TFMAdapter: Lightweight Instance-Level Adaptation of Foundation Models for Forecasting with Covariates
Afrin Dange, Sunita Sarawagi
CIKM 2025 · 2025 Oral presentation
Hi, I'm Afrin
I am a Research Scientist at Amazon Lab126, where I work on long-term memory for language agents with Dr. Deepak Gupta and Sankalp Dayal. My research focuses on understanding and improving how language agents store, retrieve, and manage information over long horizons. Specifically, I worked on analyzing failure modes in episodic memory systems. Through systematic analysis, I developed a taxonomy of failure modes. I used these findings to design process reward models for training policies based on small open-weight language models for memory retrieval and management.
Prior to this, I completed my MS by Research at CMInDS, IIT Bombay, where I was fortunate to work with Prof. Sunita Sarawagi on in-context learning and adaptation of foundation models for time series. I was also fortunate to collaborate with Dr. Praneeth Netrapalli, where I studied in-context learning in transformers under controlled numerical settings.
I am broadly interested in how agents can learn from experience and adapt over time. I am particularly interested in understanding how different components of agent systems, including memory, learning and feedback mechanisms, and reasoning interact to enable continual improvement. I am also interested in topics related to improving the capabilities of small open-weight language models.
Research interests: Agent Memory, Continual Learning, Reasoning, Learning from Feedback
Publications
Transformer Designs for In-Context Learning in Foundation Models for Time Series Forecasting with Covariates
Afrin Dange, Vaibhav Raj, Praneeth Netrapalli, Sunita Sarawagi
ICML 2024 Workshop on Theoretical Foundations of Foundation Models · 2024
Code
Manuscript
Updates
Attending COLM 2026 in San Francisco, US.
Our position on task-triggered evolution of memory was accepted at the COLM 2026 Workshop on Context Beyond Window.
Received a reviewer acknowledgment from the ICML 2026 Workshop on Foundation Models for Structured Data.
Attended ICML 2026 in Seoul, Korea.
Served as a reviewer for NeurIPS 2026.
Served as a reviewer for ICML 2026 Workshop on Foundation Models for Structured Data.
Served as a reviewer for ICLR 2026 Workshop on Time Series in the Age of Language Models.
Attended CIKM 2025 in Seoul, Korea.
Served as a reviewer for ICLR 2026.
Joined Amazon Lab126 as a Research Scientist, working on Long-Term Memory for Language Agents.
Our work on instance-level adaptation of time series foundation models was accepted for an oral presentation at CIKM 2025.
Attended the LLM Summer School in Yerevan, Armenia.
Completed my MS thesis defense.
Served as a reviewer for the ICML 2025 Workshop on Foundation Models for Structured Data.
Attended the Google DeepMind Research Symposium in Bangalore, India.
Attended the reinforcement learning workshop at IISc Bangalore, India.
Worked as a teaching assistant for CS726: Advanced Machine Learning.
Served as a reviewer for ICLR 2025.
Gave a talk titled "Conditional Generation Using GANs and Diffusion Models."
Our work on transformer designs for in-context learning in time series foundation models was accepted at the ICML 2024 TF2M Workshop.
Gave a talk titled "Generative AI for Vision."