NASA's AI Rover Drives on Mars, Breakthrough AI In Brain MRI Scans, Apple's Xcode Partners with OpenAI and Anthropic

This Week In AI

Over the past week, artificial intelligence continued its rapid shift from assistive software toward autonomous, real-world decision-making systems, highlighting how quickly AI is moving beyond screens and into physical, medical, and developer-centric environments. A standout milestone came from space exploration, where NASA’s Perseverance rover completed its first Mars drive planned entirely by AI, demonstrating that generative and agentic systems are now capable of navigating uncertain, high-stakes environments with minimal human intervention.

At the same time, AI’s impact on healthcare accelerated meaningfully. Researchers unveiled a new AI model capable of analyzing brain MRI scans in seconds with near-clinical accuracy, a development that could dramatically shorten diagnosis times for strokes and other neurological emergencies. This signals a broader trend: AI systems are increasingly being positioned not just as decision aids, but as frontline tools for time-critical domains where speed directly affects outcomes.

Meanwhile, the software ecosystem continued its evolution toward agentic workflows. Apple announced that AI agents from OpenAI and Anthropic are now deeply integrated into Xcode, transforming the IDE into a more autonomous development environment where agents can write, test, and debug code with minimal oversight. Together, these developments reinforce a central theme of the week: AI is no longer defined by conversation alone, but by its growing ability to act independently across science, medicine, and production-grade engineering systems.

This Week In AI Research

Do MLLMs Really See It: Reinforcing Visual Attention in Multimodal LLMs

Image from arXiv paper.

What’s the research question?
Can reinforcement learning improve the visual attention focus of multimodal large language models (MLLMs) during reasoning tasks?

What did the authors do?
The authors introduced SAYO, a new multimodal large language model trained with a specialized reinforcement learning (RL) framework that emphasizes visual focus:

• Region-level visual attention reward: The model receives a reward based on how well its attention aligns with relevant visual regions, measured by the attention weights on object bounding boxes within images.

• Data alignment: Textual questions are paired with detailed visual token information, including object locations, to guide the model’s visual focus.

• Selective reward application: The attention-based reward is applied to high-entropy (uncertain and informative) tokens to encourage focus during critical reasoning steps.

• Training method: SAYO is trained using Group Relative Policy Optimization (GRPO), combining the attention reward with a format reward to optimize both visual focus and answer correctness.

• Evaluation: The model is tested on various visual reasoning benchmarks, including structured image reasoning (ChartQA, CharXiv), mathematical reasoning (We-Math), and real-world visual tasks (MME-RealWorld).

What did they find?
SAYO showed notable improvements over baseline models:

• Enhanced visual focus: Higher target attention scores indicate better alignment with relevant visual regions.

• Better reasoning accuracy: Outperformed models without the attention-based reward on structured image reasoning tasks like ChartQA and CharXiv.

• Generalization: Improved performance on dense visual tasks such as We-Math and MME-RealWorld, despite not being explicitly trained on math datasets.

• Ablation insights: Removing the attention reward led to significant drops in performance, confirming its importance.

• Visual confirmation: Attention weight analyses showed SAYO consistently attended to relevant regions even with longer output sequences.

Why does this matter?
This work advances the understanding of how to effectively train multimodal models to focus on the right visual information during complex reasoning. By integrating a region-level attention reward into reinforcement learning, SAYO demonstrates that targeted incentives can significantly improve the alignment between visual perception and language-based reasoning. This has broad implications for applications requiring precise visual understanding, such as medical imaging diagnostics, scientific document analysis, and advanced visual question answering. Moreover, the methodology provides a new blueprint for future research aiming to enhance perceptual and reasoning capabilities in multimodal AI systems, bridging the gap between visual grounding and language understanding.

Key Points

• Introduces SAYO, a multimodal LLM trained with reinforcement learning using a visual attention reward.

• Reward encourages the model to focus on relevant visual regions, improving reasoning accuracy.

• Outperforms baselines on structured image reasoning and dense visual tasks.

• Provides a new approach to aligning visual attention with reasoning in multimodal models.

Read on arXiv

Time Series Reasoning via Process-Verifiable Thinking Data Synthesis and Scheduling for Tailored LLM Reasoning

Image from arXiv paper.

What’s the research question?
How can large language models (LLMs) be enhanced for effective reasoning in time series analysis?

What did the authors do?
The authors introduced VeriTime, a comprehensive framework designed to improve LLM reasoning on time series tasks through three main components:

• TSRgen: A data synthesis pipeline that generates a multimodal dataset with process-verifiable annotations for diverse time series tasks such as anomaly detection, scenario attribution, and inferential calculation. It combines a rule-based extractor with an LLM (DeepSeek-R1) to produce structured reasoning trajectories aligned with time series properties.

