Publications
Our long-horizon research focuses on foundational models and a harder question than most: whether AI can represent the structure of the world, not merely its patterns.
Enterprise scale reveals difficult problems. Research lets us ask better questions of them.
Catalog
Peer-reviewed workshop papers — linked to venue pages and preprints where available.
NeurIPS · 2025
The paper introduces activation transplantation, which intervenes on a model's internal hidden states to impose the statistical signature of one event onto another — letting researchers steer forecasts toward or away from crash-like behavior.
Debdeep Sanyal · Aaryan Nagpal · Dhruv Kumar · Murari Mandal · Saurabh Deshpande
Read →ICML · 2026
A lightweight framework that fixes a frozen time-series forecasting model's blind spots at inference time — without retraining — by detecting and softening anomalous stretches of input data.
Manya Pandey · Dhruv Kumar · Murari Mandal · Saurabh Deshpande
Read →ICML · 2026
Shows that picking a single “best” synthetic data generator for pretraining is the wrong approach — an equal-weight mix of many generators consistently matches or beats the best individual one.
Aaryan Nagpal · Debdeep Sanyal · Murari Mandal · Dhruv Kumar · Saurabh Deshpande
Read →ICML · 2026
A generative method that uses a small number of real reference sequences as structural templates to synthesize realistic multivariate time series — useful where real data is scarce.
Moulik Gupta · Aaditya Jain · Dhruv Kumar · Murari Mandal · Saurabh Deshpande
Read →ICML · 2026
Tests whether forecasting models that combine text with numbers actually respond to what the text says, finding that swapping text for unrelated text barely changes predictions.
Karthik Sridhar · Atharva Gupta · Nishant Pradhan · Murari Mandal · Dhruv Kumar · Saurabh Deshpande
Read →