Nirmalendu Prakash

Nirmalendu Prakash, Narmeen Oozeer, Michael Lan et al.

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. To address this gap, we introduce DreamReader: a unified framework that formalizes diffusion interpretability as composable representation operators spanning activation extraction, causal patching, structured ablations, and activation steering across modules and timesteps. DreamReader provides a model-agnostic abstraction layer enabling systematic analysis and intervention across diffusion architectures. Beyond consolidating existing methods, DreamReader introduces three novel intervention primitives for diffusion models: (1) representation fine-tuning (LoReFT) for subspace-constrained internal adaptation; (2) classifier-guided gradient steering using MLP probes trained on activations; and (3) component-level cross-model mapping for systematic study of transferability of representations across modalities. These mechanisms allows us to do lightweight white-box interventions on T2I models by drawing inspiration from interpretability techniques on LLMs. We demonstrate DreamReader through controlled experiments that (i) perform activation stitching between two models, and (ii) apply LoReFT to steer multiple activation units, reliably injecting a target concept into the generated images. Experiments are specified declaratively and executed in controlled batched pipelines to enable reproducible large-scale analysis. Across multiple case studies, we show that techniques adapted from language model interpretability yield promising and controllable interventions in diffusion models. DreamReader is released as an open source toolkit for advancing research on T2I interpretability.

Nirmalendu Prakash, Narmeen Fatimah Oozeer, Xin Su et al.

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 arise from local geometric inconsistencies: nearby items are incorrectly ordered, leading to systematic confusions (e.g., pentagon vs. hexagon) and produces diffuse, weakly controlled result sets. Prior work largely optimizes for point wise relevance or finetuning to mitigate these problems. We instead view retrieval as a problem of neighborhood alignment. Our work introduces (1) neighborhood-level re-ranking via Hungarian matching, which rewards structural consistency; (2) query-conditioned local steering, where directions derived from contrastive neighborhoods around the query reshape retrieval. We show that these techniques improve retrieval performance on attribute-binding and compositional retrieval tasks. Together, these methods operate on local neighborhoods but serve different roles: re-ranking rewards alignment whereas local steering controls neighborhood structure. This shows that retrieval quality and controllability depend critically on local structure, which can be exploited at inference time without retraining.

Wei Jie Yeo, Nirmalendu Prakash, Clement Neo et al.

Refusal is a key safety behavior in aligned language models, yet the internal mechanisms driving refusals remain opaque. In this work, we conduct a mechanistic study of refusal in instruction-tuned LLMs using sparse autoencoders to identify latent features that causally mediate refusal behaviors. We apply our method to two open-source chat models and intervene on refusal-related features to assess their influence on generation, validating their behavioral impact across multiple harmful datasets. This enables a fine-grained inspection of how refusal manifests at the activation level and addresses key research questions such as investigating upstream-downstream latent relationship and understanding the mechanisms of adversarial jailbreaking techniques. We also establish the usefulness of refusal features in enhancing generalization for linear probes to out-of-distribution adversarial samples in classification tasks. We open source our code in https://github.com/wj210/refusal_sae.

Ming Shan Hee, Aditi Kumaresan, Nguyen Khoi Hoang et al.

The rise of social media platforms has brought about a new digital culture called memes. Memes, which combine visuals and text, can strongly influence public opinions on social and cultural issues. As a result, people have become interested in categorizing memes, leading to the development of various datasets and multimodal models that show promising results in this field. However, there is currently a lack of a single library that allows for the reproduction, evaluation, and comparison of these models using fair benchmarks and settings. To fill this gap, we introduce the Meme Analytical Tool Kit (MATK), an open-source toolkit specifically designed to support existing memes datasets and cutting-edge multimodal models. MATK aims to assist researchers and engineers in training and reproducing these multimodal models for meme classification tasks, while also providing analysis techniques to gain insights into their strengths and weaknesses. To access MATK, please visit \url{https://github.com/Social-AI-Studio/MATK}.

Han Wang, Nirmalendu Prakash, Nguyen Khoi Hoang et al.

