A new review by researchers from the Beijing University of Posts and Telecommunications, CETC 54 (54th Research Institute of Electronics Technology Group Corporation), Sun Yat-sen University, Shenzhen University, and the University of Electronic Science and Technology of China surveys the latest developments in 3D printing for microelectronic and microfluidic applications. The paper released on Springer Nature Link highlights how additive manufacturing methods have reached sub-micron precision, allowing the production of devices previously limited to traditional cleanroom fabrication. High-resolution techniques like two-photon polymerization (2PP), electrohydrodynamic jet printing, and computed axial lithography (CAL) are now being used to create structures with feature sizes down to 100 nanometers. These capabilities have broad implications for biomedical sensors, flexible electronics, and microfluidic systems used in diagnostics and environmental monitoring. Overview of 3D printing applications for microelectronic and microfluidic device fabrication. Image via Springer Nature. Classification of High-Precision Additive Processes Seven categories of additive manufacturing, as defined by the American Society for Testing and Materials (ASTM) serve as the foundation for modern 3D printing workflows: binder jetting, directed energy deposition (DED), material extrusion (MEX), material jetting, powder bed fusion (PBF), sheet lamination (SHL), and vat photopolymerization (VP). Among these, 2PP provides the finest resolution, enabling the fabrication of nanoscale features for optical communication components and MEMS support structures. Inkjet-based material jetting and direct ink writing (DIW) allow patterned deposition of conductive or biological materials, including stretchable gels and ionic polymers. Binder jetting, which operates by spraying adhesives onto powdered substrates, is particularly suited for large-volume structures using metals or ceramics with minimal thermal stress. Fused deposition modeling, a form of material extrusion, continues to be widely used for its low cost and compatibility with thermoplastics. Although limited in resolution, it remains practical for building mechanical supports or sacrificial molds in soft lithography. Various micro-scale 3D printing strategies. Image via Springer Nature. 3D Printing in Microelectronics, MEMS, and Sensing Additive manufacturing is now routinely used to fabricate microsensors, microelectromechanical system (MEMS) actuators, and flexible electronics. Compared to traditional lithographic processes, 3D printing reduces material waste and bypasses the need for masks or etching steps. In one example cited by the review, flexible multi-directional sensors were printed directly onto skin-like substrates using a customized FDM platform. Another case involved a cantilever support for a micro-accelerometer produced via 2PP and coated with conductive materials through evaporation. These examples show how additive techniques can fabricate both support and functional layers with high geometric complexity. MEMS actuators fabricated with additive methods often combine printed scaffolds with conventional micromachining. A 2PP-printed spiral structure was used to house liquid metal in an electrothermal actuator. Separately, FDM was used to print a MEMS switch, combining conductive PLA and polyvinyl alcohol as the sacrificial layer. However, achieving the mechanical precision needed for switching elements remains a barrier for fully integrated use. 3D printing material and preparation methods. Image via Springer Nature. Development of Functional Inks and Composite Materials Microelectronic applications depend on the availability of printable materials with specific electrical, mechanical, or chemical properties. MXene-based conductive inks, metal particle suspensions, and piezoelectric composites are being optimized for use in DIW, inkjet, and light-curing platforms. Researchers have fabricated planar asymmetric micro-supercapacitors using ink composed of nickel sulfide on nitrogen-doped MXene. These devices demonstrate increased voltage windows (up to 1.5 V) and volumetric capacitance, meeting the demands of compact power systems. Other work involves composite hydrogels with ionic conductivity and high tensile stretch, used in flexible biosensing applications. PEDOT:PSS, a common conductive polymer, has been formulated into a high-resolution ink using lyophilization and re-dispersion in photocurable matrices. These formulations are used to create electrode arrays for neural probes and flexible circuits. Multiphoton lithography has also been applied to print complex 3D structures from organic semiconductor resins. Bioelectronic applications are driving the need for biocompatible inks that can perform reliably in wet and dynamic environments. One group incorporated graphene nanoplatelets and carbon nanotubes into ink for multi-jet fusion, producing pressure sensors with high mechanical durability and signal sensitivity. 3D printed electronics achieved through the integration of active initiators into printing materials. Image via Springer Nature. Microfluidic Devices Fabricated via Direct and Indirect Methods Microfluidic systems have traditionally relied on soft lithography techniques using polydimethylsiloxane (PDMS). Additive manufacturing now offers alternatives through both direct printing of fluidic chips and indirect fabrication using 3D printed molds. Direct fabrication using SLA, DLP, or inkjet-based systems allows the rapid prototyping of chips with integrated reservoirs and channels. However, achieving sub-100 µm channels requires careful calibration. One group demonstrated channels as small as 18 µm × 20 µm using a customized DLP printer. Indirect fabrication relies on printing sacrificial or reusable molds, followed by casting and demolding. PLA, ABS, and resin-based molds are commonly used, depending on whether water-soluble or solvent-dissolvable materials are preferred. These techniques are compatible with PDMS and reduce reliance on photolithography equipment. Surface roughness and optical transparency remain concerns. FDM-printed molds often introduce layer artifacts, while uncured resin in SLA methods can leach toxins or inhibit PDMS curing. Some teams address these issues by polishing surfaces post-print or chemically treating molds to improve release characteristics. Integration and Future Directions for Microdevices 3D printed microfluidic devices in biology and chemistry.Image via Springer Nature. 3D printing is increasingly enabling the integration of structural, electrical, and sensing components into single build processes. Multi-material printers are beginning to produce substrates, conductive paths, and dielectric layers in tandem, although component embedding still requires manual intervention. Applications in wearable electronics, flexible sensors, and soft robotics continue to expand. Stretchable conductors printed onto elastomeric backings are being used to simulate mechanoreceptors and thermoreceptors for electronic skin systems. Piezoelectric materials such as BaTiO₃-PVDF composites are under investigation for printed actuators and energy harvesters. MEMS fabrication remains constrained by the mechanical limitations of printable materials. Silicon continues to dominate high-performance actuators due to its stiffness and precision. Additive methods are currently better suited for producing packaging, connectors, and sacrificial scaffolds within MEMS systems. Multi-photon and light-assisted processes are being explored for producing active devices like microcapacitors and accelerometers. Recent work demonstrated the use of 2PP to fabricate nitrogen-vacancy center–based quantum sensors, capable of detecting thermal and magnetic fluctuations in microscopic environments. As materials, resolution, and system integration improve, 3D printing is poised to shift from peripheral use to a central role in microsystem design and production.  3D printing micro-nano devices. Image via Springer Nature. Ready to discover who won the 20243D Printing Industry Awards? Subscribe to the 3D Printing Industry newsletter to stay updated with the latest news and insights. Take the 3DPI Reader Survey — shape the future of AM reporting in under 5 minutes. Featured image shows an Overview of 3D printing applications for microelectronic and microfluidic device fabrication. Image via Springer Nature. Anyer Tenorio Lara Anyer Tenorio Lara is an emerging tech journalist passionate about uncovering the latest advances in technology and innovation. With a sharp eye for detail and a talent for storytelling, Anyer has quickly made a name for himself in the tech community. Anyer's articles aim to make complex subjects accessible and engaging for a broad audience. In addition to his writing, Anyer enjoys participating in industry events and discussions, eager to learn and share knowledge in the dynamic world of technology.

