Rishabh Maheshwary

Corentin Dancette, Spencer Whitehead, Rishabh Maheshwary et al.

Despite advances in Visual Question Answering (VQA), the ability of models to assess their own correctness remains underexplored. Recent work has shown that VQA models, out-of-the-box, can have difficulties abstaining from answering when they are wrong. The option to abstain, also called Selective Prediction, is highly relevant when deploying systems to users who must trust the system's output (e.g., VQA assistants for users with visual impairments). For such scenarios, abstention can be especially important as users may provide out-of-distribution (OOD) or adversarial inputs that make incorrect answers more likely. In this work, we explore Selective VQA in both in-distribution (ID) and OOD scenarios, where models are presented with mixtures of ID and OOD data. The goal is to maximize the number of questions answered while minimizing the risk of error on those questions. We propose a simple yet effective Learning from Your Peers (LYP) approach for training multimodal selection functions for making abstention decisions. Our approach uses predictions from models trained on distinct subsets of the training data as targets for optimizing a Selective VQA model. It does not require additional manual labels or held-out data and provides a signal for identifying examples that are easy/difficult to generalize to. In our extensive evaluations, we show this benefits a number of models across different architectures and scales. Overall, for ID, we reach 32.92% in the selective prediction metric coverage at 1% risk of error (C@1%) which doubles the previous best coverage of 15.79% on this task. For mixed ID/OOD, using models' softmax confidences for abstention decisions performs very poorly, answering <5% of questions at 1% risk of error even when faced with only 10% OOD examples, but a learned selection function with LYP can increase that to 25.38% C@1%.

Vivek Kumar, Rishabh Maheshwary, Vikram Pudi

Existing Math Word Problem (MWP) solvers have achieved high accuracy on benchmark datasets. However, prior works have shown that such solvers do not generalize well and rely on superficial cues to achieve high performance. In this paper, we first conduct experiments to showcase that this behaviour is mainly associated with the limited size and diversity present in existing MWP datasets. Next, we propose several data augmentation techniques broadly categorized into Substitution and Paraphrasing based methods. By deploying these methods we increase the size of existing datasets by five folds. Extensive experiments on two benchmark datasets across three state-of-the-art MWP solvers show that proposed methods increase the generalization and robustness of existing solvers. On average, proposed methods significantly increase the state-of-the-art results by over five percentage points on benchmark datasets. Further, the solvers trained on the augmented dataset perform comparatively better on the challenge test set. We also show the effectiveness of proposed techniques through ablation studies and verify the quality of augmented samples through human evaluation.

Pulkit Pattnaik, Rishabh Maheshwary, Kelechi Ogueji et al.

Direct Preference Optimization (DPO) is an effective technique that leverages pairwise preference data (usually one chosen and rejected response pair per user prompt) to align LLMs to human preferences. In practice, multiple responses can exist for a given prompt with varying quality relative to each other. With availability of such quality ratings for multiple responses, we propose utilizing these responses to create multiple preference pairs for a given prompt. Our work focuses on systematically using the constructed multiple preference pair in DPO training via curriculum learning methodology. In particular, we order these multiple pairs of preference data from easy to hard (emulating curriculum training) according to various criteria. We show detailed comparisons of our proposed approach to the standard single-pair DPO setting. Our method, which we call Curry-DPO consistently shows increased performance gains on MTbench, Vicuna, WizardLM, and the UltraFeedback test set, highlighting its effectiveness. More specifically, Curry-DPO achieves a score of 7.43 on MT-bench with Zephy-7B model outperforming majority of existing LLMs with similar parameter size. Curry-DPO also achieves the highest adjusted win rates on Vicuna, WizardLM, and UltraFeedback test datasets (90.7%, 87.1%, and 87.9% respectively) in our experiments, with notable gains of upto 7.5% when compared to standard DPO technique. We release the preference pairs used in alignment at: https://huggingface.co/datasets/ServiceNow-AI/Curriculum_DPO_preferences

Raghu Vamshi Hemadri, Humaira Firdowse Mohammed, Rishabh Maheshwary et al.