Topic modeling is a widely used technique for revealing underlying thematic structures within textual data. However, existing models have certain limitations, particularly when dealing with short text datasets that lack co-occurring words. Moreover, these models often neglect sentence-level semantics, focusing primarily on token-level semantics. In this paper, we propose PromptTopic, a novel topic modeling approach that harnesses the advanced language understanding of large language models (LLMs) to address these challenges. It involves extracting topics at the sentence level from individual documents, then aggregating and condensing these topics into a predefined quantity, ultimately providing coherent topics for texts of varying lengths. This approach eliminates the need for manual parameter tuning and improves the quality of extracted topics. We benchmark PromptTopic against the state-of-the-art baselines on three vastly diverse datasets, establishing its proficiency in discovering meaningful topics. Furthermore, qualitative analysis showcases PromptTopic's ability to uncover relevant topics in multiple datasets.

Nirmalendu Prakash, Han Wang, Nguyen Khoi Hoang et al.

The proliferation of social media has given rise to a new form of communication: memes. Memes are multimodal and often contain a combination of text and visual elements that convey meaning, humor, and cultural significance. While meme analysis has been an active area of research, little work has been done on unsupervised multimodal topic modeling of memes, which is important for content moderation, social media analysis, and cultural studies. We propose \textsf{PromptMTopic}, a novel multimodal prompt-based model designed to learn topics from both text and visual modalities by leveraging the language modeling capabilities of large language models. Our model effectively extracts and clusters topics learned from memes, considering the semantic interaction between the text and visual modalities. We evaluate our proposed model through extensive experiments on three real-world meme datasets, which demonstrate its superiority over state-of-the-art topic modeling baselines in learning descriptive topics in memes. Additionally, our qualitative analysis shows that \textsf{PromptMTopic} can identify meaningful and culturally relevant topics from memes. Our work contributes to the understanding of the topics and themes of memes, a crucial form of communication in today's society.\\ \red{\textbf{Disclaimer: This paper contains sensitive content that may be disturbing to some readers.}}

Ri Chi Ng, Nirmalendu Prakash, Ming Shan Hee et al.

To address the limitations of current hate speech detection models, we introduce \textsf{SGHateCheck}, a novel framework designed for the linguistic and cultural context of Singapore and Southeast Asia. It extends the functional testing approach of HateCheck and MHC, employing large language models for translation and paraphrasing into Singapore's main languages, and refining these with native annotators. \textsf{SGHateCheck} reveals critical flaws in state-of-the-art models, highlighting their inadequacy in sensitive content moderation. This work aims to foster the development of more effective hate speech detection tools for diverse linguistic environments, particularly for Singapore and Southeast Asia contexts.

Narmeen Oozeer, Dhruv Nathawani, Nirmalendu Prakash et al.

The study of representation universality in AI models reveals growing convergence across domains, modalities, and architectures. However, the practical applications of representation universality remain largely unexplored. We bridge this gap by demonstrating that safety interventions can be transferred between models through learned mappings of their shared activation spaces. We demonstrate this approach on two well-established AI safety tasks: backdoor removal and refusal of harmful prompts, showing successful transfer of steering vectors that alter the models' outputs in a predictable way. Additionally, we propose a new task, \textit{corrupted capabilities}, where models are fine-tuned to embed knowledge tied to a backdoor. This tests their ability to separate useful skills from backdoors, reflecting real-world challenges. Extensive experiments across Llama, Qwen and Gemma model families show that our method enables using smaller models to efficiently align larger ones. Furthermore, we demonstrate that autoencoder mappings between base and fine-tuned models can serve as reliable ``lightweight safety switches", allowing dynamic toggling between model behaviors.

Nirmalendu Prakash, Yeo Wei Jie, Amir Abdullah et al.

