The variables of the low-level model (left) are divided into partitions (center) such that each low-level partition corresponds to a high level variable from the high-level model (right). The circles represent variables and the arrows represent causal dependencies. Blue circles are variables that are not being intervened on and red circles are variables that are being intervened on. Observe that a low-level causal dependence between partitions does not necessarily result in a high-level causal dependence between variables and that not every low-level intervention results in a high level intervention.

Reveals Large Language Models' Strengths and Biases In this supplementary materials we show additional results mentioned in the main paper. First, we give experimental details in Section A. Next, we show results for Llama 2 on the bandit task in Section B. Afterwards, we show in Section C.1 additional quantitative results for the expertise-based Section D provides additional details about the vision and language tasks. For more details on the code please refer to the README.md Section A.1) and the amount of compute required to reproduce our experiments (Section Section A.2) A.1 Prompt variations generated by meta-prompting Work done whilst visiting University of Tübingen 37th Conference on Neural Information Processing Systems (NeurIPS 2023). For all Vicuna-13B based experiments (bandit, reasoning and vision) we used a single Nvidia A100-40GB GPU.

Reveals Large Language Models' Strengths and Biases In this supplementary materials we show additional results mentioned in the main paper. First, we give experimental details in Section A. Next, we show results for Llama 2 on the bandit task in Section B. Afterwards, we show in Section C.1 additional quantitative results for the expertise-based Section D provides additional details about the vision and language tasks. For more details on the code please refer to the README.md Section A.1) and the amount of compute required to reproduce our experiments (Section Section A.2) A.1 Prompt variations generated by meta-prompting Work done whilst visiting University of Tübingen 37th Conference on Neural Information Processing Systems (NeurIPS 2023). For all Vicuna-13B based experiments (bandit, reasoning and vision) we used a single Nvidia A100-40GB GPU.

We introduce Blueprint-Bench, a benchmark designed to evaluate spatial reasoning capabilities in AI models through the task of converting apartment photographs into accurate 2D floor plans. While the input modality (photographs) is well within the training distribution of modern multimodal models, the task of spatial reconstruction requires genuine spatial intelligence: inferring room layouts, understanding connectivity, and maintaining consistent scale. We evaluate leading language models (GPT-5, Claude 4 Opus, Gemini 2.5 Pro, Grok-4), image generation models (GPT-Image, NanoBanana), and agent systems (Codex CLI, Claude Code) on a dataset of 50 apartments with approximately 20 interior images each. Our scoring algorithm measures similarity between generated and ground-truth floor plans based on room connectivity graphs and size rankings. Results reveal a significant blind spot in current AI capabilities: most models perform at or below a random baseline, while human performance remains substantially superior. Image generation models particularly struggle with instruction following, while agent-based approaches with iterative refinement capabilities show no meaningful improvement over single-pass generation. Blueprint-Bench provides the first numerical framework for comparing spatial intelligence across different model architectures. We will continue evaluating new models as they are released and welcome community submissions, monitoring for the emergence of spatial intelligence in generalist AI systems.

Multi-digit addition is a clear probe of the computational power of large language models. To dissect the internal arithmetic processes in LLaMA-3-8B-Instruct, we combine linear probing with logit-lens inspection. Inspired by the step-by-step manner in which humans perform addition, we propose and analyze a coherent four-stage trajectory in the forward pass:Formula-structure representations become linearly decodable first, while the answer token is still far down the candidate list.Core computational features then emerge prominently.At deeper activation layers, numerical abstractions of the result become clearer, enabling near-perfect detection and decoding of the individual digits in the sum.Near the output, the model organizes and generates the final content, with the correct token reliably occupying the top rank.This trajectory suggests a hierarchical process that favors internal computation over rote memorization. We release our code and data to facilitate reproducibility.

