Research Highlights

How AI systems learn to reason, and how that process can be measured and shaped.

01 Recursive Self-Improvement · MILO

When Search Stalls, the Orchestrator Steps In

An agent is a model plus a harness: the prompts, tools, memory and control flow around it. MILO searches for better harnesses with islands that evolve in parallel, and an orchestrator agent that watches the whole search, diagnoses stalls and intervenes.

Interactive demo from the MILO project page.

Paper Project page

02 Test-Time Scaling · Hermes · co-first author

The Model Decides How to Use Each Fresh Context

Spreading test-time compute over many context windows means deciding what each fresh context should do and what it hands back. Hermes leaves those decisions to the model, and Hermes-Learn (SFT, then RL) teaches a 4B model to make them.

Interactive overview from the Hermes project page.

Paper Project page

03 On-Policy Distillation · SCOUT

The Teacher Learns to Continue the Student’s Prefixes

In on-policy distillation the student writes the response and a stronger teacher scores every token. But the teacher was trained to continue its own reasoning, and the deeper into a student’s response it reads, the worse it does. SCOUT trains the teacher to continue student-written prefixes.

Interactive method walkthrough from the SCOUT project page.

Paper Project page

Publications

At ICML, ICLR, NeurIPS, EMNLP, AAAI, KDD, and other venues. Full list on Google Scholar and Semantic Scholar.

Post-Training & Test-Time Scaling

SCOUT: On the Off-Policy Teacher in On-Policy Distillation

Langlin Huang, Hao Liu, Mononito Goswami, Xinyu Li, Prithwish Jana, Nikos Kanakaris, Patrick Blöbaum, Purak Jain

ArXiv Preprint, 2026

Identifies teacher-side off-policy degradation in on-policy distillation and mitigates it by adapting the teacher to student-generated prefixes with RL.

PDF Project page

Hermes: Learning Contextual Reasoning Unlocks Test-Time Scaling

Xinyu Li^*, Mononito Goswami^*, Hao Liu, Nikos Kanakaris, Langlin Huang, Prithwish Jana, Patrick Blöbaum, Purak Jain

ArXiv Preprint, 2026

Flexible context management for long-horizon test-time scaling: a parameterized delegation/digestion harness, and an SFT+RL recipe (Hermes-Learn) that teaches a 4B model adaptive context reuse.

PDF Project page

Agents

SpIDER: Spatially Informed Dense Embedding Retrieval for Software Issue Localization

Shravan Chaudhari, Rahul Thomas Jacob, Jiajun Cao, Shihab Rashid, Mononito Goswami, Christian Bock

Conference on Empirical Methods in Natural Language Processing (EMNLP), 2026

Retrieval and reasoning framework for software issue localization, combining structural graph exploration with LLM-based reasoning to improve localization accuracy.

PDF

MILO: Automated Harness Discovery via Orchestrated Multi-Agent Evolution

Prithwish Jana, Mononito Goswami, Hao Liu, Xinyu Li, Langlin Huang, Zhehui Huang, Zhishen Huang, Patrick Blöbaum, Anoop Deoras, Purak Jain, Nikos Kanakaris, Sahika Genc

ArXiv Preprint, 2026

Automated harness discovery through orchestrated multi-agent evolution: the search co-evolves the harness and its own strategy, reaching state-of-the-art results on Terminal-Bench 2.1, PaperBench, and DeepSWE.

PDF Project page

TimeSeriesGym: A Scalable Benchmark for (Time Series) Machine Learning Engineering Agents

Yifu Cai, Xinyu Li, Mononito Goswami, Michał Wiliński, Gus Welter, Artur Dubrawski

ArXiv Preprint, 2025

A benchmarking environment for evaluating ML engineering agents on the full time-series modeling stack, from data exploration and modeling to testing and deployment.

PDF Code

Foundation Models

Chronos-2: From Univariate to Universal Forecasting

Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, Michael Bohlke-Schneider

ArXiv Preprint, 2025

Extends the Chronos forecasting model to handle multivariate, covariate-conditioned, and probabilistic forecasting in a single unified architecture.

