Dip Roy

Dip Roy, Rajiv Misra, Sanjay Kumar Singh et al.

Although mechanism-based interpretability has generated an abundance of insight for discriminative network analysis, generative models are less understood -- particularly outside of image-related applications. We investigate how much of the causal circuitry found within image-related variational autoencoders (VAEs) will generalize to tabular data, as VAEs are increasingly used for imputation, anomaly detection, and synthetic data generation. In addition to extending a four-level causal intervention framework to four tabular and one image benchmark across five different VAE architectures (with 75 individual training runs per architecture and three random seed values for each run), this paper introduces three new techniques: posterior-calibration of Causal Effect Strength (CES), path-specific activation patching, and Feature-Group Disentanglement (FGD). The results from our experiments demonstrate that: (i) Tabular VAEs have circuits with modularity that is approximately 50% lower than their image counterparts. (ii) $β$-VAE experiences nearly complete collapse in CES scores when applied to heterogeneous tabular features (0.043 CES score for tabular data compared to 0.133 CES score for images), which can be directly attributed to reconstruction quality degradation (r = -0.886 correlation coefficient between CES and MSE). (iii) CES successfully captures nine of eleven statistically significant architecture differences using Holm--Šidák corrections. (iv) Interventions with high specificity predict the highest downstream AUC values (r = 0.460, p < .001). This study challenges the common assumption that architectural guidance from image-related studies can be transferred to tabular datasets.

Dip Roy, Rajiv Misra, Sanjay Kumar Singh et al.

When do large language models decide to hallucinate? Despite serious consequences in healthcare, law, and finance, few formal answers exist. Recent work shows autoregressive models maintain internal representations distinguishing factual from fictional outputs, but when these representations peak as a function of model scale remains poorly understood. We study the temporal dynamics of hallucination-indicative internal representations across 7 autoregressive transformers (117M--7B parameters) using three fact-based datasets (TriviaQA, Simple Facts, Biography; 552 labeled examples). We identify a scale-dependent phase transition: models below 400M parameters show chance-level probe accuracy at every generation position (AUC = 0.48--0.67), indicating no reliable factuality signal. Above $\sim$1B parameters, a qualitatively different regime emerges where peak detectability occurs at position zero -- before any tokens are generated -- then declines during generation. This pre-generation signal is statistically significant in both Pythia-1.4B (p = 0.012) and Qwen2.5-7B (p = 0.038), spanning distinct architectures and training corpora. At the 7B scale, we observe a striking dissociation: Pythia-6.9B (base model, trained on The Pile) produces a flat temporal profile ($Δ$ = +0.001, p = 0.989), while instruction-tuned Qwen2.5-7B shows a dominant pre-generation effect. This indicates raw scale alone is insufficient -- knowledge organization through instruction tuning or equivalent post-training is required for pre-commitment encoding. Activation steering along probe-derived directions fails to correct hallucinations across all models, confirming the signal is correlational rather than causal. Our findings provide scale-calibrated detection protocols and a concrete hypothesis on instruction tuning's role in developing knowledge circuits supporting factual generation.

Dip Roy, Rajiv Misra, Sanjay Kumar Singh

Extreme neural network sparsification (90% activation reduction) presents a critical challenge for mechanistic interpretability: understanding whether interpretable features survive aggressive compression. This work investigates feature survival under severe capacity constraints in hybrid Variational Autoencoder--Sparse Autoencoder (VAE-SAE) architectures. We introduce an adaptive sparsity scheduling framework that progressively reduces active neurons from 500 to 50 over 50 training epochs, and provide empirical evidence for fundamental limits of the sparsification-interpretability relationship. Testing across two benchmark datasets -- dSprites and Shapes3D -- with both Top-k and L1 sparsification methods, our key finding reveals a pervasive paradox: while global representation quality (measured by Mutual Information Gap) remains stable, local feature interpretability collapses systematically. Under Top-k sparsification, dead neuron rates reach $34.4\pm0.9\%$ on dSprites and $62.7\pm1.3\%$ on Shapes3D at k=50. L1 regularization -- a fundamentally different "soft constraint" paradigm -- produces equal or worse collapse: $41.7\pm4.4\%$ on dSprites and $90.6\pm0.5\%$ on Shapes3D. Extended training for 100 additional epochs fails to recover dead neurons, and the collapse pattern is robust across all tested threshold definitions. Critically, the collapse scales with dataset complexity: Shapes3D (RGB, 6 factors) shows $1.8\times$ more dead neurons than dSprites (grayscale, 5 factors) under Top-k and $2.2\times$ under L1. These findings establish that interpretability collapse under sparsification is intrinsic to the compression process rather than an artifact of any particular algorithm, training duration, or threshold choice.

