The core idea of Multimodal Large Language Models (MLLMs) is to create models that can combine the richness of visual content with the logic of language. However, despite advances in this field, many models struggle to connect the two domains effectively, leading to limited performance in complex reasoning tasks that involve visual components. A major challenge in building such models is their limited ability to combine visual understanding with logical thinking. Current systems often produce textual outputs that explain reasoning but fail to reference the specific parts of an image they rely on. This creates a gap where models may arrive at an answer without clearly showing how the visual evidence contributed to their decision. It’s also difficult to ensure that models generate visual reasoning steps directly connecting to their answers. The fundamental problem lies in how to naturally train models to interleave text and image reasoning without needing large datasets annotated with visual references, which are scarce and expensive to produce. Existing methods try to address this by using reinforcement learning or prompting strategies. Some systems generate bounding box coordinates as answers, while others produce step-by-step textual reasoning chains. However, these approaches have limitations. Models that only produce bounding boxes lack explanation, while those generating only text risk ignoring visual evidence. Previous methods often separate visual grounding and reasoning, making it hard for models to explain why a particular visual element leads to a certain conclusion. While some models use dense supervision data or additional tools, they generally require heavy annotation and do not scale well. This makes it difficult for developers to create models that can explain their reasoning transparently and handle various visual tasks with minimal data. Researchers from UC Santa Cruz and eBay introduced a new method called Grounded Reasoning with Images and Text (GRIT) that allows MLLMs like Qwen 2.5-VL and InternVL 3 to generate reasoning chains that mix natural language with explicit bounding box coordinates pointing to relevant image regions. This unified approach enables models to reason about and visually ground their answers without requiring dense annotations or labeled reasoning chains. GRIT also uses a lightweight reinforcement learning algorithm called GRPO-GR, which optimizes both the accuracy of the final answer and the structure of the reasoning, encouraging models to include specific tokens like <think> and <rethink>, as well as bounding box formats. This design eliminates the need for costly annotated data while ensuring that models learn to reference visual content meaningfully within their logical steps. The methodology in GRIT focuses on generating outputs that combine textual reasoning and visual grounding seamlessly. Instead of requiring models to process cropped images or additional visual data after generating bounding boxes, GRIT teaches models to use their internal understanding of the image. Bounding boxes are generated during the reasoning process, and models learn to reflect on these coordinates within their logical reasoning. The reinforcement learning framework rewards the correct use of bounding box formats and reasoning structure, and it guides models to produce coherent, grounded reasoning chains. GRIT demonstrates remarkable data efficiency by using only 20 image-question-answer triplets sourced from Visual Spatial Reasoning and TallyQA datasets. The model training was conducted on NVIDIA A100 GPUs, with optimization techniques like AdamW and a cosine scheduler applied over 200 training steps, which shows the method’s scalability despite limited data. Performance evaluations revealed that GRIT-trained models outperform several baselines in reasoning and grounding accuracy. For example, Qwen 2.5-VL trained with GRIT achieved 72.9% answer accuracy on Visual Spatial Reasoning, 47.8% on TallyQA, and 62.8% on GQA datasets. It also reached a grounding IoU score of 0.325 on VSR and 0.447 on TallyQA. In contrast, baseline models like Direct Query or Chain-of-Thought often performed significantly lower, showing limited ability to unify reasoning with visual grounding. GRIT models demonstrated a strong correlation between visual regions and textual reasoning, producing outputs that reflected a meaningful connection between image evidence and logical thought. GRIT also showed improvements on out-of-domain benchmarks, though gains were more pronounced on in-domain data, highlighting the importance of training data diversity for broader generalization. In conclusion, the research addressed the problem of disconnected reasoning and visual grounding in MLLMs by introducing GRIT. The method allows models to reason with images through a simple, efficient approach that requires minimal data. GRIT successfully teaches MLLMs to combine visual evidence with logical reasoning in a unified output, achieving strong performance across multiple benchmarks and demonstrating a promising step toward more interpretable AI systems. Check out the Paper, Project, and GitHub Page. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 95k+ ML SubReddit and Subscribe to our Newsletter. NikhilNikhil is an intern consultant at Marktechpost. He is pursuing an integrated dual degree in Materials at the Indian Institute of Technology, Kharagpur. Nikhil is an AI/ML enthusiast who is always researching applications in fields like biomaterials and biomedical science. With a strong background in Material Science, he is exploring new advancements and creating opportunities to contribute.Nikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces Group Think: A Token-Level Multi-Agent Reasoning Paradigm for Faster and Collaborative LLM InferenceNikhilhttps://www.marktechpost.com/author/nikhil0980/Researchers from the National University of Singapore Introduce ‘Thinkless,’ an Adaptive Framework that Reduces Unnecessary Reasoning by up to 90% Using DeGRPONikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces MathCoder-VL and FigCodifier: Advancing Multimodal Mathematical Reasoning with Vision-to-Code AlignmentNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces PARSCALE (Parallel Scaling): A Parallel Computation Method for Efficient and Scalable Language Model Deployment

