FrugalRAG: Less is More in RL Finetuning for Multi-Hop Question Answering

Abhinav Java, Srivathsan Koundinyan, Nagarajan Natarajan, Amit Sharma

NeurIPS Efficient Reasoning Workshop 2025

[arXiv], [Code]

• We present a post-training framework for SLM-based multi-hop question answering RAG Agents.
• We find that a two-stage curriculum that teaches policies to first explore exhaustively to maximize document recall and then learns when to stop is both data and compute efficient.

Characterizing Deep Research: A Benchmark and Formal Definition

Abhinav Java*, Ashmit Khandelwal*, Sukruta Midigeshi*, Aaron Halfaker, Amit Deshpande, Navin Goyal, Ankur Gupta, Nagarajan Natarajan, Amit Sharma

NeurIPS Scaling Environments for Agents Workshop 2025

[arXiv], [Dataset]

• We present a formal definition of deep research and an objective benchmark comprising 8 different domains (Science, Everyday Use, Prior Art, etc).
• We find that even frontier models like OpenAI DR, Gemini DR, and Perplexity achieve a modest F1 score of <=55% on our benchmark.

Understanding Task Transfer in VLMs

Bhuvan Sachdeva*, Karan Uppal*, Abhinav Java*, Vineeth N Balasubramanian

NeurIPS UniReps Workshop 2025

[arXiv will be released soon!], [Code will be released soon!]

• We present a comprehensive analysis of zero-shot task-transfer in VLMs across 10+ datasets and three models.
• We introduce a novel metric to quantify cross-task influence in VLM finetuning.
• Our analysis reveals actionable insights – beneficial task cliques and transfer trends for effective finetuning strategies.

Towards Operationalizing Right to Data Protection

Abhinav Java*, Simra Shahid*, Chirag Agarwal

NAACL (Main) 2025

[arXiv], [Code]

• We present a framework for generating unlearnable text datasets. These datasets prevent data misuse by failing on test datasets despite converging normally during training.
• We show that imperceptible perturbations can expose vulnerabilities of strong models like GPT-4, dropping their performance on sentiment analysis tasks below the zero-shot baseline.

ReEdit: Multimodal Exemplar-Based Image Editing

Ashutosh Srivastava*, Tarun Menta*, Abhinav Java*, Avadhoot Jadhav, Silky Singh, Surgan Jandial, Balaji Krishnamurthy

WACV 2025

[arXiv], [Code]

• We present an efficient inference time approach for image editing using edit exemplars.
• We find that textual descriptions are not enough to perform faithful edits, incorporating edits in both image and text modality.

All Should Be Equal in the Eyes of LMs: Counterfactually Aware Fair Text Generation

Pragyan Banerjee*, Abhinav Java*, Surgan Jandial*, Simra Shahid*, Shaz Furniturewala, Balaji Krishnamurthy, Sumit Bhatia

AAAI 2024

[arXiv], [Code]

• We present a framework to steer models towards unbiased generations using counterfactual contexts.
• We find that re-weighing logits using probabilities distributions can encourage equitable responses across several benchmarks for varying model sizes.

One Shot Doc Snippet Detection: Towards Search in Document Beyond Text

Abhinav Java, Shripad Deshmukh, Milan Aggarwal, Surgan Jandial, Mausoom Sarkar, Balaji Krishnamurthy

WACV 2023

[arXiv]

• We highlight a new problem in document object detection of searching for arbitrary snippets using a novel example snippet.
• To solve this, we generate a large scale training dataset using rudimentary object structures and propose a new architecture that fuses various modalities.

Learning to Censor by Noisy Sampling

Ayush Chopra, Abhinav Java, Abhishek Singh, Vivek Sharma, Ramesh Raskar

ECCV 2022

[arXiv], [Code]

• We present an end-to-end differentiable sampler to protect 3D point clouds from perception attacks.
• Our key insight is to leverage the localized saliency of perception tasks on point clouds to provide good privacy-utility trade-offs .

Rethinking Neural Networks With Benford's Law

Surya Kant Sahu, Abhinav Java, Arshad Shaikh, Yannic Kilcher

NeurIPS ML4 Physical Sciences 2022

[arXiv], [Code]

• We study: Is the distribution of the Neural Network parameters related to the network’s generalization capability?
• We define a new metric called Model Enthalpy which measures the closeness of NN parameters to the ideal Benford's distribution and provide experimental evidence that this metric can be used for Early Stopping for training shallow networks.