Designing human-centered AI products can be arduous.Keeping up with the overall pace of change isn’t easy. But here’s a bigger challenge:The wildly different paces of change attached to the key elements of AI product strategy, design, and development can make managing those elements — and even thinking about them — overwhelming.Yesterday’s design processes and frameworks offer priceless guidance that still holds. But in many spots, they just don’t fit today’s environment.For instance, designers used to map out and user-test precise, predictable end-to-end screen flows. But flows are no longer precisely predictable. AI generates dynamic dialogues and custom-tailored flows on the fly, rendering much of the old practice unhelpful and infeasible.It’s easy for product teams to feel adrift nowadays — we can hoist the sails, but we’re missing a map and a rudder. We need frameworks tailored to the traits that fundamentally set AI apart from traditional software, including:its capabilities for autonomy and collaboration,its probabilistic nature,its early need for quality data, andits predictable unpredictability. Humans tend to be perpetually surprised by its abilities — and its inabilities.AI pace layers: design for resilienceHere’s a framework to address these challenges.Building on Stewart Brand’s “Shearing Layers” framework, AI Pace Layers helps teams grow thriving AI products by framing them as layered systems with components that function and evolve at different timescales.It helps anticipate points of friction and create resilient and humane products.Each layer represents a specific domain of activity and responsibility, with a distinct pace of change.* Unlike the other layers, Services cuts across multiple layers rather than sitting between them, and its pace of change fluctuates erratically.Boundaries between layers call for special attention and care — friction at these points can produce destructive shearing and constructive turbulence.I’ll dive deeper into this framework with some practical examples showing how it works. But first, a brief review of the precursors that inspired this framework will help you put it to good use.The foundationsThis model builds on the insights of several influential design frameworks from the professions of building architecture and traditional software design.Shearing layers (Duffy and Brand)In his 1994 book How Buildings Learn, Stewart Brand expanded on architect Frank Duffy’s concept of shearing layers. The core insight: buildings consist of components that change at different rates.Shell, Services, Scenery, and Sets. (Frank Duffy, 1992).“…there isn’t any such thing as a building. A building properly conceived is several layers of longevity of built components.” — Frank DuffyShearing Layers of Change, from How Buildings Learn: What Happens after they’re built (Stewart Brand, 1994).Expanding on Duffy’s work, Brand identified six layers, from the slow-changing “Site” to the rapidly evolving “Stuff.”As the layers move at different speeds, friction forms where they meet. Buildings designed without mindful consideration of these different velocities tear themselves apart at these “shearing” points. Before long, they tend to be demolished and replaced.Buildings designed for resiliency allow for “slippage” between the moving layers — flexibility for the different rates of change to unfold with minimal conflict. Such buildings can thrive and remain useful for hundreds of years.Pace layers (Brand)In 1999, Brand drew insights from ecologists to expand this concept beyond buildings and encompass human society. In The Clock Of The Long Now: Time And Responsibility, he proposed “Pace Layers” — six levels ranging from rapid fashion to glacially-slow nature.Brand’s Pace Layers (1999) as sketched by Jono Hey.Brand again pointed out the boundaries, where the most intriguing and consequential changes emerge. Friction at the tension points can tear a building apart — or spur a civilization’s collapse–when we try to bind the layers too tightly together. But with mindful design and planning for slippage, activity along these boundary zones can also generate “constructive turbulence” that keeps systems balanced and resilient.The most successful systems survive and thrive through times of change through resiliency, by absorbing and incorporating shocks.“…a few scientists (such as R. V. O’Neill and C. S. Holling) have been probing the same issue in ecological systems: how do they manage change, how do they absorb and incorporate shocks? The answer appears to lie in the relationship between components in a system that have different change-rates and different scales of size. Instead of breaking under stress like something brittle, these systems yield as if they were soft. Some parts respond quickly to the shock, allowing slower parts to ignore the shock and maintain their steady duties of system continuity.” — Stewart BrandRoles and tendencies of the fast (upper) and slow (lower) layers. (Brand).Slower layers provide constraints and underpinnings for the faster layers, while faster layers induce adaptations in the slower layers that evolve the system.Elements of UX (Garrett)Jesse James Garrett’s classic The Elements of User Experience (2002) presents a five-layer model for digital design:Surface (visual design)Skeleton (interface design, navigation design, information design)Structure (interaction design, information architecture)Scope (functional specs, content requirements)Strategy (user needs, site objectives)Structure, Scope, and Strategy. Each layer answers a different set of questions, with the questions answered at each level setting constraints for the levels above. Lower layers set boundaries and underpinnings that help define the more concrete layers.Jesse James Garrett’s 5 layers from The Elements of User Experience Design (2002)This framework doesn’t focus on time, or on tension points resulting from conflicting velocities. But it provides a comprehensive structure for shaping different aspects of digital product design, from abstract strategy to concrete surface elements.AI Pace Layers: diving deeperBuilding on these foundations, the AI Pace Layers framework adapts these concepts specifically for AI systems design.Let’s explore each layer and understand how design expertise contributes across the framework.SessionsPace of change: Very fast (milliseconds to minutes)Focus: Performance of real-time interactions.This layer encompasses real-time dialogue, reasoning, and processing. These interplays happen between the user and AI, and between AI agents and other services and people, on behalf of the user. Sessions draw on lower-layer capabilities and components to deliver the “moments of truth” where product experiences succeed or fail. Feedback from the Sessions layer is crucial for improving and evolving the lower layers.Key contributors: Users and AI agents — usually with zero direct human involvement backstage.Example actions/decisions/artifacts: User/AI dialogue. Audio, video, text, images, and widgets are rendered on the fly (using building blocks provided by lower levels). Real-time adaptations to context.SkinPace of change: Moderately fast (days to months)Focus: Design patterns, guidelines, and assetsSkin encompasses visual, interaction, and content design.Key contributors: Designers, content strategists, front-end developers, and user researchers.Design’s role: This is where designers’ traditional expertise shines. They craft the interface elements, establish visual language, define interaction patterns, and create the design systems that represent the product’s capabilities to users.Example actions/decisions/artifacts: UI component libraries, brand guidelines, prompt templates, tone of voice guidelines, navigation systems, visual design systems, patterns (UI, interaction, and conversation), content style guides.ServicesPace of change: Wildly variable (slow to moderately fast)Focus: AI computation capabilities, data systems orchestration, and operational intelligenceThe Services layer provides probabilistic AI capabilities that sometimes feel like superpowers — and like superpowers, they can be difficult to manage. It encompasses foundation models, algorithms, data pipelines, evaluation frameworks, business logic, and computing resources.Services is an outlier that behaves differently from the other layers:• It’s more prone to “shocks” and surprises that can ripple across the rest of the system.• It varies wildly in pace of change. (But its components rarely change faster than Skin, or slower than Skeleton.)