As International Women's Day 2025 approaches, calls grow for a human-centric approach to make cybersecurity more appealing to women, who currently comprise just 17% of the workforce.

IWD 2025: Attracting Women to Cybersecurity with a Human-Centric Approach

Subscribe to our daily newsletter to keep in touch with the subjects shaping our future.

Humanoid Robots Are Starting to Work Human Warehouse Jobs

Y Combinator Pulls Support for AI Startup After Video of Boss Barking at Human Worker, Calling Him "Number 17"

Proper GenAI governance will control and manage the risks associated with NHI growth, bringing equilibrium and balance between security and AI innovation to IT ecosystems.

The Hidden Cybersecurity Crisis: How GenAI is Fueling the Growth of Unchecked Non-Human Identities - Security Boulevard

Bengaluru-based startup Topmate is redefining quick commerce—delivering human expertise in just 10 minutes.

Bengaluru-Based Startup Topmate Promises 10-Minute Human Delivery, Declares It’s Over for Blinkit and Zepto

Andrew G. Nixon, a spokesman for the Department of Health and Human Services, said in an email Monday that “our Administration wants to help America have the best research in the world, and we believe that by ensuring that more cents on every dollar go directly to science and not to ...

Trump’s NIH challenges the model that underlies U.S. scientific dominance - The Washington Post

Scientists discovered a fungus that caused spiders to "exhibit behavioral changes similar to those reported for zombie ants."

Scientists Discover "Zombie" Fungus That Seizes Control of Spiders, Suggest It Be Used for Human Medicine

Human

human

AI’s ability to mimic human behavior, emotional intelligence, creativity, and morality while exploring the future of artificial intelligence Artificial intelligence, especially that afforded by ChatGPT, has changed our manner of dealing with machines. AI now puts us in a state of conversation, learns from us, and even attempts to understand the context behind the questions we ask. However, the significant issue that people keep asking is whether ChatGPT can be trained to become human. At the core of AIs like ChatGPT is a neural network, a complex web of algorithms meant to mimic the functioning of the human brain. The algorithms learn about patterns, make predictions, and improve over time through a process known as machine learning. However, there is quite a gap between teaching patterns and being human. But being human isn’t just about patterns or predictions and it is about emotions, experience, understanding of culture, making moral judgments, and the ability for truly original creativity. Interacting with AI today may often generate human-like responses that are interesting and even insightful at times. Yet, what generates those responses — patterns found in large datasets, not quite the sort of personal contemplation or emotive depth carried into every encounter by humans. To answer if ChatGPT can be trained to be human, we must consider what exactly it means for something to be human. For this argument, let’s break it down into several such components: emotional intelligence, creativity, morality, memory, and learning from experiences. Some of these are important facets of human behavior and cognition. Can AI, particularly ChatGPT, ever be imbued with these so that we consider the AI “human” to that degree? Emotional Intelligence Emotional intelligence is the ability to sense how others feel and manage one’s own emotions. Our words usually carry emotional intensity from subtle sarcasm to overt anger or happiness. Humans use tone, context, and body language to communicate and interpret emotions. ChatGPT will simulate aspects of emotional intelligence because it can generate text that carries an emotional flavor for a conversation. If you express frustration, then it produces a response that would appear empathetic; if you are ebullient, it can echo some lighthearted repartee. Does this mean it “knows” or understands emotions? No. Instead, what ChatGPT does is recognize patterns in language as they reflect emotion, but it does not feel emotions. It doesn’t have the physiological and psychological mechanisms that underlie human emotion. Training ChatGPT to “be human” in this way would take more than feeding it datasets on emotional expression. It would rather require an architecture that allowed for some authentic experiences of emotions — that is, something beyond the scope of current developments in AI. Even if we can, in some way, simulate emotional reactions in AI, they will still be fundamentally different from human emotions, which are tied closely with consciousness, biology, and experience. Creativity Creativity- the mark of human intelligence: the ability to come up with new ideas, solve problems in innovative ways, and create art that vibrates at a deeply emotional or intellectual level. Is ChatGPT creative? In this respect, yes. ChatGPT can write, for instance, poems, short stories, code- whatever you want it to do. All the same, this creativity is based on the pattern rather than on actual thinking outside of the box that may reflect human creativity. Some of the most creative solutions or artistic expressions are often born out of personal experience, emotional ups and downs, moments of inspiration, those things that cannot easily be expressed and replicated in an algorithm. Although AI can generate creative content, it remains without experiential or emotional depth to inform human creativity. This means it offers content based on learned patterns rather than any personal experience or emotional journey. That is the main point of explaining why, despite impressive outputs, ChatGPT is fundamentally different from a human mind. Morality The other very important element of being human is morality, for which decisions are generally guided by the moral compass of and through upbringing, culture, experiences, and societal norms. More often than not, humans struggle with conflicting values, weigh the consequences of their deeds, and sometimes, change their beliefs over time. Could ChatGPT make moral decisions? Not at all. While it may be able to be programmed to adhere to certain guidelines of morality or avoid harmful content, its understanding of morality is superficial. It does not possess personal values or understand right and wrong in the way humans do. Instead, it responds based on patterns in data that often simply mimic moral reasoning without truly engaging in the complex, human process of ethical decision-making. It is a huge task to train ChatGPT to be human in this regard since even humans are unable to accept one sense of moral principle. Morality also relates to consciousness, which AI lacks. Though AI systems can follow rules, they are not able to understand the sense of moral dilemmas that define human experience. Memory and Experience Humans are a product of their memories and experiences. Every conversation, every relationship, and triumph and every failure leave an impression on who we are and how we approach the world. AI does not have any kind of real memory of past interactions unless it has been explicitly programmed to retain specific data. Each interaction with ChatGPT is unique, and while it can process historical data, it doesn’t “remember” interactions the same way a human does. It would give this form so that ChatGPT can learn through experiences, creating a sense of identity and personal history over time. Now, it does not necessarily imply that it should remember anything but interpret it to affect the subsequent actions; the same way humans do in life, growing and changing according to their past. Current AI models, including ChatGPT, do not have this ability. Though they might learn from big datasets, they lack personal experiences and do not change according to individual interaction. Because of this, for ChatGPT to be closer to “being human,” it requires something more than the capability to store and accumulate information but must also develop the capability to reflect on that same information and to change in meaningful ways relative to past experiences. Learning by Experience Human intelligence is reinforced by the ability to reflect upon mistakes and utilize those as learning experiences. We sometimes reflect on our actions, question our decisions, and make alterations in behavior based on our experiences. ChatGPT, though, can be trained through machine learning models, though it learns very differently compared to humans. Improvement of AI in machine learning is through wide training on various datasets and adjustment of weights and biases for better outcomes. However, the methodology itself is not like that of human learning based on trial and error and is contextual, emotionally and socially tuned, and based on personal experience and interactions with other humans. In contrast, the learning process of ChatGPT is abstract and data-driven. While AI may learn to make improvements in its responses, it is without a human ability to reflect on itself. Human learning isn’t just adding new information, but building that new information off of what we already know and thinking about what the implications are. Sometimes, what we learn is questioned. ChatGPT learns to do this task better at getting statistically likely responses without any understanding or reflection about what it is learning. The Future of AI So, can ChatGPT be made human? Currently, as of the technology, the answer is no. In full, that is. While still able to mimic and even generate human-like conversations, produce creative content, and maintain adherence to all sorts of ethical directives, AI continues to lack depth, consciousness, and emotional complexities of what it means to be human. However, the future of AI is no longer so static. Researchers are more and more pushing the frontiers of what could be done with AI: find new approaches to making machines more intelligent, adaptive, and even emotionally intelligent. Some believe that with domains such as neuro-symbolic AI-machine learning coupled with symbolic reasoning or even quantum computation, we may eventually build one day the essence of human-like cognition in an AI. But yet, all of this raises a basic question: Do we want AI to be human? The development of machines that could think and feel like humans raises profound questions in the realms of ethics and philosophy. Where do we draw that line between the machine and the person if AI becomes too human?
Should AI have rights when it can feel? These are questions that society will have to grapple with as AI continues its evolution. AI’s ability to mimic human behavior, emotional intelligence, creativity, and morality while exploring the future of artificial intelligence Artificial intelligence, especially that afforded by ChatGPT, has changed our manner of dealing with machines. AI now puts us in a state of conversation, learns from us, and even attempts to understand the context behind the questions we ask. However, the significant issue that people keep asking is whether ChatGPT can be trained to become human. However, the significant issue that people keep asking is whether ChatGPT can be trained to become human. However, the significant issue that people keep asking is whether ChatGPT can be trained to become human. At the core of AIs like ChatGPT is a neural network, a complex web of algorithms meant to mimic the functioning of the human brain. The algorithms learn about patterns, make predictions, and improve over time through a process known as machine learning. However, there is quite a gap between teaching patterns and being human. But being human isn’t just about patterns or predictions and it is about emotions, experience, understanding of culture, making moral judgments, and the ability for truly original creativity. Interacting with AI today may often generate human-like responses that are interesting and even insightful at times. Yet, what generates those responses — patterns found in large datasets, not quite the sort of personal contemplation or emotive depth carried into every encounter by humans. To answer if ChatGPT can be trained to be human, we must consider what exactly it means for something to be human. For this argument, let’s break it down into several such components: emotional intelligence, creativity, morality, memory, and learning from experiences. Some of these are important facets of human behavior and cognition. Can AI, particularly ChatGPT, ever be imbued with these so that we consider the AI “human” to that degree? Can AI, particularly ChatGPT, ever be imbued with these so that we consider the AI “human” to that degree? Can AI, particularly ChatGPT, ever be imbued with these so that we consider the AI “human” to that degree? Emotional Intelligence Emotional intelligence is the ability to sense how others feel and manage one’s own emotions. Our words usually carry emotional intensity from subtle sarcasm to overt anger or happiness. Humans use tone, context, and body language to communicate and interpret emotions. ChatGPT will simulate aspects of emotional intelligence because it can generate text that carries an emotional flavor for a conversation. If you express frustration, then it produces a response that would appear empathetic; if you are ebullient, it can echo some lighthearted repartee. Does this mean it “knows” or understands emotions? No. Instead, what ChatGPT does is recognize patterns in language as they reflect emotion, but it does not feel emotions. It doesn’t have the physiological and psychological mechanisms that underlie human emotion. Training ChatGPT to “be human” in this way would take more than feeding it datasets on emotional expression. It would rather require an architecture that allowed for some authentic experiences of emotions — that is, something beyond the scope of current developments in AI. Even if we can, in some way, simulate emotional reactions in AI, they will still be fundamentally different from human emotions, which are tied closely with consciousness, biology, and experience. Creativity Creativity- the mark of human intelligence: the ability to come up with new ideas, solve problems in innovative ways, and create art that vibrates at a deeply emotional or intellectual level. Is ChatGPT creative? In this respect, yes. ChatGPT can write, for instance, poems, short stories, code- whatever you want it to do. All the same, this creativity is based on the pattern rather than on actual thinking outside of the box that may reflect human creativity. Some of the most creative solutions or artistic expressions are often born out of personal experience, emotional ups and downs, moments of inspiration, those things that cannot easily be expressed and replicated in an algorithm. Although AI can generate creative content, it remains without experiential or emotional depth to inform human creativity. This means it offers content based on learned patterns rather than any personal experience or emotional journey. That is the main point of explaining why, despite impressive outputs, ChatGPT is fundamentally different from a human mind. Morality The other very important element of being human is morality, for which decisions are generally guided by the moral compass of and through upbringing, culture, experiences, and societal norms. More often than not, humans struggle with conflicting values, weigh the consequences of their deeds, and sometimes, change their beliefs over time. Could ChatGPT make moral decisions? Not at all. While it may be able to be programmed to adhere to certain guidelines of morality or avoid harmful content, its understanding of morality is superficial. It does not possess personal values or understand right and wrong in the way humans do. Instead, it responds based on patterns in data that often simply mimic moral reasoning without truly engaging in the complex, human process of ethical decision-making. It is a huge task to train ChatGPT to be human in this regard since even humans are unable to accept one sense of moral principle. Morality also relates to consciousness, which AI lacks. Though AI systems can follow rules, they are not able to understand the sense of moral dilemmas that define human experience. Memory and Experience Humans are a product of their memories and experiences. Every conversation, every relationship, and triumph and every failure leave an impression on who we are and how we approach the world. AI does not have any kind of real memory of past interactions unless it has been explicitly programmed to retain specific data. Each interaction with ChatGPT is unique, and while it can process historical data, it doesn’t “remember” interactions the same way a human does. It would give this form so that ChatGPT can learn through experiences, creating a sense of identity and personal history over time. Now, it does not necessarily imply that it should remember anything but interpret it to affect the subsequent actions; the same way humans do in life, growing and changing according to their past. Current AI models, including ChatGPT, do not have this ability. Though they might learn from big datasets, they lack personal experiences and do not change according to individual interaction. Because of this, for ChatGPT to be closer to “being human,” it requires something more than the capability to store and accumulate information but must also develop the capability to reflect on that same information and to change in meaningful ways relative to past experiences. Learning by Experience Human intelligence is reinforced by the ability to reflect upon mistakes and utilize those as learning experiences. We sometimes reflect on our actions, question our decisions, and make alterations in behavior based on our experiences. ChatGPT, though, can be trained through machine learning models, though it learns very differently compared to humans. Improvement of AI in machine learning is through wide training on various datasets and adjustment of weights and biases for better outcomes. However, the methodology itself is not like that of human learning based on trial and error and is contextual, emotionally and socially tuned, and based on personal experience and interactions with other humans. In contrast, the learning process of ChatGPT is abstract and data-driven. While AI may learn to make improvements in its responses, it is without a human ability to reflect on itself. Human learning isn’t just adding new information, but building that new information off of what we already know and thinking about what the implications are. Sometimes, what we learn is questioned. ChatGPT learns to do this task better at getting statistically likely responses without any understanding or reflection about what it is learning. The Future of AI So, can ChatGPT be made human? Currently, as of the technology, the answer is no. So, can ChatGPT be made human? Currently, as of the technology, the answer is no. So, can ChatGPT be made human? Currently, as of the technology, the answer is no. In full, that is. While still able to mimic and even generate human-like conversations, produce creative content, and maintain adherence to all sorts of ethical directives, AI continues to lack depth, consciousness, and emotional complexities of what it means to be human. However, the future of AI is no longer so static. Researchers are more and more pushing the frontiers of what could be done with AI: find new approaches to making machines more intelligent, adaptive, and even emotionally intelligent. Some believe that with domains such as neuro-symbolic AI-machine learning coupled with symbolic reasoning or even quantum computation, we may eventually build one day the essence of human-like cognition in an AI. But yet, all of this raises a basic question: Do we want AI to be human? The development of machines that could think and feel like humans raises profound questions in the realms of ethics and philosophy. Where do we draw that line between the machine and the person if AI becomes too human? Should AI have rights when it can feel? Where do we draw that line between the machine and the person if AI becomes too human? Should AI have rights when it can feel? These are questions that society will have to grapple with as AI continues its evolution.

