10 papers
Reasoning Errors Have a Region and a Direction in the Residual-Stream Trajectory of LLMs
Hamed Damirchi, Ignacio Meza De la Jara, Damith Ranasinghe +2
As language models are increasingly used for tasks that require verifiable reasoning, reliably distinguishing sound reasoning from flawed reasoning has become an important practica…
Representation Trajectories Matters: Complementary Evidence for OOD Detection and Image Classification
Ignacio M. De la Jara, Cristian Rodriguez-Opazo, Hamed Damirchi +2
Vision models do not form a representation at once; each block revises it. We ask whether the resulting computation path contains evidence that the final representation discards, a…
Vertical Fusion: Condensing Internal Representations for Robust ViT Classification
Francesco Di Salvo, Shyam Nandan Rai, Hamed Damirchi +4
Despite exposing rich intermediate representations, Vision Transformers (ViTs) are almost exclusively utilized as black-box feature extractors, where only the last layer is conside…
Truth as a Trajectory: What Internal Representations Reveal About Large Language Model Reasoning
Hamed Damirchi, Ignacio Meza De la Jara, Ehsan Abbasnejad +3
Existing explainability methods for Large Language Models (LLMs) typically treat hidden states as static points in activation space, assuming that correct and incorrect inferences…
Decomposing Task Vectors for Refined Model Editing
Hamed Damirchi, Ehsan Abbasnejad, Zhen Zhang +1
Large pre-trained models have transformed machine learning, yet adapting these models effectively to exhibit precise, concept-specific behaviors remains a significant challenge. Ta…
The Quest for Winning Tickets in Low-Rank Adapters
Hamed Damirchi, Cristian Rodriguez-Opazo, Ehsan Abbasnejad +2
The Lottery Ticket Hypothesis (LTH) suggests that over-parameterized neural networks contain sparse subnetworks ("winning tickets") capable of matching full model performance when…
Synergy and Diversity in CLIP: Enhancing Performance Through Adaptive Backbone Ensembling
Cristian Rodriguez-Opazo, Ehsan Abbasnejad, Damien Teney +3
Contrastive Language-Image Pretraining (CLIP) stands out as a prominent method for image representation learning. Various architectures, from vision transformers (ViTs) to convolut…
Unveiling Backbone Effects in CLIP: Exploring Representational Synergies and Variances
Cristian Rodriguez-Opazo, Edison Marrese-Taylor, Ehsan Abbasnejad +4
Contrastive Language-Image Pretraining (CLIP) stands out as a prominent method for image representation learning. Various neural architectures, spanning Transformer-based models li…
Zero-shot Retrieval: Augmenting Pre-trained Models with Search Engines
Hamed Damirchi, Cristian Rodríguez-Opazo, Ehsan Abbasnejad +4
Large pre-trained models can dramatically reduce the amount of task-specific data required to solve a problem, but they often fail to capture domain-specific nuances out of the box…
Independent Modular Networks
Hamed Damirchi, Forest Agostinelli, Pooyan Jamshidi
Monolithic neural networks that make use of a single set of weights to learn useful representations for downstream tasks explicitly dismiss the compositional nature of data generat…