Robust low-rank training via approximate orthonormal constraints
Robust low-rank training via approximate orthonormal constraints Read Paper As models and datasets grow, pruning techniques using low-rank matrix factorizations have become popular for reducing resource demands while maintaining accuracy. However, we find that these methods often degrade robustness against adversarial attacks due to exploding singular values in the low-rank matrices. To address this, we propose a robust low-rank training algorithm that keeps weights on the low-rank manifold while enforcing approximate orthonormal constraints. This approach reduces both training and inference costs while improving model conditioning and adversarial robustness, all without compromising accuracy. Our theoretical analysis and experimental results confirm that the robust low-rank network closely approximates the performance of full models when effective low-rank sub-networks are available. To address this, we introduce a new error taxonomy and create MMLU-Redux—a refined subset of 3,000 manually re-annotated questions across 30 subjects. Our experiments with MMLU-Redux reveal notable discrepancies in previously reported model performance metrics, underscoring the need for revising MMLU’s flawed questions. We invite the community to contribute to further annotations to enhance the reliability of this important benchmark.
Are We Done with MMLU?
Read Paper Analyzing Flaws in MMLU Our analysis uncovers significant issues with the Massive Multitask Language Understanding (MMLU) benchmark, which is widely used to assess LLMs. We found numerous ground truth errors, with 57% of the Virology subset’s questions containing inaccuracies, obscuring the true capabilities of models. To address this, we introduce a new error taxonomy and create MMLU-Redux—a refined subset of 3,000 manually re-annotated questions across 30 subjects. Our experiments with MMLU-Redux reveal notable discrepancies in previously reported model performance metrics, underscoring the need for revising MMLU’s flawed questions. We invite the community to contribute to further annotations to enhance the reliability of this important benchmark.
Enhancing AI Model Robustness with Natural Language Explanations
Enhancing AI Model Robustness with Natural Language Explanations Read Paper In this paper, we explore how natural language explanations (NLEs) can improve the robustness of large language models (LLMs) in tasks like natural language inference and paraphrase detection. By prompting LLMs with a mix of human-generated and AI-produced NLEs, we observed notable improvements in handling adversarial inputs. Our findings indicate that this method consistently outperforms traditional approaches, offering a more effective way to enhance model accuracy in challenging scenarios.
Probing the Emergence of Cross-lingual Alignment during LLM Training
Probing the Emergence of Cross-lingual Alignment during LLM Training Read Paper Multilingual LLMs excel at zero-shot cross-lingual transfer, likely by aligning languages without parallel sentence supervision. This study uses intrinsic probing to analyze neuron overlap encoding linguistic features, correlating it with transfer performance. By examining BLOOM checkpoints across training steps and model scales, a strong link between neuron overlap and downstream performance is identified. The findings also reveal phases in pre-training where alignment and multilingual abilities degrade, offering new insights into multilingual training dynamics.
Using Natural Language Explanations to Improve Robustness of In-context Learning
Using Natural Language Explanations to Improve Robustness of In-context Learning Read Paper This work explores improving the robustness of LLMs against adversarial inputs by augmenting in-context learning (ICL) with natural language explanations (NLEs). Prompting models to generate NLEs from a small set of human-crafted examples yields better results than zero-shot ICL and using only human-generated NLEs. Evaluated across five LLMs, the approach delivers a 6% improvement on eight adversarial datasets. Additionally, while prompt selection strategies boost ICL on standard tests, they prove less effective for robustness, showing an 8% accuracy drop compared to this method.
SPARSEFIT: Few-shot Prompting with Sparse Fine-tuning
SPARSEFIT: Few-shot Prompting with Sparse Fine-tuning for Jointly Generating Predictions and Natural Language Explanations Read Paper This work introduces SparseFit, a sparse few-shot fine-tuning strategy for generating natural language explanations (NLEs) with pre-trained language models (PLMs). SparseFit uses discrete prompts to jointly generate predictions and NLEs while fine-tuning only 6.8% of the model’s parameters, making it more efficient than full fine-tuning. Tested on three T5 model sizes and four datasets, SparseFit achieves competitive task performance and NLE quality, outperforming other parameter-efficient fine-tuning (PEFT) methods on average in both predictive accuracy and explanation quality.
