Is the Model Even Thinking?bookmark_border

Nemo Explain Visualizer (nev) is an interactive visual query tracer and builder for Nemo, a powerful Datalog reasoner with extended features. Our tools were developed with and for expert users. However, considering the lack of resources to learn Datalog and similar declarative logic programming languages, we conducted a qualitative user study to assess how our tools might help students. The study, interviewing 14 participants with varying levels of involvement with the content of a university course on knowledge graphs, revealed a very positive assessment of our tools, which strengthens the value of visual explanation tools beyond their intended use.

Diffusion-based text-to-image models can synthesize complex and highly structured visual content, yet the emergence and evolution of semantic structure remain difficult to interpret. Many existing workflows rely on aggregated attention or scalar summaries that separate temporal change from image-space evidence. To address this gap, we present a visual analytics framework for exploring attention dynamics in diffusion models: the step-indexed evolution of token-level cross-attention maps, their temporal concentration, and their spatial relationships. Our approach enables structured analysis of attention behavior across generation steps by integrating quantitative measures with data-driven stage identification in an interactive workflow. Case studies on a structured 60-prompt Stable-Diffusion-class benchmark illustrate recurring, interpretable patterns within this setting and show how linked temporal and spatial views facilitate the observation and discussion of generative processes, supporting more effective human–AI collaboration.

Rendering time series as chart images for CNN-based classification has become increasingly common in time-series classification (TSC). However, it remains unclear whether models learn underlying temporal patterns or rely on encoding-specific visual cues introduced by the chart design. We present VEIL: a systematic study that examines how different chart encodings influence learned representations using complementary analyses of similarity, transferability, and attribution. Attention-guided training appears to mitigate this effect when encoding sensitivity is consistently identified across multiple diagnostics, but provides limited or negative benefit when such signals are absent. These findings position VEIL within a broader question of how machines perceive visualizations---extending graphical perception from human readers to vision models---and show that visualization design choices shape learned representations in ways that warrant the treatment of chart-based TSC as a representation and measurement problem rather than a simple modeling decision.

As language models (LMs) rise in prominence, there is interest in making them more transparent in order to better understand their internal behavior. Recent interpretability work has focused on using sparse autoencoders (SAEs) to break down neuron activations at a given layer in the LM into human-understandable features, where each feature represents a concept that the model has learned. In this paper, we share work on using SAEs to analyze the behavior of text classification LMs. We present techniques for exploring the relationships between the SAE's features and the model's predictions and errors. We integrate these techniques into SAEfarer, a tool for analyzing concepts learned by text classification LMs. We assess SAEfarer in an expert pilot evaluation with five Ph.D. students.

Reinforcement learning has emerged as a dominant technique for fine-tuning the behavior of large language models, with policy optimization (PO) algorithms such as GRPO, DAPO, and Dr. GRPO emerging in rapid succession to advance state-of-the-art reasoning and alignment performance. However, the modular differences between these algorithms, including targeted improvements to clipping, advantage estimation, and reward aggregation, are introduced across separate papers with inconsistent notation, making them difficult to compare and intimidating to the non-expert community. We present UNIPO, to our knowledge the first interactive visualization tool that exposes the token-level training dynamics of RL fine-tuning algorithms through a unified design. UNIPO connects three complementary views, a high-level training overview, a step-level prompt and response inspector, and a side-by-side algorithm comparison, allowing learners to observe how individual design decisions propagate through training. Through two usage scenarios, we demonstrate how UNIPO supports both classroom instruction for non-experts and algorithm selection for AI practitioners. Our tool is open-source and publicly available at https://poloclub.github.io/unipo/.

Large language models are increasingly used to answer questions over annual reports, earnings decks, and analyst notes, yet their outputs remain difficult to verify in high-stakes financial workflows. A fluent answer can blend directly grounded statements, weak synthesis, and unsupported claims across narrative text, tables, and charts. We present EvidenceLens, a visual analytics prototype that treats financial question answering as a claim-evidence alignment problem. The system decomposes an answer into atomic claims, summarizes support composition and confidence--support gaps, and coordinates claim-level inspection with source passages, table cells, and chart regions. Its core representation is a multimodal claim-evidence matrix that exposes coverage, contradiction, and modality imbalance. To support reproducibility, we specify a JSON-based artifact schema, an inspectable multimodal alignment pipeline, and a deterministic review-priority ranking that maps backend signals into an auditable visual structure. Through two report-auditing scenarios, we illustrate how the interface makes grounded, weakly supported, and contradicted claims easier to inspect than in a linear chat transcript.

While attention heatmaps are the standard interface for inspect- ing multi-head attention, they primarily expose pairwise token-to- token weights rather than supporting comparison of how multiple tokens attend over the full context or how those attention profiles evolve across layers. We present AttentionRadViz, a visual analyt- ics tool that maps each token’s attention distribution to a point in a RadViz layout whose anchors are the input tokens. In this token- anchored projection, spatial proximity indicates similarity between tokens as attenders, while anchor positions preserve a semantic ref- erence to the input sequence. The system combines single-head inspection with layer-by-layer small multiples and radial token tra- jectories, helping researchers compare attention profiles across to- kens, follow how those profiles change across depth, and triage heads by profile structure, readily distinguishing structured heads from collapsed or sink-like ones. We demonstrate AttentionRadViz on Pythia-410M attention exports with focused examples of inter- active coreference-profile inspection, attention-sink detection, and transient cross-layer profile emergence. We also report a projection sanity check showing that RadViz preserves local attention-profile neighborhoods while retaining interpretable token anchors. Atten- tionRadViz is released as an open-source library.

Understanding how vision-language models (VLMs) interpret data visualizations remains an open problem, and is increasingly important as these models are used for analytical tasks where reliable reasoning is essential. We introduce a lightweight, diagnostic saliency map method tailored for text generation over images using transformer models, the current state-of-the-art models in visualization interpretation. Our approach aggregates the language model's attention over the visual tokens across all heads and layers, then maps this attention back onto the vision encoder's patch grid to localise it over the image, producing a direct correspondence between each generated answer token and the image regions it attended to. This yields fast, gradient-free saliency maps that expose how VLMs allocate focus across visual elements during answer generation, enabling inspection of whether model attention aligns with semantically relevant components. We evaluate our approach using a deletion metrics which validates the causal faithfulness of our saliency maps to the model’s behavior.