Anisha Roy, Dip Roy

Large language models (LLMs) are increasingly used to generate financial alpha signals, yet growing evidence shows that LLMs memorize historical financial data from their training corpora, producing spurious predictive accuracy that collapses out-of-sample. This memorization-induced look-ahead bias threatens the validity of LLM-based quantitative strategies. Prior remedies -- model retraining and input anonymization -- are either prohibitively expensive or introduce significant information loss. No existing method offers practical, zero-cost signal-level filtering for real-time trading. We introduce MemGuard-Alpha, a post-generation framework comprising two algorithms: (i) the MemGuard Composite Score (MCS), which combines five membership inference attack (MIA) methods with temporal proximity features via logistic regression, achieving Cohen's d = 18.57 for contamination separation (d = 0.39-1.37 using MIA features alone); and (ii) Cross-Model Memorization Disagreement (CMMD), which exploits variation in training cutoff dates across LLMs to separate memorized signals from genuine reasoning. Evaluated across seven LLMs (124M-7B parameters), 50 S&P 100 stocks, 42,800 prompts, and five MIA methods over 5.5 years (2019-2024), CMMD achieves a Sharpe ratio of 4.11 versus 2.76 for unfiltered signals (49% improvement). Clean signals produce 14.48 bps average daily return versus 2.13 bps for tainted signals (7x difference). A striking crossover pattern emerges: in-sample accuracy rises with contamination (40.8% to 52.5%) while out-of-sample accuracy falls (47% to 42%), providing direct evidence that memorization inflates apparent accuracy at the cost of generalization.

Dip Roy

We present a quantitative circuit-level analysis of diffusion models, establishing computational pathways and mechanistic principles underlying image generation processes. Through systematic intervention experiments across 2,000 synthetic and 2,000 CelebA facial images, we discover fundamental algorithmic differences in how diffusion architectures process synthetic versus naturalistic data distributions. Our investigation reveals that real-world face processing requires circuits with measurably higher computational complexity (complexity ratio = 1.084 plus/minus 0.008, p < 0.001), exhibiting distinct attention specialization patterns with entropy divergence ranging from 0.015 to 0.166 across denoising timesteps. We identify eight functionally distinct attention mechanisms showing specialized computational roles: edge detection (entropy = 3.18 plus/minus 0.12), texture analysis (entropy = 4.16 plus/minus 0.08), and semantic understanding (entropy = 2.67 plus/minus 0.15). Intervention analysis demonstrates critical computational bottlenecks where targeted ablations produce 25.6% to 128.3% performance degradation, providing causal evidence for identified circuit functions. These findings establish quantitative foundations for algorithmic understanding and control of generative model behavior through mechanistic intervention strategies.

Dip Roy

Mechanistic interpretability of deep learning models has emerged as a crucial research direction for understanding the functioning of neural networks. While significant progress has been made in interpreting discriminative models like transformers, understanding generative models such as Variational Autoencoders (VAEs) remains challenging. This paper introduces a comprehensive causal intervention framework for mechanistic interpretability of VAEs. We develop techniques to identify and analyze "circuit motifs" in VAEs, examining how semantic factors are encoded, processed, and disentangled through the network layers. Our approach uses targeted interventions at different levels: input manipulations, latent space perturbations, activation patching, and causal mediation analysis. We apply our framework to both synthetic datasets with known causal relationships and standard disentanglement benchmarks. Results show that our interventions can successfully isolate functional circuits, map computational graphs to causal graphs of semantic factors, and distinguish between polysemantic and monosemantic units. Furthermore, we introduce metrics for causal effect strength, intervention specificity, and circuit modularity that quantify the interpretability of VAE components. Experimental results demonstrate clear differences between VAE variants, with FactorVAE achieving higher disentanglement scores (0.084) and effect strengths (mean 4.59) compared to standard VAE (0.064, 3.99) and Beta-VAE (0.051, 3.43). Our framework advances the mechanistic understanding of generative models and provides tools for more transparent and controllable VAE architectures.

Dip Roy

In medical imaging, anomaly detection is a vital element of healthcare diagnostics, especially for neurological conditions which can be life-threatening. Conventional deterministic methods often fall short when it comes to capturing the inherent uncertainty of anomaly detection tasks. This paper introduces a Bayesian Variational Autoencoder (VAE) equipped with multi-head attention mechanisms for detecting anomalies in brain magnetic resonance imaging (MRI). For the purpose of improving anomaly detection performance, we incorporate both epistemic and aleatoric uncertainty estimation through Bayesian inference. The model was tested on the BraTS2020 dataset, and the findings were a 0.83 ROC AUC and a 0.83 PR AUC. The data in our paper suggests that modeling uncertainty is an essential component of anomaly detection, enhancing both performance and interpretability and providing confidence estimates, as well as anomaly predictions, for clinicians to leverage in making medical decisions.