Can We Train AI to Become More Human-Like?

Humans have proved to be not only biologically complex creatures, but also one with uncanny, complex intellectual capabilities; they have successfully created robots after their own likeness. Robots now do repetitive work in which human labor becomes too costly and ineffective to justify. In fact, there is hope that these robots will surpass many human limitations–many of which have already been met. To researchers in the extensive field of robotics development – the speed, accuracy and intelligence these robots currently exhibit are monumental achievements – while to others who have a modicum of knowledge of robots, these exhibitions by robots are cause for concern. They worry, like Polish playwright Karel Capek painted in his play “ ”, that robots will eventually displace humans from their jobs, or worse still, like presented in “Terminator” series, that robots might become future overlords and go to war with humans. Rossum’s Universal Robots How much human abilities are robots able to exhibit? To a large extent, contemporary robots have replicated human dexterity, perception and intelligence. For example, , a robot developed at the Massachusetts Institute of Technology, owing to sophisticated hardware and high-level programming, works like the human nervous system, which consists of both voluntary and involuntary functionality. , a robotic avatar developed by Toyota, mimics the movement of its human operator, and will be able to perform surgeries in the future while being controlled from another world by a doctor. , an AI-powered humanoid robot, talks, draws, sings and shows some emotions–and , another AI-powered humanoid robot, mimics human-like facial expressions. Despite all these, robots are yet to perform tasks in which complex, rigorous human intellectual capabilities are required. This is a consequence of the fact that researchers in the field of robotics development have not understood the intricacies of human intelligence. Kismet T-HR3 Sophia Ameca Why do robots lack unique human intellectual abilities? What is known is that manifests in the elaborate and interconnected networks of neurons within the human brain and is a function of learning and experience. Neurons in the human brain form electrical connections with one another, but it is rather unclear how they collectively cultivate brain activities like thinking and reasoning. Consequently, researchers in robotics development have been unable to download into robots this wondrous human intelligence–and hence robots lack the intellectual independence that they will need to beat humans at their own game. Furthermore, robots are biologically limited in their make-up and do not have the natural power-house like in humans–therefore, availability of power for use will continue to limit the functionality of even the most sophisticated robot for quite a long time. human intelligence Moreover, the human sensory system–vision, hearing, touch, smell and taste–are naturally complex. However, they are somehow being implemented into robots through elegant technology. For instance, robots like Ameca and Sophia are already doing well at responding to these human senses, but their ability therein is still considered to be in a nascent stage like that of a newborn baby that lacks the experience to understand what it senses. The recognition of sound, for instance, by a robot is intricately dependent on the extent of its “lexicon”, where the sound is compared among similar sound waves–and the lexicon does not expand unless through high-level programming, which needs to be carried out regularly. Besides, robots lack empathy, which is the ability to understand and share the feelings of another–and it has been established that empathy is a powerful driver of many human decisions and actions. In essence, robots can replicate emotions but cannot feel them; they can show sadness but do not feel sad. What are the jobs robots cannot do because of lack of unique human intellectual abilities? Robots are unable to undertake tasks in which special human intellectual capabilities are needed. who has not only medical education, but also the experience in prescribing medicines, applying operation techniques, and using medical apparatus. Robots cannot perform scientific analysis at particularly complex levels that usually bring about novel epiphany that engenders scientific discoveries; although they can be employed to do certain tasks, robots cannot be depended on for new ways of completing those tasks. Robots lack emotions and hence are useless in creative writing, which is a work of feelings and emotions brought on paper through words; they can only write what humans instruct them. Robots cannot teach like humans because they lack the understanding power to engage in interactive exchange as demonstrated between a teacher and the students; a teacher understands the teaching materials as well as the students on an individual level. Robots cannot perform the duties of a doctor Conclusion Robots have done exceedingly very well in performing jobs humans considered drudgery, in which health risks and hazards are high and human labor is mostly costly. They have allowed humans to improve on their living standard and become more precise and accurate in carrying out their jobs, especially those that are done casually that do not need intellectual capabilities but precision and accuracy. This is because robots are currently as intelligent as to only follow rigidly and precisely the instructions laid down for them by humans. Therefore, they have enabled humans to upgrade their skills and go for those jobs that maximize their capabilities. Although robots are seemingly intelligent–especially when considering the speed and accuracy with which they perform tasks–they are not any more intelligent than the human that controls them. If ever they have any intelligence of their own at all, it is still not up to that of a housefly. advanced intellectual

