Vaibhav Singh

Soumya Suvra Ghosal, Souradip Chakraborty, Vaibhav Singh et al.

Reinforcement learning (RL) based post-training for explicit chain-of-thought (e.g., GRPO) improves the reasoning ability of multimodal large-scale reasoning models (MLRMs). But recent evidence shows that it can simultaneously degrade safety alignment and increase jailbreak success rates. We propose SafeThink, a lightweight inference-time defense that treats safety recovery as a satisficing constraint rather than a maximization objective. SafeThink monitors the evolving reasoning trace with a safety reward model and conditionally injects an optimized short corrective prefix ("Wait, think safely") only when the safety threshold is violated. In our evaluations across six open-source MLRMs and four jailbreak benchmarks (JailbreakV-28K, Hades, FigStep, and MM-SafetyBench), SafeThink reduces attack success rates by 30-60% (e.g., LlamaV-o1: 63.33% to 5.74% on JailbreakV-28K, R1-Onevision: 69.07% to 5.65% on Hades) while preserving reasoning performance (MathVista accuracy: 65.20% to 65.00%). A key empirical finding from our experiments is that safety recovery is often only a few steering steps away: intervening in the first 1-3 reasoning steps typically suffices to redirect the full generation toward safe completions.

Tatiana Chakravorti, Robert Fraleigh, Timothy Fritton et al.

We present a prototype hybrid prediction market and demonstrate the avenue it represents for meaningful human-AI collaboration. We build on prior work proposing artificial prediction markets as a novel machine-learning algorithm. In an artificial prediction market, trained AI agents buy and sell outcomes of future events. Classification decisions can be framed as outcomes of future events, and accordingly, the price of an asset corresponding to a given classification outcome can be taken as a proxy for the confidence of the system in that decision. By embedding human participants in these markets alongside bot traders, we can bring together insights from both. In this paper, we detail pilot studies with prototype hybrid markets for the prediction of replication study outcomes. We highlight challenges and opportunities, share insights from semi-structured interviews with hybrid market participants, and outline a vision for ongoing and future work.

Tatiana Chakravorti, Robert Fraleigh, Timothy Fritton et al.

Determining whether published scientific findings can successfully be replicated is a long-standing challenge in the empirical sciences. Existing approaches for replicability assessment typically rely either on human judgment, i.e., creative assembly of human experts, or on machine learning models trained on paper content metadata. While both approaches have demonstrated value, each also has important limitations. Human forecasts can be influenced by cognitive biases and narrow exposure to the research literature, while automated assessments often struggle to capture contextual cues and subtle signals of credibility. In this paper, we examine a hybrid approach. Specifically, we introduce a hybrid prediction market in which algorithmic agents trade alongside human participants to jointly estimate the likelihood that a published scientific finding will be corroborated via the outcome of a controlled replication study. Agents are trained on outcomes from hundreds of prior replication studies while human participants contribute domain knowledge through real-time trading. We evaluate this hybrid approach through multiple live experiments involving participants from different academic disciplines and compare its performance to artificial-only and human-only baselines. Our results show that, except for a few cases, hybrid markets match or outperform artificial prediction markets, producing more accurate and reliable replication forecasts.

Vaibhav Singh, Eugene Belilovsky, Rahaf Aljundi

In this paper, we investigate the phenomenon of grokking, where models exhibit delayed generalization following overfitting on training data. We focus on data-scarce regimes where the number of training samples falls below the critical threshold, making grokking unobservable, and on practical scenarios involving distribution shift. We first show that Knowledge Distillation (KD) from a model that has already grokked on a distribution (p1) can induce and accelerate grokking on a different distribution (p2), even when the available data lies below the critical threshold. This highlights the value of KD for deployed models that must adapt to new distributions under limited data. We then study training on the joint distribution (p1, p2) and demonstrate that while standard supervised training fails when either distribution has insufficient data, distilling from models grokked on the individual distributions enables generalization. Finally, we examine a continual pretraining setup, where a grokked model transitions from p1 to p2, and find that KD both accelerates generalization and mitigates catastrophic forgetting, achieving strong performance even with only 10% of the data. Together, our results provide new insights into the mechanics of grokking under knowledge transfer and underscore the central role of KD in enabling generalization in low-data and evolving distribution settings.

