Anirudh Lakhotia

At Couchbase, I build SDKs and connectors between applications and production data systems. That work has made me think about representations as more than model internals. They also have to be cheap to store and fast to serve at query time.

My recent work studies this trade-off in multi-vector retrieval. In Most MaxSim Winners Flip, Retrieval Survives (ReNeuIR @ SIGIR 2026), we show that sign coding changes most token-level MaxSim winners while largely preserving retrieval effectiveness. I am also working on compositionality in learned representations and how to separate learned structure from dataset-level effects.

Earlier, I worked on efficient multitask learning and conditional computation. UnoLoRA (FITML @ NeurIPS 2024) shares one low-rank adapter across tasks. Our task-sampling study examines how sampling strategies change a transformer’s internal representations. I also built Baraat, a multilingual mixture-of-experts system for code-switched Indic inputs.

I studied Computer Science at PES University, where I was advised by Dr. Gowri Srinivasa. Before university, I interned at Cisco, developing an event platform with a personalized recommendation system. I later built geospatial data infrastructure for Nokia’s cellular network data.

selected publications

ReNeuIR @ SIGIR

Most MaxSim Winners Flip, Retrieval Survives: Low-Margin Substitution in Sign-Coded Late Interaction

Anirudh Lakhotia and Nischal Helagally Shantharaju

5th Workshop on Reaching Efficiency in Neural Information Retrieval (ReNeuIR), co-located with ACM SIGIR, 2026

Sign coding flips most token-level MaxSim winners, yet retrieval survives because the substitutes are low-margin.

Abs PDF Listen Code

Late-interaction retrieval appears unusually robust to extreme compression. We study a sub-bit sign-coded ColBERT representation that stores each 128-dimensional document token as the 64 signs of a projected vector (0.5 bits per original dimension, 8 B/token). On a 100k-passage MS MARCO diagnostic slice, sign coding changes the MaxSim winner for roughly 70% of query tokens, yet relevant-document recall remains near the fp32 ceiling and MRR@10 is recoverable by a short fp32 rerank. We find that sign coding substantially disrupts token-level structure while leaving retrieval effectiveness largely intact. We trace this robustness to low-margin substitution: although sign coding frequently changes the winning document token, the replacement usually has a similar fp32 score, so the resulting loss in the MaxSim sum remains small. At full 8.8M-passage MS MARCO scale, a trained projection provides no statistically reliable advantage in MRR@10 or Recall@1000 over a random orthogonal projection at matched storage. A projection-training-free recipe using a random 64-dimensional projection, 8 B/token sign-coded document storage, and fp32 reranking of the top 100 candidates recovers fp32 MRR@10.
FITML @ NeurIPS

UnoLoRA: Single Low-Rank Adaptation for Efficient Multitask Fine-tuning

Anirudh Lakhotia , Akash Kamalesh , Nischal Helagally Shantharaju , Prerana Sanjay Kulkarni , and Gowri Srinivasa

Workshop on Fine-Tuning in Machine Learning, co-located with NeurIPS, 2024

One shared low-rank adapter supports multitask fine-tuning without duplicating a separate adapter for every task.

Abs PDF Listen

Recent advances in Parameter-Efficient Fine-Tuning (PEFT) have shown Low-Rank Adaptation (LoRA) to be an effective implicit regularizer for large language models. Building on these findings, we propose UnoLoRA, a novel approach that leverages a single shared LoRA module for efficient multi-task learning. While existing methods typically use separate LoRA adaptations for each task, our approach demonstrates that a single shared adapter can effectively capture both task-specific and task-agnostic knowledge. We further introduce UnoLoRA*, an enhanced variant that employs a shared hypernetwork to generate task-specific embeddings, improving convergence and task adaptation. Our method significantly reduces trainable parameters to just 0.05% per task while maintaining competitive performance on the GLUE benchmark. Our analysis reveals that the A and B matrices in our shared LoRA adapter naturally develop complementary roles: A matrices capture generalizable features across tasks, while B matrices specialize in task-specific representations. Our results show that sharing a single LoRA adapter can achieve efficient multi-task learning while significantly reducing memory requirements, making it particularly valuable for resource-constrained applications.

news

Jun 24, 2026	Our work on extreme compression for late-interaction retrieval was accepted to ReNeuIR 2026, co-located with SIGIR. The paper, code, and audio overview are available in publications.
Jul 01, 2025	Started full-time at Couchbase as a Software Engineer on the SDKs and Connectors team.
Jan 06, 2025	Joined Couchbase as a Software Engineering Intern on the SDKs and Connectors team, building developer libraries and integrations.

… all news →