Sparsh Mittal

Author directory

2026

While Large Language Models (LLMs) perform well on arithmetic word problems, they remain brittle to adversarial perturbations like irrelevant distractors or contradictory facts. To address this, we introduce ReThink, a training-free neuro-symbolic framework that formulates adversarial arithmetic reasoning as a Weighted MaxSAT/MaxSMT problem. ReThink uses a four-stage pipeline: it explicitly decomposes problems to strip noise, encodes the mathematical core as confidence-weighted soft constraints in the Z3 theorem prover, and cross-verifies the formal solution against a parallel Chain-of-Thought (CoT) path. This formulation can sacrifice lower-weight conflicting statements while satisfying higher-weight constraints. Evaluated on a balanced 2,400-example subset of GSM-Plus, ReThink with GPT-4.1 achieves 81.62% adversarial accuracy, outperforming CoT, PAL, and Self-Verification. It narrows the robustness gap to 14.69 percentage points while achieving 96.31% clean accuracy.

2024

Zero-shot temporal action localization (TAL) aims to temporally localize actions in videos without prior training examples. To address the challenges of TAL, we offer GRIZAL, a model that uses multimodal embeddings and dynamic motion cues to localize actions effectively. GRIZAL achieves sample diversity by using large-scale generative models such as GPT-4 for generating textual augmentations and DALL-E for generating image augmentations. Our model integrates vision-language embeddings with optical flow insights, optimized through a blend of supervised and self-supervised loss functions. On ActivityNet, Thumos14 and Charades-STA datasets, GRIZAL greatly outperforms state-of-the-art zero-shot TAL models, demonstrating its robustness and adaptability across a wide range of video content. We will make all the models and code publicly available by open-sourcing them.