Naman Gupta, Vaibhav Singh, Arun Iyer et al.

Sequential multi-agent reasoning frameworks such as Chain-of-Agents (CoA) handle long-context queries by decomposing inputs into chunks and processing them sequentially using LLM-based worker agents that read from and update a bounded shared memory. From a probabilistic perspective, CoA aims to approximate the conditional distribution corresponding to a model capable of jointly reasoning over the entire long context. CoA achieves this through a latent-state factorization in which only bounded summaries of previously processed evidence are passed between agents. The resulting bounded-memory approximation introduces a lossy information bottleneck, making the final evidence state inherently dependent on the order in which chunks are processed. In this work, we study the problem of chunk ordering for long-context reasoning. We use the well-known Chow-Liu trees to learn a dependency structure that prioritizes strongly related chunks. Empirically, we show that a breadth-first traversal of the resulting tree yields chunk orderings that reduce information loss across agents and consistently outperform both default document-chunk ordering and semantic score-based ordering in answer relevance and exact-match accuracy across three long-context benchmarks.

Swayam Rajat Mohanty, Moin Uddin Maruf, Vaibhav Singh et al.

Solar photovoltaic (PV) modules are prone to damage during manufacturing, installation and operation which reduces their power conversion efficiency. This diminishes their positive environmental impact over the lifecycle. Continuous monitoring of PV modules during operation via unmanned aerial vehicles is essential to ensure that defective panels are promptly replaced or repaired to maintain high power conversion efficiencies. Computer vision provides an automatic, non-destructive and cost-effective tool for monitoring defects in large-scale PV plants. We review the current landscape of deep learning-based computer vision techniques used for detecting defects in solar modules. We compare and evaluate the existing approaches at different levels, namely the type of images used, data collection and processing method, deep learning architectures employed, and model interpretability. Most approaches use convolutional neural networks together with data augmentation or generative adversarial network-based techniques. We evaluate the deep learning approaches by performing interpretability analysis on classification tasks. This analysis reveals that the model focuses on the darker regions of the image to perform the classification. We find clear gaps in the existing approaches while also laying out the groundwork for mitigating these challenges when building new models. We conclude with the relevant research gaps that need to be addressed and approaches for progress in this field: integrating geometric deep learning with existing approaches for building more robust and reliable models, leveraging physics-based neural networks that combine domain expertise of physical laws to build more domain-aware deep learning models, and incorporating interpretability as a factor for building models that can be trusted. The review points towards a clear roadmap for making this technology commercially relevant.

Soumya Suvra Ghosal, Vaibhav Singh, Akash Ghosh et al.

Reward models are essential for aligning large language models (LLMs) with human preferences. However, most open-source multilingual reward models are primarily trained on preference datasets in high-resource languages, resulting in unreliable reward signals for low-resource Indic languages. Collecting large-scale, high-quality preference data for these languages is prohibitively expensive, making preference-based training approaches impractical. To address this challenge, we propose RELIC, a novel in-context learning framework for reward modeling in low-resource Indic languages. RELIC trains a retriever with a pairwise ranking objective to select in-context examples from auxiliary high-resource languages that most effectively highlight the distinction between preferred and less-preferred responses. Extensive experiments on three preference datasets- PKU-SafeRLHF, WebGPT, and HH-RLHF-using state-of-the-art open-source reward models demonstrate that RELIC significantly improves reward model accuracy for low-resource Indic languages, consistently outperforming existing example selection methods. For example, on Bodo-a low-resource Indic language-using a LLaMA-3.2-3B reward model, RELIC achieves a 12.81% and 10.13% improvement in accuracy over zero-shot prompting and state-of-the-art example selection method, respectively.