A prominent area of exploration involves enabling large language models (LLMs) to function collaboratively. Multi-agent systems powered by LLMs are now being examined for their potential to coordinate challenging problems by splitting tasks and working simultaneously. This direction has gained attention due to its potential to increase efficiency and reduce latency in real-time applications. A common issue in collaborative LLM systems is agents’ sequential, turn-based communication. In such systems, each agent must wait for others to complete their reasoning steps before proceeding. This slows down processing, especially in situations demanding rapid responses. Moreover, agents often duplicate efforts or generate inconsistent outputs, as they cannot see the evolving thoughts of their peers during generation. This latency and redundancy reduce the practicality of deploying multi-agent LLMs, particularly when time and computation are constrained, such as edge devices. Most current solutions have relied on sequential or independently parallel sampling techniques to improve reasoning. Methods like Chain-of-Thought prompting help models to solve problems in a structured way but often come with increased inference time. Approaches such as Tree-of-Thoughts and Graph-of-Thoughts expand on this by branching reasoning paths. However, these approaches still do not allow for real-time mutual adaptation among agents. Multi-agent setups have explored collaborative methods, but mostly through alternating message exchanges, which again introduces delays. Some advanced systems propose complex dynamic scheduling or role-based configurations, which are not optimized for efficient inference. Research from MediaTek Research introduced a new method called Group Think. This approach enables multiple reasoning agents within a single LLM to operate concurrently, observing each other’s partial outputs at the token level. Each reasoning thread adapts to the evolving thoughts of the others mid-generation. This mechanism reduces duplication and enables agents to shift direction if another thread is better positioned to continue a specific line of reasoning. Group Think is implemented through a token-level attention mechanism that lets each agent attend to previously generated tokens from all agents, supporting real-time collaboration. The method works by assigning each agent its own sequence of token indices, allowing their outputs to be interleaved in memory. These interleaved tokens are stored in a shared cache accessible to all agents during generation. This design allows efficient attention across reasoning threads without architectural changes to the transformer model. The implementation works both on personal devices and in data centers. On local devices, it effectively uses idle compute by batching multiple agent outputs, even with a batch size of one. In data centers, Group Think allows multiple requests to be processed together, interleaving tokens across agents while maintaining correct attention dynamics. Performance tests demonstrate that Group Think significantly improves latency and output quality. In enumeration tasks, such as listing 100 distinct names, it achieved near-complete results more rapidly than conventional Chain-of-Thought approaches. The acceleration was proportional to the number of thinkers; for example, four thinkers reduced latency by a factor of about four. In divide-and-conquer problems, using the Floyd–Warshall algorithm on a graph of five nodes, four thinkers reduced the completion time to half that of a single agent. Group Think solved code generation challenges in programming tasks more effectively than baseline models. With four or more thinkers, the model produced correct code segments much faster than traditional reasoning models. This research shows that existing LLMs, though not explicitly trained for collaboration, can already demonstrate emergent group reasoning behaviors under the Group Think setup. In experiments, agents naturally diversified their work to avoid redundancy, often dividing tasks by topic or focus area. These findings suggest that Group Think’s efficiency and sophistication could be enhanced further with dedicated training on collaborative data. Check out the Paper. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 95k+ ML SubReddit and Subscribe to our Newsletter. NikhilNikhil is an intern consultant at Marktechpost. He is pursuing an integrated dual degree in Materials at the Indian Institute of Technology, Kharagpur. Nikhil is an AI/ML enthusiast who is always researching applications in fields like biomaterials and biomedical science. With a strong background in Material Science, he is exploring new advancements and creating opportunities to contribute.Nikhilhttps://www.marktechpost.com/author/nikhil0980/Researchers from the National University of Singapore Introduce ‘Thinkless,’ an Adaptive Framework that Reduces Unnecessary Reasoning by up to 90% Using DeGRPONikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces MathCoder-VL and FigCodifier: Advancing Multimodal Mathematical Reasoning with Vision-to-Code AlignmentNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces PARSCALE (Parallel Scaling): A Parallel Computation Method for Efficient and Scalable Language Model DeploymentNikhilhttps://www.marktechpost.com/author/nikhil0980/Google AI Releases Standalone NotebookLM Mobile App with Offline Audio and Seamless Source Integration

