Memorized facts in fine-tuned LLMs often sit off the mid-layer reasoning path; relocating those representations recovers most multi-hop generalization failures.
Canonical reference
Towards vision-language mechanistic interpretability: A causal tracing tool for BLIP
Canonical reference. 100% of citing Pith papers cite this work as background.
21
Pith papers citing it
10
external citations · external index
Background
100% of classified citations
citation-role summary
background 5
citation-polarity summary
roles
background 5polarities
background 5representative citing papers
Face-Feature Tuning is a label-free logit remapping method that reduces FPR/TPR gaps across groups in deepfake detection while preserving overall accuracy.
PluRule is a new multimodal multilingual benchmark showing that state-of-the-art vision-language models perform only marginally better than a trivial baseline at detecting specific rule violations in pluralistic online communities.
SARR modifies trigonometric rotation encodings with object symmetry orders to produce unique continuous poses, enabling standard CNNs to outperform existing methods on symmetry-aware 6D pose estimation without custom losses or 3D models.
A Debye CZT-based wave-optics pipeline generates lens-diverse synthetic defocus blur datasets that improve cross-device deblurring generalization over existing real and synthetic data.
SENTRY is a plug-and-play module that replaces confidence-based memory writes with neighbor-aware cycle-consistent validation in SAM2 trackers, yielding new zero-shot SOTA results on LaSOT, GOT-10k and other benchmarks.
AtlasGS uses shared subject-specific Gaussian geometry learned from isotropic scans to achieve through-plane super-resolution and multi-modal harmonization in brain MRI with reported state-of-the-art fidelity on UK Biobank, GBM, and ABCD datasets.
SegRAG is a training-free retrieval-augmented framework that extracts class-specific point prompts from a filtered DINOv3 feature bank to boost SAM3 semantic segmentation performance on standard and agricultural benchmarks.
Complex multimodal architectures do not reliably outperform unimodal baselines or a simple multimodal baseline under standardized evaluation.
A PRISMA-based survey of 158 computational works on toxic meme detection introduces a new toxicity taxonomy and a framework linking target, intent, and conveyance tactics while noting trends in LLMs and cross-modal methods.
Learned priors (population, consensus, auto-decoder, meta-learning) all cut early tracking error of cardiac motion INRs versus random init; meta-learning keeps the best 50-step trajectory.
μMatch applies student-teacher semi-supervised methods with foundation models to improve segmentation of mitochondria, nuclei, and neurites in EM images over strong baselines.
Model interpretation methods are reformulated to emphasize baselines; gradient-based methods, IG, and Taylor expansion are unified with explicit baselines identified, and a revised IG is developed for improved results from any layer.
A new Gaussian Mixture Descriptor models 3D fragment surfaces by fitting GMMs to concave and convex patches and uses L2 distance for pair matching.
Token compression in ViT segmentation degrades sharply at high ratios due to information loss while structural pruning degrades smoothly; a moderate prune-then-merge pipeline improves the trade-off on ADE20K and Cityscapes under corruption.
Grad-ECLIP is an equivalent but flawed variant of attention-based interpretation, with two principles proposed to ensure model explanations reflect the original model.
XiYOLO uses iterative energy-aware neural architecture search and scaling to produce object detectors with stronger accuracy-energy tradeoffs than YOLO baselines on GPUs and NPUs.
Quantum circuits using FRQI and QPIE encodings implement Sobel edge and Harris corner detection, yielding outputs consistent with classical methods in noiseless simulations, with QPIE showing greater stability under limited shots.
A DenseNet201 base model trained on a constructed plant leaf disease dataset outperforms baselines and enables faster, more robust transfer learning with less data than general models.
A literature survey on abstract concept recognition in videos that catalogs prior tasks and datasets while advocating for foundation models and reuse of decades of community experience.
A literature review that categorizes deep learning approaches for visual hand gesture recognition, summarizes state-of-the-art methods across tasks, reviews datasets and metrics, and identifies challenges and future directions.
citing papers explorer
-
Towards Mechanistically Understanding Why Memorized Knowledge Fails to Generalize in Large Language Model Finetuning
Memorized facts in fine-tuned LLMs often sit off the mid-layer reasoning path; relocating those representations recovers most multi-hop generalization failures.
-
Toward Calibrated, Fair, and accurate Deepfake Detection
Face-Feature Tuning is a label-free logit remapping method that reduces FPR/TPR gaps across groups in deepfake detection while preserving overall accuracy.
-
PluRule: A Benchmark for Moderating Pluralistic Communities on Social Media
PluRule is a new multimodal multilingual benchmark showing that state-of-the-art vision-language models perform only marginally better than a trivial baseline at detecting specific rule violations in pluralistic online communities.