The Lack of Unique Human Abilities in Robots

The iconic figure in SDG’s logo — the silhouette of a man with two pairs of legs and a mystical third eye — was inspired by Leonardo da Vinci’s Vitruvian Man. The original drawing, created in the 15th century, symbolizes the perfect harmony between art and science, the blend of intellect and aesthetics, and the endless curiosity that defines human potential. in SDG’s logo in SDG’s logo The added elements in our version — the extra pair of legs and a third all-seeing eye — symbolize our forward-thinking nature, nomadic freedom, and constant drive for innovation. So it’s no coincidence that we celebrate Vitruvian Man Day on Leonardo da Vinci’s birthday, April 15th, which is also recognized globally as 🎨 World Art Day. It’s a day to honor creativity, curiosity, and the human spirit. How the Vitruvian Man Tradition Begins at SDG Vitruvian Man Day was officially added to the company´s Digital Nomad Calendar in 2023, along with other unique SDG holidays. Since then, it’s grown into an annual tradition that employees worldwide look forward to with excitement. SDG SDG Now in its third year of celebration, 🌟Vitruvian Man Day is more than a day — it’s a week-long celebration of culture, innovation, and connection. Each year, teams across the globe put their creative spin on the festivities, making it a true reflection of SDG’s diverse and global community. How We Celebrate: Company Traditions and Activities Vitruvian Man Day at SDG is all about celebrating creativity, which is reflected in our week-long festivities. There’s something for everyone, no matter if you’re a fan of art, a lover of learning, or just here for the good vibes and great food. 🎭 Some of our favorite traditions include: Some of our favorite traditions include: Local art-inspired parties that mix culture and fun


Museum visits and guided tours that feed our curiosity


Art lectures that dive into the minds of great masters like Da Vinci and Bosch


Creative workshops, from resin art to candle making


Catered brunches and dinners at local gems, where teams reconnect and celebrate together Local art-inspired parties that mix culture and fun Local art-inspired parties that mix culture and fun Museum visits and guided tours that feed our curiosity Museum visits and guided tours that feed our curiosity Art lectures that dive into the minds of great masters like Da Vinci and Bosch Art lectures that dive into the minds of great masters like Da Vinci and Bosch Creative workshops, from resin art to candle making Creative workshops, from resin art to candle making Catered brunches and dinners at local gems, where teams reconnect and celebrate together Catered brunches and dinners at local gems, where teams reconnect and celebrate together And of course, the launch of our special Vitruvian Day merch collection — featuring stylish travel organizers and other nomad-friendly goodies exclusively available at the Taco Shop — our internal corporate merch shop. These celebrations bring the SDG community together, fostering a shared sense of inspiration and pride in who we are and what we stand for. SDG community SDG community 🌏 Where We Celebrate? Spoiler: all around the Globe! From the heart of Europe and around the globe, SDG employees brought the Vitruvian Man Day spirit to life in their creative ways. Here’s a peek at how this year’s celebration unfolded across key SDG locations: Batumi, Georgia: 🏙️ The local team kicked off the week with a Leonardo da Vinci-themed lecture, breakfast at the stylish Kultura café, and an inspiring creative workshop (participants chose between epoxy resin art and candle making).


Belgrade, Serbia: Our engineers explored the city through a unique lens with a “Belgrade through the eyes of an engineer” tour, followed by a cozy dinner at the artistic Plato café.


Cyprus: Art met relaxation in a painting masterclass with catering, giving the team a chance to unwind and create.


Lisbon, Portugal: A thoughtfully curated guided museum tour sparked fresh inspiration and brought colleagues together around shared passions.


Tbilisi, Georgia: A deep dive into the fantastical world of Hieronymus Bosch set the tone for the day, followed by a charming brunch with snacks and wine.


Poland: Team members took in cutting-edge creativity at a new contemporary art museum, then gathered for a stylish dinner at a cozy Italian restaurant.
And way more! ☕


Yerevan, Armenia: 🎨The team explored the world of Armenian legend Sergei Parajanov with a museum tour and collage workshop.


Ubud, Bali: SDGangers crafted silver jewelry by hand before sharing a delicious dinner together.


Phuket, Thailand: The team spirit flourished during a therapeutic mandala-making class at the Mandala Art Therapy & Reiki Center.


Each celebration was united by a common thread: a love for learning, creating, and connecting — the cornerstones of SDG culture.


At its heart, Vitruvian Man Day is a celebration of the values that drive SDG forward: curiosity, creativity, innovation, and a deep appreciation for the world around us. It reminds us that behind every line of code, every product decision, and every global meeting, there’s a human spirit striving for balance, beauty, and brilliance. Batumi, Georgia: 🏙️ The local team kicked off the week with a Leonardo da Vinci-themed lecture, breakfast at the stylish Kultura café, and an inspiring creative workshop (participants chose between epoxy resin art and candle making). Batumi, Georgia: 🏙️ The local team kicked off the week with a Leonardo da Vinci-themed lecture, breakfast at the stylish Kultura café, and an inspiring creative workshop (participants chose between epoxy resin art and candle making). Batumi, Georgia Belgrade, Serbia: Our engineers explored the city through a unique lens with a “Belgrade through the eyes of an engineer” tour, followed by a cozy dinner at the artistic Plato café. Belgrade, Serbia: Our engineers explored the city through a unique lens with a “Belgrade through the eyes of an engineer” tour, followed by a cozy dinner at the artistic Plato café. Belgrade, Serbia Cyprus: Art met relaxation in a painting masterclass with catering, giving the team a chance to unwind and create. Cyprus: Art met relaxation in a painting masterclass with catering, giving the team a chance to unwind and create. Cyprus Lisbon, Portugal: A thoughtfully curated guided museum tour sparked fresh inspiration and brought colleagues together around shared passions. Lisbon, Portugal: A thoughtfully curated guided museum tour sparked fresh inspiration and brought colleagues together around shared passions. Lisbon, Portugal Tbilisi, Georgia: A deep dive into the fantastical world of Hieronymus Bosch set the tone for the day, followed by a charming brunch with snacks and wine. Tbilisi, Georgia: A deep dive into the fantastical world of Hieronymus Bosch set the tone for the day, followed by a charming brunch with snacks and wine. Tbilisi, Georgia Poland: Team members took in cutting-edge creativity at a new contemporary art museum, then gathered for a stylish dinner at a cozy Italian restaurant.
And way more! ☕


Yerevan, Armenia: 🎨The team explored the world of Armenian legend Sergei Parajanov with a museum tour and collage workshop.


Ubud, Bali: SDGangers crafted silver jewelry by hand before sharing a delicious dinner together.


Phuket, Thailand: The team spirit flourished during a therapeutic mandala-making class at the Mandala Art Therapy & Reiki Center.


Each celebration was united by a common thread: a love for learning, creating, and connecting — the cornerstones of SDG culture. Poland: Team members took in cutting-edge creativity at a new contemporary art museum, then gathered for a stylish dinner at a cozy Italian restaurant. Poland And way more! ☕ Yerevan, Armenia: 🎨The team explored the world of Armenian legend Sergei Parajanov with a museum tour and collage workshop.


Ubud, Bali: SDGangers crafted silver jewelry by hand before sharing a delicious dinner together.


Phuket, Thailand: The team spirit flourished during a therapeutic mandala-making class at the Mandala Art Therapy & Reiki Center. Yerevan, Armenia: 🎨The team explored the world of Armenian legend Sergei Parajanov with a museum tour and collage workshop. Yerevan, Armenia: 🎨The team explored the world of Armenian legend Sergei Parajanov with a museum tour and collage workshop. Yerevan, Armenia Ubud, Bali: SDGangers crafted silver jewelry by hand before sharing a delicious dinner together. Ubud, Bali: SDGangers crafted silver jewelry by hand before sharing a delicious dinner together. Ubud, Bali Phuket, Thailand: The team spirit flourished during a therapeutic mandala-making class at the Mandala Art Therapy & Reiki Center. Phuket, Thailand: The team spirit flourished during a therapeutic mandala-making class at the Mandala Art Therapy & Reiki Center. Phuket, Thailand: Each celebration was united by a common thread: a love for learning, creating, and connecting — the cornerstones of SDG culture. At its heart, Vitruvian Man Day is a celebration of the values that drive SDG forward: curiosity, creativity, innovation, and a deep appreciation for the world around us. It reminds us that behind every line of code, every product decision, and every global meeting, there’s a human spirit striving for balance, beauty, and brilliance. At its heart, Vitruvian Man Day is a celebration of the values that drive SDG forward: curiosity, creativity, innovation, and a deep appreciation for the world around us. It reminds us that behind every line of code, every product decision, and every global meeting, there’s a human spirit striving for balance, beauty, and brilliance. 🏛️ As we continue to build a future shaped by technology and human connection, Vitruvian Man Day grounds us in what truly matters: our shared humanity and our potential to create meaningful, lasting impact. Join the Celebration Next Year! Whether you’re part of the SDGangster crew or someone inspired by our unique company culture, Vitruvian Man Day is a tradition worth celebrating. Mark your calendars for April 15, 2026, and unleash your inner Da Vinci. Until then, keep creating, keep exploring, and keep living the SDG way. April 15, 2026 Join us next year to celebrate!

Celebrating Vitruvian Man Day at SDG: 
A Tribute to Arts, Science, & Human Potential

Vitruvian Man Day op SDG: 'n eerbetoon aan kuns, wetenskap en menslike potensiaal

Celebración del Día Vitruviano del Hombre en el SDG: un homenaje a las artes, la ciencia y el potencial humano

SDG में Vitruvian Man Day का जश्न मनाना: कला, विज्ञान और मानव क्षमता का सम्मान

Obilježavanje Vitruvijskog dana čovjeka na SDG-u: počast umjetnosti, znanosti i ljudskom potencijalu

Հիմնական հոդված՝ SDG- ի Vitruvian Man Day: A Tribute to Arts, Science, and Human Potential

SDGでヴィトルヴィア人デーを祝う:芸術、科学、そして人間の潜在能力への敬意

សព្វថ្ងៃបុរស Vitruvian នៅទីក្រុង SDG: ការអរគុណដល់សិល្បៈវិទ្យាសាស្រ្តនិងសមត្ថភាពបុរស

SDG-д Vitruvian Man's Day - Хувийн, шинжлэх ухааны, хүний хүчин чадалтай ач холбогдол

Прослава Витрувијског дана човека на СДГ-у: Похвала уметности, науци и људском потенцијалу

Святкування Вітрувійського дня людини на SDG: Повага мистецтву, науці та людському потенціалу

Ukukhanyisa iVitruvian Man Day kwi-SDG: I-Tribute to Arts, Science, and Human Potential

在SDG庆祝维特鲁维亚人日:对艺术,科学和人类潜力的敬意

Training a large language model is like teaching someone how to speak and behave. Imagine a kid learning a language and being gradually guided to answer questions appropriately. We can break this process into three main stages, each with everyday examples. speak behave Step 1: Learning by Reading — Self-Supervised Pre-Training Step 1: Learning by Reading — Self-Supervised Pre-Training Think of a child growing up by reading every book available in the home library. Instead of having a teacher always explain things, the child learns the language naturally by listening, reading, and trying to understand context, grammar, and vocabulary. How It Works in LLMs: How It Works in LLMs: Learning from All Kinds of Text:
The model "reads" a huge amount of text from books, websites, and online forums. Just as a child picks up language patterns and common phrases, the model learns how words connect to form meaningful sentences.


