Aditi Gupta
I am a Research Intern at Adobe Research, Bengaluru, and an MS by Research student in Computer Science at IIIT Hyderabad, where I worked in CSTAR Lab and was advised by Dr. Pawan Kumar.
My research focuses on making large language models more efficient and more capable as agents; spanning LLM inference optimization, model compression, and agentic systems. At Adobe, I am working on context management in agentic LLM harnesses, using agent reasoning traces to improve retrieval and reduce wasted context.
I came to research through curiosity: in my undergrad at IISER Bhopal I kept following whatever problem looked interesting; computer vision, NLP, systems until it stopped being a habit and became a direction. I am currently looking for full-time roles in AI/ML research and engineering.
Research Interests
My primary interest is the intersection of LLM efficiency and agentic capability, two goals I think are converging rather than competing. A model that wastes its context window on noise cannot reason well over long horizons. My recent work spans training-free decoding strategies (hybrid autoregressive-diffusion generation), rank-structured model compression, and context management in agentic harnesses. I am drawn to ideas that are clever rather than just expensive — getting more from what is already there.
News
| Jun 2026 | Paper accepted at ACL 2026 (GEM Workshop) — Speculative Refinement: A Hybrid Autoregressive Diffusion Decoding Strategy. |
| May 2025 | Started Research Internship at Adobe Research, Bengaluru — context rot in agentic LLM harnesses. |
| Jan 2025 | Paper accepted at CODS 2025 (ACM IKDD) — Hierarchical Sparse Plus Low-Rank Compression of LLM. |
| Jan 2025 | Placed 2nd at IndoML Datathon 2025. |
| Jul 2024 | Started MS by Research at IIIT Hyderabad, joining CSTAR under Dr. Pawan Kumar. |
| 2024 | Selected as Amazon ML Summer School 2024 mentee. |
| 2023 | Top-15 finalist, Smart India Hackathon 2023. |
Publications
- Speculative Refinement: A Hybrid Autoregressive Diffusion Decoding Strategy and Its Behavior Across Benchmarks ACL 2026 · GEM Workshop, Association for Computational Linguistics [PDF]·
- Hierarchical Sparse Plus Low-Rank Compression of LLM CODS 2025 · 13th ACM IKDD International Conference on Data Science [PDF]
Projects
- Agentic RAG System for Legal QA LLM-Compiler-based multi-agent pipeline with DAG-based execution and tool orchestration. Voyage-law embeddings and Cohere reranking achieved 86% on CUAD and 84% on AILA legal benchmarks. LLMs · RAG · Agentic Systems · Reranking
- Decoupled Online Speculative Decoding Implemented decoupled speculative decoding (ICLR 2024) with MPI-based cross-process communication. Achieved >13s latency reduction via inference-time distillation from Qwen2.5-7B to 0.5B on GSM8K. LLM Inference · Distributed Systems · MPI · PyTorch
- SiamBERT — Word Sense Disambiguation Siamese encoder with contrastive and triplet loss for WSD. Outperformed GlossBERT by 22% accuracy with 7× faster inference. Trained on 200k+ SemCor sentence pairs. BERT · PyTorch · Contrastive Learning · NLP
- Smart Analog Water Meter (SPCRC) Retrofitted analog water meters with real-time monitoring using Raspberry Pi. CNN digit recognition pipeline achieved 97.7% accuracy; interactive dashboard with regional usage maps via Leaflet.js. TensorFlow · OpenCV · Raspberry Pi · Leaflet.js