Loading ePaper
Deutsches Forschungszentrum für Künstliche Intelligenz German Research Center for Artificial Intelligence next How AI orients to the real world From perception to action d e t a r e n e g I A
About us Human-Centric AI Intelligent solutions for the knowledge society DFKI has more than 35 years of experience in human-centric AI, conducting research in key areas of basic research and future- oriented applications, while always maintaining a focus on social relevance and scientific excellence in the field of artificial intelligence. The German Research Center for Artificial Intelligence (DFKI) was founded in 1988 as a non-profit public-private partnership. It operates facilities in Kaiserslautern, Saarbrücken, Bremen, Lower Saxony (Osnabrück and Oldenburg), Darmstadt, with laboratories in Berlin and Lübeck, and a branch office in Trier. It supports 29 research areas, thirteen competence centers, and eight living labs. Based on application-oriented research, it develops product functions, prototypes, and patentable solutions in the information and communication technology (ICT) sector. Funding is provided by public funding sources and industrial development contracts. Project results and milestones are periodically reviewed institutionally and by a Scientific Advisory Board, an international panel of experts. In addition to the German Federal Ministry of Research, Technology and Space (BMFTR) and the federal states of Rhineland-Palatinate, Saarland, Bremen, Lower Saxony and Hesse, several leading German and international high-tech companies from a wide range of industries also have seats on the board. www.dfki.de 2
Editorial Think about AI in Federal and European terms International cooperation, scientific visibility, and strategic sovereignty Advances in artificial intelligence are not made in isolation. The work of CERTAIN and other collaborations with strong European and international partners such as DFKI and INRIA show that scientific excellence is achieved when research is brought together across institute and national boundaries. The network of German AI Competence Centers serves to streng- then the international reputation of German AI research while hel- ping to establish AI made in Europe as a hallmark of excellence and trustworthy AI. Associations like the Robotics Institute of Germany (RIG) pool expertise, foster connectivity between disciplines, and demonstrate that AI research is often at its best where perception, orientation, and action are considered holistically, not separately. The research now being discussed on major international stages – from CVPR, ICML, and other IEEE conferences to IJCAI – is shaping the next major advance in AI. Germany’s High-Tech Agenda establishes the political framework. The goal is clear: strengthen technological sovereignty, improve the accessibility of AI resources, and ensure that Germany remains a global leader for the next generation of AI. This issue treats that as its starting point. It demonstrate how excellent research, international cooperation, and domain-specific networking can give rise to an outstanding AI landscape. 3
AI-Orientation How AI orients to the real world From perception to action Artificial intelligence has reached a point where performance improvements alone are no longer enough. The crucial question is not how systems can become even bigger, faster, or more ubiquitous. The crucial question is what kind of AI Europe wants to develop to ensure future scientific autonomy, societal reliability, and technological agency. This is the true challenge researchers face today: not signed for maximum scale, but for resilience, security, merely to make AI more powerful, but to enable it to efficiency, and controllability (scientific and industrial). navigate the complexities of an open and human-centric world. Even more fundamental is the question of which mod- el of AI we actually want to continue developing. The Germany and Europe are well-positioned for this chal- mounting success of large generative systems is undeni- lenge: scientific excellence, growing political aware- able. Equally undeniable, however, are their limitations: ness, and a public ever more alert to the fact that AI is enormous data requirements, opaque decision-making not some distant future technology, but is already a crit- processes, high energy consumption, and significant ical structural issue for economic, scientific, and soci- levels of unreliability in safety-critical or highly special- etal success. Consequently, now is the time to translate ized systems. Many applications in industry, medicine, these insights into concrete goals. Where AI is viewed mobility, or the energy sector demand more than a sys- merely as a global race for scale, it will ultimately be tem that merely appears powerful. It must be transpar- subject to the development logic of others. However, ent, verifiable, and capable of explaining its decisions those who take AI seriously as a research subject will in robust contexts. This is precisely why one field of re- be able to set their own standards – regarding transpar- search is gaining traction, because it represents much ency, energy efficiency, controls, and real-world inte- more for Europe than just another technical alternative: gration. Technological sovereignty does not equate with hybrid, neuro-explicit AI. It combines learning methods market success, but with the ability to determine the with symbolic, knowledge-based, and analytical com- foundation and direction of future systems. ponents. Europe must invest precisely in these areas when de- This aligns nicely with the shift away from the simplis- veloping the next generation of AI if it wishes to reduce tic “bigger is better” mindset. Small, specialized lan- dependence on closed, non-European systems. This guage models and frugal AI architectures demonstrate includes new computer architectures, energy-efficient that efficiency itself can be a step forward. Such models processing paradigms, and infrastructures – not de- require less energy, less special hardware, and can be 4