The effectiveness of language models relies on their ability to simulate human-like step-by-step deduction. However, these reasoning sequences are resource-intensive and can be wasteful for simple questions that do not require elaborate computation. This lack of awareness regarding the complexity of the task is one of the core challenges in these models. They often default to detailed reasoning even for queries that could be answered directly. Such an approach increases token usage, extends response time, and increases system latency and memory usage. As a result, there’s a pressing need to equip language models with a mechanism that allows them to make autonomous decisions about whether to think deeply or respond succinctly. Current tools attempting to solve this issue either rely on manually set heuristics or prompt engineering to switch between short and long responses. Some methods use separate models and route questions based on complexity estimates. Still, these external routing systems often lack insight into the target model’s strengths and fail to make optimal decisions. Other techniques fine-tune models with prompt-based cues like “reasoning on/off,” but these rely on static rules rather than dynamic understanding. Despite some improvements, these approaches fail to enable fully autonomous and context-sensitive control within a single model. Researchers from the National University of Singapore introduced a new framework called Thinkless, which equips a language model with the ability to dynamically decide between using short or long-form reasoning. The framework is built on reinforcement learning and introduces two special control tokens—<short> for concise answers and <think> for detailed responses. By incorporating a novel algorithm called Decoupled Group Relative Policy Optimization (DeGRPO), Thinkless separates the training focus between selecting the reasoning mode and improving the accuracy of the generated response. This design prevents the model from falling into one-dimensional behavior and enables adaptive reasoning tailored to each query. The methodology involves two stages: warm-up distillation and reinforcement learning. In the distillation phase, Thinkless is trained using outputs from two expert models—one specializing in short responses and the other in detailed reasoning. This stage helps the model establish a firm link between the control token and the desired reasoning format. The reinforcement learning stage then fine-tunes the model’s ability to decide which reasoning mode to use. DeGRPO decomposes the learning into two separate objectives: one for training the control token and another for refining the response tokens. This approach avoids the gradient imbalances in earlier models, where longer responses would overpower the learning signal, leading to a collapse in reasoning diversity. Thinkless ensures that both <short> and <think> tokens receive balanced updates, promoting stable learning across response types. When evaluated, Thinkless significantly reduced long-form reasoning while preserving high accuracy. On the Minerva Algebra benchmark, the model used the <think> token in only 25.88% of cases while achieving 94.59% accuracy. In contrast, conventional reasoning models had to use extended chains of thought much more frequently. On the AIME 2024 dataset, Thinkless reached a 27.33% accuracy rate with 100% usage of the reasoning mode, showing that it could maintain performance when full reasoning was necessary. On the GSM8K dataset, it utilized <think> only 13.31% of the time, yet still achieved 84.18% accuracy. These results reflect the model’s ability to handle simple and complex queries with appropriate reasoning depth, cutting down on unnecessary token generation by as much as 90% in some tasks. Overall, this study from the National University of Singapore researchers presents a compelling solution to the inefficiencies of uniform reasoning in large language models. By introducing a mechanism that enables models to judge task complexity and adjust their inference strategy accordingly, Thinkless optimizes both accuracy and efficiency. The method balances depth of reasoning and response precision without relying on fixed rules, offering a data-driven approach to more intelligent language model behavior. Check out the Paper and GitHub Page. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 95k+ ML SubReddit and Subscribe to our Newsletter. NikhilNikhil is an intern consultant at Marktechpost. He is pursuing an integrated dual degree in Materials at the Indian Institute of Technology, Kharagpur. Nikhil is an AI/ML enthusiast who is always researching applications in fields like biomaterials and biomedical science. With a strong background in Material Science, he is exploring new advancements and creating opportunities to contribute.Nikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces MathCoder-VL and FigCodifier: Advancing Multimodal Mathematical Reasoning with Vision-to-Code AlignmentNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces PARSCALE (Parallel Scaling): A Parallel Computation Method for Efficient and Scalable Language Model DeploymentNikhilhttps://www.marktechpost.com/author/nikhil0980/Google AI Releases Standalone NotebookLM Mobile App with Offline Audio and Seamless Source IntegrationNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper from Microsoft Introduces a DiskANN-Integrated System: A Cost-Effective and Low-Latency Vector Search Using Azure Cosmos DB

