arXiv subjects

Lachin Naghashyar

Publications and source records attributed to Lachin Naghashyar.

4 recordsLinked to original sources

arXiv2026

Multimodal Model Diffing for Feature Discovery and Control

Multimodal Large Language Models (MLLMs) exhibit strong visual understanding, yet the internal features that cause these behaviors remain difficult to identify, audit, or control. While applicable to post-hoc inspection, hidden states that are decomposed into interpretable feature directions using sparse autoencoders (SAEs) neither readily isolate which features are changed by multimodal training, nor are they directly useful for targeted control. We introduce MMDiff, a multimodal model-diffing framework that trains multimodal SAEs and turns them into feature-level interfaces for discovering and controlling multimodal behavior. MMDiff supports three uses: (i) feature isolation, by diffing a base-LM SAE against its multimodal-adapted counterpart to identify features altered by multimodal training; (ii) task-specific feature detection, via per-token contrastive firing analysis that isolates causal features; and (iii) feature-level control, by causally removing or steering the discovered feature directions. We train multimodal SAEs for three MLLM families, LLaVA-MORE, PaliGemma 2, and InternVL3.5, and evaluate on visual-spatial understanding, multimodal safety, and OCR. MMDiff discovers sparse, causally specific features whose removal selectively degrades target behaviors by an average of 12% on spatial tasks and 17% on OCR, and reduces attack success rate by 24% on multimodal safety attacks, with no impact on VQA performance. Steering these features improves spatial and OCR accuracy by +3.6% and +1.8% on average over a standard single-layer steering baseline. These results show that multimodal SAEs can serve not only as interpretability tools, but as mechanisms for auditing, steering, and controlling MLLMs behavior toward safer and more capable generations.

cs.CV↗

arXiv2026

Towards Understanding Multimodal Fine-Tuning: Spatial Features

Contemporary Vision-Language Models (VLMs) achieve strong performance on a wide range of tasks by pairing a vision encoder with a pre-trained language model, fine-tuned for visual-text inputs. Yet despite these gains, it remains unclear how language backbone representations adapt during multimodal training and when vision-specific capabilities emerge. In this work, we present the first mechanistic analysis of VLM adaptation. Using stage-wise model diffing, a technique that isolates representational changes introduced during multimodal fine-tuning, we reveal how a language model learns to "see". We first identify vision-preferring features that emerge or reorient during fine-tuning. We then show that a selective subset of these features reliably encodes spatial relations, revealed through controlled shifts to spatial prompts. Finally, we trace the causal activation of these features to a small group of attention heads. Our findings show that stage-wise model diffing reveals when and where spatially grounded multimodal features arise. It also provides a clearer view of modality fusion by showing how visual grounding reshapes features that were previously text-only. This methodology enhances the interpretability of multimodal training and provides a foundation for understanding and refining how pretrained language models acquire vision-grounded capabilities.

cs.CV↗

arXiv2025

Improving Myocardial Infarction Detection via Synthetic ECG Pretraining

Myocardial infarction is a major cause of death globally, and accurate early diagnosis from electrocardiograms (ECGs) remains a clinical priority. Deep learning models have shown promise for automated ECG interpretation, but require large amounts of labeled data, which are often scarce in practice. We propose a physiology-aware pipeline that (i) synthesizes 12-lead ECGs with tunable MI morphology and realistic noise, and (ii) pre-trains recurrent and transformer classifiers with self-supervised masked-autoencoding plus a joint reconstruction-classification objective. We validate the realism of synthetic ECGs via statistical and visual analysis, confirming that key morphological features are preserved. Pretraining on synthetic data consistently improved classification performance, particularly in low-data settings, with AUC gains of up to 4 percentage points. These results show that controlled synthetic ECGs can help improve MI detection when real clinical data is limited.

eess.IV↗

arXiv2023

Enhanced Synthetic MRI Generation from CT Scans Using CycleGAN with Feature Extraction

In the field of radiotherapy, accurate imaging and image registration are of utmost importance for precise treatment planning. Magnetic Resonance Imaging (MRI) offers detailed imaging without being invasive and excels in soft-tissue contrast, making it a preferred modality for radiotherapy planning. However, the high cost of MRI, longer acquisition time, and certain health considerations for patients pose challenges. Conversely, Computed Tomography (CT) scans offer a quicker and less expensive imaging solution. To bridge these modalities and address multimodal alignment challenges, we introduce an approach for enhanced monomodal registration using synthetic MRI images. Utilizing unpaired data, this paper proposes a novel method to produce these synthetic MRI images from CT scans, leveraging CycleGANs and feature extractors. By building upon the foundational work on Cycle-Consistent Adversarial Networks and incorporating advancements from related literature, our methodology shows promising results, outperforming several state-of-the-art methods. The efficacy of our approach is validated by multiple comparison metrics.

eess.IV↗