Sandeep Mishra

Software Engineer, Google
Integrated M.Sc. in Mathematics and Computing, IIT Kharagpur ’25

I study how language models spend computation, especially whether they can reason in continuous latent space instead of through tokens. My earlier work on adaptive inference appeared at EACL 2026, covering dynamic routing between vision-language models and entropy-based early stopping. I am applying to PhD programs for Fall 2027.

Email/GitHub/Scholar/CV

sandeepmishraismyname [at] gmail [dot] com

New on the blog · Sep 2026

The Geometry of a Latent Thought

A controlled study of the geometry of latent thoughts in models like Coconut.

Read the post →

Research Interests

Latent reasoning Can language models reason in continuous latent space instead of through tokens, and what do their internal representations reveal about that reasoning?
Adaptive and efficient inference How can a model spend compute only where it is needed, through routing, early exit, or latent computation instead of longer chains of thought?
Multimodal and vision-language models How can language models ground their reasoning in images, especially in interactive settings like visually-grounded dialog?

News

Sep 2026 New blog post: The Geometry of a Latent Thought.

Mar 2026 Two papers at EACL 2026: Router-Suggest in the Industry Track and Chat-Ghosting in the Main Conference.

2025 Graduated from IIT Kharagpur.

Publications

Mar 2026

Router-Suggest: Dynamic Routing for Multimodal Auto-Completion in Visually-Grounded Dialogs

Sandeep Mishra, D. Budagam, A. Mandal, B. Santra, P. Goyal, M. Gupta

EACL 2026 · Industry Track

Defined the Multimodal Auto-Completion (MAC) vision-language task from scratch — task definition, dataset curation, and evaluation protocol. Benchmarked state-of-the-art VLMs (Qwen2-VL, MiniCPM-V, PaliGemma) and proposed Router-Suggest, a dynamic routing framework achieving 2.3×–10× speedup over the best VLM baseline while maintaining near-parity accuracy.

PDF/ACL Anthology

Mar 2026

Chat-Ghosting: Methods for Auto-Completion in Dialog Systems

A. Mandal, Sandeep Mishra, B. Santra, T. Abhishek, P. Goyal, M. Gupta

EACL 2026 · Main Conference

Introduced the first comprehensive benchmark for chat-ghosting (predictive text completion in conversations) across 4 public dialog datasets. Proposed an entropy-based dynamic early-stopping mechanism that significantly improves Partial-Precision and TES, making auto-completion practical for real dialog applications.

PDF/ACL Anthology

Projects

Apr 2023

Zero-Shot Cross-Lingual QA via Prefix Tuning on mT5

Research project · advised by Prof. Pawan Goyal, CSE, IIT Kharagpur

Investigated parameter-efficient fine-tuning using prefix tuning on mT5 for zero-shot cross-lingual question answering. Demonstrated competitive transfer performance with under 1% trainable parameters compared to full fine-tuning.

Report PDF

Apr 2023

Efficient Product Embeddings & Semantic Search Engine

Research project · advised by Prof. B. Adhikari, Mathematics, IIT Kharagpur

Showed BERT contextual embeddings outperform Word2Vec and GloVe by 10× on product retrieval across 14,000 Amazon products. Built a semantic search engine with 3-D TensorFlow Embedding Projector visualisation for interpretability.

Report PDF