AI-Orientation tailored more precisely to specific domains – from in- These challenges for technology and science converge dustrial processes to scientific data spaces or sensitive in this issue of ai next. The real question for the future knowledge environments. Especially for a research and is not only what AI can recognize but also how it ori- economic landscape that relies heavily on highly spe- ents itself to real-world, dynamic, and ambiguous en- cialized, regulated, and SME-dominated systems, this vironments. How does perception transform to reliable is not a minor consideration, but a strategic advantage. knowledge? How is knowledge used to create a robust Reliability under real-world constraints is more impor- framework of classification? And how to take action that tant than maximum generality at any price. is not merely efficient, but also justifiable? This move Add to this another ongoing shift: AI no longer gener- only a plausible narrative theme for this publication, but ates only textual and visual content; it can now plan, co- also a precise research horizon for the coming years. from perception to orientation to action presents not ordinate, and manage expanded workloads. The focus is shifting once again with the transition to agents. As Europe has no need to copy foreign models in pursuit of systems enter contexts of related actions, transparency, this horizon. It can rely on its own strength in developing intervention options, and human oversight become key. AI research centered on reliability, clarity of orientation, and real-world integration – open to international col- The next step in AI research is not to unleash the next laboration yet steadfast in its own scientific standards. level of autonomy, but rather to ensure the societal and The next step will be defined not by the loudest system, technical integration of systems that are controllable but by the one that combines transparency, efficiency, and transparent. Only then does raw computing power and accountability with performance. That is precisely transform into reliable agency. where AI begins to operate in the real world, not within isolated data spaces. « The promise of hybrid, neuro- explicit AI lies in scaling up existing models, rather than in merely increasing their size, as it involves a synergis- tic concept of machine intelligence that combines the strengths of probabilistic and deductive AI approaches. Systems of this kind do more than recognize patterns: they explicitly represent knowledge, structure their decisions, and allow auditing under defined con- ditions. This is a scientific view that does not pit performance against trustworthiness but systematically brings these two together. » Prof. Dr. Antonio Krüger, CEO DFKI
AI-Perception Papers and authors: From raw geometry to meaningful scenes Three approaches to 3D scene under- standing, attention, and reconstruction Understanding 3D scenes begins where geometry, relationships, and perspective meet. Researchers at DFKI recognized this dynamic and made it a priority of their work, presenting their findings at the June Conference on Computer Vision and Pattern Recognition (CVPR) 2026: Spatial reconstruction is no longer viewed as the end goal, but rather as a starting point for semantic interpretation, human-centric orientation, and targeted intervention within complex 360-degree environments. ReLaGS represents a fundamental advance in 3D scene objects. The process turns a raw reconstruction into a understanding. The method combines Gaussian splat- meaningful scene. ting with hierarchical scene representation and rela- Spatial constructs, part-whole relationships, and lan- tional logic in a way that enables machines to capture guage descriptions of relationships are explicitly repre- and model the spatial and semantic relationships among sented by geometric points and surfaces in a structured ReLaGS Making scenes readable Pure reconstruction becomes a structured environment — with objects, layers, and relationships. DriverGaze360 Understanding orientation Where do people direct their attention — all around, not just straight ahead? Inpaint360GS Intervening in scenes Missing elements are added without destroying the inner structure of the scene. 4
AI-Perception environment. ReLaGS achieves a scientific innovation matters most here is how something is reconstructed in by joining hierarchical semantics and relational 3D rep- the scene. Inpaint360GS fills in missing areas in a way resentation in a unified, scene-agnostic approach. This that preserves object logic, spatial structure, and seman- approach creates an open 3D Scene Graph without the tic coherence. need for scene-specific training, while at the same time The overarching scientific link to the ReLaGS and improving spatial and semantic stability through prun- DriverGaze360 projects is clear. While ReLaGS demon- ing and robust attribute aggregation. In a world where strates how scenes become interpretable and Driver- structured 3D models of environments are gaining inter- Gaze360 reveals how attention is distributed within national importance, experts are teaching machines not them, Inpaint360GS shows how, on this basis, it is pos- only what to recognize in a scene but also the relation- sible to intervene in scenes without destroying their in- ships among objects. DriverGaze360 explores this issue ternal structure. Perception does not automatically lead in another, equally important direction: How does a per- to action, but it does provide the basis for editable nodes son look at a complex scene? that integrate geometric accuracy, semantic understand- DFKI researchers have compiled a dataset of approxi- ing, and contextual awareness. mately one million gaze-annotated frames from a realistic The three studies together demonstrate more than 360-degree driving simulator to demonstrate how driver a methodological improvement. They mark an advance attention can be captured under controlled yet realistic toward AI systems that are able not only to reconstruct conditions. Such measurements are relevant not only for spatial environments, but also to understand and pro- analyzing human perception, but also for assistance sys- cess their structure, relationships, and gaps. By integrat- tems that need to understand what drivers are actually ing scenes, attention, and reconstruction, the foundation focusing