Multimodal mathematical reasoning enables machines to solve problems involving textual information and visual components like diagrams and figures. This requires combining language understanding and visual interpretation to make sense of complex mathematical contexts. Such capabilities are vital in education, automated tutoring, and document analysis, where problems are often presented with a blend of text and images. A major obstacle in this area is the lack of high-quality, precise alignment between math images and their textual or symbolic representations. Most datasets used to train large multimodal models are derived from image captions in natural settings, which often miss the detailed elements essential for mathematical accuracy. This creates problems for models that rely on these data sources, making them unreliable when dealing with geometry, figures, or technical diagrams. A model’s performance in mathematical reasoning depends heavily on its ability to correctly interpret and link these visual details with mathematical expressions or instructions. In the past, some approaches tried to address this by either enhancing the visual encoders or using manually crafted datasets. However, these methods tend to produce low image diversity, relying on hand-coded or template-based generation, which limits their applicability. Some efforts, like Math-LLaVA and MAVIS, developed synthetic datasets and used templates or predefined categories. Still, they could not dynamically create a wide variety of math-related visuals. This shortfall restricts the learning scope of models and leaves them struggling with more complex or less structured mathematical problems. Researchers from the Multimedia Laboratory at The Chinese University of Hong Kong and CPII under InnoHK introduced a novel approach called MathCoder-VL. This method combines a vision-to-code model named FigCodifier and a synthetic data engine. They constructed the ImgCode-8.6M dataset using a model-in-the-loop strategy, which allowed them to build the largest image-code dataset to date iteratively. Further, they developed MM-MathInstruct-3M, a multimodal instruction dataset enriched with newly synthesized images. The MathCoder-VL model is trained in two stages: mid-training on ImgCode-8.6M to improve visual-text alignment and fine-tuning on MM-MathInstruct-3M to strengthen reasoning abilities. The FigCodifier model works by translating mathematical figures into code that can recreate those figures exactly. This code-image pairing ensures strict alignment and accuracy, unlike caption-based datasets. The process begins with 119K image-code pairs from DaTikZ and expands through iterative training using images collected from textbooks, K12 datasets, and arXiv papers. The final dataset includes 8.6 million code-image pairs and covers various mathematical topics. FigCodifier also supports Python-based rendering, which adds variety to image generation. The system filters low-quality data by checking code validity and removing redundant or unhelpful visuals, resulting in 4.3M high-quality TikZ and 4.3M Python-based pairs. Performance evaluations show that MathCoder-VL outperforms multiple open-source models. The 8B version achieved 73.6% accuracy on the MathVista Geometry Problem Solving subset, surpassing GPT-4o and Claude 3.5 Sonnet by 8.9% and 9.2%, respectively. It also scored 26.1% on MATH-Vision and 46.5% on MathVerse. In Chinese-language benchmarks, it achieved 51.2% on GAOKAO-MM. On the We-Math benchmark, it solved two-step problems at 58.6%, outperforming GPT-4o’s 58.1%. Its performance on three-step problems reached 52.1%, again exceeding GPT-4o’s 43.6%. Compared to its base model InternVL2-8B, it showed gains of 6.1% on MATH-Vision and 11.6% on MathVista. This work clearly defines the problem of insufficient visual-textual alignment in multimodal math reasoning and provides a scalable and innovative solution. The introduction of FigCodifier and synthetic datasets allows models to learn from accurate, diverse visuals paired with exact code, significantly boosting their reasoning abilities. MathCoder-VL represents a practical advancement in this field, demonstrating how thoughtful model design and high-quality data can overcome longstanding limitations in mathematical AI. Check out the Paper and GitHub Page. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 95k+ ML SubReddit and Subscribe to our Newsletter. NikhilNikhil is an intern consultant at Marktechpost. He is pursuing an integrated dual degree in Materials at the Indian Institute of Technology, Kharagpur. Nikhil is an AI/ML enthusiast who is always researching applications in fields like biomaterials and biomedical science. With a strong background in Material Science, he is exploring new advancements and creating opportunities to contribute.Nikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper Introduces PARSCALE (Parallel Scaling): A Parallel Computation Method for Efficient and Scalable Language Model DeploymentNikhilhttps://www.marktechpost.com/author/nikhil0980/Google AI Releases Standalone NotebookLM Mobile App with Offline Audio and Seamless Source IntegrationNikhilhttps://www.marktechpost.com/author/nikhil0980/This AI Paper from Microsoft Introduces a DiskANN-Integrated System: A Cost-Effective and Low-Latency Vector Search Using Azure Cosmos DBNikhilhttps://www.marktechpost.com/author/nikhil0980/LLMs Struggle to Act on What They Know: Google DeepMind Researchers Use Reinforcement Learning Fine-Tuning to Bridge the Knowing-Doing Gap