The idea is to turn a raw reasoning trace into a Thought Graph and then read structure off that graph directly, without asking another model to judge it. This non-generative analysis avoids the cost and circularity of LLM-as-a-judge setups: instead of a subjective score, we derive a compact cognitive profile, a set of structural metrics organized into a handful of interpretable dimensions.
What surprised us most is how consistent these profiles are. Across thousands of traces spanning several models and question sets, reasoning structure behaves as a stable, model-level property: two models can reach similar accuracy while thinking in visibly different shapes, and the variation between models is larger than the variation across problem domains. In other words, how a model reasons is closer to a fingerprint than to a per-task artifact.
I think structural profiling is a useful complement to accuracy for understanding, comparing, and eventually improving reasoning systems. This work is currently under review for EMNLP 2026.
See the paper for the full methodology and results.
]]>The context is that point-based scanpath comparisons are brittle. Small, perceptually meaningless offsets between two fixations can be treated as large disagreements, while the shared area of visual intake around each point is ignored. This mismatch between the metric and how the eye actually samples a scene distorts how we score similarity between viewing behaviors.
Our idea is simple and grounded in perception: represent each fixation as a disc approximating its foveal extent, and measure agreement through the intersection over union of these discs along the scanpath. Overlap in attended area, rather than exact coincidence of points, becomes the basis for comparison, which tolerates natural jitter while still rewarding genuinely shared attention.
By closing this foveal gap, Disc IoU aims to give more faithful and robust scanpath comparisons for both human studies and model evaluation. This was joint work with M. Tliba, Z. Sellam, C. Distante, A. Bruno, and A. Chetouani.
The reference implementation is open-sourced as FDISS, so the metric can drop straight into existing scanpath-evaluation pipelines.
See the paper for the full methodology and results.
]]>The problem is that two scanpaths can trace very different spatial trajectories while landing on the same objects and meanings, or follow similar routes while attending to entirely different content. Purely geometric comparisons miss this. They reward spatial coincidence and penalize legitimate variation in viewing strategy, which limits how well they reflect shared understanding of a scene.
Our approach describes what lies under each fixation using a vision-language model, then measures agreement between these descriptions with established NLP similarity metrics. This shifts scanpath comparison from coordinates toward semantics, letting us quantify whether observers converged on the same interpretation of an image even when their eyes took different paths.
This matters for evaluating attention models, studying expertise, and comparing human and machine viewing behavior in settings where meaning matters more than pixel-level overlap. This was joint work with M. Tliba, B. Wang, A. Chetouani, U. Bagci, and A. Bruno.
The similarity tools are open-sourced as scanpath_nlp_metrics so others can apply semantic comparison to their own gaze data.
See the paper for the full methodology and results.
]]>The core difficulty is data. Rich eye-tracking corpora exist for natural images, but labeled gaze on paintings is scarce, and art invites a broader, more idiosyncratic range of exploration than everyday photographs. Training a scanpath model directly on paintings is therefore hard, and models trained only on natural images transfer poorly to the aesthetics and composition of art.
Our approach bridges this gap. Unsupervised domain adaptation transfers knowledge from natural-image gaze data toward paintings, while the stochastic formulation lets SPGen sample multiple distinct yet coherent scanpaths rather than collapsing to a single average trajectory. The aim is generated gaze behavior that looks and varies like that of real observers.
This opens practical uses in digital humanities, cultural heritage, and interactive museum experiences, where modeling how people look at art is as interesting as modeling what they look at. This was joint work with M. Tliba, A. Chetouani, and A. Bruno.
The code is released as SP_Gen for generating and experimenting with stochastic scanpaths on paintings.
See the paper for the full methodology and results.
]]>The challenge is that severity grading depends on structural cues, such as joint space narrowing and the arrangement of bony features, that a purely appearance-based model may capture only implicitly. Standard vision backbones learn powerful features but do not natively represent the morphology of the joint, which can make grading less robust and harder to ground in anatomy.
Our idea is to make the model morphology-aware by introducing graph priors alongside the visual representation. Graphs are well suited to encoding the structural relationships between anatomical regions, and coupling them with a vision model lets appearance and structure inform one another during classification, aligning the model more closely with how the joint is actually read.
