Searcharxiv⌕ Search

arXiv subjects

Stephan Olbrich

Publications and source records attributed to Stephan Olbrich.

2 recordsLinked to original sources

Programs-of-Layers in LLMs through the Lens of Cortical Areas

Inference in LLMs is conventionally a fixed-depth, fixed-order forward pass through every layer, regardless of how difficult the input is. The human brain does not work this way. Using the thalamus as a central hub, it routes information flexibly to all regions of the cortex according to demand. Li et al. (2026) recently showed, with program-of-layers (PoLar), that transformers can be given an analogous flexibility if their layers are treated as a library of functions rather than a fixed sequence. Performance improves over the standard forward pass when each input is dynamically routed through an adaptive sequence of skipped or repeated contiguous layer blocks. Reconstructing PoLar's diagnostic Monte Carlo Tree Search in more detail across 6 models, we reproduce several of its findings: skipping outperforms the standard pass, repeating outperforms skipping, and combining both outperforms either alone. Shorter programs suffice for easier questions, while harder questions require more layer repeats. For their main claim, the learned router's gain, we find what drives it. Under their setup, the router gains by repairing the format of the model's answers, which a single fixed program, selected on validation questions, does better. In our setup, with a prompt that already fixes the format, a larger data split and our search, its top prediction collapses to the standard pass on every model, although its top-5 programs together still gain. Beyond reproduction, we find in our setup that a small fixed menu of programs solves far more questions than the standard pass, and that programs which correct errors are brittle: undoing a single edit typically breaks the correction. PoLar mirrors principles of thalamo-cortical coordination between cortical-area-like transformer layers, although a working thalamus analogue, a router that picks a better program per input, remains to be found.

cs.AI↗

Efficient Analysis and Visualization of High-Resolution Computed Tomography Data for the Exploration of Enclosed Cuneiform Tablets

Cuneiform is the earliest known system of writing, first developed for the Sumerian language of southern Mesopotamia in the second half of the 4th millennium BC. Cuneiform signs are obtained by impressing a stylus on fresh clay tablets. For certain purposes, e.g. authentication by seal imprint, some cuneiform tablets were enclosed in clay envelopes, which cannot be opened without destroying them. The aim of our interdisciplinary project is the non-invasive study of clay tablets. A portable X-ray micro-CT scanner is developed to acquire density data of such artifacts on a high-resolution, regular 3D grid at collection sites. The resulting volume data is processed through feature-preserving denoising, extraction of high-accuracy surfaces using a manifold dual marching cubes algorithm and extraction of local features by enhanced curvature rendering and ambient occlusion. For the non-invasive study of cuneiform inscriptions, the tablet is virtually separated from its envelope by curvature-based segmentation. The computational- and data-intensive algorithms are optimized or near-real-time offline usage with limited resources at collection sites. To visualize the complexity-reduced and octree-based compressed representation of surfaces, we develop and implement an interactive application. To facilitate the analysis of such clay tablets, we implement shape-based feature extraction algorithms to enhance cuneiform recognition. Our workflow supports innovative 3D display and interaction techniques such as autostereoscopic displays and gesture control.

cs.GR↗