on at any given moment. This is lacking in pre- is set for reliable robots, more robust assistance sys- vious research where traditional datasets primarily cap- tems, more precise digital twins, and visual AI that does ture frontal views and underestimate lateral or rearward not settle for what is simply visible but also understands attention dynamics, for example, when changing lanes, context. merging onto a highway, or checking behind the vehicle. DriverGaze360 is more than just a topic related to traffic. The work demonstrates that scene understand- CVPR 2026: A Stage for Visual AI ing is not merely a matter of machine segmentation and geometric relationships, but also involves situated, hu- The IEEE/CVF Conference on Computer Vision man orientation under realistic conditions. Tracking ad- and Pattern Recognition (CVPR) is one of the ditional objects is key to the system’s ability to learn not most important international conferences in only gaze distribution, but also which object is the focus computer vision research. This year‘s conference of attention at any critical moment. As a result, attention took place in Denver from June 3-7, 2026. For becomes semantically connectable and describable as DFKI, it was a key event to discuss current work a structured relationship between the person and their on 3D scene understanding, relational logic, at- environment. tention modeling, and object-aware reconstruc- The third approach, as introduced by Inpaint360GS, tion. This issue features three contributions from is active scene editing. The core challenge of any realistic DFKI‘s Augmented Reality Research Department: 3D representation is that scenes are often incomplete, ReLaGS, DriverGaze360, and Inpaint360GS. To- obscured, or have gaps. Our method addresses this by gether, they show how geometry, attention, and filling in missing scene and spatial elements in 360-de- scene editing can be viewed as part of a related gree environments in an object-aware manner. What research movement. 7
. d e t a r e n e g - I A d n a l a n o i t c fi s i i d e t c p e d n o s r e p e h T . i n a m o D c i l b u P , d r a w o H a s i u o L : y p o c s o r c M i AI-Perception - Papers and authors: AI + µSAM - + + The bottleneck in microscopy How AI foundation models are reshaping image analysis in the life sciences Modern microscopes provide image data at resolutions and volumes that are revolu- tionizing biological research. Consequently, the bottlenecks are encountered less and less at image acquisition and more often with the analysis: cells, nuclei, and organelles must be precisely delineated and consistently annotated across large datasets. Segmentation is what turns microscopic images into reliable research data. A recent study published in Nature Methods, “Segment The scientific significance of this work is found not only Anything for Microscopy,” investigates whether a foun- in the improved results. A crucial aspect is adapting a dation model developed for everyday images can be generated AI image model to a domain where the visual adapted to microscopic data without losing its flexibility. logic differs fundamentally from that of the everyday Researchers succeeded in demonstrating that the „Seg- visual world. Object boundaries, contrasts, shapes, and ment Anything Model“ (µSAM) can be adapted for light spatial relationships follow different rules in microscopy; and electron microscopes to more reliably capture bio- it is particularly important to determine the true extent logical structures while simultaneously supporting both of modern AI’s generalization capabilities. interactive and automated segmentation tasks. 8
« The µSAM project demonstrates the productive use of foundation models in highly specialized scientific imaging domains – pro- vided they are methodically and consistently adapted to the specific domain. Based on this experience, we are preparing a spin-off to perform cell culture analysis that should effectively translate the potential of this research into a concrete application. » Prof. Dr. Andreas Dengel, Head of Smart Data & Knowledge Services Research Department, Executive Director, DFKI Kaiserslautern AI-Perception µSAM transforms a slow, routine task into a collaborative workflow that accelerates annotation without sacrificing scientific oversight. For the life sciences, this is more than just a tech- nological innovation. Fields such as cell biology, can- cer research, and developmental biology generate vast amounts of data that can only be analyzed manually at considerable cost in terms of time and effort. The re- search results demonstrate how this bottleneck can be overcome, not by replacing scientific work, but through a more precise integration of the models, interfaces, and domain knowledge. This study also contributes to the international debate surrounding foundation models. It states that the case This approach is especially relevant where human exper- for such systems lies not in abstract, universal applica- tise and model performance intersect. µSAM is designed bility but in their precise translation into actual scientific to be an interactive working environment: researchers applications. That is where it is determined whether it is provide a few prompts; the system fills in the contours just an exciting new model or whether it delivers a genu- and incorporates corrections into the ongoing workflow. ine impact on scientific practice. a t a d y p o c s o r c i m Raw 00:03 Automatic segmentation The foundation model generates initial structure predictions, instantly identifying base biological boundaries without human intervention. Example: two cells merged. Independent input + - + Add Selection - Subtract Selection ++ 00:06 Interactive correction Researchers initialize the selection with a single point or bounding box, adding optional positive or negative clicks to refine the mask (e.g., fixing a merged pair). 7 00:09 Refined mask The system refines the output in real-time, delivering a highly precise, reviewed segmentation mask of the correct object. Corrected masks retrain the model . s e s o p r u p e v i t a r t s u l l i r o f d e t a r e n e g - I A , l l e c c i t o y r a k u e a f o n o i t a l u m S i