Hongbo Lan, Zhenlin An, Haoyu Li et al.

This paper introduces \sysname, a system that accelerates vision-guided physical property reasoning to enable augmented visual cognition. \sysname minimizes the run-time latency of this reasoning pipeline through a combination of both algorithmic and systematic optimizations, including rapid geometric 3D reconstruction, efficient semantic feature fusion, and parallel view encoding. Through these simple yet effective optimizations, \sysname reduces the end-to-end latency of this reasoning pipeline from 10--20 minutes to less than 6 seconds. A head-to-head comparison on the ABO dataset shows that \sysname achieves this 62.9$\times$--287.2$\times$ speedup while not only reaching on-par (and sometimes slightly better) object-level physical property estimation accuracy(e.g. mass), but also demonstrating superior performance in material segmentation and voxel-level inference than two SOTA baselines. We further combine gaze-tracking with \sysname to localize the object of interest in cluttered, real-world environments, streamlining the physical property reasoning on smart glasses. The case study with Meta Aria Glasses conducted at an IKEA furniture store demonstrates that \sysname achives consistently high performance compared to controlled captures, providing robust property estimations even with fewer views in real-world scenarios.

Soumya Suvra Ghosal, Souradip Chakraborty, Vaibhav Singh et al.

With the widespread deployment of Multimodal Large Language Models (MLLMs) for visual-reasoning tasks, improving their safety has become crucial. Recent research indicates that despite training-time safety alignment, these models remain vulnerable to jailbreak attacks. In this work, we first highlight an important safety gap to describe that alignment achieved solely through safety training may be insufficient against jailbreak attacks. To address this vulnerability, we propose Immune, an inference-time defense framework that leverages a safe reward model through controlled decoding to defend against jailbreak attacks. Additionally, we provide a mathematical characterization of Immune, offering insights on why it improves safety against jailbreaks. Extensive evaluations on diverse jailbreak benchmarks using recent MLLMs reveal that Immune effectively enhances model safety while preserving the model's original capabilities. For instance, against text-based jailbreak attacks on LLaVA-1.6, Immune reduces the attack success rate by 57.82% and 16.78% compared to the base MLLM and state-of-the-art defense strategy, respectively.

Vaibhav Singh, Paul Janson, Paria Mehrbod et al.

The ever-growing availability of unlabeled data presents both opportunities and challenges for training artificial intelligence systems. While self-supervised learning (SSL) has emerged as a powerful paradigm for extracting meaningful representations from vast amounts of unlabeled data, existing methods still struggle to adapt to the non-stationary, non-IID nature of real-world data streams without forgetting previously learned knowledge. Recent works have adopted a repeated cosine annealing schedule for large-scale continual pre-training; however, these schedules (1) inherently cause forgetting during the re-warming phase and (2) have not been systematically compared to existing continual SSL methods. In this work, we systematically compare the widely used cosine schedule with the recently proposed infinite learning rate schedule and empirically find the latter to be a more effective alternative. Our extensive empirical evaluation across diverse image and language datasets demonstrates that the infinite learning rate schedule consistently enhances continual pre-training performance compared to a repeated cosine decay without being restricted to a fixed iteration budget. For instance, in a small-scale MAE pre-training setup, it outperforms several strong baselines from the literature. We then scale up our experiments to larger MAE pre-training and autoregressive language model pre-training. Our results show that the infinite learning rate schedule remains effective at scale, surpassing repeated cosine decay for both MAE pre-training and zero-shot LM benchmarks.

Vaibhav Singh, Zafir Khalid, Edouard Oyallon et al.