This points toward KOA grading that is both more accurate and more interpretable, grounded in the anatomy clinicians rely on. This was joint work with M. Tliba, Y. Nasser, N. Aburaed, A. Chetouani, U. Bagci, and colleagues.
See the paper for the full methodology and results.
]]>The context is that osteoarthritis severity is expressed through structural changes in the knee, and grading these changes is inherently relational: what matters is how regions of the joint relate to one another, not only their isolated appearance. Models that treat the image as an unstructured whole can miss these relationships and the diagnostic patterns they encode.
Our approach uses graphs to make these structural relationships explicit. By modeling the joint as connected components rather than a flat image, the method can capture dependencies relevant to grading and use them to inform the severity prediction, aligning the representation with the anatomy behind the diagnosis.
This contributes to a broader line of work on structure-aware medical image analysis, where encoding relationships explicitly helps grading models generalize and remain interpretable. This was joint work with M. Tliba, Y. Nasser, A. Chetouani, and R. Jennane.
This paper appeared at EUSIPCO 2025 and is listed on my Publications page.
See the paper for the full methodology and results.
]]>The problem with collapsing quality into a single MOS is that it discards almost everything about why something looks good or bad. A number cannot say which distortion is present, whether it matters for the task at hand, or how the intended use changes what “good” means. As quality assessment moves into richer settings, this loss of context and explanation becomes a real limitation.
Our position is that quality models should reason about content and context, draw on multiple modalities including language, and produce assessments that are interpretable rather than opaque scores. Recent advances in multimodal models and language-based reasoning make this shift feasible, allowing quality to be described and justified, not merely predicted.
We hope this reframes quality assessment as a reasoning problem about perception and utility, pointing toward models that explain their judgments and adapt to context. This was joint work with M. Tliba, A. Chetouani, N. Aburaed, and A. Bruno.
See the paper for the full methodology and results.
]]>The context is that AI-generated imagery is entering medical workflows and research, yet we understand little about how experts perceive it. If a synthetic image subtly departs from real anatomy or texture, it may change how a clinician’s attention is deployed, even when the difference is hard to name explicitly. Gaze offers a window into that perceptual response.
Our study records and compares clinicians’ eye movements as they view real and generated images, looking for systematic differences in how attention is distributed. Such shifts would suggest that experts respond, consciously or not, to cues that separate synthetic from authentic medical content.
This is relevant both to trust in AI-generated imagery within clinical settings and to the possibility of using gaze as a signal for detecting synthetic content. This was joint work with D.C. Wong, B. Wang, G. Durak, M. Tliba, A. Chetouani, A.E. Cetin, and colleagues.
Presented at ETRA 2025; see my Publications page.
See the paper for the full methodology and results.
]]>APEX is an experimental agentic pipeline for editing portraits from natural-language instructions. The idea is to let an AI agent sit between a user’s plain-language request and the image editing operations that carry it out, so that “soften the lighting” or “change the background” becomes something the system can plan and execute. It is very much a prototype for exploring that connection between agents and image editing.
The repository experiments with how an agent can decompose an editing instruction, choose the right tools, and apply them to a portrait. Because it is exploratory, the emphasis is on trying out the workflow and understanding where it breaks rather than on delivering a polished editor.
This ties together two threads of my work: computer vision on one side and the reasoning capabilities of LLM-based agents on the other. Portraits are a nice testbed because the edits are visually intuitive, which makes it easy to judge whether the agent understood what was actually asked.
Explore the code on GitHub.
]]>llm-tuning-lab is my personal lab for experimenting with the fine-tuning of small language models. It is meant for practitioners and researchers who want a place to collect working recipes, sanity-check ideas, and iterate quickly without committing to a heavyweight framework. Think of it as a notebook of experiments rather than a finished product: an experimental and evolving repository.
The goal is practical: gather reusable recipes and scripts for adapting compact models to specific tasks, so that lessons learned in one experiment carry over to the next. Because it is a living workspace, expect the contents to shift as I test new approaches and prune the ones that do not hold up.
This project sits close to my current interest in LLMs and their practical adaptation. Working with small models keeps the iteration loop short and the compute footprint modest, which makes them a good vehicle for understanding what actually moves the needle during fine-tuning before scaling any of it up.
Explore the code on GitHub.
]]>