Miniml – Turning AI Potential into Business Reality
The AI Implementation Gap While 92% of enterprises are investing in AI initiatives, only 31% have successfully moved their projects from pilot to production. This stark “implementation gap” isn’t just a technology challenge—it’s a business problem with significant consequences for competitiveness and growth. Why does this gap exist? Through our work with dozens of SMEs and enterprise clients across the UK, US, and Europe, we’ve identified three critical barriers: The Miniml Approach: Production-First AI At Miniml, we’ve built our entire methodology around solving these implementation challenges. Founded in Edinburgh with operations in San Francisco, our team combines deep technical expertise with practical business acumen. Unlike traditional consultancies that focus primarily on strategy, or development shops that deliver code without business context, Miniml specializes in the critical middle ground: transforming AI potential into operational reality. What We Mean by “Production-First” Production-first means every solution we build is designed from day one to operate in real business environments. This includes: Core Capabilities: Where We Excel Our services span the full AI implementation lifecycle, with particular strength in domains requiring domain-specific knowledge and data security: Custom Large Language Models (LLMs) Generic AI models like ChatGPT have captured public imagination, but businesses often need models that understand their unique terminology, processes, and data. Our custom LLM development creates tailored models that: Intelligent Workflow Automation Many business processes contain repetitive, high-volume tasks that are too complex for traditional automation but perfect for AI-enhanced solutions. Our workflow automation practice: Generative AI for Business Beyond the consumer applications, generative AI offers transformative potential for internal business processes, content creation, and decision support. Our generative AI solutions: Why Organizations Choose Miniml 1. We’re engineers first, consultants second Our founding team comes from engineering backgrounds at leading AI companies and research institutions. We value working solutions over perfect slide decks. 2. We understand regulated industries We’ve built our security practices and development methodology specifically for industries where data protection and regulatory compliance are non-negotiable. 3. We focus on business metrics that matter Every project begins with clear definitions of success tied to operational KPIs—whether cost reduction, throughput improvement, or enhanced customer experience. 4. We bridge technical and operational reality Our teams combine AI expertise with practical business knowledge, ensuring solutions that work within your operational constraints and organizational culture. Starting Your AI Implementation Journey If you’re looking to move beyond AI experimentation to real business impact, we offer several engagement models: Each engagement follows our proven methodology that emphasizes early validation, iterative development, and clear success metrics. Ready to Bridge Your AI Implementation Gap? AI adoption doesn’t have to be high-risk or disruptive. With the right partner, you can move confidently from ambition to implementation—transforming AI potential into business reality. Whether you’re exploring AI for the first time or scaling existing initiatives, we’re ready to help you move forward with confidence. Book a Consultation Miniml is a specialist AI consultancy and development firm headquartered in Edinburgh, Scotland, with operations in San Francisco. We support organizations across the UK, US, and Europe in building and deploying bespoke AI systems that deliver real operational impact.
The Future of Large Language Models (LLMs): Opportunities for Enterprises
What Are LLMs & Why Should Enterprises Care? In a business landscape where technology adoption defines market leadership, Large Language Models (LLMs) have emerged as the most transformative AI technology of the decade. As we witness the rapid evolution of these systems from research curiosities to business-critical tools, forward-thinking enterprises are no longer asking if they should integrate LLMs into their operations, but how and where they’ll deliver the greatest value. Large Language Models are AI systems trained on vast amounts of text data that can recognize patterns and relationships in language. Think of them as having “read” millions of books, websites, documents, and conversations, allowing them to develop a deep understanding of how human language works. Unlike traditional business intelligence tools that require structured data in specific formats, LLMs can work with language as it naturally occurs across your organization—in emails, documents, customer support logs, social media, and more. According to a recent MIT Technology Review report, 71% of enterprises are planning to build their own custom LLMs or other generative AI models. This signals the growing recognition that LLMs represent a new paradigm in how enterprises can process, analyze, and leverage their information assets. How LLMs Transform Enterprise Operations When properly implemented, LLMs serve as cognitive assistants that augment human capabilities across virtually every business function. The potential applications span all departments and functions within an enterprise. Here are the key areas where we’re seeing the most significant impact today: 1. Knowledge Management and Accessibility with LLMs Many enterprises struggle with information siloing—valuable knowledge trapped in documents, systems, or individual employees’ expertise. LLMs can transform how organizations access and leverage their institutional knowledge by: A global professional services firm we worked with at miniml reduced research time by 67% after implementing an LLM-powered knowledge system customized to their proprietary data and domain expertise. This demonstrates how large language models for enterprises can deliver measurable ROI through improved knowledge accessibility. 2. Customer Experience Enhancement Through LLM Implementation Today’s consumers expect personalized, responsive interactions across every touchpoint. Large language models are redefining what’s possible in customer experience through: One financial services client saw a 40% reduction in support ticket escalations after deploying an LLM-powered support system that could understand and respond to complex product questions. This illustrates how enterprise LLM solutions can simultaneously improve customer satisfaction while reducing operational costs. 3. Workflow Automation and Process Intelligence Using LLMs Beyond simple robotic process automation, large language models can transform how complex cognitive tasks are performed: A healthcare provider we partnered with automated 85% of their post-consultation documentation process using a domain-specific LLM, freeing up valuable clinical time while improving consistency. Enterprise LLM implementation in this context demonstrates the potential for significant time savings in document-intensive industries. 