Predicting the Next Word:
Imagine a fill-in-the-blank game where the child must guess the missing word in a sentence. Over time, by guessing repeatedly, the child starts to know which words fit best. The model is trained to predict the next word in a sentence, helping it learn the structure and flow of language. Learning from All Kinds of Text:
The model "reads" a huge amount of text from books, websites, and online forums. Just as a child picks up language patterns and common phrases, the model learns how words connect to form meaningful sentences. Learning from All Kinds of Text: Learning from All Kinds of Text: The model "reads" a huge amount of text from books, websites, and online forums. Just as a child picks up language patterns and common phrases, the model learns how words connect to form meaningful sentences. Predicting the Next Word:
Imagine a fill-in-the-blank game where the child must guess the missing word in a sentence. Over time, by guessing repeatedly, the child starts to know which words fit best. The model is trained to predict the next word in a sentence, helping it learn the structure and flow of language. Predicting the Next Word: Predicting the Next Word: Imagine a fill-in-the-blank game where the child must guess the missing word in a sentence. Over time, by guessing repeatedly, the child starts to know which words fit best. The model is trained to predict the next word in a sentence, helping it learn the structure and flow of language. Step 2: Learning What’s Right and Wrong — Supervised Fine-Tuning Step 2: Learning What’s Right and Wrong — Supervised Fine-Tuning Now, consider that the same child has a teacher who guides them on proper behavior and acceptable language. The teacher shows the child examples of good answers and explains why certain responses are more appropriate than others. The same concept works in LLMs: The same concept works in LLMs: Curated Examples:
In this stage, engineers compile examples of questions and correct answers. It’s like a teacher offering a set of “model answers” to common questions.


Adjusting the Response:
The model is fine-tuned to mimic these good answers. For instance, if the model had previously been able to generate harmful or inappropriate responses (just as a child might mimic questionable language they overheard), now it’s corrected with examples of what to say. Curated Examples:
In this stage, engineers compile examples of questions and correct answers. It’s like a teacher offering a set of “model answers” to common questions. Curated Examples: Curated Examples: In this stage, engineers compile examples of questions and correct answers. It’s like a teacher offering a set of “model answers” to common questions. Adjusting the Response:
The model is fine-tuned to mimic these good answers. For instance, if the model had previously been able to generate harmful or inappropriate responses (just as a child might mimic questionable language they overheard), now it’s corrected with examples of what to say. Adjusting the Response: Adjusting the Response: The model is fine-tuned to mimic these good answers. For instance, if the model had previously been able to generate harmful or inappropriate responses (just as a child might mimic questionable language they overheard), now it’s corrected with examples of what to say. Real-Life Analogy: Real-Life Analogy: Consider a scenario where a teacher explains why certain instructions—like “don’t run in the hallways”—are important even though a child might know how to run. The teacher’s guidance shapes the child’s understanding of the proper way to act or answer. Step 3: Learning by Feedback — Reinforcement Learning from Human Feedback (RLHF) Step 3: Learning by Feedback — Reinforcement Learning from Human Feedback (RLHF) Imagine the child now taking part in a class debate. They speak, and the teacher and classmates provide immediate praise for good points and gentle correction for weak arguments. Over time, the child refines how they articulate ideas based on this feedback. How It Works in LLMs: How It Works in LLMs: Multiple Attempts and Feedback:The model generates several responses for a given question, much like a student offering different answers during a discussion.


Reward System:
Human reviewers (acting as the teacher) score these responses. Good answers earn “rewards” while poor answers receive lower scores. Think of it as a gold star system where each star helps the student understand what they did well.


Adjusting Future Responses:
Based on these scores, the model learns to favor responses similar to the highly scored ones. This is similar to the child learning which arguments or explanations are most persuasive or correct.


Two Methods – PPO and DPO:


PPO (Proximal Policy Optimization):
Imagine a situation where the teacher not only gives stars but also keeps track of how closely the student's answers match the ideal answers. This keeps the child’s responses balanced between creativity and correctness.


DPO (Direct Preference Optimization):
In another classroom, instead of tracking each detail, the teacher simply praises or gives feedback based on overall performance. This streamlines the process and works well when there isn’t time for detailed monitoring. Multiple Attempts and Feedback:The model generates several responses for a given question, much like a student offering different answers during a discussion. Multiple Attempts and Feedback:The model generates several responses for a given question, much like a student offering different answers during a discussion. Multiple Attempts and Feedback Reward System:
Human reviewers (acting as the teacher) score these responses. Good answers earn “rewards” while poor answers receive lower scores. Think of it as a gold star system where each star helps the student understand what they did well. Reward System: Reward System: Human reviewers (acting as the teacher) score these responses. Good answers earn “rewards” while poor answers receive lower scores. Think of it as a gold star system where each star helps the student understand what they did well. Adjusting Future Responses:
Based on these scores, the model learns to favor responses similar to the highly scored ones. This is similar to the child learning which arguments or explanations are most persuasive or correct. Adjusting Future Responses: Adjusting Future Responses: Based on these scores, the model learns to favor responses similar to the highly scored ones. This is similar to the child learning which arguments or explanations are most persuasive or correct. Two Methods – PPO and DPO:


PPO (Proximal Policy Optimization):
Imagine a situation where the teacher not only gives stars but also keeps track of how closely the student's answers match the ideal answers. This keeps the child’s responses balanced between creativity and correctness.


DPO (Direct Preference Optimization):
In another classroom, instead of tracking each detail, the teacher simply praises or gives feedback based on overall performance. This streamlines the process and works well when there isn’t time for detailed monitoring. Two Methods – PPO and DPO: Two Methods – PPO and DPO: PPO (Proximal Policy Optimization):
Imagine a situation where the teacher not only gives stars but also keeps track of how closely the student's answers match the ideal answers. This keeps the child’s responses balanced between creativity and correctness.


DPO (Direct Preference Optimization):
In another classroom, instead of tracking each detail, the teacher simply praises or gives feedback based on overall performance. This streamlines the process and works well when there isn’t time for detailed monitoring. PPO (Proximal Policy Optimization):
Imagine a situation where the teacher not only gives stars but also keeps track of how closely the student's answers match the ideal answers. This keeps the child’s responses balanced between creativity and correctness. PPO (Proximal Policy Optimization): PPO (Proximal Policy Optimization) Imagine a situation where the teacher not only gives stars but also keeps track of how closely the student's answers match the ideal answers. This keeps the child’s responses balanced between creativity and correctness. DPO (Direct Preference Optimization):
In another classroom, instead of tracking each detail, the teacher simply praises or gives feedback based on overall performance. This streamlines the process and works well when there isn’t time for detailed monitoring. DPO (Direct Preference Optimization): DPO (Direct Preference Optimization) In another classroom, instead of tracking each detail, the teacher simply praises or gives feedback based on overall performance. This streamlines the process and works well when there isn’t time for detailed monitoring. Further Reading for Advanced Users: Further Reading for Advanced Users: For those interested in delving deeper into the methodologies behind InstructGPT and RLHF (the main papers I was focused on in this article), the following papers are recommended: InstructGPT Paper: "Training language models to follow instructions with human feedback" by Ouyang et al. (2022). This paper outlines the process of aligning language models with user intent through fine-tuning with human feedback. Link to paper
PPO and DPO Techniques:

"Policy Optimization with RLHF — PPO/DPO/ORPO" by Sulbha Jindal (2024). This article provides an overview of various policy optimization techniques in RLHF, including PPO and DPO. Link to article
"Unpacking DPO and PPO: Disentangling Best Practices for Learning from Preference Feedback" (2024). This paper discusses the differences and best practices between DPO and PPO in the context of RLHF. Link to paper InstructGPT Paper: "Training language models to follow instructions with human feedback" by Ouyang et al. (2022). This paper outlines the process of aligning language models with user intent through fine-tuning with human feedback. Link to paper InstructGPT Paper: Link to paper PPO and DPO Techniques:

"Policy Optimization with RLHF — PPO/DPO/ORPO" by Sulbha Jindal (2024). This article provides an overview of various policy optimization techniques in RLHF, including PPO and DPO. Link to article
"Unpacking DPO and PPO: Disentangling Best Practices for Learning from Preference Feedback" (2024). This paper discusses the differences and best practices between DPO and PPO in the context of RLHF. Link to paper PPO and DPO Techniques: "Policy Optimization with RLHF — PPO/DPO/ORPO" by Sulbha Jindal (2024). This article provides an overview of various policy optimization techniques in RLHF, including PPO and DPO. Link to article
"Unpacking DPO and PPO: Disentangling Best Practices for Learning from Preference Feedback" (2024). This paper discusses the differences and best practices between DPO and PPO in the context of RLHF. Link to paper "Policy Optimization with RLHF — PPO/DPO/ORPO" by Sulbha Jindal (2024). This article provides an overview of various policy optimization techniques in RLHF, including PPO and DPO. Link to article Link to article "Unpacking DPO and PPO: Disentangling Best Practices for Learning from Preference Feedback" (2024). This paper discusses the differences and best practices between DPO and PPO in the context of RLHF. Link to paper Link to paper These resources offer a comprehensive understanding of the techniques and advancements in training LLMs to align with human preferences.