Efficient agentic systems should incur expensive frontier-model costs only on decisions where a cheaper local model is likely to fail. Existing LLM cascades usually route whole queries before execution, but task difficulty shifts mid-trajectory - after flaky tool calls, truncated observations, or compounding local errors - making pre-execution routing brittle. We introduce \textbf{R2V-Agent}, a risk-calibrated SLM-LLM routing framework for interactive agents. R2V combines four components: a distilled small language model (SLM) policy, a stronger teacher LLM, a lightweight process verifier that scores candidate actions at each step, and a calibrated step-level router. The router is our central contribution: after the SLM is trained, it estimates residual failure risk at each step and escalates only when teacher intervention is warranted. To make the routing problem well-defined, we first train a stable local SLM using a standard offline pipeline: behavioral cloning (BC) on teacher trajectories, followed by verifier-guided Direct Preference Optimization (DPO) with consistency regularization. The router is then trained on this fixed policy's residual failures using Brier-calibrated probability estimation and a Conditional Value-at-Risk (CVaR)-constrained objective that penalizes worst-case failures across perturbation seeds. Across HumanEval+, TextWorld, and TerminalBench with four SLM backbones, R2V improves the reliability-cost frontier: it achieves $94.3\%$ HumanEval+ success with $0.60\%$ LLM escalation, recovers TextWorld from $64.6\%$ SLM-only success to $98.2\%$ at $41.7\%$ escalation, and reaches $93.3\%$ TerminalBench success at $33.9\%$ LLM calls, roughly half the heuristic-router cost.

Jishnu Sethumadhavan Nair, Patrice Bechard, Rishabh Maheshwary et al.

World models enable agents to anticipate the effects of their actions by internalizing environment dynamics. In enterprise systems, however, these dynamics are often defined by tenant-specific business logic that varies across deployments and evolves over time, making models trained on historical transitions brittle under deployment shift. We ask a question the world-models literature has not addressed: when the rules can be read at inference time, does an agent still need to learn them? We argue, and demonstrate empirically, that in settings where transition dynamics are configurable and readable, runtime discovery complements offline training by grounding predictions in the active system instance. We propose enterprise discovery agents, which recover relevant transition dynamics at runtime by reading the system's configuration rather than relying solely on internalized representations. We introduce CascadeBench, a reasoning-focused benchmark for enterprise cascade prediction that adopts the evaluation methodology of World of Workflows on diverse synthetic environments, and use it together with deployment-shift evaluation to show that offline-trained world models can perform well in-distribution but degrade as dynamics change, whereas discovery-based agents are more robust under shift by grounding their predictions in the current instance. Our findings suggest that, in configurable enterprise environments, agents should not rely solely on fixed internalized dynamics, but should incorporate mechanisms for discovering relevant transition logic at runtime.

Shruthan Radhakrishna, Aman Tiwari, Aanjaneya Shukla et al.

We present Apriel-1.5-15B-Thinker, a 15-billion parameter open-weights multimodal reasoning model that achieves frontier-level performance through training design rather than sheer scale. Starting from Pixtral-12B, we apply a progressive three-stage methodology: (1) depth upscaling to expand reasoning capacity without pretraining from scratch, (2) staged continual pre-training that first develops foundational text and vision understanding, then enhances visual reasoning through targeted synthetic data generation addressing spatial structure, compositional understanding, and fine-grained perception, and (3) high-quality text-only supervised fine-tuning on curated instruction-response pairs with explicit reasoning traces spanning mathematics, coding, science, and tool use. Notably, our model achieves competitive results without reinforcement learning or preference optimization, isolating the contribution of our data-centric continual pre-training approach. On the Artificial Analysis Intelligence Index, Apriel-1.5-15B-Thinker attains a score of 52, matching DeepSeek-R1-0528 despite requiring significantly fewer computational resources. Across ten image benchmarks, its performance is on average within five points of Gemini-2.5-Flash and Claude Sonnet-3.7, a key achievement for a model operating within single-GPU deployment constraints. Our results demonstrate that thoughtful mid-training 2 design can close substantial capability gaps without massive scale, making frontier-level multimodal reasoning accessible to organizations with limited infrastructure. We release the model checkpoint, all training recipes, and evaluation protocols under the MIT license to to advance open-source research.