-
Towards Symmetry-sensitive Pose Estimation: A Rotation Representation for Symmetric Object Classes
SARR modifies trigonometric rotation encodings with object symmetry orders to produce unique continuous poses, enabling standard CNNs to outperform existing methods on symmetry-aware 6D pose estimation without custom losses or 3D models.
-
Realistic Compound-Lens Defocus Blur Synthesis
A Debye CZT-based wave-optics pipeline generates lens-diverse synthetic defocus blur datasets that improve cross-device deblurring generalization over existing real and synthetic data.
-
SENTRY: SAM2-Enhanced Neighbor-Aware and Temporally Reasoned Memory for Visual Tracking
SENTRY is a plug-and-play module that replaces confidence-based memory writes with neighbor-aware cycle-consistent validation in SAM2 trackers, yielding new zero-shot SOTA results on LaSOT, GOT-10k and other benchmarks.
-
AtlasGS: Brain MRI Spatial Resolution Harmonization With Shared Gaussian Geometry
AtlasGS uses shared subject-specific Gaussian geometry learned from isotropic scans to achieve through-plane super-resolution and multi-modal harmonization in brain MRI with reported state-of-the-art fidelity on UK Biobank, GBM, and ABCD datasets.
-
SegRAG: Training-Free Retrieval-Augmented Semantic Segmentation
SegRAG is a training-free retrieval-augmented framework that extracts class-specific point prompts from a filtered DINOv3 feature bank to boost SAM3 semantic segmentation performance on standard and agricultural benchmarks.
-
Fusion or Confusion? Multimodal Complexity Is Not All You Need
Complex multimodal architectures do not reliably outperform unimodal baselines or a simple multimodal baseline under standardized evaluation.
-
Toxic Memes: A Survey of Computational Perspectives on the Detection and Explanation of Meme Toxicities
A PRISMA-based survey of 158 computational works on toxic meme detection introduces a new toxicity taxonomy and a framework linking target, intent, and conveyance tactics while noting trends in LLMs and cross-modal methods.
-
Learning Cardiac Motion Priors for Implicit Neural Representations
Learned priors (population, consensus, auto-decoder, meta-learning) all cut early tracking error of cardiac motion INRs versus random init; meta-learning keeps the best 50-step trajectory.
-
$\mu$Match: Foundation Models for Semi-supervised Learning and Domain Adaptation in EM
μMatch applies student-teacher semi-supervised methods with foundation models to improve segmentation of mitochondria, nuclei, and neurites in EM images over strong baselines.
-
The Neglected Baseline in Model Interpretation
Model interpretation methods are reformulated to emphasize baselines; gradient-based methods, IG, and Taylor expansion are unified with explicit baselines identified, and a revised IG is developed for improved results from any layer.
-
Gmd: Gaussian mixture descriptor for pair matching of 3D fragments
A new Gaussian Mixture Descriptor models 3D fragment surfaces by fitting GMMs to concave and convex patches and uses L2 distance for pair matching.
-
When Token Compression Breaks: Structural Pruning vs. Token Reduction for Robust ViT Segmentation under High Compression
Token compression in ViT segmentation degrades sharply at high ratios due to information loss while structural pruning degrades smoothly; a moderate prune-then-merge pipeline improves the trade-off on ADE20K and Cityscapes under corruption.
-
Debunking Grad-ECLIP: A Comprehensive Study on Its Incorrectness and Fundamental Principles for Model Interpretation
Grad-ECLIP is an equivalent but flawed variant of attention-based interpretation, with two principles proposed to ensure model explanations reflect the original model.
-
XiYOLO: Energy-Aware Object Detection via Iterative Architecture Search and Scaling
XiYOLO uses iterative energy-aware neural architecture search and scaling to produce object detectors with stronger accuracy-energy tradeoffs than YOLO baselines on GPUs and NPUs.
-
Quantum Gradient-Based Approach for Edge and Corner Detection Using Sobel Kernels
Quantum circuits using FRQI and QPIE encodings implement Sobel edge and Harris corner detection, yielding outputs consistent with classical methods in noiseless simulations, with QPIE showing greater stability under limited shots.
-
Developing a Strong Pre-Trained Base Model for Plant Leaf Disease Classification
A DenseNet201 base model trained on a constructed plant leaf disease dataset outperforms baselines and enables faster, more robust transfer learning with less data than general models.
-
Looking Beyond the Obvious: A Survey on Abstract Concept Recognition for Video Understanding
A literature survey on abstract concept recognition in videos that catalogs prior tasks and datasets while advocating for foundation models and reuse of decades of community experience.
-
Visual Hand Gesture Recognition with Deep Learning: A Comprehensive Review of Methods, Datasets, Challenges and Future Research Directions
A literature review that categorizes deep learning approaches for visual hand gesture recognition, summarizes state-of-the-art methods across tasks, reviews datasets and metrics, and identifies challenges and future directions.