Decoding the Magic: How Machines Master Human Language

Table of Links Abstract and 1 Introduction 2 COCOGEN: Representing Commonsense structures with code and 2.1 Converting (T,G) into Python code 2.2 Few-shot prompting for generating G 3 Evaluation and 3.1 Experimental setup 3.2 Script generation: PROSCRIPT 3.3 Entity state tracking: PROPARA 3.4 Argument graph generation: EXPLAGRAPHS 4 Analysis 5 Related work 6 Conclusion, Acknowledgments, Limitations, and References A Few-shot models size estimates B Dynamic prompt Creation C Human Evaluation D Dataset statistics E Sample outputs F Prompts G Designing Python class for a structured task H Impact of Model size I Variation in prompts C Human Evaluation Out of the four tasks used in this work, PROSCRIPT edge prediction and PROPARA have only one possible correct value. Thus, following prior work, we report the automated, standard metrics for these tasks. For EXPLAGRAPHS, we use model-based metrics proposed by Saha et al. (2021), which were found to have a high correlation with human judgments. For PROSCRIPT graph generation, we conducted an exhaustive automated evaluation that separately scores the correctness of the nodes and the correctness of the edges. However, automated metrics are limited in their ability to evaluate model-generated output. Thus, to further investigate the quality of outputs, we conduct a human evaluation to compare the outputs generated by COCOGEN and DAVINCI. We sampled 20 examples, and three of the authors performed the evaluation. Annotators were shown two graphs (generated by COCOGEN and DAVINCI) and were asked to select one they thought was better regarding relevance and correctness. The selection for each criterion was made independently: the same graph could The annotations were done separately: the same graph could have more relevant nodes (higher relevance) but may not be correct. The identity of the model that generated each graph (COCOGEN or DAVINCI) was shuffled and unknown to the evaluators. The results in Table 11 indicate that human evaluation is closely correlated with the automated metrics: for EXPLAGRAPHS, annotators found the graphs generated by COCOGEN to be more relevant and correct. We find that DAVINCI often fails to recover semantic relations between nodes in the argument graphs. For example, consider a belief (B) urbanization harms natural habitats for the animals in the world. We want to generate a graph that can counter this belief with the argument (A) urbanization causes increase in jobs. For the same prompt, COCOGEN generated (urbanization; causes; increase in jobs); (increase in jobs; has context; good); (good; not capable of; harms) whereas DAVINCI generated (jobs; not harms; natural habitats) → (natural habitats; not part of; animals). Note that DAVINCI successfully recovered relevant events (“natural habitat” “animals”) but arranged them in incorrect relations. For PROSCRIPT, the human evaluation shows that COCOGEN and DAVINCI have complementary strengths, while COCOGEN generally produces more relevant and correct outputs. D Dataset statistics Dataset statistics are shown in Table 12. The test split for EXPLAGRAPHS is not available, so we evaluate on the validation split. For PROSCRIPT, we obtained the test splits from the authors. E Sample outputs Sample outputs from COCOGEN for all the tasks are located at https://github.com/ madaan/CoCoGen/tree/main/outputs. Representative examples from each task are presented in Figure 5. Surprisingly, COCOGEN (CODEX with a Python prompt) generates syntactically valid Python graphs that are similar to the task graphs/tables in nearly 100% of the cases. F Prompts The prompts for each tasks are present at this anonymous URL: PROSCRIPT script-generation: https://github.com/madaan/CoCoGen/tree/main/data/proscript_ script_generation/prompt.txt


PROSCRIPT edge-prediction: https://github.com/madaan/CoCoGen/tree/main/data/proscript_edge_ prediction/prompt.txt


PROPARA: https://github.com/madaan/CoCoGen/tree/main/data/ explagraphs/prompt.txt


EXPLAGRAPHS: https://github.com/madaan/CoCoGen/tree/main/data/ explagraphs/prompt.txt These prompts are also present in the attached supplementary material, and can be found in the data folder under respective task sub-directories. This paper is available on arxiv under CC BY 4.0 DEED license. Authors:
(1) Aman Madaan, Language Technologies Institute, Carnegie Mellon University, USA (amadaan@cs.cmu.edu);
(2) Shuyan Zhou, Language Technologies Institute, Carnegie Mellon University, USA (shuyanzh@cs.cmu.edu);
(3) Uri Alon, Language Technologies Institute, Carnegie Mellon University, USA (ualong@cs.cmu.edu);
(4) Yiming Yang, Language Technologies Institute, Carnegie Mellon University, USA (yiming@cs.cmu.edu);
(5) Graham Neubig, Language Technologies Institute, Carnegie Mellon University, USA (gneubig@cs.cmu.edu). Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 COCOGEN: Representing Commonsense structures with code and 2.1 Converting (T,G) into Python code 2 COCOGEN: Representing Commonsense structures with code and 2.1 Converting (T,G) into Python code 2.2 Few-shot prompting for generating G 2.2 Few-shot prompting for generating G 3 Evaluation and 3.1 Experimental setup 3 Evaluation and 3.1 Experimental setup 3.2 Script generation: PROSCRIPT 3.2 Script generation: PROSCRIPT 3.3 Entity state tracking: PROPARA 3.3 Entity state tracking: PROPARA 3.4 Argument graph generation: EXPLAGRAPHS 3.4 Argument graph generation: EXPLAGRAPHS 4 Analysis 4 Analysis 5 Related work 5 Related work 6 Conclusion, Acknowledgments, Limitations, and References 6 Conclusion, Acknowledgments, Limitations, and References A Few-shot models size estimates A Few-shot models size estimates B Dynamic prompt Creation B Dynamic prompt Creation C Human Evaluation C Human Evaluation D Dataset statistics D Dataset statistics E Sample outputs E Sample outputs F Prompts F Prompts G Designing Python class for a structured task G Designing Python class for a structured task H Impact of Model size H Impact of Model size I Variation in prompts I Variation in prompts C Human Evaluation Out of the four tasks used in this work, PROSCRIPT edge prediction and PROPARA have only one possible correct value. Thus, following prior work, we report the automated, standard metrics for these tasks. For EXPLAGRAPHS, we use model-based metrics proposed by Saha et al. (2021), which were found to have a high correlation with human judgments. For PROSCRIPT graph generation, we conducted an exhaustive automated evaluation that separately scores the correctness of the nodes and the correctness of the edges. However, automated metrics are limited in their ability to evaluate model-generated output. Thus, to further investigate the quality of outputs, we conduct a human evaluation to compare the outputs generated by COCOGEN and DAVINCI. We sampled 20 examples, and three of the authors performed the evaluation. Annotators were shown two graphs (generated by COCOGEN and DAVINCI) and were asked to select one they thought was better regarding relevance and correctness. The selection for each criterion was made independently: the same graph could The annotations were done separately: the same graph could have more relevant nodes (higher relevance) but may not be correct. The identity of the model that generated each graph (COCOGEN or DAVINCI) was shuffled and unknown to the evaluators. The results in Table 11 indicate that human evaluation is closely correlated with the automated metrics: for EXPLAGRAPHS, annotators found the graphs generated by COCOGEN to be more relevant and correct. We find that DAVINCI often fails to recover semantic relations between nodes in the argument graphs. For example, consider a belief (B) urbanization harms natural habitats for the animals in the world. We want to generate a graph that can counter this belief with the argument (A) urbanization causes increase in jobs. For the same prompt, COCOGEN generated (urbanization; causes; increase in jobs); (increase in jobs; has context; good); (good; not capable of; harms) whereas DAVINCI generated (jobs; not harms; natural habitats) → (natural habitats; not part of; animals). Note that DAVINCI successfully recovered relevant events (“natural habitat” “animals”) but arranged them in incorrect relations. For PROSCRIPT, the human evaluation shows that COCOGEN and DAVINCI have complementary strengths, while COCOGEN generally produces more relevant and correct outputs. D Dataset statistics Dataset statistics are shown in Table 12. The test split for EXPLAGRAPHS is not available, so we evaluate on the validation split. For PROSCRIPT, we obtained the test splits from the authors. E Sample outputs Sample outputs from COCOGEN for all the tasks are located at https://github.com/ madaan/CoCoGen/tree/main/outputs. Representative examples from each task are presented in Figure 5. Surprisingly, COCOGEN (CODEX with a Python prompt) generates syntactically valid Python graphs that are similar to the task graphs/tables in nearly 100% of the cases. F Prompts The prompts for each tasks are present at this anonymous URL: PROSCRIPT script-generation: https://github.com/madaan/CoCoGen/tree/main/data/proscript_ script_generation/prompt.txt PROSCRIPT edge-prediction: https://github.com/madaan/CoCoGen/tree/main/data/proscript_edge_ prediction/prompt.txt PROPARA: https://github.com/madaan/CoCoGen/tree/main/data/ explagraphs/prompt.txt EXPLAGRAPHS: https://github.com/madaan/CoCoGen/tree/main/data/ explagraphs/prompt.txt PROSCRIPT script-generation: https://github.com/madaan/CoCoGen/tree/main/data/proscript_ script_generation/prompt.txt PROSCRIPT script-generation: https: //github.com/madaan/CoCoGen/ tree/main/data/proscript_ script_generation/prompt.txt //github.com/madaan/CoCoGen/ PROSCRIPT edge-prediction: https://github.com/madaan/CoCoGen/tree/main/data/proscript_edge_ prediction/prompt.txt PROSCRIPT edge-prediction: https: //github.com/madaan/CoCoGen/ tree/main/data/proscript_edge_ prediction/prompt.txt //github.com/madaan/CoCoGen/ PROPARA: https://github.com/madaan/CoCoGen/tree/main/data/ explagraphs/prompt.txt PROPARA: https://github.com/ madaan/CoCoGen/tree/main/data/ explagraphs/prompt.txt https://github.com/ EXPLAGRAPHS: https://github.com/madaan/CoCoGen/tree/main/data/ explagraphs/prompt.txt EXPLAGRAPHS: https://github.com/ madaan/CoCoGen/tree/main/data/ explagraphs/prompt.txt https://github.com/ These prompts are also present in the attached supplementary material, and can be found in the data folder under respective task sub-directories. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv Authors: (1) Aman Madaan, Language Technologies Institute, Carnegie Mellon University, USA (amadaan@cs.cmu.edu); (2) Shuyan Zhou, Language Technologies Institute, Carnegie Mellon University, USA (shuyanzh@cs.cmu.edu); (3) Uri Alon, Language Technologies Institute, Carnegie Mellon University, USA (ualong@cs.cmu.edu); (4) Yiming Yang, Language Technologies Institute, Carnegie Mellon University, USA (yiming@cs.cmu.edu); (5) Graham Neubig, Language Technologies Institute, Carnegie Mellon University, USA (gneubig@cs.cmu.edu). Authors: Authors: (1) Aman Madaan, Language Technologies Institute, Carnegie Mellon University, USA (amadaan@cs.cmu.edu); (2) Shuyan Zhou, Language Technologies Institute, Carnegie Mellon University, USA (shuyanzh@cs.cmu.edu); (3) Uri Alon, Language Technologies Institute, Carnegie Mellon University, USA (ualong@cs.cmu.edu); (4) Yiming Yang, Language Technologies Institute, Carnegie Mellon University, USA (yiming@cs.cmu.edu); (5) Graham Neubig, Language Technologies Institute, Carnegie Mellon University, USA (gneubig@cs.cmu.edu).