Refusal on harmful prompts is a key safety behaviour in instruction-tuned large language models (LLMs), yet the internal causes of this behaviour remain poorly understood. We study two public instruction-tuned models, Gemma-2-2B-IT and LLaMA-3.1-8B-IT, using sparse autoencoders (SAEs) trained on residual-stream activations. Given a harmful prompt, we search the SAE latent space for feature sets whose ablation flips the model from refusal to compliance, demonstrating causal influence and creating a jailbreak. Our search proceeds in three stages: (1) Refusal Direction: find a refusal-mediating direction and collect SAE features near that direction; (2) Greedy Filtering: prune to a minimal set; and (3) Interaction Discovery: fit a factorization machine (FM) that captures nonlinear interactions among the remaining active features and the minimal set. This pipeline yields a broad set of jailbreak-critical features, offering insight into the mechanistic basis of refusal. Moreover, we find evidence of redundant features that remain dormant unless earlier features are suppressed. Our findings highlight the potential for fine-grained auditing and targeted intervention in safety behaviours by manipulating the interpretable latent space.

Narmeen Oozeer, Nirmalendu Prakash, Michael Lan et al.

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. However, this approach is inefficient, as some tokens carry more information than others. BatchTopK addresses this limitation by selecting top activations across a batch of tokens. This improves average reconstruction but risks an "activation lottery," where rare high-magnitude features crowd out more informative but lower-magnitude ones. To address this issue, we introduce Sampled-SAE: we score the columns (representing features) of the batch activation matrix (via $L_2$ norm or entropy), forming a candidate pool of size $Kl$, and then apply Top-$K$ to select tokens across the batch from the restricted pool of features. Varying $l$ traces a spectrum between batch-level and token-specific selection. At $l=1$, tokens draw only from $K$ globally influential features, while larger $l$ expands the pool toward standard BatchTopK and more token-specific features across the batch. Small $l$ thus enforces global consistency; large $l$ favors fine-grained reconstruction. On Pythia-160M, no single value optimizes $l$ across all metrics: the best choice depends on the trade-off between shared structure, reconstruction fidelity, and downstream performance. Sampled-SAE thus reframes BatchTopK as a tunable, distribution-aware family.

Nirmalendu Prakash, Lee Ka Wei Roy

Large Language Models (LLMs) such as Mistral and LLaMA have showcased remarkable performance across various natural language processing (NLP) tasks. Despite their success, these models inherit social biases from the diverse datasets on which they are trained. This paper investigates the propagation of biases within LLMs through a novel feature-based analytical approach. Drawing inspiration from causal mediation analysis, we hypothesize the evolution of bias-related features and validate them using interpretability techniques like activation and attribution patching. Our contributions are threefold: (1) We introduce and empirically validate a feature-based method for bias analysis in LLMs, applied to LLaMA-2-7B, LLaMA-3-8B, and Mistral-7B-v0.3 with templates from a professions dataset. (2) We extend our method to another form of gender bias, demonstrating its generalizability. (3) We differentiate the roles of MLPs and attention heads in bias propagation and implement targeted debiasing using a counterfactual dataset. Our findings reveal the complex nature of bias in LLMs and emphasize the necessity for tailored debiasing strategies, offering a deeper understanding of bias mechanisms and pathways for effective mitigation.

Nirmalendu Prakash, Ming Shan Hee, Roy Ka-Wei Lee

Total Defence is a defence policy combining and extending the concept of military defence and civil defence. While several countries have adopted total defence as their defence policy, very few studies have investigated its effectiveness. With the rapid proliferation of social media and digitalisation, many social studies have been focused on investigating policy effectiveness through specially curated surveys and questionnaires either through digital media or traditional forms. However, such references may not truly reflect the underlying sentiments about the target policies or initiatives of interest. People are more likely to express their sentiment using communication mediums such as starting topic thread on forums or sharing memes on social media. Using Singapore as a case reference, this study aims to address this research gap by proposing TotalDefMeme, a large-scale multi-modal and multi-attribute meme dataset that captures public sentiments toward Singapore's Total Defence policy. Besides supporting social informatics and public policy analysis of the Total Defence policy, TotalDefMeme can also support many downstream multi-modal machine learning tasks, such as aspect-based stance classification and multi-modal meme clustering. We perform baseline machine learning experiments on TotalDefMeme and evaluate its technical validity, and present possible future interdisciplinary research directions and application scenarios using the dataset as a baseline.