18 papers
What Accuracy Hides In Learned Transformer Programs
Jonathan Lee, Narmeen Oozeer, Philip Quirke
Transformer Programs (TP) (Friedman et al., 2023) constrain a Transformer's architecture so that its learned computation can be extracted as readable Python code. TP expose learned…
Judging by the Cover: Cleaning LLM Truthfulness Benchmarks to Avoid Surface-Level Feature Leakage
Foad Namjoo, Remy Ogasawara, Amirali Abdullah +3
Binary-choice truth benchmarks ask models to choose between a correct and an incorrect answer, but if the two answers differ systematically in surface-level features, models can ex…
The Capability Frontier: Benchmarks Miss 82% of Model Performance
Bradley Fowler, Ryan Smith, Daniel Thi Graviet +8
Existing benchmarks typically report accuracy for a single model on a single run. This systematically understates real-world LLM capabilities, particularly under heterogeneous data…
Riemannian-Manifold Steering: Geometry-Aware Generative Autoencoders for Label-Free Steering
Narmeen Oozeer, Shivam Raval, Philip Quirke +4
Steering a language model - intervening on its internal activations to change downstream behaviour - has recently expanded beyond linear interpolation to nonlinear methods such as…
Make Mechanistic Interpretability Auditable: A Call to Develop Guidelines via Continuous Collaborative Reviewing
Michael Lan, Narmeen Fatimah Oozeer, Chaithanya Bandi +4
While mechanistic interpretability (MI) has produced important insights into neural network internals, the field has yet to establish a standardized system to audit experiments. As…
Geometry-Aware CLIP Retrieval via Local Cross-Modal Alignment and Steering
Nirmalendu Prakash, Narmeen Fatimah Oozeer, Xin Su +8
CLIP retrieval is typically framed as a pointwise similarity problem in a shared embedding space. While CLIP achieves strong global cross-modal alignment, many retrieval failures a…
Understanding and Mitigating Dataset Corruption in LLM Steering
Cullen Anderson, Narmeen Oozeer, Foad Namjoo +3
Contrastive steering has been shown as a simple and effective method to adjust the generative behavior of LLMs at inference time. It uses examples of prompt responses with and with…
DreamReader: An Interpretability Toolkit for Text-to-Image Models
Nirmalendu Prakash, Narmeen Oozeer, Michael Lan +6
Despite the rapid adoption of text-to-image (T2I) diffusion models, causal and representation-level analysis remains fragmented and largely limited to isolated probing techniques.…
Spectral Superposition: A Theory of Feature Geometry
Georgi Ivanov, Narmeen Oozeer, Shivam Raval +3
Neural networks represent more features than they have dimensions via superposition, forcing features to share representational space. Current methods decompose activations into sp…
Are LLM Evaluators Really Narcissists? Sanity Checking Self-Preference Evaluations
Dani Roytburg, Matthew Bozoukov, Matthew Nguyen +3
Recent research has shown that large language models (LLMs) favor their own outputs when acting as judges, undermining the integrity of automated post-training and evaluation workf…
Approximating Human Preferences Using a Multi-Judge Learned System
Eitán Sprejer, Fernando Avalos, Augusto Bernardi +3
Aligning LLM-based judges with human preferences is a significant challenge, as they are difficult to calibrate and often suffer from rubric sensitivity, bias, and instability. Ove…
Position: Require Frontier AI Labs To Release Small "Analog" Models
Shriyash Upadhyay, Chaithanya Bandi, Narmeen Oozeer +1
Recent proposals for regulating frontier AI models have sparked concerns about the cost of safety regulation, and most such regulations have been shelved due to the safety-innovati…
Breaking the Mirror: Activation-Based Mitigation of Self-Preference in LLM Evaluators
Dani Roytburg, Matthew Bozoukov, Matthew Nguyen +3
Large language models (LLMs) increasingly serve as automated evaluators, yet they suffer from "self-preference bias": a tendency to favor their own outputs over those of other mode…
Distribution-Aware Feature Selection for SAEs
Narmeen Oozeer, Nirmalendu Prakash, Michael Lan +2
Sparse autoencoders (SAEs) decompose neural activations into interpretable features. A widely adopted variant, the TopK SAE, reconstructs each token from its K most active latents.…
Beyond Monoliths: Expert Orchestration for More Capable, Democratic, and Safe Language Models
Philip Quirke, Narmeen Oozeer, Chaithanya Bandi +8
This position paper argues that the prevailing trajectory toward ever larger, more expensive generalist foundation models controlled by a handful of companies limits innovation and…
Beyond Linear Steering: Unified Multi-Attribute Control for Language Models
Narmeen Oozeer, Luke Marks, Shreyans Jain +2
Controlling multiple behavioral attributes in large language models (LLMs) at inference time is a challenging problem due to interference between attributes and the limitations of…
Activation Space Interventions Can Be Transferred Between Large Language Models
Narmeen Oozeer, Dhruv Nathawani, Nirmalendu Prakash +3
The study of representation universality in AI models reveals growing convergence across domains, modalities, and architectures. However, the practical applications of representati…
Bilinear Convolution Decomposition for Causal RL Interpretability
Narmeen Oozeer, Sinem Erisken, Alice Rigg
Efforts to interpret reinforcement learning (RL) models often rely on high-level techniques such as attribution or probing, which provide only correlational insights and coarse cau…