Sample, Align, Synthesize: Graph-Based Response Synthesis with ConGrsPaper

COLM 2026 · Prof. Swabha SwayamdiptaS. Ghosh, S.S. Warraich, D. Tarsadiya, G. Yauney, S. Swayamdipta2025 – 2026

Introduced Consensus Graphs (ConGrs), a flexible DAG-based data structure that captures shared information and semantic variation across multiple sampled LM responses to the same prompt, enabling task-dependent decoding into a single, more reliable response.

• Designed Consensus Graphs (ConGrs), built using a lightweight lexical sequence alignment algorithm from bioinformatics, supplemented by targeted use of a secondary LM judge

How Reliable is Language Model Micro-Benchmarking?Paper

ICLR 2026 (Oral) · Prof. Swabha SwayamdiptaG. Yauney, S.S. Warraich, S. Swayamdipta2025 – 2026

Introduced a meta-evaluation measure for LLM micro-benchmarking that quantifies how reliably a micro-benchmark can rank two models as a function of their performance gap on the full benchmark, revealing when small evaluation subsets fail to preserve model rankings.

• Proposed a meta-evaluation measure for micro-benchmarking that analyzes ranking reliability as a function of the performance difference between model pairs on the full benchmark

Proposed RBANet, a robust and lightweight CNN for real-time traffic scene semantic segmentation, introducing a novel balanced attention module and a novel residual module to jointly optimize precision and computational efficiency for intelligent transportation systems.

• Designed RBANet, a lightweight CNN combining a novel balanced attention module and a novel residual module for real-time traffic scene semantic segmentation