Pre-training large neural networks at scale imposes heavy memory demands on accelerators and often requires costly communication. We introduce Subnetwork Data Parallelism (SDP), a distributed training framework that partitions a model into structured subnetworks trained across workers without exchanging activations. We study two complementary masking regimes: backward masking, which applies sparsity only in the backward step to retain unbiased gradients, and forward masking, which also removes parameters in the forward pass to deliver stronger efficiency gains while providing additional regularization. We further explore two subnetwork construction strategies: neuron level and block level, applied across both CNNs and transformers. In experiments spanning CNNs and transformers on CIFAR and ImageNet, as well as LLM pre-training on FineWeb, SDP reduces per-device memory usage by 30%-75% while maintaining or improving performance. Notably, in FLOP-matched settings, forward masking can sometimes achieve better performance.

Yashwanth M., Vaibhav Singh, Ayush Maheshwari et al. · berkeley, uw

We propose ARISE, a framework that iteratively induces rules and generates synthetic data for text classification. We combine synthetic data generation and automatic rule induction, via bootstrapping, to iteratively filter the generated rules and data. We induce rules via inductive generalisation of syntactic n-grams, enabling us to capture a complementary source of supervision. These rules alone lead to performance gains in both, in-context learning (ICL) and fine-tuning (FT) settings. Similarly, use of augmented data from ARISE alone improves the performance for a model, outperforming configurations that rely on complex methods like contrastive learning. Further, our extensive experiments on various datasets covering three full-shot, eight few-shot and seven multilingual variant settings demonstrate that the rules and data we generate lead to performance improvements across these diverse domains and languages.

Nitin Priyadarshini Shankar, Vaibhav Singh, Sheetal Kalyani et al.

We introduce BERTO, a BERT-based framework for traffic prediction and energy optimization in cellular networks. Built on transformer architectures, BERTO delivers high prediction accuracy, while its Balancing Loss Function and prompt-based customization allow operators to adjust the trade-off between power savings and performance. Natural language prompts guide the model to manage underprediction and overprediction in accordance with the operator's intent. Experiments on real-world datasets show that BERTO improves upon existing models with a $4.13$\% reduction in MSE while introducing the feature of balancing competing objectives of power saving and performance through simple natural language inputs, operating over a flexible range of $1.4$ kW in power and up to $9\times$ variation in service quality, making it well suited for intelligent RAN deployments.

Vaibhav Singh, Oleksiy Ostapenko, Pierre-André Noël et al.

Diffusion-based language models have recently emerged as a promising alternative to autoregressive generation, yet their reliance on Transformer backbones limits inference efficiency due to quadratic attention and KV-cache overhead. In this work, we introduce DiffuApriel, a masked diffusion language model built on a bidirectional Mamba backbone that combines the diffusion objective with linear-time sequence modeling. DiffuApriel matches the performance of Transformer-based diffusion models while achieving up to 4.4x higher inference throughput for long sequences with a 1.3B model. We further propose DiffuApriel-H, a hybrid variant that interleaves attention and mamba layers, offering up to 2.6x throughput improvement with balanced global and local context modeling. Our results demonstrate that bidirectional state-space architectures serve as strong denoisers in masked diffusion LMs, providing a practical and scalable foundation for faster, memory-efficient text generation.

Vaibhav Singh, Soumya Suvra Ghosal, Kapu Nirmal Joshua et al.

In-context learning (ICL) has emerged as a powerful paradigm for adapting large language models (LLMs) to new and data-scarce tasks using only a few carefully selected task-specific examples presented in the prompt. However, given the limited context size of LLMs, a fundamental question arises: Which examples should be selected to maximize performance on a given user query? While nearest-neighbor-based methods like KATE have been widely adopted for this purpose, they suffer from well-known drawbacks in high-dimensional embedding spaces, including poor generalization and a lack of diversity. In this work, we study this problem of example selection in ICL from a principled, information theory-driven perspective. We first model an LLM as a linear function over input embeddings and frame the example selection task as a query-specific optimization problem: selecting a subset of exemplars from a larger example bank that minimizes the prediction error on a specific query. This formulation departs from traditional generalization-focused learning theoretic approaches by targeting accurate prediction for a specific query instance. We derive a principled surrogate objective that is approximately submodular, enabling the use of a greedy algorithm with an approximation guarantee. We further enhance our method by (i) incorporating the kernel trick to operate in high-dimensional feature spaces without explicit mappings, and (ii) introducing an optimal design-based regularizer to encourage diversity in the selected examples. Empirically, we demonstrate significant improvements over standard retrieval methods across a suite of classification tasks, highlighting the benefits of structure-aware, diverse example selection for ICL in real-world, label-scarce scenarios.