AI-Orientation Papers and authors: Trained only in English, AI will remain exclusive Synthetic training data helps to create models for low-resource languages Large language models (LLMs) perform extremely well in those languages for which a wealth of digital data is available. Vast quantities of text are available for high-resource languages like English and German. However, many other languages are barely repre- sented in the training of modern AI systems. Currently, there are no high-performance AI applications available for many of the minority language communities. This dispari- ty leaves the potential economic and social benefits of modern language technologies largely untapped. Can synthetically generated training data bridge the Researchers trained new AI models to verify the quali- gap? A research team from DFKI, the Brno University of ty of the synthetically generated data. The results were Technology, and the Kempelen Institute of Intelligent compared to results from models trained on real, hu- Technologies in Bratislava thinks they have the answer. man-annotated “gold standard” data. In this way, a pre- Their work explored the central issue: under what con- cise determination was made of how reliable the syn- ditions can large language models (LLMs) generate suffi- thetic datasets actually are. cient high-quality training data to be suitable for training additional AI models? The team found that the performance gap to the models based on real training data had significantly narrowed. The researchers used four open-source LLMs of vary- The new models nearly matched the performance levels ing sizes, including models from the Llama and Gem- achieved with the gold standard data, particularly in the ma families. These models generated training data for task of user intent recognition. eleven typologically diverse languages, ranging from high-resource languages such as English and German to It also became clear that intelligent prompting and re- languages with a significantly smaller digital presence, vision techniques can partially offset the advantages such as Welsh, Romanian, Azerbaijani, Slovenian, and offered by the large models. This introduces practical Telugu. The team then used the generated data in their possibilities to generate high-quality data even with investigation of three key language processing tasks: smaller, more cost-effective open-source models. user intent recognition, topic classification, and senti- ment analysis (positive or negative opinions). 10
« A common assumption is that more parameters automatically yield better results. Our study shows that the approach taken is equally important. With the right prompting and revision techniques, high- quality training data can be generated even for low-resource languages. This is crucial for promoting inclusivity and access to modern AI techno- logy for minority languages. » Tatiana Anikina, lead author of the study Low-resource languages Welsh, Telugu, or Azerbaijani are hardly represented in AI training. avansate de învățare automată ajută cercet să descopere noi modele în știință, dar lips telor de antrenament în limbi regionale cr dezechilibre majore. Pentru a dezvolta alg mi echitabili, este esențial să colectăm res lingvistice de înaltă calitate din toate reg le globului. कृत्रिम बुद्धिमत्ता और डγटा विज्ञान ं। आधुनिक एल्गोरिदम अनुसंधान के नए द्वार खोल रहγ हδ ं और जटिल विशाल डγटासेट का विश्लेषण करते हδ प्रणालियों को समझने मे ं मदद करते हδ ं। वैज्ञानिक प्रगति तेज़ी से हो रही हδ, लेकिन ሰራሽ አስተዋይነት እና የሳይንስ ምርምር ከፍተኛ መረጃዎችን ይመረምራሉ። ዘመናዊ ስልተ-ቀመሮች አዳዲስ ሳይንሳዊ ግኝቶችን ለማግኘት ይረዳሉ። ነገር ግን በቂ የሥልጠና መረጃ አለመኖሩ በዲጂታል ቴክኖሎጂ ውስጥ የቋንቋ እኩልነት እንዳይኖር ያደርጋል፡ ይህም ለወደፊቱ ትልቅ ፈተና ነው። High-Resource Languages Huge amounts of text exist for English or German. and researcher at the Multilinguality and Language Technology Research Department, DFKI The results show that the size of a language model is not the sole determinant of the quality of the generated data; a decisive factor is also the generation strategy. The most successful approach used a combination of examples in the respective target language followed by an automated quality check performed by the model itself. The AI reviews its own output and removes erro- neous or unsuitable examples. Despite the promising results, synthetic data does not achieve the same levels of quality across all languag- es and tasks. Significant performance gaps persist, particularly for severely underrepresented languages. However, the study demonstrates that this gap narrows when data generation is systematically evaluated and subject to appropriate quality control mechanisms. Synthetic data LLMs generate their own training examples in the target language. Self-revision The model checks its own data and removes faulty examples. 11 Close to the gold standard The generation strategy matters more than the model size. ß'sधዳ
AI-Orientation Papers and authors: The end of the all-rounder How AI builds customized high-speed databases Have you ever wondered why software is often so unwieldy? The reason usually lies in a long-held tenet of software development: „one size fits all.“ Traditional databases like PostgreSQL or DuckDB are like The AI system then constructs an autonomous mini-da- Swiss Army knives – they can handle almost any query. tabase in C++ – exclusively for this one task. Since the But this flexibility comes at the cost of performance. You mini-database does not need to support other types of pay an invisible “flexibility tax” in the form of process- queries, the overhead of generality is eliminated. Data ing time and storage consumption for features you don’t types, storage layouts, and runtime paths are hard-cod- even need in your specific case. General-purpose analyt- ed into the database engine. ical databases are designed for arbitrary workloads and support any query and any database schema. This ver- satility is both a strength and a weakness: Runtime in- terpretation, levels of indirection, and abstraction bound- Up to 60-times faster than the industry competition The results of initial scientific testing (using the stand- aries all introduce additional overhead – even in highly ardized TPC-H benchmark) have sent shockwaves optimized engines. through the tech world: The