COCOGEN vs DAVINCI: A Human Evaluation of Structured Commonsense Graph Generation

As AI becomes more advanced and capable, it’s being applied to a growing number of use cases. There are more and more instances of AI being used as a tool by software developers and the recent emergence of ‘vibe coding’ has certainly been grabbing a lot of media attention. Just a few months ago, Andrej Karpathy coined this term to describe the process of asking LLMs such as ChatGPT and Claude to write code by giving it simple, natural language instructions. While this approach is capable of generating working code, it's not yet able to produce advanced software programs unassisted – not yet anyway. But LLMs are being used as tools by developers as they build software intended for real-world use cases. Andrej Karpathy Andrej Karpathy are being used are being used The key here is that even though the LLMs are writing the code, the people commanding it to do so are still capable of understanding what's been written. If they couldn't, then debugging, maintaining and securing the codebase effectively would be impossible. AI is an assistant that is being used to make suggestions, rather than provide answers that can be trusted 100%. Like in other parts of software development, human oversight still plays an important role. It's especially vital for humans to be involved in the testing of usability and accessibility in software – but it’s also critical that humans remain the master of all software testing processes. LLMs can write code – but are poor usability testers LLMs can write code – but are poor usability testers LLMs have been trained on the vast reams of data, images and text from a magnitude of sources around the internet and are designed to behave in an expected manner when carrying out tasks such as writing an email, for example. They’ve been trained to code by ingesting simple programs written by human software engineers. And LLMs have also been trained to understand imperfect input. They’re designed to try and understand the user at all times, so they will attempt to fill in the gaps when the user skips words or doesn’t express themselves clearly. And this ‘understand at all costs’ approach is in conflict with the idea of usability testing. Usability testing requires developers to switch their focus to specific frames of reference where they ask if the behaviour of software is clear, or work with a focus group to establish that the product is intuitively usable. LLMs are not well-positioned to test for any flaws in usability as they are trained to understand everything – for them, there’s no difference between bad and good. But for human users, usability is judged in terms of whether software works in line with their preconceptions, which are often based on previous experience. You can’t ask the humans to re-learn their approach to using software – the program has to meet their expectations. Even if software comes with documentation, you can’t rely on users to read these instructions. That’s why a significant part of usability testing is making sure that software behaves with an implied knowledge of how things should work. should Accessibility testing requires human oversight Accessibility testing requires human oversight Poor accessibility is something that impacts many different groups of people — not just those with permanent disabilities. It also includes temporary impairments (like a broken arm) and situational ones (like bright sunlight or background noise). These concepts, popularized by Microsoft’s Inclusive Design work in the mid-2010s, show that accessibility isn’t just about edge cases — it’s about building software that works for everyone, in any context. Currently, LLMs are unlikely to have been trained to recognise poor colour contrast that could affect users with visual impairments – because they are trained to understand everything, they will see every pixel as perfect. Nor will they pick up issues with bad keyboard navigation, text with poor readability and a lack of zoom functionality. However, it is entirely possible to train models to check for accessibility issues and offer advice to developers. But LLMs could only ever provide assistance in these scenarios – human oversight would be necessary to understand what accessibility aspects we are looking to address. AI cannot manage risk or be held accountable AI cannot manage risk or be held accountable With large, complex software applications, systematic testing is necessary prior to release. However, exhaustive testing isn't possible – there are just too many potential use cases and inputs that the application could encounter. That means decisions have to be made about what to test – and what tests don't need to be carried out. This is a vital part of QA – and it cannot be left to AI. QA teams must draw on their experience, knowledge, and judgement to make informed decisions. They must plan, prioritize and make risk assessments. They will understand that there are trade-offs that need to be made – but they can make accountable decisions and justify them to relevant stakeholders. If AI makes these decisions, then there's no accountability if something goes wrong. Where AI  can streamline software testing and test management Where AI  can streamline software testing and test management There are many areas of the software testing process where AI can be of assistance. There’s a lot of heavy lifting involved in software testing such as preparing test data, suggesting the specific scenarios that need to be tested, and writing automated tests. AI agents are already capable of performing some basic tests today, and it’s likely they’ll take on even more complex testing tasks in the near future. And AI also has utility for software test management. It can be used to help with writing test documentation, and to suggest ideas for key areas to test. It can be used to research other, similar software projects to surface relevant data and insights that could assist the testing. And it can also perform data analysis of the test results, creating reports for relevant stakeholders quickly so they don't have to wait for human operatives to compile their reports. Humans are the end users – so human expertise is needed Humans are the end users – so human expertise is needed Quality in software is non-negotiable. We’ve all heard stories about disastrous software launches where products have been shipped with massive issues that have made them completely unusable for a certain group of people. For now, usability and accessibility are human domains so it is natural to use humans for testing these aspects. In the era of AI, it's important to maintain human oversight. If AI is used in software creation, testing or test management it is vital to check the results. QA teams need to be the ones that take accountability for software test management processes. Only they have the domain knowledge and understanding of business priorities to determine test coverage and acceptable levels of risk. While AI can support these processes with data and recommendations, it cannot currently make business critical decisions on its own.

The Human Testing Element is Still Important Despite Advances in AI 

It’s 2025, and AI has officially checked into every part of healthcare—from the front desk scheduler to the operating room. It reminds your patient to refill their medication and helps flag cancer on a slide. It doesn’t complain, take breaks or call in sick. It’s fast, tireless and… maybe not enough? Because for all its progress, we’re still seeing AI systems stumble in real clinical settings. It might not handle edge cases as effectively as we’d like. It misses context that experienced clinicians could catch before their first cup of coffee. And no, this isn’t about fearing automation or fighting the march of progress. I’ve spent years building and implementing these systems, and I know firsthand that AI can be an incredibly effective collaborator… but only when it’s built to work with people. with It might be true that in some industries, you can plug AI in and get pretty incredible outcomes, with a few duds here and there. But healthcare doesn’t have much margin for error. What matters is good judgment and context, and that takes human input. Let’s talk about what it means to build AI infrastructure that puts humans in the loop by design. AI Isn’t Failing Because It’s Wrong—It Just Needs Help Starting with data, because it’s where your AI will: AI models want structure. Healthcare gives them chaos. Any clinician is familiar with the digital mess behind the scenes, ranging from poorly scanned PDFs of half-filled forms to different EHRs calling the same condition by different names. An algorithm might see an “abnormal” lab and sound the alarm, not realizing the patient just had surgery last week. Or maybe there’s a coffee stain over a 0. According to ECRI’s 2025 safety report, AI-driven clinical errors, ranging from misdiagnoses to inappropriate treatment recommendations, are now considered one of the top patient risks. And to be clear, the models aren’t malicious. But letting them fly blind and build a foundation on human error is asking for trouble. We need to provide a scaffolding for uncertainty and ambiguity. And coffee stains. ECRI’s 2025 safety report ECRI’s 2025 safety report This is a huge context problem, and wouldn’t you know it: humans are very good at context. AI sees the signal. Humans know the story. The best results happen when AI proposes and humans decide. False Positives Create More Work Let’s follow that same example further. What happens when AI goes off on its own? A radiology AI cries wolf and flags nodules of interest in every single scan, including the ones that are just shadows or scar tissue. Trust falls off, teams start to ignore it. Then usage drops. At some point, the system gets shelved. So much for a decision support system. decision support system This is a classic tech-in-medicine problem. A 2023 NPJ Digital Medicine study flagged this exact pattern. AI without oversight created so many false positives that it added to staff burden, increasing alert fatigue and driving its users to disengage. 2023 NPJ Digital Medicine study 2023 NPJ Digital Medicine study We’ve seen in the field again and again that a model can’t triage itself, and it certainly won’t know whether its prediction is helpful or harmful. That’s the role of a human reviewer: someone needs to be partnered and trained to verify and decide on what action needs to happen next. It’s an important layer of checks and balances that should drive every AI-assisted medical decision. Black-Box Models Don’t Fly in Healthcare This shouldn’t come as a surprise: Even when the model is right, doctors need to know why. If your AI tells a provider that patient X is at high risk of deterioration, they’re going to ask: based on what evidence? Is it recent vitals? A pattern? If the model can’t explain itself, or if it dumps raw scores with no context, it’s not providing meaningful or actionable output. A 2024 American Medical Association survey found that nearly half of physicians rank transparency and human oversight as top criteria for trusting an AI tool. Accuracy is one thing, but someone has to be accountable, and it will always be the practicing physician. 2024 American Medical Association survey 2024 American Medical Association survey That’s where human-in-the-loop design really counts. Reviewers vet the alert, while clinical leads can escalate or override. Everyone sees what changed and why, and knows who to turn to for questions. Introducing HITL turns your AI from a black box into a system with feedback and contextual memory. Regulations Are Catching Up The FDA, WHO, and pretty much every major regulatory body now agrees: AI in healthcare needs oversight, even in day-to-day ops. The FDA’s 2025 guidance emphasizes traceability and documented human oversight at each major decision. The World Health Organization warns against deploying AI without rigorous evaluation and expert supervision. A very recent U.S. Senate investigation found that fully autonomous AI systems in insurance led to care denials 16 times higher than expected, calling for immediate policy intervention. I won’t mince words: If you’re not building with humans in the loop, you’re inviting much more scrutiny in an already audit-heavy industry. But Humans Are Slow... Right? Not If You Design for It There’s that persistent myth that adding people into an AI system slows things down. But that’s only true if the system wasn’t designed to support collaboration in the first place. A few ways to do it without bogging down your ops: Set confidence thresholds. Only flag low-confidence cases, where human input adds value.
Build effective tooling. Reviewers need intuitive and actionable dashboards, not a 50-column CSV.
Track outcomes. Log every review, override and resolution. You’ll be ready for the next audit and simultaneously build a goldmine of data for model improvement. Set confidence thresholds. Only flag low-confidence cases, where human input adds value. Build effective tooling. Reviewers need intuitive and actionable dashboards, not a 50-column CSV. Track outcomes. Log every review, override and resolution. You’ll be ready for the next audit and simultaneously build a goldmine of data for model improvement. It’s easy to start thinking it’s bureaucratic overhead, but a well-designed HITL system becomes a learning system. The more it runs, the better it gets. If You Want to Scale AI in Healthcare, Start With Humans Let’s recap. We don’t need to guess where things go wrong. AI models are left in a vacuum, trusted to succeed with poor inputs, then start eroding trust through noisy outputs. Clinicians tune out and the systems get shut off. But healthcare has never been about hitting 99% accuracy in a sandbox. We focus on making the right call when everything’s messy to begin with. That’s what makes human judgment irreplaceable, and why we have the responsibility to design with the people who still carry the weight of patient care. By all means, let AI handle the routine stuff. Automate the admin, surface insights, reduce friction where it makes sense. But when it comes to escalation and practicing good judgment, build systems that know when to hand it back to a human. In the future, we might start realizing the smartest thing an AI can do is know when to ask for help.

