NVIDIA GTC 2026, OpenAI GPT-5.4 Mini & Nano, xAI API Launch, Google Gemini Expansion

This Week In AI

Over the past week, the biggest AI developments focused on AI agents, smaller and faster models, developer platforms, and more personalized consumer AI.

NVIDIA led the week at GTC 2026, where it made a strong push toward agentic AI. The company introduced an open stack for building AI agents and made it clear that it sees autonomous software systems as the next major step after chatbots. This matters because the industry is moving beyond tools that only answer questions and toward systems that can take actions, complete tasks, and operate across business workflows.

OpenAI also made an important move by releasing GPT-5.4 mini and nano on March 17. These smaller models show that the race in AI is no longer just about building the largest model possible. It is also about delivering strong performance at lower cost and lower latency. OpenAI says these models are built for coding, tool use, multimodal reasoning, and high-volume workloads, which makes them especially useful for real-world applications at scale.

On the platform side, xAI opened its API to the public, giving developers direct access to its models. This is a meaningful step because it shifts xAI from being mainly a consumer chatbot brand into a more serious model provider for builders and businesses. As more labs compete for developers, API access is becoming just as important as headline model quality.

Google also expanded Gemini-powered personalization and health AI, showing how quickly AI is being woven into daily products and high-value industries. Its new “Personal Intelligence” features connect Gemini more deeply with tools like Search, Chrome, Gmail, and Photos to deliver more tailored responses. At the same time, Google announced new healthcare-focused AI efforts, signaling that it wants Gemini to play a larger role not just in consumer software, but also in regulated areas where trust and usefulness matter most.

This Week In AI Research

Thinking Deeper, Not Longer: Depth-Recurrent Transformers for Compositional Generalization

What’s the research question?
Can Transformers perform variable-depth reasoning—i.e., adaptively think through problems step-by-step—without increasing their parameter count or requiring larger input contexts?

What did the authors do?
The authors proposed a novel Transformer architecture called Depth-Recurrent Transformers that unrolls reasoning depth through recurrence rather than expanding model size or input length. Key design features include:

• Shared-weight Transformer core: A single Transformer block applied repeatedly in latent space T times, with shared weights across steps.

• Stability mechanisms: LayerScale initializes sub-layer outputs near zero to prevent early training instability.

• Gated recurrence: An identity-biased gating mechanism combines the new candidate state with the previous state, creating a gradient highway for stable deep unrolling.

• Depth embeddings: Encode the iteration index to give the model awareness of its reasoning depth.

• Perception interfaces: Encode inputs with task-specific structural biases, such as topological masking for graphs, relative positional encoding for hierarchies, or unstructured input for language.

• Task-specific readouts: Extract final predictions from the last recurrent state, tailored to each input modality.

What did they find?
The Depth-Recurrent Transformer was tested on three challenging compositional reasoning tasks:

• Graph reachability: Determining if a node is reachable within a graph, requiring variable reasoning steps.

• Nested boolean logic: Evaluating deeply nested logical expressions.

• Relational text: Understanding and reasoning over relationships in language.

Results showed a clear computational frontier:

• Performance sharply transitioned from chance to near-perfect as the number of reasoning steps increased.

• The model generalized well beyond training depths, maintaining stable accuracy up to 8 hops in graphs, 14 nested logic levels, and 6–7 relational steps in text.

• Ablation studies revealed that intermediate supervision (training with step-by-step targets) caused the model to learn shallow heuristics, whereas silent thinking (supervision only at the final step) encouraged genuine multi-step reasoning.

Limitations include potential challenges in scaling to even more complex tasks or modalities, and the need to carefully design perception interfaces for different input types.

Why does this matter?
This work introduces a powerful new approach for enabling Transformers to perform variable-depth reasoning without increasing their parameter count or input context size. By decoupling reasoning depth from model size, it allows for more scalable and efficient deep reasoning in language models and other AI systems. The architecture provides a mechanistic perspective on vertical chain-of-thought reasoning, complementing traditional horizontal token-by-token generation. This advancement could improve the robustness and generalization of AI systems tackling complex, compositional problems across language, graphs, and hierarchies, with potential applications in reasoning-heavy AI tasks, multi-modal understanding, and autonomous agents.

Key Points

• Introduces Depth-Recurrent Transformers that unroll reasoning depth via recurrence with shared weights.

• Uses stability mechanisms (LayerScale, gated recurrence, depth embeddings) to enable deep, stable unrolling.

• Demonstrates strong generalization on graph, logic, and language reasoning tasks beyond training depths.

• Decouples reasoning depth from parameter count and input context size, enhancing scalability.

Read on arXiv

Mind over Space: Can Multimodal Large Language Models Mentally Navigate?

What’s the research question?
Can multimodal large language models (MLLMs) develop the ability to mentally navigate using structured spatial representations?

