Current comparisons of large reasoning models often stop at aggregate metrics such as accuracy or chain length. We profile 12 open-source LRMs across science, math, and coding tasks to uncover their distinctive reasoning signatures.

Our automated method clusters reasoning styles, revealing families of models that hedge, self-correct, or boldly extrapolate. The taxonomy lets practitioners pick LRMs whose thinking style aligns with their safety or creativity needs.

Could models better at reducing its own undersiable behaviors if they have better understanding of what the bad behaviors are? We explore the possibility that pre-training on more toxic data can lead to better control in post-training, ultimately decreasing a model’s output toxicity.

We found empirical evidence that assistant LLMs maintain implicit user models that shape downstream answers. TalkTuner, our prototype dashboard, surfaces this hidden profile in real time and lets people recalibrate it.

The design study shows that exposing internal models improves trust calibration while revealing latent biases. Participants preferred the transparent workflow over today’s black-box chat interfaces.

Guardrails rarely behave uniformly across user demographics. We designed an evaluation framework that perturbs the conversational context with different gender, racial, and political cues, and uncovered refusal disparities in ChatGPT.

Younger, female, and Asian-American personas were disproportionately blocked—even when making the same request. The analysis urges safety teams to audit contextual sensitivity before deploying refusal policies at scale.

Building on correlational evidence from prior work, we intervene on spatial traces inside transformer activations. Editing latent directions that encode north/south cues causally improves performance on spatial reasoning probes.

The results suggest that causal representation analysis can bridge the gap between probing and control, offering a recipe for steering how LLMs internalize the physical world.

We asked whether 2D diffusion models secretly learn 3D priors. By identifying and steering geometry-sensitive directions, we gain precise control over rotations and camera elevations without retraining.

The intervention produces view-consistent edits, enabling controllable generation pipelines for film and XR teams.

Collaborating with Catherine Yeh, I designed visual analytics that expose learned attention flows within ViTs. The system lets researchers scrub through heads, layers, and patches to see how patterns align with human perception.

Insightful attention narratives improve model debugging and storytelling for broader audiences.

KALMUS is an open-source Python package for quantitative film color analysis. It computes palette statistics, contrast timelines, and interactive swatches so filmmakers can study authorship styles or plan grading.

The toolkit is now used in classroom critiques and was supported by the Mellon Fellowship.

We introduced a sliding-window augmentation strategy that learns motion-aware segmentation from sparsely annotated echocardiography videos. The approach generalizes multi-heartbeat sequences beyond traditional ED/ES frames.

Clinicians receive both segmentation and volumetric trend estimates, streamlining downstream cardiac assessments.

Training a 3D U-Net jointly for segmentation and motion estimation yielded superior generalization from CAMUS to EchoNet. The model handles sparsely annotated videos and improves tracking of cardiac structures in new hospitals.

The study offers a pathway for video-native clinical AI without dense frame labels.