new AI-generated “One- Realistically, most systems are not asked to execute Size-Fits-One” databases easily outperformed es- arbitrary queries and do not require such flexibility. They tablished systems like DuckDB, Clickhouse, or Umbra execute your query: a known set of SQL templates against that have been optimized by human experts for many a known schema, over and over again. A research team years. The custom AI engines were 10 to 60 times faster at TU Darmstadt (Johannes Wehrstein, Timo Eckmann, [2603.02001]. Matthias Jasny, and Carsten Binnig) is breaking with this principle. Their revolutionary approach is called “Be- spoke OLAP” (Custom Data Analytics) [2603.02001] – and it proposes a radically minimalist future for software development. The trick: A custom-made digital solution instead of an off-the-shelf product Imagine: what if there were no longer any need to install a one-size-fits-all database? Instead, an artificial intel- ligence system (like Claude or GPT) analyzes the exact data and search queries your company actually uses. 12 The custom AI engines C++ were 10 to 60 times float fast_inverse_sqrt(float number) { long i; float x2, y; const float threehalfs = 1.5F; faster x2 = number * 0.5F; y = number; i = * ( long * ) &y; // evil floating point bit level hacking i = 0x5f3759df - ( i >> 1 ); // calculate inverse sqrt y = * ( float * ) &i; y = y * ( threehalfs - ( x2 * y * y ) ); // 1st iteration return y; } 60x Speedup #pragma omp parallel for aligned(data:32) for(size_t i = 0; i < buffer_size; i+=8) { __m256 v_data = _mm256_load_ps(&buffer[i]); __m256 v_mask = _mm256_cmp_ps(v_data, v_threshold, _CMP_GT_OQ); __m256 v_result = _mm256_mul_ps(v_data, v_multiplier); _mm256_store_ps(&output[i], _mm256_and_ps(v_result, v_mask)); } inline void fast_memory_purge(void* dest, size_t n) { auto d = static_cast<uint64_t*>(dest); while (n >= 8) { *d++ = 0x00000000; n -= 8; } } </>
Where is the turbo boost seen? This technology is not a purely theoretical project. “Be- spoke OLAP” is a total game changer wherever compa- The highlight: Continuous testing of AI (The feedback loop) Historically, AI models fail to generate error-free nies crunch massive volumes of data every night: and extremely fast system code for complex data E-Commerce & Personalization: structures. Researchers solved this problem with an ingenious, closed feedback loop that operates en- Every night, an e-commerce giant analyzes the click and tirely without human intervention. purchase behavior of millions of customers to personal- ize the homepage for the following day. Instead of run- ning a standard system at full capacity for hours, an AI system builds a mini-engine in minutes that performs this specific nightly task in a fraction of the time. Financial Sector & Risk Analysis: Banks are required to generate regulatory risk and fraud reports on a daily basis. The mathematical queries re- quired for this rarely change. A custom AI database can perform these standard calculations in real time, rather than tying up IT infrastructure for hours. IoT & Industry 4.0: In modern factories, thousands of sensors collect ma- chine data every minute for predictive maintenance. A tiny, AI-generated database fits directly onto the on-site control devices and analyzes the data in real time, with- out having to send it to the cloud first. What is the current status of the project? The research is already making waves: following a peer review, the research papers on Bespoke OLAP have been accepted for publication in the IEEE Data Engi- neering Bulletin (3/2026), one of the most important GENERATE CODE 1 The AI system writes an initial creative program draft for the mini-database. 2 THE ULTIMATE TEST An automated system checks the code for errors and measures the exact speed. 3 SYSTEM LEVEL REVIEW The system analyzes each error message and detects any drag on performance. FINE TUNING VIA PROMPT 4 The AI receives precise feedback: “Correct error X in line 42 and optimize.” (The cycle restarts until everything runs perfectly) What’s next? SynnoDB (https://synnodb.com/) is the spin-off com- international journals. It is also accepted for presenta- pany where Prof. Carsten Binnig’s team aims to exploit tion at VLDB (9/2026, Boston), one of the world’s lead- the full potential of the technology and bring the innova- ing database conferences. Furthermore, the project was tive Bespoke-OLAP technology to market. Conventional featured in April 2026 on UC Berkeley’s renowned Sky databases are like all-purpose vehicles. SynnoDB con- Computing blog, invited to the exclusive HPTS Work- structs a custom-built racing car for each customer, one shop, and discussed in guest lectures at companies in- precisely tuned to the data and queries required, while cluding Google and SAP. Future outlook “Bespoke OLAP” gives us a glimpse of a new era in soft- ware architecture. In the future, AI will no longer mere- delivering significantly faster performance. The compa- ny’s website currently features a “try-it-yourself” demo that allows users to create synthetic mini-databases, as well as a live playground where you can: ly assist programmers with coding. It will take on the • Run queries on a real engine generated by SynnoDB. role of chief architect, creating customized digital sys- • Compare performance against DuckDB and Umbra. tems in a matter of seconds, something human devel- • View the generated C++-source code. opers could previously only dream about. • See how the engine was built. 13