• It cuts across multiple layers rather than sitting between two of them. That produces more cross-layer boundaries, more tension points, more risks of destructive friction, and more opportunities for constructive turbulence.Key contributors: Data scientists, engineers, service designers, ethicists, product teamsDesign’s role: Designers partner with technical teams on evaluation frameworks, helping define what “good” looks like from a human experience perspective. They contribute to guardrails, monitoring systems, and multi-agent collaboration patterns, ensuring technical capabilities translate to meaningful human experiences. Service design expertise helps orchestrate complex, multi-touchpoint AI capabilities.Example actions/decisions/artifacts: Foundation model selection, changes, and fine-tuning. Evals, monitoring systems, guardrails, performance metrics. Business rules, workflow orchestration. Multiagent collaboration and use of external tools (APIs, A2A, MCP, etc.) Continual appraisal and adoption of new tools, protocols, and capabilities.SkeletonPace of change: Moderately slow (months) Focus: Fundamental structure and organizationThis layer establishes the foundational architecture — the core interaction models, information architecture and organizing principles.Key contributors: Information architects, information designers, user researchers, system architects, engineersDesign’s role: Designers with information architecture expertise are important in this layer. They design taxonomies, knowledge graphs, and classification systems that make complex AI capabilities comprehensible and usable. UX researchers help ensure these structures fit the audience’s mental models, contexts, and expectations.Example actions/decisions/artifacts: Taxonomies, knowledge graphs, data models, system architecture, classification systems.ScopePace of change: Slow (months to years)Focus: Product requirementsThis layer defines core functional, content, and data requirements, accounting for the probabilistic nature of AI and defining acceptable levels of performance and variance.Key contributors: Product managers, design strategists, design researchers, business stakeholders, data scientists, trust & safety specialistsDesign’s role: Design researchers and strategists contribute to requirements through generative and exploratory research. They help define error taxonomies and acceptable failure modes from a user perspective, informing metrics that capture technical performance and human experience quality. Design strategists balance technical possibilities with human needs and ethical considerations.Example actions/decisions/artifacts: Product requirements documents specifying reliability thresholds, data requirements (volume, diversity, quality standards), error taxonomies and acceptable failure modes, performance metrics frameworks, responsible AI requirements, risk assessment, core user stories and journeys, documentation of expected model variance and handling approaches.StrategyPace of change: Very slow (years)Focus: Long-term vision and business goalsThis foundation layer defines audience needs, core problems to solve, and business goals. In AI products, data strategy is central.Key contributors: Executive leadership, design leaders, product leadership, business strategists, ethics boardsDesign’s role: Design leaders define problem spaces, identify opportunities, and plan roadmaps. They deliver a balance of business needs with human values in strategy development. Designers with expertise in responsible AI help establish ethical frameworks and guiding principles that shape all other layers.Example actions/decisions/artifacts: Problem space and opportunity assessments, market positioning documents, long-term product roadmaps, comprehensive data strategy planning, user research findings on core needs, ethical frameworks and guiding principles, business model documentation, competitive/cooperative AI ecosystem mapping.Practical examples: tension points between layersTension point example 1: Bookmuse’s timeline troubles(Friction between Sessions and Skin)Bookmuse is a promising new AI tool for novelists. Samantha, a writer, tries it out while hashing out the underpinnings of her latest time-travel historical fiction thriller. The Bookmuse team planned for plenty of Samantha’s needs. At first, she considers Bookmuse a handy assistant. It supplements chats with tailored interactive visualizations that efficiently track character personalities, histories, relationships, and dramatic arcs.But Samantha is writing a story about time travelers interfering with World War I events, so she’s constantly juggling dates and timelines. Bookmuse falls short. It’s a tiny startup, and Luke, the harried cofounder who serves as a combination designer/researcher/product manager, hasn’t carved out any date-specific timeline tools or date calculators. He forgot to provide even a basic date picker in the design system.Problem: Bookmuse does its best to help Samantha with her story timeline (Sessions layer). But it lacks effective tools for the job (Skin layer). Its date and time interactions feel confusing, clumsy, and out of step with the rest of its tone, look, and feel. Whenever Samantha consults the timeline, it breaks her out of her creative flow.Constructive turbulence opportunities:a) Present feedback mechanisms that ensure this sort of “missing piece” event results in the product team learning about the type of interaction pothole that appeared — without revealing details or content that compromise Samantha’ privacy and her work. (For instance, a session tagging system can flag all interaction dead-ends during date choice interactions.)b) Improve timeline/date UI and interaction patterns. Table stakes: Standard industry-best-practice date picker components that suit Bookmuse’s style, tone, and voice. Game changers: Widgets, visualizations, and patterns tailored to the special time-tracking/exploration challenges that fiction writers often wrestle with.c) Update the core usability heuristics and universal interaction design patterns baked into the evaluation frameworks (in the Services layer), as part of regular eval reviews and updates. Result: When the team learns about a friction moment like this, they can prevent a host of future similar issues before they emerge.These improvements will make Bookmuse more resilient and useful.Tension point example 2: MedicalMind’s diagnostic dilemma(Friction between Services and Skeleton)Thousands of healthcare providers use MedicalMind, an AI-powered clinical decision support tool. Dr. Rina Patel, an internal medicine physician at a busy community hospital, relies on it to stay current with rapidly evolving medical research while managing her patient load.Thanks to a groundbreaking update, a MedicalMind AI model (Services layer) is familiar with new medical research data and can recognize newly discovered connections between previously unrelated symptoms across different medical specialties. For example, it identified patterns linking certain dermatological symptoms to early indicators of cardiovascular issues — connections not yet widely recognized in standard medical taxonomies.But MedicalMind’s information architecture (Skeleton layer) was tailored to traditional medical classification systems, so it’s organized by body system, conditions by specialty, and treatments by mechanism of action. The MedicalMind team constructed this structure based on how doctors were traditionally trained to approach medical knowledge.Problem: When Dr. Patel enters a patient’s constellation of symptoms (Sessions layer), MedicalMind’s AI can recognize potentially valuable cross-specialty patterns (Services layer). But these insights can’t be optimally organized and presented because the underlying information architecture (Skeleton layer) doesn’t easily accommodate the new findings and relationships. The AI either forces the insights into ill-fitting categories or presents them as disconnected “additional notes” that tend to be overlooked. That reduces their clinical utility and Dr. Patel’s trust in the system.Constructive turbulence opportunities:a) Create an “emerging patterns” framework within the information architecture (Skeleton layer) that can accommodate new AI-identified patterns in ways that augment, rather than disrupt, the familiar classification systems that doctors rely on.b) Design flexible visualization components and interaction patterns and styles (in the Skin layer) specifically for exploring, discussing, and documenting cross-category relationships. Let doctors toggle between traditional taxonomies and newer, AI-generated knowledge maps depending on their needs and comfort level.c) Implement a clinician feedback loop where specialists can validate and discuss new AI-surfaced relationships, gradually promoting validated patterns into the main classification system.These improvements will make MedicalMind more adaptive to emerging medical knowledge while maintaining the structural integrity that healthcare professionals rely on for critical decisions. This provides more efficient assistants for clinicians and better health for patients.Tension point example 3: ScienceSeeker’s hypothesis bottleneck(Friction between Skin and Services)ScienceSeeker is an AI research assistant used by scientists worldwide. Dr. Elena Rodriguez, a molecular biologist, uses it to investigate protein interactions for targeted cancer drug delivery.The AI engine (Services layer) recently gained the ability to generate sophisticated hypothesis trees with multiple competing explanations, track confidence levels for each branch, and identify which experiments would most efficiently disambiguate between theories. It can reason across scientific domains, connecting molecular biology with physics, chemistry, and computational modeling.But the interface (Skin layer) remains locked