Balasubramaniam
Srinivasan

Casts workflow generation as Bayesian inference over a posterior distribution on workflows: a sampling framework that builds workflows step-by-step with parallel look-ahead rollouts and a sequential in-loop refiner. Improves accuracy by up to 9 points over state-of-the-art and 65 points over zero-shot prompting.

MEDAL uses Monte Carlo Tree Search at inference time to explore promising unmasking trajectories in diffusion language models, delivering up to 22% improvement over existing inference strategies without additional training.

Routes queries to the optimal LLM within a model family using lightweight quality estimators trained on 1.5M annotated prompts across 11 candidate LLMs. Powers Bedrock Intelligent Prompt Routing in production, delivering up to 43.9% cost reduction while matching the quality of the strongest model in the family.

Unified framework for automatic prompt optimization spanning iteration depth, filtering strategies, prompt generation, and feedback mechanisms. Defines formal APO problem formulations and identifies axes of meaningful advancement.

Multi-task fine-tuning framework for LLM-based data discovery (joinable and unionable table search) that models how auxiliary supervisions such as entity resolution and column type classification complement or conflict with each other. Plug-and-play with open-domain auxiliary tasks; outperforms the best baseline by up to 7% in F1.

Selects representative in-context examples by building semantic graphs over unlabeled candidates and solving a maximum-coverage problem, augmented with LLM uncertainty. Beats baselines by 2 to 4.6% across ten datasets and seven LLMs with 2× better annotation-budget efficiency.

Score-based diffusion in a VAE-learned latent space for mixed-type tabular data. Cuts column distribution and pair-wise correlation estimation error by 86% and 67% vs. competitive baselines.

Parameter-efficient framework using knowledge graphs to learn transformations between independently trained biomedical foundation models without fine-tuning. Outperforms KG embedding baselines by ~76% on cross-modal retrieval and enables multimodal biomedical question answering.

Open-domain table reasoning framework powered by LLMs: a table retriever fetches relevant tables, the LLM generates SQL programs to parse them efficiently, and grounded inference over the intermediate SQL execution results produces the answer. Outperforms baselines in both open- and closed-domain settings, with up to 21.5% higher accuracy.

Captures tabular invariances with hypergraphs: cells become nodes, row/column/table co-occurrences become hyperedges. Maximally invariant under natural table permutations; consistently outperforms baselines on four downstream tasks with minimal pretraining.

Frames abbreviated column-name expansion as conditional natural-language generation on 384K synthetic and 9.2K human-annotated pairs. Conditioning LLMs on table content and headers resolves polysemy and matches human performance on real-world table schemas.

Addresses GNN expressiveness limits by representing each graph as a set of subgraphs processed with equivariant architectures. Proves strictly higher expressiveness via WL variants, introduces tractable subgraph-sampling, and improves both theory and empirical performance.

Hypergraph neural network that uses the incidence structure to represent variable-sized hyperedges, invariant to the ordering of vertices within a hyperedge and preserving local isomorphism in the hypergraph's line graph. Targets hyperedge classification and expansion of partially observed hyperedges; consistently improves on prior methods across real-world hypergraph datasets.

Unifies positional and structural node embeddings under a common theoretical framework, proving their equivalence under mild conditions and providing a principled basis for combining both approaches.

Introduces Relational Pooling via random injective mappings, achieving maximum expressiveness under permutation invariance. Generalizes the Weisfeiler-Lehman test and many popular GNN architectures within a single framework.

Expresses permutation-invariant functions of sequences as the average of a permutation-sensitive function over all reorderings, unlocking the full literature of RNNs and CNNs for invariant tasks. Introduces tractable approximations via canonical orderings, k-ary interactions, and stochastic optimization.