Publications.

Dropping the Anchor: Statistical Context Summarization for Distributed Systems via Pulsar Attention

Aryan Sood, Shantanu Acharya, Gaurav Kumar Nayak

Findings of EMNLP 2026

Developed Pulsar Attention for distributed long-context inference, replacing static anchor duplication with content-aware statistical summaries and reporting up to 4.7% accuracy gain over dense attention, 6.1% over Star Attention, and 3.3x lower Phase-1 FLOPs. Also accepted at the AdaptFM Workshop on Resource-Adaptive Foundation Model Inference, ICML 2026.

@misc{sood2026droppinganchorstatisticalcontext,
      title={Dropping the Anchor: Statistical Context Summarization for Distributed Systems via Pulsar Attention},
      author={Aryan Sood and Shantanu Acharya and Gaurav Kumar Nayak},
      year={2026},
      eprint={2607.20457},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2607.20457},
}

Are You Sure You're Sure? Two Confounds in a Sycophancy Benchmark

Atharv Gupta*, Akshat Jindal*, Lavanya Nigam†, Aryan Sood†

* equal contribution (first authors) · † equal contribution (second authors)

First Workshop on Reliable Evaluation for Language Models (JUDGe), NeurIPS 2026 · Oral

Audits SycEval and shows two template confounds: naming a target answer raises the follow rate by 14.1–49.5 pp and, once controlled, reverses the reported timing effect on three of five model conditions; relocating the output-format instruction flips its effect across models. Closes with three checks benchmark authors can run on their own templates.

@inproceedings{
gupta2026are,
title={Are You Sure You{\textquoteright}re Sure? Two Confounds in a Sycophancy Benchmark},
author={Atharv Gupta and Akshat Jindal and Lavanya Nigam and Aryan Sood},
booktitle={First Workshop on Reliable Evaluation for Language Models (JUDGe)},
year={2026},
url={https://openreview.net/forum?id=tySSyo3JbS}
}