in a traditional chatbot paradigm — a single-threaded exchange with responses appearing sequentially in a scrolling window.Problem: The AI engine and the problem space are natively multithreaded and multimodal, but the UI is limited to single-threaded conversation. When Dr. Rodriguez inputs her experimental results (Sessions layer), the AI generates a rich, multidimensional analysis (Services layer), but must flatten this complex reasoning into linear text (Skin layer). Critical relationships between hypotheses become buried in paragraphs, probability comparisons are difficult, and the holistic picture of how variables influence multiple hypotheses is lost. Dr. Rodriguez resorts to taking screenshots and manually drawing diagrams to reconstruct the reasoning that the AI possesses but cannot visually express.Constructive turbulence opportunities:a) Develop an expandable, interactive, infinite-canvas “hypothesis tree” visualization (Skin) that helps the AI dynamically represent multiple competing explanations and their relationships. Scientists can interact with this to explore different branches spatially rather than sequentially.b) Create a dual-pane interface that maintains the chat for simple queries but provides the infinite canvas for complex reasoning, transitioning seamlessly based on response complexity.c) Implement collaborative, interactive node-based diagrams for multi-contributor experiment planning, where potential experiments appear as nodes showing how they would affect confidence in different hypothesis branches.This would transform ScienceSeeker’s limited text assistant into a scientific reasoning partner. It would help researchers visualize and interact with complex possibilities in ways that better fit how they tackle multidimensional problems.Navigating the future with AI Pace LayersAI Pace Layers offers product teams a new framework for seeing and shaping the bewildering structures and dynamics that power AI products.By recognizing the evolving layers and heeding and designing for their boundaries, AI design teams can:Transform tension points into constructive innovationAnticipate friction before it damages the product experienceGrow resilient and humane AI systems that absorb and integrate rapid technological change without losing sight of human needs.The framework’s value isn’t in rigid categorization, but in recognizing how components interact across timescales. For AI product teams, this awareness enables more thoughtful design choices that prevent destructive shearing that can tear apart an AI system.This framework is a work in progress, evolving alongside the AI landscape it describes.I’d love to hear from you, especially if you’ve built successful AI products and have insights on how this model could better reflect your experience. Please drop me a line or add a comment. Let’s develop more effective approaches to creating AI systems that enhance human potential while respecting human agency.Part of the Mindful AI Design series. Also see:The effort paradox in AI design: Why making things too easy can backfireBlack Mirror: “Override”. Dystopian storytelling for humane AI designStay updatedSubscribe to be notified when new articles in the series are published. Join our community of designers, product managers, founders and ethicists as we shape the future of mindful AI design.AI Pace Layers: a framework for resilient product design was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.

Large Reasoning Models (LRMs) like OpenAI’s o1 and o3, DeepSeek-R1, Grok 3.5, and Gemini 2.5 Pro have shown strong capabilities in long CoT reasoning, often displaying advanced behaviors such as self-correction, backtracking, and verification—collectively known as “aha moments.” These behaviors have been observed to emerge through outcome-driven RL without the need for supervised fine-tuning. Models like DeepSeek-R1 and its open-source replications (e.g., TinyZero and Logic-RL) have demonstrated that carefully designed RL pipelines—using rule-based rewards, curriculum learning, and structured training—can induce such reflective reasoning abilities. However, these emergent behaviors tend to be unpredictable and inconsistent, limiting their practical reliability and scalability. To address this, researchers have explored structured RL frameworks that target specific reasoning types, such as deduction, abduction, and induction. These approaches involve aligning specialist models, merging them in parameter space, and applying domain-specific continual RL. Tools like Logic-RL use rule-conditioned RL to solve logic puzzles, improving transferability to tasks like math reasoning. Meanwhile, other works propose mechanisms to enhance reasoning robustness, such as training models to reason both forwards and backwards, or iteratively self-critiquing their outputs. Studies analyzing “aha moments” suggest that these behaviors stem from internal shifts in uncertainty, latent representation, and self-assessment, offering new insights into engineering more reliable reasoning models.  Researchers from the National University of Singapore, Tsinghua University, and Salesforce AI Research address the limitations of relying on spontaneous “aha moments” in large language models by explicitly aligning them with three core reasoning abilities: deduction, induction, and abduction. They introduce a three-stage pipeline—individual meta-ability alignment, parameter-space merging, and domain-specific reinforcement learning—significantly enhancing model performance. Using a programmatically generated, self-verifiable task suite, their approach boosts accuracy over instruction-tuned baselines by over 10%, with further gains from domain-specific RL. This structured alignment framework offers a scalable, generalizable method for improving reasoning across math, coding, and science domains.  The researchers designed tasks aligned with deduction, induction, and abduction by using a structured “given two, infer the third” format based on hypothesis (H), rule (R), and observation (O). Deduction is framed as satisfiability checking, induction as masked-sequence prediction, and abduction as reverse rule-graph inference. These tasks are synthetically generated and automatically verified. The training pipeline includes three stages: (A) independently training models for each reasoning type using REINFORCE++ with structured rewards, (B) merging models through weighted parameter interpolation, and (C) fine-tuning the unified model on domain-specific data via reinforcement learning, isolating the benefit of meta-ability alignment.  The study evaluates models aligned with meta-abilities—deduction, induction, and abduction—using a curriculum learning setup across difficulty levels. Models trained on synthetic tasks strongly generalize to seven unseen math, code, and science benchmarks. At both 7B and 32B scales, meta-ability–aligned and merged models consistently outperform instruction-tuned baselines, with the merged model offering the highest gains. Continued domain-specific RL from these merged checkpoints (Domain-RL-Meta) leads to further improvements over standard RL finetuning (Domain-RL-Ins), especially in math benchmarks. Overall, the alignment strategy enhances reasoning abilities, and its benefits scale with model size, significantly boosting performance ceilings across tasks.  In conclusion, the study shows that large reasoning models can develop advanced problem-solving skills without depending on unpredictable “aha moments.” By aligning models with three core reasoning abilities—deduction, induction, and abduction—using self-verifiable tasks, the authors create specialist agents that can be effectively combined into a single model. This merged model outperforms instruction-tuned baselines by over 10% on diagnostic tasks and up to 2% on real-world benchmarks. When used as a starting point for domain-specific reinforcement learning, it raises performance by another 4%. This modular, systematic training approach offers a scalable and controllable foundation for building reliable, interpretable reasoning systems.  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. Sana HassanSana Hassan, a consulting intern at Marktechpost and dual-degree student at IIT Madras, is passionate about applying technology and AI to address real-world challenges. With a keen interest in solving practical problems, he brings a fresh perspective to the intersection of AI and real-life solutions.Sana Hassanhttps://www.marktechpost.com/author/sana-hassan/RXTX: A Machine Learning-Guided Algorithm for Efficient Structured Matrix MultiplicationSana Hassanhttps://www.marktechpost.com/author/sana-hassan/From Protocol to Production: How Model Context Protocol (MCP) Gateways Enable Secure, Scalable, and Seamless AI Integrations Across EnterprisesSana Hassanhttps://www.marktechpost.com/author/sana-hassan/Researchers from Renmin University and Huawei Propose MemEngine: A Unified Modular AI Library for Customizing Memory in LLM-Based AgentsSana Hassanhttps://www.marktechpost.com/author/sana-hassan/Meta Introduces KernelLLM: An 8B LLM that Translates PyTorch Modules into Efficient Triton GPU Kernels

A man tore his windpipe, in part, due to hay fever.