AI in Healthcare Works—But Only When Humans Stay in the Loop

Table of Links Abstract and 1 Introduction 2 Related Work 3 Method and 3.1 Phase 1: Taxonomy Generation 3.2 Phase 2: LLM-Augmented Text Classification 4 Evaluation Suite and 4.1 Phase 1 Evaluation Strategies 4.2 Phase 2 Evaluation Strategies 5 Experiments and 5.1 Data 5.2 Taxonomy Generation 5.3 LLM-Augmented Text Classification 5.4 Summary of Findings and Suggestions 6 Discussion and Future Work, and References A. Taxonomies B. Additional Results C. Implementation Details D. Prompt Templates 4.2 Phase 2 Evaluation Strategies To quantitatively evaluate text classification, we create a benchmark dataset with reliable ground-truth annotations as follows: Task and Annotation Reliability. We first assess the reliability of the label assignment task and the human annotations by involving multiple human annotators and calculating the interrater agreement (Cohen’s Kappa [6] between two raters and Fleiss’ Kappa [7] among multiple raters). We then resolve disagreements between human annotations by either voting or deliberation, and obtain a consensus human annotation for each instance. Then we use an LLM as an additional annotator to perform the same label assignment task, and measure the agreement between the LLM annotation and the consensus human label. Intuitively, this agreement captures how well the LLM is aligned with (the majority of) human annotators and how reliable it is for this label assignment task. Classification Metrics. We apply both human and LLM annotations on a small-scale corpus sample and calculate the conventional multiclass and multilabel classification metrics (e.g., Accuracy, F1) with human annotations as the ground truth. These metrics evaluate how the label classifier is aligned with human preferences on a small subset of the corpus. We then apply the LLM annotator on a larger-scale corpus sample and leverage the resulting annotations as the oracle to calculate the same classification metrics. These metrics enable a comprehensive diagnosis of the label classifier performance at scale on different aspects of the corpus, such as domains, languages, and time ranges. In practice, we recommend leveraging both human evaluation and LLM-based metrics as a holistic evaluation suite, while also taking into account the task and annotation reliability. This approach can help us identify and mitigate the possible bias that may arise from either method or be affected by the task complexity, and enable us to scale up the evaluation and annotation to a large corpus sample with confidence, thus obtaining more robust and informative evaluation results. This paper is available on arxiv under CC BY-NC-ND 4.0 DEED license. Authors:
(1) Mengting Wan, Microsoft Corporation and Microsoft Corporation;
(2) Tara Safavi (Corresponding authors), Microsoft Corporation;
(3) Sujay Kumar Jauhar, Microsoft Corporation;
(4) Yujin Kim, Microsoft Corporation;
(5) Scott Counts, Microsoft Corporation;
(6) Jennifer Neville, Microsoft Corporation;
(7) Siddharth Suri, Microsoft Corporation;
(8) Chirag Shah, University of Washington and Work done while working at Microsoft;
(9) Ryen W. White, Microsoft Corporation;
(10) Longqi Yang, Microsoft Corporation;
(11) Reid Andersen, Microsoft Corporation;
(12) Georg Buscher, Microsoft Corporation;
(13) Dhruv Joshi, Microsoft Corporation;
(14) Nagu Rangan, Microsoft Corporation. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Work 2 Related Work 3 Method and 3.1 Phase 1: Taxonomy Generation 3 Method and 3.1 Phase 1: Taxonomy Generation 3.2 Phase 2: LLM-Augmented Text Classification 3.2 Phase 2: LLM-Augmented Text Classification 4 Evaluation Suite and 4.1 Phase 1 Evaluation Strategies 4 Evaluation Suite and 4.1 Phase 1 Evaluation Strategies 4.2 Phase 2 Evaluation Strategies 4.2 Phase 2 Evaluation Strategies 5 Experiments and 5.1 Data 5 Experiments and 5.1 Data 5.2 Taxonomy Generation 5.2 Taxonomy Generation 5.3 LLM-Augmented Text Classification 5.3 LLM-Augmented Text Classification 5.4 Summary of Findings and Suggestions 5.4 Summary of Findings and Suggestions 6 Discussion and Future Work, and References 6 Discussion and Future Work, and References A. Taxonomies A. Taxonomies B. Additional Results B. Additional Results C. Implementation Details C. Implementation Details D. Prompt Templates D. Prompt Templates 4.2 Phase 2 Evaluation Strategies To quantitatively evaluate text classification, we create a benchmark dataset with reliable ground-truth annotations as follows: Task and Annotation Reliability. We first assess the reliability of the label assignment task and the human annotations by involving multiple human annotators and calculating the interrater agreement (Cohen’s Kappa [6] between two raters and Fleiss’ Kappa [7] among multiple raters). We then resolve disagreements between human annotations by either voting or deliberation, and obtain a consensus human annotation for each instance. Then we use an LLM as an additional annotator to perform the same label assignment task, and measure the agreement between the LLM annotation and the consensus human label. Intuitively, this agreement captures how well the LLM is aligned with (the majority of) human annotators and how reliable it is for this label assignment task. Task and Annotation Reliability. Classification Metrics. We apply both human and LLM annotations on a small-scale corpus sample and calculate the conventional multiclass and multilabel classification metrics (e.g., Accuracy, F1) with human annotations as the ground truth. These metrics evaluate how the label classifier is aligned with human preferences on a small subset of the corpus. We then apply the LLM annotator on a larger-scale corpus sample and leverage the resulting annotations as the oracle to calculate the same classification metrics. These metrics enable a comprehensive diagnosis of the label classifier performance at scale on different aspects of the corpus, such as domains, languages, and time ranges. Classification Metrics. In practice, we recommend leveraging both human evaluation and LLM-based metrics as a holistic evaluation suite, while also taking into account the task and annotation reliability. This approach can help us identify and mitigate the possible bias that may arise from either method or be affected by the task complexity, and enable us to scale up the evaluation and annotation to a large corpus sample with confidence, thus obtaining more robust and informative evaluation results. This paper is available on arxiv under CC BY-NC-ND 4.0 DEED license. This paper is available on arxiv under CC BY-NC-ND 4.0 DEED license. available on arxiv Authors: (1) Mengting Wan, Microsoft Corporation and Microsoft Corporation; (2) Tara Safavi (Corresponding authors), Microsoft Corporation; (3) Sujay Kumar Jauhar, Microsoft Corporation; (4) Yujin Kim, Microsoft Corporation; (5) Scott Counts, Microsoft Corporation; (6) Jennifer Neville, Microsoft Corporation; (7) Siddharth Suri, Microsoft Corporation; (8) Chirag Shah, University of Washington and Work done while working at Microsoft; (9) Ryen W. White, Microsoft Corporation; (10) Longqi Yang, Microsoft Corporation; (11) Reid Andersen, Microsoft Corporation; (12) Georg Buscher, Microsoft Corporation; (13) Dhruv Joshi, Microsoft Corporation; (14) Nagu Rangan, Microsoft Corporation. Authors: Authors: (1) Mengting Wan, Microsoft Corporation and Microsoft Corporation; (2) Tara Safavi (Corresponding authors), Microsoft Corporation; (3) Sujay Kumar Jauhar, Microsoft Corporation; (4) Yujin Kim, Microsoft Corporation; (5) Scott Counts, Microsoft Corporation; (6) Jennifer Neville, Microsoft Corporation; (7) Siddharth Suri, Microsoft Corporation; (8) Chirag Shah, University of Washington and Work done while working at Microsoft; (9) Ryen W. White, Microsoft Corporation; (10) Longqi Yang, Microsoft Corporation; (11) Reid Andersen, Microsoft Corporation; (12) Georg Buscher, Microsoft Corporation; (13) Dhruv Joshi, Microsoft Corporation; (14) Nagu Rangan, Microsoft Corporation.

