17 papers
Compositional Performance Modeling of LLM Inference Clusters with Hierarchical KV Caching
Ruiyang Ma, Sheng Qi, DaiKang Kuang +2
Design-space exploration of LLM inference clusters requires predicting how hardware and serving policies interact with hierarchical key-value (KV) caching. Isolated operation profi…
Removing Information Content Does Not Certify Tamper Resistance in Open-Weight Models
Domenic Rosati, Alessa Carbo, Ali Dadsetan +5
Does removing harmful information make open-weight models resistant to fine-tuning attacks? We show that mutual information at release alone cannot universally certify slow recover…
GAPS: Dimension-Level Gates for Conditional Activation Steering
Moghis Fereidouni, Muhammad Umair Haider, Hassan Sajjad +1
Activation steering suppresses undesired behaviors in language models by adding a steering vector to the hidden state during generation. Recent conditional methods such as CAST and…
Distribution-Specific Curvature Control with Finite-Sample Guarantees for Open-Weight Safety
Domenic Rosati, Ali Dadsetan, Hong Huang +5
A short fine-tuning run can undo the safety guards of an open-weight model---retraining a refusal-trained assistant to aid weapons development or produce hate speech. Preventing su…
Limits of Convergence-Rate Control for Open-Weight Safety
Domenic Rosati, Xijie Zeng, Hong Huang +4
Open-weight foundation models can be fine-tuned for harmful purposes after release, yet no existing training resistance methods provide theoretical guarantees. Treating these inter…
Vector Quantized Latent Concepts: A Scalable Alternative to Clustering-Based Concept Discovery
Xuemin Yu, Ankur Garg, Samira Ebrahimi Kahou +1
Large language models (LLMs) encode rich semantic information in their hidden states, yet it remains difficult to understand what information these internal representations capture…
Multi-Granular Node Pruning for Causal Circuit Discovery
Muhammad Umair Haider, Hammad Rizwan, Hassan Sajjad +1
Circuit discovery aims to identify minimal subnetworks that are responsible for specific behaviors in large language models (LLMs). Existing approaches primarily rely on iterative…
Dependency Parsing is More Parameter-Efficient with Normalization
Paolo Gajo, Domenic Rosati, Hassan Sajjad +1
Dependency parsing is the task of inferring natural language structure, often approached by modeling word interactions via attention through biaffine scoring. This mechanism works…
Neurons Speak in Ranges: Breaking Free from Discrete Neuronal Attribution
Muhammad Umair Haider, Hammad Rizwan, Hassan Sajjad +2
Pervasive polysemanticity in large language models (LLMs) undermines discrete neuron-concept attribution, posing a significant challenge for model interpretation and control. We sy…
Instance-Level Difficulty: A Missing Perspective in Machine Unlearning
Hammad Rizwan, Mahtab Sarvmaili, Hassan Sajjad +1
Current research on deep machine unlearning primarily focuses on improving or evaluating the overall effectiveness of unlearning methods while overlooking the varying difficulty of…
Evaluating Defences against Unsafe Feedback in RLHF
Domenic Rosati, Giles Edkins, Harsh Raj +5
While there has been progress towards aligning Large Language Models (LLMs) with human values and ensuring safe behaviour at inference time, safety guards can easily be removed whe…
Resolving Lexical Bias in Model Editing
Hammad Rizwan, Domenic Rosati, Ga Wu +1
Model editing aims to modify the outputs of large language models after they are trained. Previous approaches have often involved direct alterations to model weights, which can res…
Explaining the role of Intrinsic Dimensionality in Adversarial Training
Enes Altinisik, Safa Messaoud, Husrev Taha Sencar +2
Adversarial Training (AT) impacts different architectures in distinct ways: vision models gain robustness but face reduced generalization, encoder-based models exhibit limited robu…
Representation Noising: A Defence Mechanism Against Harmful Finetuning
Domenic Rosati, Jan Wehner, Kai Williams +7
Releasing open-source large language models (LLMs) presents a dual-use risk since bad actors can easily fine-tune these models for harmful purposes. Even without the open release o…
Long-form evaluation of model editing
Domenic Rosati, Robie Gonzales, Jinkun Chen +6
Evaluations of model editing currently only use the `next few token' completions after a prompt. As a result, the impact of these methods on longer natural language generation is l…
Data-centric Prediction Explanation via Kernelized Stein Discrepancy
Mahtab Sarvmaili, Hassan Sajjad, Ga Wu
Existing example-based prediction explanation methods often bridge test and training data points through the model's parameters or latent representations. While these methods offer…
Immunization against harmful fine-tuning attacks
Domenic Rosati, Jan Wehner, Kai Williams +4
Large Language Models (LLMs) are often trained with safety guards intended to prevent harmful text generation. However, such safety training can be removed by fine-tuning the LLM o…