Venkatesh Saligrama

@Saligrama_Lab

AI Prof. and Amazon Scholar focused on learning under constraints of all forms - feature costs, communication, computation, limited/biased supervision.

venkatesh-saligrama.github.io

Joined September 2020

Following

120

Followers

Replies Reposts Media

Venkatesh Saligrama
@Saligrama_Lab
Aug 16
Audio access is not evidence of audio use. Holding words fixed, we vary affect, prosody, or timing. Across 11 models, contrastive success can overstate native reliability; similar scores can hide different failures. With Kevin Miller & Arjun Chandra arxiv.org/abs/2608.06718
Venkatesh Saligrama
@Saligrama_Lab
Jul 6
What algorithm does a transformer run in-context? Our #ICML2026 spotlight shows that layerwise symmetry makes the computation readable: not hidden gradient descent, but a geometry-driven mean-shift classifier. Paper: arxiv.org/abs/2604.11613
1
Venkatesh Saligrama
@Saligrama_Lab
Jul 4
DeepFact: Co-Evolving Benchmarks and Agents for Deep Research Factuality will be presented as an oral paper at ACL 2026 tomorrow. @YukunHuang9 will present the work: Sunday, July 5, 11:50 AM–12:00 PM Harbor B-C Oral Session A: Resources and Evaluation 1 We argue that
Ground truth is a process, not a dataset
From amazon.science
Venkatesh Saligrama
@Saligrama_Lab
Apr 14
We thought the harder problem was building a better AI fact-checker to verify complex research reports. We were wrong. The real bottleneck is the benchmark itself. Ground truth has to become a process, not a static dataset.
Yukun Huang and Markus Dreyer
1
Venkatesh Saligrama
@Saligrama_Lab
Feb 11
“Do audio judges actually listen to audio?” 🎧 We find the answer is… no. Meet TRACE: a new judge that pairs lightweight audio tools with an LLM to make genuine audio-aware judgments — at ~3× lower cost ⚡ Excited to share our work at #EACL2026 Findings (Rabat🇲🇦, March 25th)!
1