Moulik Choraria

Moulik Choraria, Xinbo Wu, Sourya Basu et al. · amazon-science

General purpose Vision Language Models (VLMs) have received tremendous interest in recent years, owing to their ability to learn rich vision-language correlations as well as their broad zero-shot competencies. One immensely popular line of work utilizes frozen unimodal models, by bridging vision representations to language using a trainable module called the QFormer. However, this method relies heavily on large-scale multimodal pretraining with huge computational overheads. To that end, we propose a more efficient framework for QFormer-based vision-language alignment. Our key idea relies on the observation that QFormer latents correspond more strongly to the frozen LLM's intermediate latent space. Consequently, instead of using QFormer latents as inputs to the LLM, we alter the framework by using the latents to directly condition the LLM latent space for image-to-text generation. We demonstrate the effectiveness of our approach against existing baselines in improving the efficiency of vision-language pretraining.

Sourya Basu, Moulik Choraria, Lav R. Varshney

We find limits to the Transformer architecture for language modeling and show it has a universal prediction property in an information-theoretic sense. We further analyze performance in non-asymptotic data regimes to understand the role of various components of the Transformer architecture, especially in the context of data-efficient training. We validate our theoretical analysis with experiments on both synthetic and real datasets.

Moulik Choraria, Ibtihal Ferwana, Ankur Mani et al.

Learning models that are robust to distribution shifts is a key concern in the context of their real-life applicability. Invariant Risk Minimization (IRM) is a popular framework that aims to learn robust models from multiple environments. The success of IRM requires an important assumption: the underlying causal mechanisms/features remain invariant across environments. When not satisfied, we show that IRM can over-constrain the predictor and to remedy this, we propose a relaxation via $\textit{partial invariance}$. In this work, we theoretically highlight the sub-optimality of IRM and then demonstrate how learning from a partition of training domains can help improve invariant models. Several experiments, conducted both in linear settings as well as with deep neural networks on tasks over both language and image data, allow us to verify our conclusions.

Moulik Choraria, Argyrios Gerogiannis, Anirban Das et al.

Compute scaling for LLM reasoning requires allocating budget between exploring solution approaches ($breadth$) and refining promising solutions ($depth$). Most methods implicitly trade off one for the other, yet why a given trade-off works remains unclear, and validation on a single model obscures the role of the model itself. We argue that $\textbf{the optimal strategy depends on the model's diversity profile, the spread of probability mass across solution approaches, and that this must be characterized before any exploration strategy is adopted.}$ We formalize this through a theoretical framework decomposing reasoning uncertainty and derive conditions under which tree-style depth refinement outperforms parallel sampling. We validate it on Qwen-3 4B and Olmo-3 7B families, showing that lightweight signals suffice for depth-based refinement on low-diversity aligned models while yielding limited utility for high-diversity base models, which we hypothesize require stronger compensation for lower exploration coverage.

Avi Bagchi, Akhil Bhimaraju, Moulik Choraria et al.

Watermarking has emerged as a promising technique to track AI-generated content and differentiate it from authentic human creations. While prior work extensively studies watermarking for autoregressive large language models (LLMs) and image diffusion models, none address discrete diffusion language models, which are becoming popular due to their high inference throughput. In this paper, we introduce the first watermarking method for discrete diffusion models by applying the distribution-preserving Gumbel-max trick at every diffusion step and seeding the randomness with the sequence index to enable reliable detection. We experimentally demonstrate that our scheme is reliably detectable on state-of-the-art diffusion language models and analytically prove that it is distortion-free with an exponentially decaying probability of false detection in the token sequence length.

Moulik Choraria, Argyrios Gerogiannis, Vidhata Jayaraman et al.

Hopfield networks and their generalizations have established deep connections among biological associative memories, statistical physics, and transformers. Yet most models treat retrieval as a fixed query-to-memory mapping, ignoring the role of external context in recall. In this work, we propose a two-stage associative memory architecture, wherein a context-gate subcircuit reshapes the retrieval energy landscape before and during recall. We show theoretically that context gating increases inter-memory separation while inducing sparsity, translating into exponential improvements in retrieval. Crucially, we prove that the system admits a unique self-consistent fixed point, revealing that the resulting retrieval state is driven by both a direct contextual bias and a second-order retrieval-gate feedback loop. We then bridge this theory to transformers; specifically, we evaluate a first-order approximation on Llama-3, confirming that in-context learning acts as context-gated retrieval. Native dynamics mirror our theory: context localizes a memory subspace, enabling the zero-shot query to cleanly discriminate. Ultimately, this framework provides a mechanistic link between associative memory theory and LLM phenomenology.

Max Hartman, Vidhata Jayaraman, Moulik Choraria et al.