Razvan-Gabriel Dumitru, Vikas Yadav, Rishabh Maheshwary et al.

We present a simple meta quantization approach that quantizes different layers of a large language model (LLM) at different bit levels, and is independent of the underlying quantization technique. Specifically, we quantize the most important layers to higher bit precision and less important layers to lower bits. We propose two effective strategies to measure the importance of layers within LLMs: the first measures the importance of a layer based on how different its output embeddings are from the input embeddings (higher is better); the second estimates the importance of a layer using the number of layer weights that are much larger than average (smaller is better). We show that quantizing different layers at varying bits according to our importance scores results in minimal performance drop with a far more compressed model size. Finally, we present several practical key takeaways from our variable layer-wise quantization experiments: (a) LLM performance under variable quantization remains close to the original model until 25-50% of layers are moved in lower quantization using our proposed ordering but only until 5-10% if moved using no specific ordering; (b) Adding layer importance to inherently dynamic quantization techniques can further improve their performance, showing that our approach is complementary to other dynamic quantization methods; (c) Quantizing LLMs to lower bits performs substantially better than pruning unless extreme quantization (2-bit) is used; and (d) Layer-wise quantization to lower bits works better in the case of larger LLMs with more layers compared to smaller LLMs with fewer layers. Our code is publicly available at https://github.com/RazvanDu/LayerwiseQuant/.

Rishabh Maheshwary, Vikas Yadav, Hoang Nguyen et al.

Instruction finetuning (IFT) is critical for aligning Large Language Models (LLMs) to follow instructions. While many effective IFT datasets have been introduced recently, they predominantly focus on high-resource languages like English. To better align LLMs across a broad spectrum of languages and tasks, we propose a fully synthetic, novel taxonomy (Evol) guided Multilingual, Multi-turn instruction finetuning dataset, called M2Lingual. It is constructed by first selecting a diverse set of seed examples and then utilizing the proposed Evol taxonomy to convert these seeds into complex and challenging multi-turn instructions. We demonstrate the effectiveness of M2Lingual by training LLMs of varying sizes and showcasing the enhanced performance across a diverse set of languages. We contribute the 2 step Evol taxonomy with the guided generation code: https://github.com/ServiceNow/M2Lingual, as well as the first fully synthetic, general and task-oriented, multi-turn, multilingual dataset built with Evol - M2Lingual: https://huggingface.co/datasets/ServiceNow-AI/ M2Lingual - containing 182K total IFT pairs, covering 70 languages and 17+ NLP tasks.

Srishti Gureja, Lester James V. Miranda, Shayekh Bin Islam et al. · cambridge

Reward models (RMs) have driven the state-of-the-art performance of LLMs today by enabling the integration of human feedback into the language modeling process. However, RMs are primarily trained and evaluated in English, and their capabilities in multilingual settings remain largely understudied. In this work, we conduct a systematic evaluation of several reward models in multilingual settings. We first construct the first-of-its-kind multilingual RM evaluation benchmark, M-RewardBench, consisting of 2.87k preference instances for 23 typologically diverse languages, that tests the chat, safety, reasoning, and translation capabilities of RMs. We then rigorously evaluate a wide range of reward models on M-RewardBench, offering fresh insights into their performance across diverse languages. We identify a significant gap in RMs' performances between English and non-English languages and show that RM preferences can change substantially from one language to another. We also present several findings on how different multilingual aspects impact RM performance. Specifically, we show that the performance of RMs is improved with improved translation quality. Similarly, we demonstrate that the models exhibit better performance for high-resource languages. We release M-RewardBench dataset and the codebase in this study to facilitate a better understanding of RM evaluation in multilingual settings.