AI-Orientation Papers and authors: X-Hacking When explainability is misleading DFKI researchers reveal how explainable AI can become a methodological trap: When AutoML selects – from among many equally good models – the very variant that provides an explanation that perfectly fits a desired narrative, transparency itself is suspect. XAI methods are intended to make machine learning The problem becomes particularly relevant with auto- systems more transparent. However, precisely because mated search and optimization methods. AutoML greatly models provide plausible explanations for their decisions, facilitates model development, while also expanding the a problem arises – described in our paper as “X-Hack- range of possible decisions. This is the scientific crux of ing.” The problem is that explanations can be deliberate- X-Hacking: When many models, preprocessing steps, and ly steered in a desired direction even though the models perform almost equally in terms of predictive accuracy. ICML: Where machine learning meets The starting point is a simple yet far-reaching observa- applications tion: multiple models achieve similar accuracy on a giv- en dataset but offer very different justifications for their The International Conference on Machine results. Anyone choosing from this set of “plausible” Learning (ICML) is one of the top three leading models can favor precisely the explanation that supports conferences in machine learning (next to Neu- a desired narrative. Our paper extends the principle of rIPS and ICLR). In scope, it is characterized by p-hacking – well-known in statistics as a way to get the theoretical analysis, algorithmic innovation, most pleasing results – to explainable AI: It optimizes not and statistical learning – with strong inter- only the prediction but also the justification for the result. faces to reinforcement learning, robotics, and optimization theory. The conference provides a central hub for researchers in AI, statistics and data science as well as application-orien- ted fields like computer vision, computational biology, speech and robotics. At ICML 2025, the DFKI team led by Prof. Sebastian Vollmer presented a paper with the title “X-Hacking: The Threat of Misguided AutoML.” It applies the principle of p-hacking to explainable AI and shows how AutoML searches pipelines systematically for models with desired, but potentially misleading explanations.
« At a time when AI explains decisions, but does not always understand them, we as scientists must take responsibility for the depth of the explanations – and recognize their limitations. The goal is a scientific culture that prioritizes not only accuracy, but also honesty DATA in explainable AI. » Prof. Dr. Sebastian Vollmer, Head of Data Science and its Applications Research Department at DFKI PREPROCESSING parameter settings are available, the search for a “suit- able” explanation can become a part of the systematic analysis process. Explainability is no longer automatical- ly seen as a corrective measure, but rather as something that must itself be methodologically tested. This takes on considerable importance for sensitive fields of application, like medicine, social research, and the jus- tice system, where explainable models are often consid- ered a prerequisite for trust, transparency, and responsi- ble decision-making. However, if a plausible explanation is selected from a range generated by models of similar high quality, a reference to interpretability is no longer adequate. This point of tension is where the Trustworthy AI debate sets in, and our contention is that what matters most is not only the accuracy of the model but also the integrity of the entire process by which explanations are generated. The paper does not conclude with a simple problem de- scription. It advocates for a more reflective scientific ap- plication: explanatory histograms that visualize the range of possible explanations, complete documentation of the pipeline, and greater awareness that automated model selection not only enhances efficiency but can also intro- duce new biases. X-hacking is not some sideshow of explainable AI, but a fundamental question of scientific integrity regarding the use of automated model search. 15 FEATURE EXTRACTION L M o t u A MODEL FITTING PERFORMANCE EXPLANATIONS
AI-Action Efficiency is the new bench- mark for language models Why size alone is not the crucial factor when choosing a language model The development of large language models has long followed a seemingly simple logic: more parameters, more computing power, and larger models mean better results. This paradigm of scale is what shaped AI research in past years. However, as more language models find their way into business and social applications, it becomes clear that size is not always better and practical utility matters more. Edge Deployment AI works on-site On devices directly in the operational environment, data privacy, energy consumption, and response time are what count — this is where efficiency becomes tangible. Size alone is not enough More parameters used to promise better results for a long time. In practice, energy, hardware, and operating conditions matter just as much. Large Language Models Model size ≠ Efficiency Finding the right model Smaller models can be just as powerful in the right context — the key is balancing quality and resource requirements. Small Language Models 18
« We are creating the scientific basis for the efficient use of language models. We will present our findings, as reported in the paper titled “Mapping the Efficiency Landscape of Small Language Models”, at IJCAI 2026. We demonstrate that it is not the size of the model that is critical, but rather how efficiently the model meets the specific requirements of the current application. » Fabian Reichwald, Research Department Intelligent Markets at DFKI Why larger is not automatically better Whether in industrial processes, on edge devices, or in sensitive data environments: language models must perform under real-world conditions and requirements. For example, the largest available model is not always the best choice for a medium-sized enterprise – such as one summarizing technical documents or automating the processing of customer inquiries. In this case, the model that performs the task on existing hardware while striking the best balance between output quality, energy consumption, and response time may be a better choice. This marks a shift in the scientific focus. Instead of eval- uating models solely on the basis of size or performance, the focus is now on practical considerations. How much computing power does a task require? When is higher energy consumption justified by a measurable gain in quality, and when does it merely drive up costs and re- source consumption? AI-Action More than performance Efficiency is no longer a secondary consideration, but a scientific variable in its own right. For a company that intends to run a language model