PDF

Exploring Representations and Interventions in Time Series Foundation Models

Michał Wiliński, Mononito Goswami, Willa Potosnak, Nina Żukowska, Artur Dubrawski

International Conference on Machine Learning (ICML), 2025

First to show that time series foundation models learn interpretable concepts (trends, seasonality) despite self-supervised training, and that these representations can be targeted for intervention.

PDF

MOMENT: A Family of Open Time-series Foundation Models

Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, Shuo Li, Artur Dubrawski

International Conference on Machine Learning (ICML), 2024

One of the first open-source time series foundation models. 2.5M+ downloads on HuggingFace, 700+ GitHub stars.

PDF Code HuggingFace Data

Evaluation Science

TimeSeriesExamAgent: Creating Time Series Reasoning Benchmarks at Scale

Małgorzata Gwiazda, Yifu Cai, Mononito Goswami, Arjun Choudhry, Artur Dubrawski

International Conference on Learning Representations (ICLR), 2026

Benchmark for temporal reasoning in LLMs, with scalable task generation via LLM agents and item response theory.

PDF

AQuA: A Benchmarking Tool for Label Quality Assessment

Mononito Goswami, Vedant Sanil, Arjun Choudhry, Arvind Srinivasan, Chalisa Udompanyawit, Artur Dubrawski

Neural Information Processing Systems (NeurIPS), 2023 Datasets and Benchmarks Track

A comprehensive benchmarking tool for evaluating label error detection methods across diverse datasets and annotation types.

PDF Code

Unsupervised Model Selection for Time-series Anomaly Detection Spotlight

Mononito Goswami, Cristian Challu, Laurent Callot, Lenon Minorics, Andrey Kan

International Conference on Learning Representations (ICLR), 2023

Shows that ensembles of unsupervised heuristics and weak supervision can accurately select anomaly detection models without any ground-truth labels.

PDF Code

Healthcare & Education

JoLT: Jointly Learned Representations of Language and Time-Series for Clinical Time-Series Interpretation Best Student Abstract

Yifu Cai, Arvind Srinivasan, Mononito Goswami, Arjun Choudhry, Artur Dubrawski

AAAI Conference on Artificial Intelligence (AAAI), 2024 Student Abstract. Also at NeurIPS 2023 DGM4H Workshop.

Learns joint language-time series embeddings so clinicians can query physiological signals with natural language descriptions.

Paper OpenReview

Using Weakly Supervised Machine Learning to Label Atrial Fibrillation in Real-World Intensive Care Unit Telemetry Data

Sydney R. Rooney, Roman Kaufman, Mononito Goswami, Michael R. Pinsky, J. Kyle Miller, Salah Al-Zaiti, Artur Dubrawski, Gilles Clermont

Circulation, 2022

Applies weak supervision to detect atrial fibrillation episodes in noisy ICU telemetry without expert-labeled training data.

Paper

Classifying Unstructured Clinical Notes via Automatic Weak Supervision

Chufan Gao, Mononito Goswami, Jieshi Chen, Artur Dubrawski

Machine Learning for Healthcare Conference (MLHC), PMLR, 2022

Automatically generates labeling functions from clinical notes to train classifiers without manual annotation.

PDF

Counterfactual Phenotyping with Censored Time-to-Events

Chirag Nagpal, Mononito Goswami, Keith Dufendach, Artur Dubrawski

ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2022

Combines causal inference with survival analysis to discover patient phenotypes under counterfactual treatment scenarios.

PDF

Weak Supervision for Affordable Modeling of Electrocardiogram Data

Mononito Goswami, Benedikt Boecking, Artur Dubrawski

AMIA Annual Symposium Proceedings, 2021

Applies weak supervision and data programming to build ECG classification models without manual expert labels.

Paper

Discriminating Cognitive Disequilibrium and Flow in Problem Solving: A Semi-Supervised Approach Using Involuntary Dynamic Behavioral Signals

Mononito Goswami, Lujie Chen, Artur Dubrawski

AAAI Conference on Artificial Intelligence (AAAI), 2020

Uses semi-supervised learning on involuntary behavioral signals (eye gaze, facial expressions) to detect cognitive states during problem solving.

Paper