Our latest and most advanced technologies — from AI to Industrial IoT, advanced robotics, and self-driving cars — share serious problems: massive energy consumption, limited on-edge capabilities, system hallucinations, and serious accuracy gaps.  One possible solution is emerging in the Netherlands. The country is developing a promising ecosystem for neuromorphic computing, which draws on neuroscience to boost IT efficiencies and performance. Billions of euros are being invested in this new form of computing worldwide. The Netherlands aims to become a leader in the market by bringing together startups, established companies, government organisations, and academics in a neuromorphic computing ecosystem. A Dutch mission to the UK In March, a Dutch delegation landed in the UK to host an “Innovation Mission” with local tech and government representatives. Top Sector ICT, a Dutch government–supported organisation, led the mission, which sought to strengthen and discuss the future of neuromorphic computing in Europe and the Netherlands.  We contacted Top Sector ICT, who connected us with one of their collaborators: Dr Johan H. Mentink, an expert in computational physics at Radboud University in the Netherlands. Dr Mentink spoke about how neuromorphic computing can solve the energy, accuracy, and efficiency challenges of our current computing architectures.  Grab that deal “Current digital computers use power-hungry processes to handle data,” Dr Mentink said.  “The result is that some modern data centres use so much energy that they even need their own power plant.”  Computing today stores data in one place (memory) and processes it in another place (processors). This means that a lot of energy is spent on transporting data, Dr Mentink explained.  In contrast, neuromorphic computing architectures are different at the hardware and software levels. For example, instead of using processors and memories, neuromorphic systems leverage new hardware components such as memristors. These act as both memory and processors.  By processing and saving data on the same hardware component, neuromorphic computing removes the energy-intensive and error-prone task of transporting data. Additionally, because data is stored on these components, it can be processed more immediately, resulting in faster decision-making, reduced hallucinations, improved accuracy, and better performance. This concept is being applied to edge computing, Industrial IoT, and robotics to drive faster real-time decision-making.  “Just like our brains process and store information in the same place, we can make computers that would combine data storage and processing in one place,” Dr Mentink explained.   Early use cases for neuromorphic computing Neuromorphic computing is far from just experimental. A great number of new and established technology companies are heavily invested in developing new hardware, edge devices, software, and neuromorphic computing applications. Big tech brands such as IBM, NVIDIA, and Intel, with its Loihi chips, are all involved in neuromorphic computing, while companies in the Netherlands, aligned with a 2024 national white paper, are taking a leading regional role.  For example, the Dutch company Innatera — a leader in ultra-low power neuromorphic processors — recently secured €15 million in Series-A funding from Invest-NL Deep Tech Fund, the EIC Fund, MIG Capital, Matterwave Ventures, and Delft Enterprises.  Innatera is just the tip of the iceberg, as the Netherlands continues to support the new industry through funds, grants, and other incentives. Immediate use cases for neuromorphic computing include event-based sensing technologies integrated into smart sensors such as cameras or audio. These neuromorphic devices only process change, which can dramatically reduce power and data load, said Sylvester Kaczmarek, the CEO of OrbiSky Systems, a company providing AI integration for space technology.   Neuromorphic hardware and software have the potential to transfer AI running on the edge, especially for low-power devices such as mobile, wearables, or IoT.  Pattern recognition, keyword spotting, and simple diagnostics — such as real-time signal processing of complex sensor data streams for biomedical uses, robotics, or industrial monitoring — are some of the leading use cases, Dr Kaczmarek explained.  When applied to pattern recognition and classification or anomaly detection, neuromorphic computing can make decisions very quickly and efficiently,  Professor Dr Hans Hilgenkamp, Scientific Director of the MESA+ Institute at the University of Twente, agreed that pattern recognition is one of the fields where neuromorphic computing excels.  “One may also think about [for example] failure prediction in industrial or automotive applications,” he said.    The gaps creating neuromorphic opportunities Despite the recent progress, the road to establishing robust neuromorphic computing ecosystems in the Netherlands is challenging. Globalised tech supply chains and the standardisation of new technologies leave little room for hardware-level innovation.  For example, optical networks and optical chips have proven to outperform traditional systems in use today, but the tech has not been deployed globally. Deploying new hardware involves strategic coordination between the public and private sectors. The global rollout of 5G technology provides a good example of the challenges. It required telcos and governments around the world to deploy not only new antennas, but also smartphones, laptops, and a lot of hardware that could support the new standard.  On the software side, meanwhile, 5G systems had a pressing need for global standards to ensure integration, interoperability, and smooth deployment. Additionally, established telcos had to move from pure competition to strategic collaboration— an unfamiliar shift for an industry long built on siloed operations. Neuromorphic computing ecosystems face similar obstacles. The Netherlands recognises that the entire industry’s success depends on innovation in materials, devices, circuit designs, hardware architecture, algorithms, and applications.  These challenges and gaps are driving new opportunities for tech companies, startups, vendors, and partners.  Dr Kaczmarek told us that neuromorphic computing requires full-stack integration. This involves expertise that can connect novel materials and devices through circuit design and architectures to algorithms and applications. “Bringing these layers together is crucial but challenging,” he said.  On the algorithms and software side of things, developing new paradigms of programming, learning rules (beyond standard deep learning backpropagation), and software tools native to neuromorphic hardware are also priorities.  “It is crucial to make the hardware usable and efficient — co-designing hardware and algorithms because they are intimately coupled in neuromorphic systems,” said Dr Kaczmarek.  Other industries which have developed or are considering research on neuromorphic computing include healthcare (brain-computer interfaces and prosthetics), agri-food, and sustainable energy.  Neuromorphic computing modules or components can also be integrated with conventional CMOS, photonics, AI, and even quantum technologies.  Long-term opportunities in the Netherlands We asked Dr Hilgenkamp what expertise or innovations are most needed and offer the greatest opportunities for contribution and growth within this emerging ecosystem. “The long-term developments involve new materials and a lot of research, which is already taking place on an academic level,” Dr Hilgenkamp said.  He added that the idea of “materials that can learn” brings up completely new concepts in materials science that are exciting for researchers.  On the other hand, Dr Mentink pointed to the opportunity to transform our economies, which rely on processing massive amounts of data.  “Even replacing a small part of that with neuromorphic computing will lead to massive energy savings,” he said.  “Moreover, with neuromorphic computing, much more processing can be done close to where the data is produced. This is good news for situations in which data contains privacy-sensitive information.”  Concrete examples, according to Dr Mentink, also include fraud detection for credit card transactions, image analysis by robots and drones, anomaly detection of heartbeats, and processing of telecom data. “The most promising use cases are those involving huge data flows, strong demands for very fast response times, and small energy budgets,” said Dr Mentink.  As the use cases for neuromorphic computing increase, Dr Mentink expects the development of software toolchains that enable quick adoption of new neuromorphic platforms to see growth. This new sector would include services to streamline deployment. “Longer-term sustainable growth requires a concerted interdisciplinary effort across the whole computing stack to enable seamless integration of foundational discoveries to applications in new neuromorphic computing systems,” Dr Mentink said.  