Angelika Romanou, Negar Foroutan, Anna Sotnikova et al.

The performance differential of large language models (LLM) between languages hinders their effective deployment in many regions, inhibiting the potential economic and societal value of generative AI tools in many communities. However, the development of functional LLMs in many languages (\ie, multilingual LLMs) is bottlenecked by the lack of high-quality evaluation resources in languages other than English. Moreover, current practices in multilingual benchmark construction often translate English resources, ignoring the regional and cultural knowledge of the environments in which multilingual systems would be used. In this work, we construct an evaluation suite of 197,243 QA pairs from local exam sources to measure the capabilities of multilingual LLMs in a variety of regional contexts. Our novel resource, INCLUDE, is a comprehensive knowledge- and reasoning-centric benchmark across 44 written languages that evaluates multilingual LLMs for performance in the actual language environments where they would be deployed.

Israfel Salazar, Manuel Fernández Burda, Shayekh Bin Islam et al. · mila

The evaluation of vision-language models (VLMs) has mainly relied on English-language benchmarks, leaving significant gaps in both multilingual and multicultural coverage. While multilingual benchmarks have expanded, both in size and languages, many rely on translations of English datasets, failing to capture cultural nuances. In this work, we propose Kaleidoscope, as the most comprehensive exam benchmark to date for the multilingual evaluation of vision-language models. Kaleidoscope is a large-scale, in-language multimodal benchmark designed to evaluate VLMs across diverse languages and visual inputs. Kaleidoscope covers 18 languages and 14 different subjects, amounting to a total of 20,911 multiple-choice questions. Built through an open science collaboration with a diverse group of researchers worldwide, Kaleidoscope ensures linguistic and cultural authenticity. We evaluate top-performing multilingual vision-language models and find that they perform poorly on low-resource languages and in complex multimodal scenarios. Our results highlight the need for progress on culturally inclusive multimodal evaluation frameworks.

Hoang H Nguyen, Khyati Mahajan, Vikas Yadav et al.

Although multilingual LLMs have achieved remarkable performance across benchmarks, we find they continue to underperform on non-Latin script languages across contemporary LLM families. This discrepancy arises from the fact that LLMs are pretrained with orthographic scripts, which are dominated by Latin characters that obscure their shared phonology with non-Latin scripts. We propose leveraging phonemic transcriptions as complementary signals to induce script-invariant representations. Our study demonstrates that integrating phonemic signals improves performance across both non-Latin and Latin script languages, with a particularly significant impact on closing the performance gap between the two. Through detailed experiments, we show that phonemic and orthographic scripts retrieve distinct examples for in-context learning (ICL). This motivates our proposed Mixed-ICL retrieval strategy, where further aggregation from both leads to our significant performance improvements for both Latin script languages (up to 12.6%) and non-Latin script languages (up to 15.1%) compared to randomized ICL retrieval.

Shruthan Radhakrishna, Soham Parikh, Gopal Sarda et al.

While large language models (LLMs) have achieved remarkable reasoning capabilities across domains like code, math and other enterprise tasks, their significant memory and computational costs often preclude their use in practical enterprise settings. To this end, we introduce Apriel-Nemotron-15B-Thinker, a 15-billion parameter model in the ServiceNow Apriel SLM series that achieves performance against medium sized state-of-the-art models such as o1-mini, QWQ32B, and EXAONE-Deep-32B while maintaining only half the memory footprint of those alternatives. Apriel-Nemotron-15B-Thinker model is trained in a four stage training pipeline including 1) Base Model upscaling, 2) Continual Pre-training 3) Supervised Fine-tuning (SFT) and 4) Reinforcement Learning using GRPO. Comprehensive evaluations across a diverse suite of benchmarks consistently demonstrate that our Apriel-Nemotron-15B-Thinker model matches or exceeds the performance of its 32-billion parameter counterparts, despite being less than half their size.