4. Innovation Acceleration with Enterprise LLM Solutions Perhaps most importantly, LLMs can accelerate the innovation cycle itself: According to Databricks research, organizations that effectively implement large language models see a marked improvement in their innovation pipelines, with new ideas moving from concept to implementation significantly faster. How to Implement LLMs: Navigating Enterprise Challenges For all their potential, implementing LLMs effectively involves addressing several important challenges. Here’s how to approach large language model implementation for enterprise use cases: Data Security and Governance for Enterprise LLMs Enterprise data is both valuable and sensitive. Using public LLM services like ChatGPT can create risks when proprietary information is involved. For many organizations, the solution lies in: Research from Master of Code Global indicates that 63.5% of enterprises cite data security and compliance as primary concerns when adopting large language models. This underscores the importance of a thoughtful approach to LLM implementation that prioritizes data protection. Addressing LLM “Hallucination” Challenges in Enterprise Settings LLMs can occasionally generate plausible-sounding but incorrect information—what AI researchers call “hallucinations.” Mitigating this risk requires: Our work at miniml has shown that domain-specific training data can reduce hallucination rates by up to 78% compared to general-purpose models, making enterprise LLM implementation more reliable and trustworthy. Integration of Large Language Models with Enterprise Systems Meaningful LLM implementation isn’t just about the models themselves but how they connect to existing systems and workflows: The Future of LLMs: Enterprise Outlook 2025-2028 As we look toward the next 3-5 years, several trends will shape how enterprises leverage LLMs: 1. From General to Domain-Specific Enterprise LLMs While general-purpose LLMs like GPT-4 have captured headlines, the real business value will increasingly come from models fine-tuned for specific industries, functions, and even individual enterprises. We’ll see the rise of specialized models for healthcare, finance, legal, manufacturing, and other sectors that incorporate domain-specific knowledge and terminology. 2. The Integration of Structured and Unstructured Data in LLM Applications Future large language model systems will increasingly bridge the gap between traditional structured data (like databases) and unstructured information (like documents and conversations). This will enable more powerful analytics and automation capabilities that leverage all enterprise information assets. 3. Multi-Modal Capabilities in Enterprise Large Language Models The next generation of language models will work seamlessly across text, images, audio, and video, enabling new applications in areas like visual inspection, multimedia content analysis, and complex document processing. Enterprise LLM implementation will expand beyond text to include all forms of business communication. 4. Enhanced Reasoning Capabilities for Enterprise Decision Support LLMs will continue to improve in logical reasoning, planning, and problem-solving, moving beyond pattern recognition to more sophisticated forms of analysis that can support complex decision-making. This evolution will make large language models increasingly valuable for strategic business applications. 5. Democratized Enterprise LLM Development The tools for customizing and deploying large language models will become increasingly accessible to business users without deep technical expertise, accelerating adoption across the enterprise. This democratization will expand the impact of LLMs beyond technical teams to all business functions. How to Implement Large Language Models: A Strategic Approach As with any transformative technology, the key to success with LLMs lies in thoughtful, strategic implementation rather than rushing to adopt the latest tools. The organizations seeing the greatest impact today are taking a
CarePoint and Miniml AI Join Forces to Revolutionize Healthcare Access in Africa
Accra, Ghana and Edinburgh, Scotland– 03/07/2025 — CarePoint, a leading healthcare provider committed to democratizing access to quality healthcare across Africa, has entered into a strategic partnership with Miniml, an AI company specializing in custom solutions across diverse sectors. This collaboration aims to leverage artificial intelligence to improve healthcare accessibility and quality across the continent. It represents a major step toward using AI to address the unique healthcare needs of African communities. Under this partnership, CarePoint and Miniml will work together to develop and implement AI-powered solutions tailored to support CarePoint’s growing network of healthcare facilities across Africa. Leveraging Miniml’s expertise in secure, scalable AI, the collaboration aims to enhance CarePoint’s operational efficiency, bringing quality healthcare within reach for millions of people. “At CarePoint, we are dedicated to transforming healthcare across Africa by making it more accessible and efficient for the communities we serve,” said Sangu Delle, CEO of CarePoint. “Our partnership with Miniml represents an exciting new chapter in this journey. We aim to enhance patient care, streamline operations, and improve health outcomes by integrating cutting-edge AI solutions into our operations. We look forward to the impact this collaboration will have on healthcare delivery throughout the continent.” John Westcott, CEO of Miniml, echoed this enthusiasm, stating, “We are thrilled to partner with CarePoint and contribute to their visionary mission of making high-quality healthcare accessible across Africa. This collaboration allows us to apply our AI expertise to one of the world’s most pressing challenges — delivering effective healthcare in underserved regions. Together, we are committed to driving meaningful improvements that will empower healthcare providers and transform patient care.” The collaboration will initially focus on developing AI-driven tools to address specific healthcare challenges within CarePoint’s facilities. These include improving operational efficiency, enhancing data accuracy, and supporting healthcare providers with actionable insights for routine care. With Miniml’s advanced AI capabilities, the partnership aims to deliver scalable solutions that are adaptable to CarePoint’s diverse healthcare environments, expanding healthcare access for communities that need it most. About CarePoint CarePoint is a technology-driven healthcare company focused on building accessible, high-quality healthcare systems across Africa. Through a network of healthcare facilities in Nigeria, Ghana, and Egypt, CarePoint leverages technology to make person-centered healthcare accessible to millions. About Miniml Headquartered in Edinburgh, Miniml develops custom AI solutions across diverse sectors, addressing complex operational and decision-making challenges. Known for its innovative research on AI reliability and security, Miniml delivers scalable, impact-driven tools that enhance capabilities and support informed outcomes. With expertise in secure, flexible deployments, Miniml empowers organizations to achieve lasting improvements with AI.