The bottom line The potential of neuromorphic computing has translated into billions of dollars in investment in the Netherlands and Europe, as well as in Asia and the rest of the world.  Businesses that can innovate, develop, and integrate hardware and software-level neuromorphic technologies stand to gain the most.   The potential of neuromorphic computing for greater energy efficiency and performance could ripple across industries. Energy, healthcare, robotics, AI, industrial IoT, and quantum tech all stand to benefit if they integrate the technology. And if the Dutch ecosystem takes off, the Netherlands will be in a position to lead the way. Supporting Dutch tech is a key mission of TNW Conference, which takes place on June 19-20 in Amsterdam. Tickets are now on sale — use the code TNWXMEDIA2025 at the checkout to get 30% off. Story by Ray Fernandez Ray Fernandez is a journalist with over a decade of experience reporting on technology, finance, science, and natural resources. His work ha (show all) Ray Fernandez is a journalist with over a decade of experience reporting on technology, finance, science, and natural resources. His work has been published in Bloomberg, TechRepublic, The Sunday Mail, eSecurityPlanet, and many others. He is a contributing writer for Espacio Media Incubator, which has reporters across the US, Europe, Asia, and Latin America. Get the TNW newsletter Get the most important tech news in your inbox each week. Also tagged with

Our latest and most advanced technologies — from AI to Industrial IoT, advanced robotics, and self-driving cars — share serious problems: massive energy consumption, limited on-edge capabilities, system hallucinations, and serious accuracy gaps.  One possible solution is emerging in the Netherlands. The country is developing a promising ecosystem for neuromorphic computing, which draws on neuroscience to boost IT efficiencies and performance. Billions of euros are being invested in this new form of computing worldwide. The Netherlands aims to become a leader in the market by bringing together startups, established companies, government organisations, and academics in a neuromorphic computing ecosystem. A Dutch mission to the UK In March, a Dutch delegation landed in the UK to host an “Innovation Mission” with local tech and government representatives. Top Sector ICT, a Dutch government–supported organisation, led the mission, which sought to strengthen and discuss the future of neuromorphic computing in Europe and the Netherlands.  We contacted Top Sector ICT, who connected us with one of their collaborators: Dr Johan H. Mentink, an expert in computational physics at Radboud University in the Netherlands. Dr Mentink spoke about how neuromorphic computing can solve the energy, accuracy, and efficiency challenges of our current computing architectures.  Grab that deal “Current digital computers use power-hungry processes to handle data,” Dr Mentink said.  “The result is that some modern data centres use so much energy that they even need their own power plant.”  Computing today stores data in one place (memory) and processes it in another place (processors). This means that a lot of energy is spent on transporting data, Dr Mentink explained.  In contrast, neuromorphic computing architectures are different at the hardware and software levels. For example, instead of using processors and memories, neuromorphic systems leverage new hardware components such as memristors. These act as both memory and processors.  By processing and saving data on the same hardware component, neuromorphic computing removes the energy-intensive and error-prone task of transporting data. Additionally, because data is stored on these components, it can be processed more immediately, resulting in faster decision-making, reduced hallucinations, improved accuracy, and better performance. This concept is being applied to edge computing, Industrial IoT, and robotics to drive faster real-time decision-making.  “Just like our brains process and store information in the same place, we can make computers that would combine data storage and processing in one place,” Dr Mentink explained.   Early use cases for neuromorphic computing Neuromorphic computing is far from just experimental. A great number of new and established technology companies are heavily invested in developing new hardware, edge devices, software, and neuromorphic computing applications. Big tech brands such as IBM, NVIDIA, and Intel, with its Loihi chips, are all involved in neuromorphic computing, while companies in the Netherlands, aligned with a 2024 national white paper, are taking a leading regional role.  For example, the Dutch company Innatera — a leader in ultra-low power neuromorphic processors — recently secured €15 million in Series-A funding from Invest-NL Deep Tech Fund, the EIC Fund, MIG Capital, Matterwave Ventures, and Delft Enterprises.  Innatera is just the tip of the iceberg, as the Netherlands continues to support the new industry through funds, grants, and other incentives. Immediate use cases for neuromorphic computing include event-based sensing technologies integrated into smart sensors such as cameras or audio. These neuromorphic devices only process change, which can dramatically reduce power and data load, said Sylvester Kaczmarek, the CEO of OrbiSky Systems, a company providing AI integration for space technology.   Neuromorphic hardware and software have the potential to transfer AI running on the edge, especially for low-power devices such as mobile, wearables, or IoT.  Pattern recognition, keyword spotting, and simple diagnostics — such as real-time signal processing of complex sensor data streams for biomedical uses, robotics, or industrial monitoring — are some of the leading use cases, Dr Kaczmarek explained.  When applied to pattern recognition and classification or anomaly detection, neuromorphic computing can make decisions very quickly and efficiently,  Professor Dr Hans Hilgenkamp, Scientific Director of the MESA+ Institute at the University of Twente, agreed that pattern recognition is one of the fields where neuromorphic computing excels.  “One may also think about [for example] failure prediction in industrial or automotive applications,” he said.    The gaps creating neuromorphic opportunities Despite the recent progress, the road to establishing robust neuromorphic computing ecosystems in the Netherlands is challenging. Globalised tech supply chains and the standardisation of new technologies leave little room for hardware-level innovation.  For example, optical networks and optical chips have proven to outperform traditional systems in use today, but the tech has not been deployed globally. Deploying new hardware involves strategic coordination between the public and private sectors. The global rollout of 5G technology provides a good example of the challenges. It required telcos and governments around the world to deploy not only new antennas, but also smartphones, laptops, and a lot of hardware that could support the new standard.  On the software side, meanwhile, 5G systems had a pressing need for global standards to ensure integration, interoperability, and smooth deployment. Additionally, established telcos had to move from pure competition to strategic collaboration— an unfamiliar shift for an industry long built on siloed operations. Neuromorphic computing ecosystems face similar obstacles. The Netherlands recognises that the entire industry’s success depends on innovation in materials, devices, circuit designs, hardware architecture, algorithms, and applications.  These challenges and gaps are driving new opportunities for tech companies, startups, vendors, and partners.  Dr Kaczmarek told us that neuromorphic computing requires full-stack integration. This involves expertise that can connect novel materials and devices through circuit design and architectures to algorithms and applications. “Bringing these layers together is crucial but challenging,” he said.  On the algorithms and software side of things, developing new paradigms of programming, learning rules (beyond standard deep learning backpropagation), and software tools native to neuromorphic hardware are also priorities.  “It is crucial to make the hardware usable and efficient — co-designing hardware and algorithms because they are intimately coupled in neuromorphic systems,” said Dr Kaczmarek.  Other industries which have developed or are considering research on neuromorphic computing include healthcare (brain-computer interfaces and prosthetics), agri-food, and sustainable energy.  Neuromorphic computing modules or components can also be integrated with conventional CMOS, photonics, AI, and even quantum technologies.  Long-term opportunities in the Netherlands We asked Dr Hilgenkamp what expertise or innovations are most needed and offer the greatest opportunities for contribution and growth within this emerging ecosystem. “The long-term developments involve new materials and a lot of research, which is already taking place on an academic level,” Dr Hilgenkamp said.  