Rishabh Maheshwary, Masoud Hashemi, Khyati Mahajan et al.

Iterative RAG for multi-hop question answering faces challenges with lengthy contexts and the buildup of irrelevant information. This hinders a model's capacity to process and reason over retrieved content and limits performance. While recent methods focus on compressing retrieved information, they are either restricted to single-round RAG, require finetuning or lack scalability in iterative RAG. To address these challenges, we propose Notes Writing, a method that generates concise and relevant notes from retrieved documents at each step, thereby reducing noise and retaining only essential information. This indirectly increases the effective context length of Large Language Models (LLMs), enabling them to reason and plan more effectively while processing larger volumes of input text. Notes Writing is framework agnostic and can be integrated with different iterative RAG methods. We demonstrate its effectiveness with three iterative RAG methods, across two models and four evaluation datasets. Notes writing yields an average improvement of 15.6 percentage points overall, with minimal increase in output tokens.

Vivek Kumar, Rishabh Maheshwary, Vikram Pudi

Standard accuracy metrics have shown that Math Word Problem (MWP) solvers have achieved high performance on benchmark datasets. However, the extent to which existing MWP solvers truly understand language and its relation with numbers is still unclear. In this paper, we generate adversarial attacks to evaluate the robustness of state-of-the-art MWP solvers. We propose two methods Question Reordering and Sentence Paraphrasing to generate adversarial attacks. We conduct experiments across three neural MWP solvers over two benchmark datasets. On average, our attack method is able to reduce the accuracy of MWP solvers by over 40 percentage points on these datasets. Our results demonstrate that existing MWP solvers are sensitive to linguistic variations in the problem text. We verify the validity and quality of generated adversarial examples through human evaluation.

Rishabh Maheshwary, Saket Maheshwary, Vikram Pudi

Existing black box search methods have achieved high success rate in generating adversarial attacks against NLP models. However, such search methods are inefficient as they do not consider the amount of queries required to generate adversarial attacks. Also, prior attacks do not maintain a consistent search space while comparing different search methods. In this paper, we propose a query efficient attack strategy to generate plausible adversarial examples on text classification and entailment tasks. Our attack jointly leverages attention mechanism and locality sensitive hashing (LSH) to reduce the query count. We demonstrate the efficacy of our approach by comparing our attack with four baselines across three different search spaces. Further, we benchmark our results across the same search space used in prior attacks. In comparison to attacks proposed, on an average, we are able to reduce the query count by 75% across all datasets and target models. We also demonstrate that our attack achieves a higher success rate when compared to prior attacks in a limited query setting.

Rishabh Maheshwary, Saket Maheshwary, Vikram Pudi

We study an important and challenging task of attacking natural language processing models in a hard label black box setting. We propose a decision-based attack strategy that crafts high quality adversarial examples on text classification and entailment tasks. Our proposed attack strategy leverages population-based optimization algorithm to craft plausible and semantically similar adversarial examples by observing only the top label predicted by the target model. At each iteration, the optimization procedure allow word replacements that maximizes the overall semantic similarity between the original and the adversarial text. Further, our approach does not rely on using substitute models or any kind of training data. We demonstrate the efficacy of our proposed approach through extensive experimentation and ablation studies on five state-of-the-art target models across seven benchmark datasets. In comparison to attacks proposed in prior literature, we are able to achieve a higher success rate with lower word perturbation percentage that too in a highly restricted setting.

Rishabh Maheshwary, Saket Maheshwary, Vikram Pudi

We study an important task of attacking natural language processing models in a black box setting. We propose an attack strategy that crafts semantically similar adversarial examples on text classification and entailment tasks. Our proposed attack finds candidate words by considering the information of both the original word and its surrounding context. It jointly leverages masked language modelling and next sentence prediction for context understanding. In comparison to attacks proposed in prior literature, we are able to generate high quality adversarial examples that do significantly better both in terms of success rate and word perturbation percentage.