Vaibhav Singh, Amrith Krishna, Karthika NJ et al.

Low-resource languages, by its very definition, tend to be under represented in the pre-training corpora of Large Language Models. In this work, we investigate three low-resource cross-lingual approaches that enable an LLM adapt to tasks in previously unseen languages. Llama-2 is an LLM where Indic languages, among many other language families, contribute to less than $0.005\%$ of the total $2$ trillion token pre-training corpora. In this work, we experiment with the English-dominated Llama-2 for cross-lingual transfer to three Indic languages, Bengali, Hindi, and Tamil as target languages. We study three approaches for cross-lingual transfer, under ICL and fine-tuning. One, we find that adding additional supervisory signals via a dominant language in the LLM, leads to improvements, both under in-context learning and fine-tuning. Two, adapting the target languages to word reordering may be beneficial under ICL, but its impact diminishes with fine tuning. Finally, continued pre-training in one low-resource language can improve model performance for other related low-resource languages.

Vaibhav Singh, Rahaf Aljundi, Eugene Belilovsky

Foundational Vision-Language Models (VLMs) excel across diverse tasks, but adapting them to new domains without forgetting prior knowledge remains a critical challenge. Continual Learning (CL) addresses this challenge by enabling models to learn sequentially from new data while mitigating the forgetting of prior information, typically under supervised settings involving label shift. Nonetheless, abrupt distribution shifts can still cause substantial forgetting, potentially nullifying the benefits of supervised updates, especially when storing or replaying past data is infeasible. In this work, we propose leveraging unlabeled testtime data in an unsupervised manner to reinforce prior task performance without requiring replay or stored examples. Unlike traditional Test Time Adaptation (TTA), which primarily focuses on domain shift or corruption, our method improves performance on earlier tasks by exploiting representative test samples encountered during deployment. We introduce a simple Teacher-Student framework with gradient-based sparse parameter updates, and show that it effectively mitigates forgetting in class-incremental CL for VLMs, offering a memory-free alternative to episodic replay with strong empirical results.

Vinayak Kumar, Vaibhav Singh, P. K. Srijith et al.

Deep Gaussian processes (DGPs) provide a Bayesian non-parametric alternative to standard parametric deep learning models. A DGP is formed by stacking multiple GPs resulting in a well-regularized composition of functions. The Bayesian framework that equips the model with attractive properties, such as implicit capacity control and predictive uncertainty, makes it at the same time challenging to combine with a convolutional structure. This has hindered the application of DGPs in computer vision tasks, an area where deep parametric models (i.e. CNNs) have made breakthroughs. Standard kernels used in DGPs such as radial basis functions (RBFs) are insufficient for handling pixel variability in raw images. In this paper, we build on the recent convolutional GP to develop Convolutional DGP (CDGP) models which effectively capture image level features through the use of convolution kernels, therefore opening up the way for applying DGPs to computer vision tasks. Our model learns local spatial influence and outperforms strong GP based baselines on multi-class image classification. We also consider various constructions of convolution kernel over the image patches, analyze the computational trade-offs and provide an efficient framework for convolutional DGP models. The experimental results on image data such as MNIST, rectangles-image, CIFAR10 and Caltech101 demonstrate the effectiveness of the proposed approaches.