly visualize states as images within reasoning chains.

  • Video Models that produce purely visual rollouts conditioned on prompts and initial frames.

• Using automated scoring methods for both text and visual outputs, including an automatic visual rater.

• Comparing model performance across all tasks and difficulty levels against chance and human benchmarks.

What did they find?
Key findings include:

• Model performance declined as task difficulty increased, indicating challenges in complex visual reasoning.

• No visual reasoning approach reliably outperformed text-only baselines, suggesting that explicit visual representations did not confer a clear advantage.

• Unified Multimodal Models (UMMs), which explicitly visualize states as images, suffered from compounding generation errors and failed to effectively leverage ground-truth visualizations.

• Video Models, which produce visual rollouts, did not outperform other approaches or human performance at higher difficulties.

• Techniques that improved language-based reasoning did not translate into better visual reasoning, highlighting a disconnect between modalities.

Why does this matter?
This work provides a rigorous framework for evaluating and understanding the capabilities and limitations of models that attempt to reason with visual mental imagery. By systematically challenging models with diverse, procedurally generated visual puzzles, MentisOculi reveals that current multimodal foundation models struggle to form, maintain, and manipulate high-fidelity visual representations in a goal-directed manner. This insight is crucial for guiding future research aimed at developing more effective visual reasoning systems, which are essential for applications requiring complex visual understanding, such as robotics, autonomous agents, and advanced AI assistants.

Key Points

• Introduces a challenging benchmark suite (MentisOculi) for visual reasoning with high-fidelity mental imagery.

• Evaluates diverse multimodal models, including text-only, latent visual, explicit visual, and video-based approaches.

• Finds that current models do not outperform text-only baselines and struggle with complex visual tasks.

• Highlights the need for improved methods to form and manipulate visual representations in AI reasoning.

Read on arXiv

Resource‑Efficient Reinforcement for Reasoning Large Language Models via Dynamic One‑Shot Policy Refinement

What’s the research question?
How can the resource demands of reinforcement learning with verifiable rewards (RLVR) be reduced while maintaining reasoning performance in large language models (LLMs)?

What did the authors do?
The authors developed a novel algorithm called Dynamic One-Shot Policy Refinement (DoPR) to improve the efficiency of RLVR for reasoning in LLMs. Key aspects include:

• Maintaining a reward history for each training sample to track its informativeness over time.

• Calculating a reward volatility score to measure how much a sample’s reward varies, indicating potential learning value.

• Computing an exploration score to encourage sampling diverse or uncertain samples.

• Combining these scores into a composite score to select the most informative sample for each policy update.

• Performing a single rollout on the selected sample to estimate its reward, then updating the reward history accordingly.

What did they find?
Empirical evaluations on mathematical reasoning benchmarks demonstrated that DoPR can achieve comparable reasoning performance to traditional RLVR methods while using significantly fewer training samples and rollouts. Notable results include:

• Matching the performance of models trained on hundreds of samples using only 16 training samples.

• Outperforming baseline methods under fixed rollout budgets, showing superior resource efficiency.

• Reducing the number of required rollouts and updates without sacrificing reasoning accuracy.

Limitations include the focus on mathematical reasoning benchmarks, which may not directly generalize to all reasoning tasks.
Why does this matter?
This work challenges the common assumption that large-scale data and compute are necessary for effective reasoning in LLMs trained with reinforcement learning. By enabling high reasoning performance with minimal resources, DoPR makes RLVR more accessible and scalable. This has important implications for:

• Reducing the cost and environmental impact of training reasoning-capable LLMs.

• Allowing researchers and practitioners to fine-tune reasoning abilities on specialized or resource-constrained settings.

• Potentially accelerating the deployment of intelligent agents that rely on reasoning in real-world applications.

Key Points

• Introduces Dynamic One-Shot Policy Refinement (DoPR) for efficient RLVR in LLMs.

• Uses reward history, volatility, and exploration scores to select the most informative training samples.

• Achieves comparable reasoning performance with fewer training samples and rollouts.

• Enables scalable reasoning training with minimal computational resources.

Read on arXiv