What did the authors do?
The authors explored whether MLLMs can perform mental navigation by constructing and reasoning over cognitive maps of complex environments. Their approach involved:

• Introducing Video2Mental, a new benchmark designed to evaluate mental navigation capabilities in MLLMs. This benchmark challenges models to process egocentric videos, generate hierarchical cognitive maps, and plan routes between landmarks.

• Creating a structured representation of spatial environments from streaming egocentric videos, capturing the scene as a cognitive map in formats like JSON.

• Designing a route planning task where models must infer paths connecting specified origin and destination landmarks based on the cognitive maps.

• Evaluating model performance using static metrics (e.g., NE/SR_t) and interactive simulation in Habitat (e.g., SR_p/SPL) to assess real-world navigation effectiveness.

• Developing NavMind, a reasoning model built on the Qwen3-VL architecture, trained through a two-stage process: (1) cognitive map construction from videos, and (2) a difficulty-stratified progressive fine-tuning paradigm that filters out simple trajectories to focus on deep spatial reasoning.

What did they find?
The study revealed several key insights:

• Existing MLLMs perform poorly on the Video2Mental benchmark, showing a significant gap compared to human spatial cognition.

• Even when provided with ground-truth cognitive maps, models struggled with route planning, indicating limitations in reasoning over spatial structures.

• NavMind, trained with the proposed paradigm, significantly outperformed previous MLLMs, demonstrating robust mental navigation abilities.

• NavMind's structured reasoning over cognitive maps led to more stable and effective global navigation planning signals, improving performance across diverse environments.

• Limitations include the challenge of accurately constructing detailed cognitive maps from streaming videos and the computational complexity of deep spatial reasoning.

Why does this matter?
This work advances the field of multimodal AI by highlighting the importance of structured spatial representations and deep spatial reasoning for long-horizon navigation tasks. By enabling MLLMs to mentally simulate and plan routes using cognitive maps, the study bridges a critical gap between AI and biological intelligence, where humans and animals excel at mental navigation. The introduction of Video2Mental and NavMind provides valuable tools and benchmarks for future research aiming to develop AI agents capable of understanding and reasoning about complex 3D environments, with potential applications in robotics, autonomous navigation, and virtual reality.

Key Points

• Introduces Video2Mental, a benchmark for evaluating mental navigation in multimodal large language models.

• Proposes NavMind, a reasoning model that constructs and operates over cognitive maps for route planning.

• Demonstrates that explicit cognitive map construction and fine-grained spatial reasoning are essential for effective mental navigation.

• Highlights the limitations of current MLLMs in spatial reasoning and navigation tasks.

Read on arXiv

A transformer architecture alteration to incentivise externalised reasoning

What’s the research question?
How can transformer architectures be modified to incentivize externalized reasoning and early exits during inference?

What did the authors do?
The authors introduced a novel architectural modification to transformer models aimed at encouraging models to externalize their reasoning process and make early exit decisions:

• Shallow early-exit heads: Added at intermediate layers to output a scalar logit representing the probability of early exit.

• Stochastic early-exit mechanism: During inference, the model probabilistically decides whether to exit at a given layer based on this logit, passing residuals unchanged if it continues.

• Two-stage training: (1) Self-distillation: The model learns to match the full-depth model’s predictions and exit probabilities by minimizing KL divergence between intermediate and final layer outputs. (2) Reinforcement learning (RL): The model is rewarded for exiting earlier while maintaining task accuracy, using a modified RLOO algorithm with tailored rewards and penalties.

What did they find?
Preliminary experiments on small reasoning models (e.g., Qwen3-4B) demonstrated promising results:

• Early exit behavior: The models learned to exit earlier without sacrificing accuracy, improving task accuracy from 47% to 55-60%.

• Compute efficiency: Average compute usage decreased from 98% to 90-95%, indicating more efficient inference.

• Adaptive depth: The models dynamically varied their depth per token, using fewer layers for predictable tokens and full depth for complex ones.

• Exit distribution alignment: The learned distribution of exit points closely matched the target distribution from training data.

Limitations include: The results are preliminary and tested on small models; scalability to larger models and diverse tasks remains to be validated.

Why does this matter?
This work advances the development of transformer models by enabling them to externalize their reasoning process and make early exit decisions more effectively. Such capabilities can improve the transparency and monitorability of large language models, which is crucial for safety-critical applications where understanding model behavior is vital. Additionally, by balancing externalization with task performance, this approach offers a tunable trade-off between efficiency and accuracy, potentially reducing computational costs and enabling more responsive AI systems.

Key Points

• Introduces stochastic early-exit heads at intermediate transformer layers to encourage externalized reasoning.

• Uses reinforcement learning incentives to reward early exits while maintaining task accuracy.

• Achieves earlier exits and improved efficiency without sacrificing performance on reasoning tasks.

• Provides a promising direction for transparent and resource-efficient large language models.

Read on arXiv