• VeriTime: A two-stage reinforcement fine-tuning (RFT) process that first trains the model using supervised learning on the synthesized TSRgen data, then applies reinforcement learning with multi-objective rewards evaluating both the correctness of final answers and the validity of intermediate reasoning steps.

• Data scheduling: A dynamic sample selection mechanism that prioritizes training samples based on task difficulty and model performance, enhancing training efficiency and effectiveness.

What did they find?
VeriTime achieved significant improvements across multiple time series reasoning tasks:

• Overall accuracy of Qwen2.5-3B-Instruct increased by 102.15%, with anomaly detection improving by 239.56% and inferential calculation by 118.17%.

• Qwen3-4B-Instruct also showed gains, with a 14.67% overall accuracy boost and 22.74% improvement in inferential calculation.

• On knowledge-based benchmarks like TimeSeriesExam and DROP, VeriTime improved performance by 23.58% and 49.79% respectively.

• Enhanced step-wise reasoning accuracy and a 71% reduction in token usage demonstrated improved reasoning efficiency and interpretability.

• Limitations include the reliance on synthesized data and the need to validate generalization to real-world, noisy time series data.

Why does this matter?
This work advances the integration of structured reasoning and reinforcement learning in time series analysis, a critical area for applications like finance, healthcare, and engineering where temporal data is abundant and complex. By generating process-verifiable reasoning data and tailoring training via dynamic scheduling, VeriTime enables LLMs to perform more accurate, interpretable, and efficient reasoning on temporal tasks. This approach offers a scalable pathway to enhance AI systems' capabilities in real-world scenarios requiring multi-step inference and cross-modal understanding of time series data.

Key Points

• Introduces VeriTime, a framework combining data synthesis, reinforcement fine-tuning, and dynamic scheduling for time series reasoning.

• Generates process-verifiable reasoning trajectories to improve model interpretability and correctness.

• Achieves over 100% accuracy improvements on multiple time series tasks with reduced token usage.

• Provides a scalable, generalizable approach to enhance LLM reasoning in temporal domains.

Read on arXiv

Joint Reward Modeling: Internalizing Chain-of-Thought for Efficient Visual Reward Models

What’s the research question?
Can discriminative reward models be enhanced with semantic understanding and reasoning capabilities without sacrificing inference efficiency?

What did the authors do?
The authors introduced a novel approach called Joint Reward Modeling (JRM) that combines preference learning and language modeling into a single, shared vision-language backbone. Key aspects include:

• Using a shared vision-language backbone with task-specific output heads for discriminative reward prediction and generative language modeling.

• Training on multimodal data with combined supervision signals: a ranking loss for preference learning and a cross-entropy loss for language modeling.

• Internalizing semantic understanding and reasoning capabilities into the shared representation space during joint training.

• Fine-tuning the model on image editing tasks emphasizing instruction following and visual quality.

• Evaluating on multiple benchmarks, including EditReward-Bench and MMRB2, as well as downstream online reinforcement learning tasks.

What did they find?
The JRM approach achieved impressive results:

• State-of-the-art accuracy on EditReward-Bench (85.1%) and MMRB2 (69.3%) benchmarks, outperforming existing models.

• Enhanced representation quality and feature space rank, indicating better semantic and reasoning internalization.

• In downstream reinforcement learning tasks, JRM-guided models showed significant performance improvements, such as scores of 1.00 and 0.50 on GEdit-Bench and ImageEdit-Bench respectively.

• A key advantage is that during inference, only the discriminative pathway is used, enabling fast evaluation without sacrificing the learned semantic and reasoning capabilities.

• Limitations include the need for joint training on large multimodal datasets and potential complexity in balancing preference and language modeling objectives.

Why does this matter?
This work demonstrates that discriminative reward models—which are crucial for aligning AI systems with human preferences—can be significantly enhanced by internalizing semantic understanding and reasoning. By jointly training preference learning with language modeling, JRM bridges the gap between efficiency and deep semantic comprehension. This enables more scalable, robust, and generalizable reward models that can better evaluate complex visual instructions and edits, with broad implications for AI alignment, multimodal understanding, and reinforcement learning applications.

Key Points

• Joint training of preference and language modeling internalizes semantic and reasoning skills into reward models.

• Shared vision-language backbone enables efficient inference by using only the discriminative pathway at test time.

• Achieves state-of-the-art results on visual reward benchmarks and improves downstream RL performance.

• Balances the complexity of multimodal supervision with the need for fast, accurate reward evaluation.

Read on arXiv