Frontier models push the boundaries of what is learnable at extreme computational costs, yet distillation via sampling reasoning traces exposes closed-source frontier models to adversarial third parties who can bypass their guardrails and misappropriate their capabilities, raising safety, security, and intellectual privacy concerns. To address this, there is growing interest in building antidistillation methods, which aim to poison reasoning traces to hinder downstream student model learning while maintaining teacher performance. However, current techniques lack theoretical grounding, requiring either heavy fine-tuning or access to student model proxies for gradient based attacks, and often lead to a significant teacher performance degradation. In this work, we present a theoretical formulation of antidistillation as a Stackelberg game, grounding a problem that has so far largely been approached heuristically. Guided by the desired design properties our formulation reveals, we propose \texttt{TraceGuard}, an efficient, post-generation black-box method to poison sentences with high importance for teacher reasoning. Our work offers a scalable solution to share model insights safely, ensuring that the advancement of reasoning capabilities does not come at the cost of intellectual privacy or AI safety alignment.

Moulik Choraria, Xinbo Wu, Akhil Bhimaraju et al. · amazon-science

The hyperscaling of data and parameter count in transformer models is yielding diminishing performance improvement, especially when weighed against training costs. Such plateauing underlines a growing need for more efficient finetuning and inference, without sacrificing performance. This is particularly pressing for multimodal learning, where the overhead of processing multimodal tokens alongside language data often limits the practical viability of these systems. In parallel, advances in representation learning and interpretability have deepened our understanding of how such models process and encode information. Notably, recent work has uncovered implicit cross-modal alignment in the deeper layers of large pretrained models. Interestingly, this aligns with our own observations that models naturally defer most cross-modal token interactions to deeper stages of computation. Building on this, we propose a simple modification. Instead of concatenation with the language prompt at the start, we insert multimodal tokens directly into the middle, allowing them to entirely bypass the early layers. Our results with diverse modalities: 1) LLaVA \& BLIP for vision, 2) LTU for audio, and 3) MoLCA for molecular data, indicate that our method reduces computational costs during both training and inference, while at the very least, preserving, if not surpassing the performance of existing baselines. Our work has important implications for scaling and composing pretrained models in a resource-efficient manner.

Max Hartman, Vidhata Jayaraman, Moulik Choraria et al.

Vision-language models (VLMs) achieve incredible performance across a wide range of tasks, but their large size makes inference costly. Recent work shows that selectively skipping VLM layers can improve efficiency with minimal performance loss or even performance improvements. However, this technique remains underused due to the limited understanding of when layer skipping is beneficial. In this paper, we develop a framework that uses information and learning theory to characterize the conditions under which layer skipping enhances efficiency without sacrificing performance. Motivated by these observations, we analyze the evolution of the VLM's hidden representations through the LLM backbone and show that layers with large redundancy as predicted by our framework coincide with those skipped by popular layer-skipping methods in practice, providing a unified theoretical scaffolding for multiple efficient inference techniques. Our experiments demonstrate that skipping such layers yields faster inference that preserves performance, and also show that applying skipping outside these conditions leads to model degradation.

Moulik Choraria, Leello Tadesse Dadi, Grigorios Chrysos et al.

Polynomial neural networks (PNNs) have been recently shown to be particularly effective at image generation and face recognition, where high-frequency information is critical. Previous studies have revealed that neural networks demonstrate a $\textit{spectral bias}$ towards low-frequency functions, which yields faster learning of low-frequency components during training. Inspired by such studies, we conduct a spectral analysis of the Neural Tangent Kernel (NTK) of PNNs. We find that the $Π$-Net family, i.e., a recently proposed parametrization of PNNs, speeds up the learning of the higher frequencies. We verify the theoretical bias through extensive experiments. We expect our analysis to provide novel insights into designing architectures and learning frameworks by incorporating multiplicative interactions via polynomials.

Moulik Choraria, Ibtihal Ferwana, Ankur Mani et al.

The Invariant Risk Minimization (IRM) framework aims to learn invariant features from a set of environments for solving the out-of-distribution (OOD) generalization problem. The underlying assumption is that the causal components of the data generating distributions remain constant across the environments or alternately, the data "overlaps" across environments to find meaningful invariant features. Consequently, when the "overlap" assumption does not hold, the set of truly invariant features may not be sufficient for optimal prediction performance. Such cases arise naturally in networked settings and hierarchical data-generating models, wherein the IRM performance becomes suboptimal. To mitigate this failure case, we argue for a partial invariance framework. The key idea is to introduce flexibility into the IRM framework by partitioning the environments based on hierarchical differences, while enforcing invariance locally within the partitions. We motivate this framework in classification settings with causal distribution shifts across environments. Our results show the capability of the partial invariant risk minimization to alleviate the trade-off between fairness and risk in certain settings.