Modern adaptive games require nuanced player understanding, yet most models use simplified 5-10 category taxonomies that fail to capture diversity. Behavioral clustering cannot distinguish players with different motivations who act similarly. We present a systematic evaluation of multi-modal classification at scale, combining behavioral telemetry with semantic context to support 36 player profiles. Using 19,413 gameplay sessions from an AI-controlled text-based RPG, we compared behavioral-only baselines with multi-modal approaches that integrate action sequences and semantic descriptions. Traditional clustering achieved only 10% accuracy for 36-category classification, limited by semantic conflation where opposite actions produced identical features. Our multi-modal LSTM processing action-text pairs improved accuracy to 21%, showing both potential and limits of non-conversational data. Analysis by behavioral complexity revealed that non-neutral profiles reached 42% accuracy (15x above random), while neutral profiles dropped to 25% (9x above random). Identical actions such as "help the merchant" cannot reveal whether a player is neutral or strategically waiting. Without access to reasoning, even multi-modal models struggle, though above-baseline results confirm a meaningful signal. Since prediction beyond 20 categories remains unexplored, our findings establish benchmarks for complex player modeling. Behavioral data alone plateaus near 10% for 36 categories, while multi-modal integration enables 25%. For designers, this shows that personality-based adaptation requires conversational interaction, as predefined choices cannot capture intent. Our evaluation at 36-category scale offers guidance for building adaptive games that better understand their players.

Accurate evaluation of event detection in time series is essential for applications such as stress monitoring with wearable devices, where ground truth is typically annotated as single-point events, even though the underlying phenomena are gradual and temporally diffused. Standard metrics like F1 and point-adjusted F1 (F1$_{pa}$) often misrepresent model performance in such real-world, imbalanced datasets. We introduce a window-based F1 metric (F1$_w$) that incorporates temporal tolerance, enabling a more robust assessment of event detection when exact alignment is unrealistic. Empirical analysis in three physiological datasets, two in-the-wild (ADARP, Wrist Angel) and one experimental (ROAD), indicates that F1$_w$ reveals meaningful model performance patterns invisible to conventional metrics, while its window size can be adapted to domain knowledge to avoid overestimation. We show that the choice of evaluation metric strongly influences the interpretation of model performance: using predictions from TimesFM, only our temporally tolerant metrics reveal statistically significant improvements over random and null baselines in the two in-the-wild use cases. This work addresses key gaps in time series evaluation and provides practical guidance for healthcare applications where requirements for temporal precision vary by context.

While large language models (LLMs) are generally considered proficient in generating language, how similar their language usage is to that of humans remains understudied. In this paper, we test whether models exhibit linguistic convergence, a core pragmatic element of human language communication, asking: do models adapt, or converge, to the linguistic patterns of their user? To answer this, we systematically compare model completions of exisiting dialogues to the original human responses across sixteen language models, three dialogue corpora, and a variety of stylometric features. We find that models strongly converge to the conversation's style, often significantly overfitting relative to the human baseline. While convergence patterns are often feature-specific, we observe consistent shifts in convergence across modeling settings, with instruction-tuned and larger models converging less than their pretrained counterparts. Given the differences between human and model convergence patterns, we hypothesize that the underlying mechanisms for these behaviors are very different.

Large Language Models (LLMs) possess intricate internal representations of the world, yet these latent structures are notoriously difficult to interpret or repurpose beyond the original prediction task. Building on our earlier work (Rothenfusser, 2025), which introduced the concept of vector ontologies as a framework for translating high-dimensional neural representations into interpretable geometric structures, this paper provides the first empirical validation of that approach. A vector ontology defines a domain-specific vector space spanned by ontologically meaningful dimensions, allowing geometric analysis of concepts and relationships within a domain. We construct an 8-dimensional vector ontology of musical genres based on Spotify audio features and test whether an LLM's internal world model of music can be consistently and accurately projected into this space. Using GPT-4o-mini, we extract genre representations through multiple natural language prompts and analyze the consistency of these projections across linguistic variations and their alignment with ground-truth data. Our results show (1) high spatial consistency of genre projections across 47 query formulations, (2) strong alignment between LLM-inferred genre locations and real-world audio feature distributions, and (3) evidence of a direct relationship between prompt phrasing and spatial shifts in the LLM's inferred vector ontology. These findings demonstrate that LLMs internalize structured, repurposable knowledge and that vector ontologies offer a promising method for extracting and analyzing this knowledge in a transparent and verifiable way.