He added that the idea of “materials that can learn” brings up completely new concepts in materials science that are exciting for researchers.  On the other hand, Dr Mentink pointed to the opportunity to transform our economies, which rely on processing massive amounts of data.  “Even replacing a small part of that with neuromorphic computing will lead to massive energy savings,” he said.  “Moreover, with neuromorphic computing, much more processing can be done close to where the data is produced. This is good news for situations in which data contains privacy-sensitive information.”  Concrete examples, according to Dr Mentink, also include fraud detection for credit card transactions, image analysis by robots and drones, anomaly detection of heartbeats, and processing of telecom data. “The most promising use cases are those involving huge data flows, strong demands for very fast response times, and small energy budgets,” said Dr Mentink.  As the use cases for neuromorphic computing increase, Dr Mentink expects the development of software toolchains that enable quick adoption of new neuromorphic platforms to see growth. This new sector would include services to streamline deployment. “Longer-term sustainable growth requires a concerted interdisciplinary effort across the whole computing stack to enable seamless integration of foundational discoveries to applications in new neuromorphic computing systems,” Dr Mentink said.  The bottom line The potential of neuromorphic computing has translated into billions of dollars in investment in the Netherlands and Europe, as well as in Asia and the rest of the world.  Businesses that can innovate, develop, and integrate hardware and software-level neuromorphic technologies stand to gain the most.   The potential of neuromorphic computing for greater energy efficiency and performance could ripple across industries. Energy, healthcare, robotics, AI, industrial IoT, and quantum tech all stand to benefit if they integrate the technology. And if the Dutch ecosystem takes off, the Netherlands will be in a position to lead the way. Supporting Dutch tech is a key mission of TNW Conference, which takes place on June 19-20 in Amsterdam. Tickets are now on sale — use the code TNWXMEDIA2025 at the checkout to get 30% off. Story by Ray Fernandez Ray Fernandez is a journalist with over a decade of experience reporting on technology, finance, science, and natural resources. His work ha (show all) Ray Fernandez is a journalist with over a decade of experience reporting on technology, finance, science, and natural resources. His work has been published in Bloomberg, TechRepublic, The Sunday Mail, eSecurityPlanet, and many others. He is a contributing writer for Espacio Media Incubator, which has reporters across the US, Europe, Asia, and Latin America. Get the TNW newsletter Get the most important tech news in your inbox each week. Also tagged with

At Google I/O 2025, Google introduced MedGemma, an open suite of models designed for multimodal medical text and image comprehension. Built on the Gemma 3 architecture, MedGemma aims to provide developers with a robust foundation for creating healthcare applications that require integrated analysis of medical images and textual data. Model Variants and Architecture MedGemma is available in two configurations: MedGemma 4B: A 4-billion parameter multimodal model capable of processing both medical images and text. It employs a SigLIP image encoder pre-trained on de-identified medical datasets, including chest X-rays, dermatology images, ophthalmology images, and histopathology slides. The language model component is trained on diverse medical data to facilitate comprehensive understanding. MedGemma 27B: A 27-billion parameter text-only model optimized for tasks requiring deep medical text comprehension and clinical reasoning. This variant is exclusively instruction-tuned and is designed for applications that demand advanced textual analysis. Deployment and Accessibility Developers can access MedGemma models through Hugging Face, subject to agreeing to the Health AI Developer Foundations terms of use. The models can be run locally for experimentation or deployed as scalable HTTPS endpoints via Google Cloud’s Vertex AI for production-grade applications. Google provides resources, including Colab notebooks, to facilitate fine-tuning and integration into various workflows. Applications and Use Cases MedGemma serves as a foundational model for several healthcare-related applications: Medical Image Classification: The 4B model’s pre-training makes it suitable for classifying various medical images, such as radiology scans and dermatological images. Medical Image Interpretation: It can generate reports or answer questions related to medical images, aiding in diagnostic processes. Clinical Text Analysis: The 27B model excels in understanding and summarizing clinical notes, supporting tasks like patient triaging and decision support. Adaptation and Fine-Tuning While MedGemma provides strong baseline performance, developers are encouraged to validate and fine-tune the models for their specific use cases. Techniques such as prompt engineering, in-context learning, and parameter-efficient fine-tuning methods like LoRA can be employed to enhance performance. Google offers guidance and tools to support these adaptation processes. Conclusion MedGemma represents a significant step in providing accessible, open-source tools for medical AI development. By combining multimodal capabilities with scalability and adaptability, it offers a valuable resource for developers aiming to build applications that integrate medical image and text analysis. Check out the Models on Hugging Face and Project 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. Asif RazzaqWebsite |  + postsBioAsif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is committed to harnessing the potential of Artificial Intelligence for social good. His most recent endeavor is the launch of an Artificial Intelligence Media Platform, Marktechpost, which stands out for its in-depth coverage of machine learning and deep learning news that is both technically sound and easily understandable by a wide audience. The platform boasts of over 2 million monthly views, illustrating its popularity among audiences.Asif Razzaqhttps://www.marktechpost.com/author/6flvq/NVIDIA Releases Cosmos-Reason1: A Suite of AI Models Advancing Physical Common Sense and Embodied Reasoning in Real-World EnvironmentsAsif Razzaqhttps://www.marktechpost.com/author/6flvq/A Step-by-Step Coding Guide to Efficiently Fine-Tune Qwen3-14B Using Unsloth AI on Google Colab with Mixed Datasets and LoRA OptimizationAsif Razzaqhttps://www.marktechpost.com/author/6flvq/Chain-of-Thought May Not Be a Window into AI’s Reasoning: Anthropic’s New Study Reveals Hidden GapsAsif Razzaqhttps://www.marktechpost.com/author/6flvq/How to Build a Powerful and Intelligent Question-Answering System by Using Tavily Search API, Chroma, Google Gemini LLMs, and the LangChain Framework 🚨 Build GenAI you can trust. ⭐️ Parlant is your open-source engine for controlled, compliant, and purposeful AI conversations — Star Parlant on GitHub! (Promoted)

The rapid proliferation and superb capabilities of widely available LLMs has ignited intense debate within the educational sector. On one side they offer students a 24/7 tutor who is always available to help; but then of course students can use LLMs to cheat! I’ve seen both sides of the coin with my students; yes, even the bad side and even at the university level. While the potential benefits and problems of LLMs in education are widely discussed, a critical need existed for robust, empirical evidence to guide the integration of these technologies in the classroom, curricula, and studies in general. Moving beyond anecdotal accounts and rather limited studies, a recent work titled “The effect of ChatGPT on students’ learning performance, learning perception, and higher-order thinking: insights from a meta-analysis” offers one of the most comprehensive quantitative assessments to date. The article, by Jin Wang and Wenxiang Fan from the Chinese Education Modernization Research Institute of Hangzhou Normal University, was published this month in the journal Humanities and Social Sciences Communications from the Nature Publishing group. It is as complex as detailed, so here I will delve into the findings reported in it, touching also on the methodology and delving into the implications for those developing and deploying AI in educational contexts. Into it: Quantifying ChatGPT’s Impact on Student Learning The study by Wang and Fan is a meta-analysis that synthesizes data from 51 research papers published between November 2022 and February 2025, examining the impact of ChatGPT on three crucial student outcomes: learning performance, learning perception, and higher-order thinking. For AI practitioners and data scientists, this meta-analysis provides a valuable, evidence-based lens through which to evaluate current LLM capabilities and