locally, simply looking at the model’s size is not enough. What matters most is which model solves the specific task with the fewest possible resources. Efficiency measurements rely on the relationship between performance, energy consump- tion, and application context. Under this scrutiny, many common assumptions prove to be invalid. Smaller models do not automatically oper- ate more efficiently than larger ones. Additional param- eters or higher energy consumption do not necessarily lead to better results. Well-informed decisions require a systematic consideration of all influencing factors. Transparency of well-informed decisions The next step in the development of language models is not simply to build more powerful models. Equally im- portant are the scientifically grounded criteria on which to base the selection of the right model for the specific task. A model’s practical value is determined not by its size alone, but by its suitability for the specific application context. Efficiency is becoming a key criterion when se- lecting language models, particularly where technical, economic, and environmental requirements converge. 17
AI-Action Bridging Generative AI and Robotics From task understanding to reliable actions How does data translate into reliable robot actions in the physical world? To achieve this, a new approach introduces an additional architectural layer between generative AI and robotics. Instead of translating language or images directly into robot move- ments, large AI models first generate structured task representations that are supple- mented with physical requirements. This creates a link between task interpretation and robust robotic actions. Language models can understand complex relation- The new approach is based on a new form of task rep- ships and generate plausible responses. Robotic sys- resentation. Instead of directly translating an instruc- tems must also contend with physical forces, dynamics, tion such as „move an object to a specific location“ into and uncertainties. The new architectural layer makes control commands, AI first analyzes the task. It cap- AI-generated task representations practical for robotic tures the objective, relevant conditions, and the char- task execution in real-world scenarios. The approach acteristics of both the robot and its environment, and also gives more flexibility to highly dynamic robotic sys- then uses this information to generate a semantic task tems to adapt to different environments. representation. . h p a r g o t o h p a n o d e s a b d e t a r e n e g - I A e g a m I The Unitree G1 is one of the research platforms used in the ActGPT project. 16
« The real challenge is not getting robots to understand instructions, but in ensuring they can execute them reliably and robustly within existing physical constraints. Generative AI helps to understand the context of natural language and formulate the problem, while physics determines whether the solution actually works. » Shubham Vyas, DFKI Robotics Innovation Center This representation is subsequently converted into a machine-readable format by a specially developed domain-specific language (DSL). The DSL serves as an interface between generative AI and the mathe- matical optimization methods used in robotics. Once provided with the movement requirements, an opti- mal control problem (OCP) is derived. This problem is then solved using established control methods to generate appropriate movements for the task, while ensuring that it is also physically possible. This approach represents a broader trend in AI re- search: What matters most now is not simply how machines understand and model their environment, but how reliably they translate that knowledge into real-world actions. Research project ActGPT The “ActGPT” research project at the DFKI Robotics Innovation Center is funded by the Federal Ministry of Research, Technology, and Space (BMFTR). Researchers are making AI-generated task representations useful for real-world application scenarios by develop- ing methods to integrate generative AI with robust robotic systems. The aim is to sup- port the reliable execution of complex robotic tasks and to simplify the development of con- trol strategies. The ActGPT Workflow 1 Input Language Images Sensor data 2 Generative AI Understands the task New Architectural Layer Domain-Specific Language (DSL) Describes: Goals Cost functions Constraints Optimal Control Problem (OCP) 3 Optimal Control Solves the OCP Computes optimal trajectories 4 Robot Executes the motion ACT GPT The new architectural layer. AI models translate tasks into structured physical problems, directly connecting semantic knowledge with robust robotics. 19 Image AI-generated based on a photograph.
AI-Action Shaping the future of AI-based robotics through collaboration DFKI combines scientific excellence, system competence, and real-world testing The next stage in the development of artificial intelligence is playing out in the real world – with robots capable of understanding their surroundings, learning from experience, and acting safely. The marriage of AI and physical action is one of the main scientific challenges in AI research today and is being pursued under the term “Embodied AI.” DFKI is making a significant contribution to the advance- ment of AI-based robotics in Germany through its work Holistic approach and system competence This approach is particularly evident at the Robotics In- at the intersection of artificial intelligence, autonomous novation Center in Bremen, where complex robotic sys- systems, and robotics. Its strengths are in basic scien- tems are developed and tested under realistic conditions. tific research, technological systems expertise, and re- al-world testing. From AI methods to intelligent robot systems At various research departments and locations, DFKI The research platforms developed here include: auton- omous underwater robot systems and space rovers, walking robots and humanoid systems, robotic exoskel- etons, and individual components like manipulators and investigates key issues in intelligent robotics – from grippers. This facilitates the study of key challenges in perception and learning, through planning and auton- intelligent