Publications | Akshit Singh

Under review

First author

Architectural Sampling: Test-Time Scaling via Computational Diversity in Frozen Vision-Language Models

Akshit Singh, Shyam Marjit, Wei Lin, Leonid Karlinsky, M. Jehanzeb Mirza

Architectural sampling improves test-time scaling by generating candidates through different computation paths, rather than only changing decoding randomness. It reuses selected decoder layers with varying locations and repetition counts, requiring no training or extra parameters. Across five Qwen checkpoints and twelve benchmarks, it improves pass@9 by 6.58 points on average. It produces more diverse candidates, with benefits persisting even under greedy decoding. These diverse rollouts also improve label-free test-time reinforcement learning.

Project page Paper Code

CVPR 2026IEEE/CVF Conference on Computer Vision and Pattern Recognition

First author

TTRV: Test-Time Reinforcement Learning for Vision Language Models

Akshit Singh, Shyam Marjit, Wei Lin, Paul Gavrikov, Serena Yeung-Levy, Hilde Kuehne, Rogerio Feris, Sivan Doveh, James Glass, M. Jehanzeb Mirza

TTRV adapts a vision-language model during inference with reinforcement learning, using rewards computed from the model’s own outputs: how often an answer appears, and how diverse the answers are. No labels are needed, it works even from a single unlabeled test sample, and across 15+ multimodal benchmarks it lifts open-source models past GPT-4o by 2%.

Project page Paper Code

ACCV 202417th Asian Conference on Computer Vision

Oral, 5.6% acceptanceFirst author

CCNDF: Curvature Constrained Neural Distance Fields from 3D LiDAR Sequences

Akshit Singh, Karan Bhakuni, Rajendra Nagar

Neural distance fields struggle on curved, complex geometry. CCNDF uses second-order derivatives of the field to estimate the radius of curvature and constrains learning with it, giving better mapping and localization from LiDAR sequences.

Paper Code

Vehicular CommunicationsElsevier, impact factor 6.5

First author

Advancing Connected Vehicle Security with Comprehensive Datasets and Novel Anomaly Detection and Recovery Techniques

Akshit Singh, Heena Rathore

A reproducible dataset built by interpolating the U.S. DOT Tampa connected-vehicle data with simulated sensor faults (white and Lévy noise, bias, spikes), plus a framework that detects and recovers from anomalies in real time with Bayesian online changepoint detection and Bayesian forecasting.

Paper

AAAI 2024Student Abstract, 38th AAAI Conference on Artificial Intelligence

Sole author

Diverse Yet Biased: Towards Mitigating Biases in Generative AI

Akshit Singh

A study of cultural bias in text-to-image generation, supported by two new datasets with more than 250,000 image-caption pairs that capture Indian context and heritage.

Paper

Under review

Equal contribution

Deep Dense Clothed Genus-0 Human 3D Shape Matching using Spherical Harmonics

Deepanshu Solanki*, Akshit Singh*, Rajendra Nagar

Clothed human meshes are mapped onto spheres with Möbius transformations, turning a non-rigid problem into a rigid one. Graph neural networks and spherical harmonics then produce features that a functional-map solver uses for dense correspondence.

Paper