inform the future development of Education technologies. The primary research question sought to determine the overall effectiveness of ChatGPT across the three key educational outcomes. The meta-analysis yielded statistically significant and noteworthy results: Regarding learning performance, data from 44 studies indicated a large positive impact attributable to ChatGPT usage. In fact it turned out that, on average, students integrating ChatGPT into their learning processes demonstrated significantly improved academic outcomes compared to control groups. For learning perception, encompassing students’ attitudes, motivation, and engagement, analysis of 19 studies revealed a moderately but significant positive impact. This implies that ChatGPT can contribute to a more favorable learning experience from the student’s perspective, despite the a priori limitations and problems associated to a tool that students can use to cheat. Similarly, the impact on higher-order thinking skills—such as critical analysis, problem-solving, and creativity—was also found to be moderately positive, based on 9 studies. It is good news then that ChatGPT can support the development of these crucial cognitive abilities, although its influence is clearly not as pronounced as on direct learning performance. How Different Factors Affect Learning With ChatGPT Beyond overall efficacy, Wang and Fan investigated how various study characteristics affected ChatGPT’s impact on learning. Let me summarize for you the core results. First, there was a strong effect of the type of course. The largest effect was observed in courses that involved the development of skills and competencies, followed closely by STEM (science/Technology) and related subjects, and then by language learning/academic writing. The course’s learning model also played a critical role in modulating how much ChatGPT assisted students. Problem-based learning saw a particularly strong potentiation by ChatGPT, yielding a very large effect size. Personalized learning contexts also showed a large effect, while project-based learning demonstrated a smaller, though still positive, effect. The duration of ChatGPT use was also an important modulator of ChatGPT’s effect on learning performance. Short durations in the order of a single week produced small effects, while extended use over 4–8 weeks had the strongest impact, which did not grow much more if the usage was extended even further. This suggests that sustained interaction and familiarity may be crucial for cultivating positive affective responses to LLM-assisted learning. Interestingly, the students’ grade levels, the specific role played by ChatGPT in the activity, and the area of application did not affect learning performance significantly, in any of the analyzed studies. Other factors, including grade level, type of course, learning model, the specific role adopted by ChatGPT, and the area of application, did not significantly moderate the impact on learning perception. The study further showed that when ChatGPT functioned as an intelligent tutor, providing personalized guidance and feedback, its impact on fostering higher-order thinking was most pronounced. Implications for the Development of AI-Based Educational Technologies The findings from Wang & Fan’s meta-analysis carry substantial implications for the design, development, and strategic deployment of AI in educational settings: First of all, regarding the strategic scaffolding for deeper cognition. The impact on the development of thinking skills was somewhat lower than on performance, which means that LLMs are not inherently cultivators of deep critical thought, even if they do have a positive global effect on learning. Therefore, AI-based educational tools should integrate explicit scaffolding mechanisms that foster the development of thinking processes, to guide students from knowledge acquisition towards higher-level analysis, synthesis, and evaluation in parallel to the AI system’s direct help. Thus, the implementation of AI tools in education must be framed properly, and as we saw above this framing will depend on the exact type and content of the course, the learning model one wishes to apply, and the available time. One particularly interesting setup would be that where the AI tool supports inquiry, hypothesis testing, and collaborative problem-solving. Note though that the findings on optimal duration imply the need for onboarding strategies and adaptive engagement techniques to maximize impact and mitigate potential over-reliance. The superior impact documented when ChatGPT functions as an intelligent tutor highlights a key direction for AI in education. Developing LLM-based systems that can provide adaptive feedback, pose diagnostic and reflective questions, and guide learners through complex cognitive tasks is paramount. This requires moving beyond simple Q&A capabilities towards more sophisticated conversational AI and pedagogical reasoning. On top, there are a few non-minor issues to work on. While LLMs excel at information delivery and task assistance (leading to high performance gains), enhancing their impact on affective domains (perception) and advanced cognitive skills requires better interaction designs. Incorporating elements that foster student agency, provide meaningful feedback, and manage cognitive load effectively are crucial considerations. Limitations and Where Future Research Should Go The authors of the study prudently acknowledge some limitations, which also illuminate avenues for future research. Although the total sample size was the largest ever, it is still small, and very small for some specific questions. More research needs to be done, and a new meta-analysis will probably be required when more data becomes available. A difficult point, and this is my personal addition, is that as the technology progresses so fast, results might become obsolete very rapidly, unfortunately. Another limitation in the studies analyzed in this paper is that they are largely biased toward college-level students, with very limited data on primary education. Wang and Fan also discuss what AI, data science, and pedagogues should consider in future research. First, they should try to disaggregate effects based on specific LLM versions, a point that is critical because they evolve so fast. Second, they should study how students and teachers typically “prompt” the LLMs, and then investigate the impact of differential prompting on the final learning outcomes. Then, somehow they need to develop and evaluate adaptive scaffolding mechanisms embedded within LLM-based educational tools. Finally, and over a long term, we need to explore the effects of LLM integration on knowledge retention and the development of self-regulated learning skills. Personally, I add at this point, I am of the opinion that studies need to dig more into how students use LLMs to cheat, not necessarily willingly but possibly also by seeking for shortcuts that lead them wrong or allow them to get out of the way but without really learning anything. And in this context, I think AI scientists are falling short in developing camouflaged systems for the detection of AI-generated texts, that they can use to rapidly and confidently tell if, for example, a homework was done with an LLM. Yes, there are some watermarking and similar systems out there (which I will cover some day!) but I haven’t seem them deployed at large in ways that educators can easily utilize. Conclusion: Towards an Evidence-Informed Integration of AI in Education The meta-analysis I’ve covered here for you provides a critical, data-driven contribution to the discourse on AI in education. It confirms the substantial potential of LLMs, particularly ChatGPT in these studies, to enhance student learning performance and positively influence learning perception and higher-order thinking. However, the study also powerfully illustrates that the effectiveness of these tools is not uniform but is significantly moderated by contextual factors and the nature of their integration into the learning process. For the AI and data science community, these findings serve as both an affirmation and a challenge. The affirmation lies in the demonstrated efficacy of LLM technology. The challenge resides in harnessing this potential through thoughtful, evidence-informed design that moves beyond generic applications towards sophisticated, adaptive, and pedagogically sound educational tools. The path forward requires a continued commitment to rigorous research and a nuanced understanding of the complex interplay between AI, pedagogy, and human learning. References Here is the paper by Wang and Fan: The effect of ChatGPT on students’ learning performance, learning perception, and higher-order thinking: insights from a meta-analysis. Jin Wang & Wenxiang Fan Humanities and Social Sciences Communications volume 12, 621 (2025) If you liked this, check out my TDS profile. The post What the Most Detailed Peer-Reviewed Study on AI in the Classroom Taught Us appeared first on Towards Data Science.