robot design – from autonomous navigation omous decision-making processes, to human-robot in- systems to safe interaction with humans. teraction and practical applications. The work concerns individual AI methods and techno- logical components as well as the development and The importance of research platforms and test facilities A key feature of robotics research is the research plat- testing of complete robotic systems. The coordinated form. In-house systems make it possible to study new interaction of different technologies is crucial: software, concepts and technologies at the system level and to sensors, controls, and physical implementation must all investigate the interaction of various components. Com- come together in a functioning overall system. Only then mercial platforms improve the research results by ena- can autonomous robots behave safely and flexibly in dy- bling reproducible experiments and comparisons under namic surroundings. standardized conditions. 20
« Robotics is a field that requires the convergence of many disciplines. We address every level of a robotic system: from hardware and electronics to software development and innovative AI methods. Few institutions offer such a broad range of expertise. » Dr. Sirko Straube, Deputy Head, DFKI Robotics Innovation Center Equally important are specialized testing environments where new methods and systems can be tested under realistic conditions. With the Maritime Exploration Hall and the Multifunctional Hall in Bremen – which, among other features, include an artificial lunar crater land- scape – and the SmartFactory in Kaiserslautern, the DFKI has unique infrastructure for various areas of ro- botics research. These facilities enable the systematic validation of new approaches and the evaluation of their performance in realistic test scenarios. 21 AI-Action Advancing Robotics Research Together DFKI is actively involved in national and international research networks. The continued development of in- telligent robotics demands such collaboration – across various scientific disciplines and technological expertise and beyond the boundaries of individual institutes and fields of study. The Robotics Institute Germany (RIG) presents a prime example of this collaboration. DFKI offers its expertise in autonomous robotics and artificial intelligence as well as its experience in the development and testing of complex systems. Networking with leading robotics hubs across Germany pools diverse competencies and creates a framework for joint research activities. Through the combination of scientific excellence, tech- nological system expertise, and interdisciplinary collab- oration, DFKI is able to make significant contributions to the study of key issues in AI-based robotics. In this way, DFKI is helping to shape the next generation of intelli- gent robotic systems. . h p a r g o t o h p a n o d e s a b d e t a r e n e g - I A e g a m I
AI-Action DFKI is a partner in the Robotics Institute Germany (RIG) The Robotics Institute Germany (RIG) is an initiative of the Federal Ministry for Research, Technology, and Space (BMFTR) for the purpose of networking the leading robotics hubs in Germany. The aim is to pursue scientific excellence, increase international visibility, attract talent, and accelerate technology transfer in AI-powered robotics. DFKI participates in the RIG through three research departments: the Robotics Innovation Center in Bremen and the de- partments of Intelligent Networks and Innovative Factory Systems in Kaiserslautern. DFKI’s contribu- tions include the provision and use of research facil- ities, training and professional development oppor- tunities for specialists and managers, participation in international robotics competitions and challeng- es, and networking with industry partners. DFKI @ IJCAI-ECAI 2026 Perspectives on the AI of Tomorrow DFKI is well represented at one of the world’s most respected conferences on Artificial Intelligence. The International Joint Conference on Artificial In- telligence (IJCAI) 2026 has accepted twelve pa- pers published by the German research center. The papers reflect the scientific diversity of the in- stitute – ranging from foundation models and com- puter vision to robotics, semantic technologies, and explainable AI as well as medical applications, envi- ronmental monitoring, and software engineering. Together, these papers highlight the topics cur- rently shaping international AI research. How can we design intelligent systems to be more efficient, more transparent, and more interactive? How can we help them to better understand complex rela- tionships and reliably support people across a wide range of application scenarios? These DFKI papers provide insights into the next generation of intelli- gent AI systems. For more information, please scan the QR code www.dfki.de 22 . e p a c s y t i c d e t a r e n e g - I A
Lübeck Oldenburg Bremen Berlin Osnabrück DFKI Site DFKI Laboratory DFKI Branch Office Trier Darmstadt Kaiserslautern Saarbrücken Imprint dfki ai next; Issue August | 2026; Published by: Deutsches Forschungszentrum für Künstliche Intelligenz GmbH (DFKI); Trippstadter Str. 122, D-67663 Kaiserslautern; Phone: +49 631 20575 0; E-Mail: news@dfki.de; Editors: Jeremy Gob (responsible), Heike Leonhard, Udo Urban, Andrea Fink, Maria Santos; Editorial review: Armindo Ribeiro; Layout: Lando Lehmann, Juan Mejia; Production: One Vision Design, Saarbrücken; Cover: AI-generated, graphic design by DFKI. [AI] Image contains AI-generated visual elements. Concept, compositing, and design: DFKI. For photorealistic images, the label is also placed directly on the image. Credits Title: DFKI, Lando M. Lehmann [AI]; Page 5: DFKI, Oliver Dietze; Page 6: DFKI, Lando M. Lehmann [AI]; Page 8: DFKI, Juan D. Mejía [AI] (Person); Louisa Howard, Public Domain (Microscopy); Page 9: DFKI, Juan D. Mejía [AI]; Page 11: DFKI, Lando M. Lehmann [AI]; Page 12: DFKI, Juan D. Mejía [AI]; Page 14: DFKI [AI]; Page 15: DFKI, Lando M. Lehmann [AI]; Page 16: DFKI, Lando M. Lehmann [AI]; Page 18: DFKI, Juan D. Mejía [AI]; Page 19: DFKI, Juan D. Mejía [AI]; Page 21: DFKI, Juan D. Mejía [AI]; Page 22, top: DFKI, Thomas Frank; DFKI, Annemarie Popp; Page 22, bottom: DFKI [AI].