Dan Zecha

Robert Dahlke, Henrik Klagges, Dan Zecha et al.

We present the Mixture-of-Tunable-Experts (MoTE), a method that extends the Mixture-of-Experts architecture of Large Language Models (LLMs). Without additional training, MoTE enables meaningful and focused behavior changes in LLMs on-the-fly during inference time. By analyzing the digital LLM brain of DeepSeek-R1 using a technique we dub 'functional Token Resonance Imaging' (fTRI) -- inspired by fMRI and using prompts designed to elicit specific behavior (e.g., 'What happened {time}{place}?') -- we empirically identify distinctive experts associated with behaviors like refusal responses. Using MoTE we are able to intervene and control such specific behavior. We switched off the top 10 most refusal-relevant experts (0.07% of R1's 14,848 routed experts), achieving a 52% refusal reduction on sensitive reference prompts without performance degradation on MT-Bench. Random expert deactivation resulted in smaller behavioral shifts with increased noise, whereas forced expert activation led to significantly higher refusal rates. Our approach shares similarities with sparse autoencoders (SAEs) in terms of explainability and steerability. Unlike SAEs, MoTE does not require large training efforts, as within MoEs with a vast number of experts, specialization already emerged naturally during pretraining. Our findings suggest that significant functional mechanisms in Mixture-of-Experts architectures can at least partially be localized in a small number of specific experts, rather than being distributed throughout the model's weights. Expert subgroups can be tuned to trigger significant behavior variations, providing insights into the inner workings of LLMs.

Henrik Klagges, Robert Dahlke, Fabian Klemm et al.

Requiring $10^{13}$-$10^{15}$ FLOPs to calculate one 8 bit weight in an LLM during pretraining is extremely expensive and seems inefficient. To better leverage the huge investments made into pretrained models, we develop the new "Assembly-of-Experts" (AoE) construction method to create capable child variants of existing Mixture-of-Experts parent models in linear time. Model weight tensors get interpolated individually, allowing to enhance or suppress semantic features of the parents. Varying the proportion of weights taken from the parent models, we observe some properties of the AoE child model changing gradually, while other behavioral traits emerge with a sharp transition. Surprisingly, nearly every generated model is functional and capable, which makes searching the model space straightforward. We construct the DeepSeek R1T "Chimera", a 671B open-weights hybrid model combining DeepSeek's V3-0324 and R1 model variants. The child inherits only the routed expert tensors of R1, but still achieves about R1-level intelligence. At the same time, it uses about 40\% fewer output tokens, close to V3 speed. Constructed without any fine-tuning or distillation, the Chimera exhibits surprisingly compact, orderly reasoning compared to its parent models.

Philipp Harzig, Dan Zecha, Rainer Lienhart et al.

Automatically generating descriptive captions for images is a well-researched area in computer vision. However, existing evaluation approaches focus on measuring the similarity between two sentences disregarding fine-grained semantics of the captions. In our setting of images depicting persons interacting with branded products, the subject, predicate, object and the name of the branded product are important evaluation criteria of the generated captions. Generating image captions with these constraints is a new challenge, which we tackle in this work. By simultaneously predicting integer-valued ratings that describe attributes of the human-product interaction, we optimize a deep neural network architecture in a multi-task learning setting, which considerably improves the caption quality. Furthermore, we introduce a novel metric that allows us to assess whether the generated captions meet our requirements (i.e., subject, predicate, object, and product name) and describe a series of experiments on caption quality and how to address annotator disagreements for the image ratings with an approach called soft targets. We also show that our novel clause-focused metrics are also applicable to other image captioning datasets, such as the popular MSCOCO dataset.

Rainer Lienhart, Moritz Einfalt, Dan Zecha

Human pose detection systems based on state-of-the-art DNNs are on the go to be extended, adapted and re-trained to fit the application domain of specific sports. Therefore, plenty of noisy pose data will soon be available from videos recorded at a regular and frequent basis. This work is among the first to develop mining algorithms that can mine the expected abundance of noisy and annotation-free pose data from video recordings in individual sports. Using swimming as an example of a sport with dominant cyclic motion, we show how to determine unsupervised time-continuous cycle speeds and temporally striking poses as well as measure unsupervised cycle stability over time. Additionally, we use long jump as an example of a sport with a rigid phase-based motion to present a technique to automatically partition the temporally estimated pose sequences into their respective phases. This enables the extraction of performance relevant, pose-based metrics currently used by national professional sports associations. Experimental results prove the effectiveness of our mining algorithms, which can also be applied to other cycle-based or phase-based types of sport.

Dan Zecha, Rainer Lienhart

In this paper we study the problem of estimating innercyclic time intervals within repetitive motion sequences of top-class swimmers in a swimming channel. Interval limits are given by temporal occurrences of key-poses, i.e. distinctive postures of the body. A key-pose is defined by means of only one or two specific features of the complete posture. It is often difficult to detect such subtle features directly. We therefore propose the following method: Given that we observe the swimmer from the side, we build a pictorial structure of poselets to robustly identify random support poses within the regular motion of a swimmer. We formulate a maximum likelihood model which predicts a key-pose given the occurrences of multiple support poses within one stroke. The maximum likelihood can be extended with prior knowledge about the temporal location of a key-pose in order to improve the prediction recall. We experimentally show that our models reliably and robustly detect key-poses with a high precision and that their performance can be improved by extending the framework with additional camera views.