PRefLexOR (Preference-based Recursive Language Modeling for Exploratory Optimization of Reasoning) combines preference optimization with concepts from Reinforcement Learning to enable models to self-teach through iterative reasoning improvements. We propose a recursive learning approach that engages the model in multi-step reasoning, revisiting, and refining intermediate steps before producing a final output in training and inference phases. Through multiple training stages, the model first learns to align its reasoning with accurate decision paths by optimizing the log odds between preferred and non-preferred responses. During this process, PRefLexOR builds a dynamic knowledge graph by generating questions from random text chunks and retrieval-augmentation to contextualize relevant details from the entire training corpus. In the second stage, preference optimization enhances model performance by using rejection sampling to fine-tune reasoning quality by continually producing in-situ training data while masking the reasoning steps. Recursive optimization within a thinking token framework introduces iterative feedback loops, where the model refines reasoning, achieving deeper coherence, consistency, and adaptability. Implemented in small language models with only 3 billion parameters, we should that even tiny models can iteratively teach themselves to reason with greater depth and reflectivity. Our implementation is straightforward and can be incorporated into any existing pretrained LLM. We focus our examples on applications in biological materials science and demonstrate the method in a variety of case studies that range from in-domain to cross-domain applications. Using reasoning strategies that include thinking and reflection modalities we build a multi-agent recursive self-improving inference approach to successively improve responses via repeated sampling in inference time.

Round particles and their properties are easy to describe mathematically. But the less round or spherical the shape, the harder it becomes to make predictions about their behavior. In his doctoral thesis at the Technical University of Kaiserslautern (TUK), Robert Hesse has trained a neural network to automatically determine the packing density and flowability of non-spherical particles. Few particles in nature or in industrial production are exactly round; instead, there are a multitude of variants and shape characteristics. This is exactly what makes it so complicated to describe non-spherical particles and optimize their handling based on the description.

Artificial intelligence enables computers to learn new skills that will allow us to find better solutions to the challenges we face, be it in the field of medicine, tackling environmental issues or overcoming social problems. It will open up huge scientific and economic potential and thus comes with great responsibility. In order to exploit the opportunities for science, business and people in Hesse, we are relying on the particular strengths of the universities in Hesse: The Technical University of Darmstadt is strong in basic research and is closely followed by other universities that are conducting specialist research into AI, while we can also call on outstanding practical research at the University of Applied Sciences. The concept was jointly developed by 13 universities and more than 40 partners from the fields of research and business already want to cooperate with the centre today. Thanks to its unique characteristics, the centre will – as confirmed by the independent evaluation commission – become highly visible internationally, acting as a beacon in this field that will reach far beyond the boundaries of our state.

We are trying to use the data that is recorded on the wind turbines to predict failures," Kalyan Veeramachaneni, principal research scientist in the Laboratory for Information and Decision Systems of the Massachusetts Institute of Technology, told DW. Ewald Hesse, CEO of Berlin-based Grid Singularity, says several countries in Africa would leapfrog the development phase of European energy systems, similar to what happened to landline phones. "In developing countries, there is no stringent regulation in the energy sector, and we don't need to convince the government of allowing a new approach to energy production and consumption. Still, local communities would benefit from one PV system in the surrounding area, which, combined with sensors to measure energy consumption, would create a localized market. "Whatever comes out in the energy field in developing countries will be by far smarter and more practical than what we have in Germany," said Hesse, adding that several companies contributed to unlocking potential markets and significant investments in developing countries.

Hesse authorities signed a joint declaration on the establishment of an AI Quality & Testing Hub with the President of the VDE (Association of Electrical, Electronic and Information Technologies). Last week the German state and the association outlined the goal of the initiative to put AI systems to the test. Research and development, standardisation and certification are combined under one roof in the hub. In this way, the hub makes an important contribution to developing and applying AI responsibly. "AI is developing into the key technology of the 21st century, as it can offer solutions to many societal challenges. The Hessian state government wants to promote the quality of AI together with the VDE and make it verifiable. We are convinced, that the high quality of AI systems is the basis for trust and use in this technology," emphasised Hesse's digital minister Prof. Dr. Kristina Sinemus.

The head of the longest-established agency personnel consultancy in the UK says caution should be exercised over the use of Artificial Intelligence in recruitment processes. Property Personnel managing director Anthony Hesse thinks that estate agency should never forget that it is primarily a people business, and technology would have stark weaknesses identifying the right person for the job. "At its heart, estate agency is and has always been about people. We may deal in property, but it is people who are doing the buying and selling. And we are at risk of missing out on some of our best potential recruits if we forget that" explains Hesse. "Artificial Intelligence is all very well; but I seriously doubt that technology will be able to identify some of the subtle factors which only come to light in a one on one relationship between two human beings.

At the same time, AI could disrupt the human workforce greatly, noted Kaku, who warned that a smarter future comes with tradeoffs. AI, he said, does "pose an existential threat." Much of that threat, Kaku said, involves the long-term impact on the human workforce. According to some estimates, about 30% of the activities in 60% of all occupations could be automated. The displacement is already taking place, noted Jeff Hesse, a PwC principal.

A study from Redwood Software and Sapio Research released October 4th revealed that IT leaders believe automation could impact 60% of businesses by 2022 and threaten jobs in the process. Now, a new, separate report from PwC, the second biggest professional services firm worldwide, suggests a similar timeline; one in which people may need to practice and learn new skills -- or be left behind as automation takes over. The report, titled Workforce of the Future, surveyed 10,000 people across China, India, Germany, the UK, and the U.S. to "better understand the future of work." Of those, nearly 37% think artificial intelligence and robotics will put their jobs at risk; in 2014, 33% had a similar concern. A startling scenario the report envisions for the future is one in which "typical" jobs -- jobs people can steadily advance in through promotions -- no longer exist, prompting the aforementioned move to develop new skills.

Hesse suggests people research what skills will be in-demand in their field. Gates says workers with skills in science, engineering and economics will soon be the most sought after. Alibaba founder and e-commerce titan Jack Ma thinks more companies will be looking for people with expertise in data analysis and collection in the future. Eric Schmidt, executive chairman of Google's parent company Alphabet, also believes data skills will be in-demand moving forward. Of course, learning new skills doesn't necessarily mean going back to school for a degree.

Machine learning has the potential to solve many of our regular human problems, like for instance having too few nightmarish, oddly cat-filled crude images to gaze upon. Luckily, Christopher Hesse created the edges2cats web-based tool to address exactly that issue. The machine learning software uses Google's Tensorflow to translate one image to another, using training data provided by a database of over 2,000 stock images of cats to identify edges and fill in simple line art with what it estimates would be the best approximation of a realistic cat coloring of what it wrongly assumes, because of the limitations set upon it by its cruel creator, must be a cat. Hesse himself understates the impact of this significantly, writing that "some of the pictures look especially creepy," and he theorizes that's because we have a very good idea of what animal faces especially should look like, anything that features eyes in particular (like my own example above, with the source material provided unwittingly by my colleague) turns out especially horrifying. We're probably only at the outer edge of the many-layered onion of fresh horrors made possible by machine learning and AI, and previously unimagined by our feeble fleshy brains, so strap in.