Evaluating TnT-LLM Text Classification: Human Agreement and Scalable LLM Metrics

Evaluating TnT-LLM: Automatic, Human, and LLM-Based Assessment

Authors:
(1) Yanpeng Ye, School of Computer Science and Engineering, University of New South Wales, Kensington, NSW, Australia, GreenDynamics Pty. Ltd, Kensington, NSW, Australia, and these authors contributed equally to this work;
(2) Jie Ren, GreenDynamics Pty. Ltd, Kensington, NSW, Australia, Department of Materials Science and Engineering, City University of Hong Kong, Hong Kong, China, and these authors contributed equally to this work;
(3) Shaozhou Wang, GreenDynamics Pty. Ltd, Kensington, NSW, Australia (shaozhou@greendynamics.com.au);
(4) Yuwei Wan, GreenDynamics Pty. Ltd, Kensington, NSW, Australia and Department of Linguistics and Translation, City University of Hong Kong, Hong Kong, China;
(5) Imran Razzak, School of Computer Science and Engineering, University of New South Wales, Kensington, NSW, Australia;
(6) Tong Xie, GreenDynamics Pty. Ltd, Kensington, NSW, Australia and School of Photovoltaic and Renewable Energy Engineering, University of New South Wales, Kensington, NSW, Australia (tong@greendynamics.com.au);
(7) Wenjie Zhang, School of Computer Science and Engineering, University of New South Wales, Kensington, NSW, Australia (wenjie.zhang@unsw.edu.au). Editor’s note: This article is part of a broader study. You’re reading Part 3 of 9. Read the rest below. Table of Links Abstract and Introduction
Methods
Data preparation and schema design
LLMs training, evaluation and inference
Entity resolution
Knowledge graph construction
Result
Discussion
Conclusion and References Data preparation and schema design Material experts undertook the annotation of nine distinct categories from 75 abstracts of research papers, act as LLM’s training dataset. These categories encompassed core labels - "Name," "Formula," or "Acronym" - which consistently represent specific materials across diverse papers, albeit in varied forms. Additionally, supplementary labels included "Descriptor," "Structure/Phase," "Application," "Property," "Synthesis," and "Characterization." The "Property" label served to delineate quantitative properties such as ’length,’ ’specific surface area,’ and ’mass,’ whereas the "Descriptor" label encapsulated qualitative attributes like ’stable,’ ’vertically,’ and ’safe.’ Since from the application aspects, developers always concern about its own properties, we also annotated the essential parameters as "Property" directly linked to specific "Application". For instance, attributes like ’specific capacity’ for lithium-ion batteries, ’energy conversion efficiency’ for solar cells, and ’H2 production rate’ for hydrogen evolution reaction were systematically annotated. As the Figure 2 shows, the central node, "Material," is pivotal and connects to nodes that describe its nomenclature, composition, and various attributes. Specifically, "Material" is consisted of "Formula", "Name" and "Acronym". "Structure/Phase" describing its physical form; "Application" denoting its practical uses; and "Property" outlining its inherent characteristics. The "Material" node is also associated with "Descriptor" node that provide additional qualitative information. Moreover, the "Application" node branches out to "Property", "Descriptor" and a "Domain" suggesting further specification and contextual relevance of application. To save the source of nodes and relations, each node is connected to a "Digital Object Identifier (DOI)" node. If we want to determine the source article from which the relation is derived, we can obtain the result by querying the intersection of "DOI" node neighbors at both ends of the relation. Given the inherent complexity of sentences and the variability in terminologies across abstracts, a normalization process was employed after the initial extraction. This normalization ensured uniform representation of entities with similar meanings. For example, terms such as ’Lithium-Ion Battery’ and ’Li-ion batteries’ were standardized to ’lithium-ion battery,’ while phrases like ’solution casting method,’ ’solvent post-treatment method,’ and ’solution-based deposition’ were simplified to ’solution-processed.’ All annotated entities underwent this normalization process, ensuring consistency and facilitating effective training of LLMs. The field of functional materials encompasses a wide range of areas. At this stage, our priority is to focus on energy materials. We download 150,000 abstracts of peer-reviewed research articles of energy material science includes battery, solar cell and catalyst from the Web of Science. Each abstract was stored in a JSON file format, structured as "DOI - text," facilitating seamless processing and analysis. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Yanpeng Ye, School of Computer Science and Engineering, University of New South Wales, Kensington, NSW, Australia, GreenDynamics Pty. Ltd, Kensington, NSW, Australia, and these authors contributed equally to this work; (2) Jie Ren, GreenDynamics Pty. Ltd, Kensington, NSW, Australia, Department of Materials Science and Engineering, City University of Hong Kong, Hong Kong, China, and these authors contributed equally to this work; (3) Shaozhou Wang, GreenDynamics Pty. Ltd, Kensington, NSW, Australia (shaozhou@greendynamics.com.au); (4) Yuwei Wan, GreenDynamics Pty. Ltd, Kensington, NSW, Australia and Department of Linguistics and Translation, City University of Hong Kong, Hong Kong, China; (5) Imran Razzak, School of Computer Science and Engineering, University of New South Wales, Kensington, NSW, Australia; (6) Tong Xie, GreenDynamics Pty. Ltd, Kensington, NSW, Australia and School of Photovoltaic and Renewable Energy Engineering, University of New South Wales, Kensington, NSW, Australia (tong@greendynamics.com.au); (7) Wenjie Zhang, School of Computer Science and Engineering, University of New South Wales, Kensington, NSW, Australia (wenjie.zhang@unsw.edu.au). Authors: Authors: (1) Yanpeng Ye, School of Computer Science and Engineering, University of New South Wales, Kensington, NSW, Australia, GreenDynamics Pty. Ltd, Kensington, NSW, Australia, and these authors contributed equally to this work; (2) Jie Ren, GreenDynamics Pty. Ltd, Kensington, NSW, Australia, Department of Materials Science and Engineering, City University of Hong Kong, Hong Kong, China, and these authors contributed equally to this work; (3) Shaozhou Wang, GreenDynamics Pty. Ltd, Kensington, NSW, Australia (shaozhou@greendynamics.com.au); (4) Yuwei Wan, GreenDynamics Pty. Ltd, Kensington, NSW, Australia and Department of Linguistics and Translation, City University of Hong Kong, Hong Kong, China; (5) Imran Razzak, School of Computer Science and Engineering, University of New South Wales, Kensington, NSW, Australia; (6) Tong Xie, GreenDynamics Pty. Ltd, Kensington, NSW, Australia and School of Photovoltaic and Renewable Energy Engineering, University of New South Wales, Kensington, NSW, Australia (tong@greendynamics.com.au); (7) Wenjie Zhang, School of Computer Science and Engineering, University of New South Wales, Kensington, NSW, Australia (wenjie.zhang@unsw.edu.au). Editor’s note: This article is part of a broader study. You’re reading Part 3 of 9. Read the rest below. Editor’s note: This article is part of a broader study. You’re reading Part 3 of 9. Read the rest below. Editor’s note: This article is part of a broader study. You’re reading Part 3 of 9. Read the rest below. Table of Links Abstract and Introduction Methods Data preparation and schema design LLMs training, evaluation and inference Entity resolution Knowledge graph construction Result Discussion Conclusion and References Abstract and Introduction Abstract and Introduction Methods Methods Data preparation and schema design Data preparation and schema design LLMs training, evaluation and inference LLMs training, evaluation and inference Entity resolution Entity resolution Knowledge graph construction Knowledge graph construction Result Result Discussion Discussion Conclusion and References Conclusion and References Data preparation and schema design Material experts undertook the annotation of nine distinct categories from 75 abstracts of research papers, act as LLM’s training dataset. These categories encompassed core labels - "Name," "Formula," or "Acronym" - which consistently represent specific materials across diverse papers, albeit in varied forms. Additionally, supplementary labels included "Descriptor," "Structure/Phase," "Application," "Property," "Synthesis," and "Characterization." The "Property" label served to delineate quantitative properties such as ’length,’ ’specific surface area,’ and ’mass,’ whereas the "Descriptor" label encapsulated qualitative attributes like ’stable,’ ’vertically,’ and ’safe.’ Since from the application aspects, developers always concern about its own properties, we also annotated the essential parameters as "Property" directly linked to specific "Application". For instance, attributes like ’specific capacity’ for lithium-ion batteries, ’energy conversion efficiency’ for solar cells, and ’H2 production rate’ for hydrogen evolution reaction were systematically annotated. As the Figure 2 shows, the central node, "Material," is pivotal and connects to nodes that describe its nomenclature, composition, and various attributes. Specifically, "Material" is consisted of "Formula", "Name" and "Acronym". "Structure/Phase" describing its physical form; "Application" denoting its practical uses; and "Property" outlining its inherent characteristics. The "Material" node is also associated with "Descriptor" node that provide additional qualitative information. Moreover, the "Application" node branches out to "Property", "Descriptor" and a "Domain" suggesting further specification and contextual relevance of application. To save the source of nodes and relations, each node is connected to a "Digital Object Identifier (DOI)" node. If we want to determine the source article from which the relation is derived, we can obtain the result by querying the intersection of "DOI" node neighbors at both ends of the relation. Given the inherent complexity of sentences and the variability in terminologies across abstracts, a normalization process was employed after the initial extraction. This normalization ensured uniform representation of entities with similar meanings. For example, terms such as ’Lithium-Ion Battery’ and ’Li-ion batteries’ were standardized to ’lithium-ion battery,’ while phrases like ’solution casting method,’ ’solvent post-treatment method,’ and ’solution-based deposition’ were simplified to ’solution-processed.’ All annotated entities underwent this normalization process, ensuring consistency and facilitating effective training of LLMs. The field of functional materials encompasses a wide range of areas. At this stage, our priority is to focus on energy materials. We download 150,000 abstracts of peer-reviewed research articles of energy material science includes battery, solar cell and catalyst from the Web of Science. Each abstract was stored in a JSON file format, structured as "DOI - text," facilitating seamless processing and analysis. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

Stories, News & Mentions about HUMAN

Categories

Trending Topics

Stories, News & Mentions about HUMAN

Human’s Stories on HackerNoon

Human's News & Mentions

Categories

Trending Topics