Parallel Scaling
Interdependent generations for LLMs
Bridge treats batched hidden states holistically to share signal across parallel generations, boosting accuracy with minimal params.
Read the paper →About
Researcher building personal superintelligence—LLMs that reason, collaborate, and act in human-centered environments. Led Meta AI post-training for the 2023 debut of Meta AI and launches of Llama 2/3/4 across the Family of Apps, improving model quality and scaling MetaAI to over 1 billion MAU.
Key areas of focus
- Frontier post-training: RL, Model behavior alignment, EQ and adaptive reasoning
- Agents that plan, use tools, and interact safely with people
- Data flywheels and online RL for continuously improving models
Earlier work: novel RL algorithms with formal guarantees, shipped across Ads, Recommendations, and Safety & Integrity to power large-scale decision systems at Meta.
Prior to Meta, Karthik completed his PhD and MS at the University of Maryland working with collaborators, focusing on mathematical foundations of algorithms & machine learning, and held intern positions at Adobe, IBM Almaden, IISc and Microsoft Research NYC. He earned a B.Tech (Hons) in Computer Science with a minor in Operations Research from IIT Madras.
Research Interests
- Large language models: RL, agents, adaptive reasoning, Human interaction
- Online learning, bandits and RL
- Online algorithms and market design
- Applied ML for advertising, recommender systems, and safety
Spotlight
Selected highlights
RLHF
The perfect blend: Mixture of Judges
CGPO balances multi-objective RLHF with mixture-of-judges to curb reward hacking and align models across tasks.
Read the paper →Inference Optimization
Adaptive reasoning under budgets
ICML 2025: IBPO learns to allocate reasoning depth based on problem difficulty, improving MATH500 with 2–4x inference budgets.
Read paper →Adversarial Bandits
Bandits with Knapsacks
Beyond worst-case analysis for adversarial bandits with knapsacks; tighter instance-dependent insights.
Read paper →Teaching
Online learning mini-series
Video series introducing adversarial and stochastic online learning foundations for upper-level undergrads.
Watch playlist →Post-training
LLM Post-training 101
Practical overview of post-training fundamentals, techniques, and workflows to align and improve LLMs.
Read post →Publications
Publications
Work spans RLHF, Inference-time Optimization, Parallel Scaling, Hallucination detection for LLMs, Bandits, Online Matching and Advertising systems. The complete conference, journal, and workshop list lives below.
Latest
- ACL 2025 – Reference-free hallucination detection with RATE-FT
- ICML 2025 – Inference Budget-Constrained Policy Optimization (IBPO)
- Bridge – Generalized Parallel Scaling with Interdependent Generations
- CGPO – The perfect blend: RLHF with Mixture of Judges
Full list of research and publications
View detailed publications
Loading publications…
Community
Teaching
Teaching and mentoring across online learning, algorithms, and programming. Instructor roles at UMD and IITM.
Online learning playlist →Teaching details
View courses & roles
Teaching Experience
- Honors
class
on
Discrete
Structures
(Spring
2019)
- Graduate
Algorithms
(Spring
2018)
Instructors: Aravind Srinivasan - Design
and
Analysis
of
Computer
Algorithms
(Fall
2016,
Spring
2017,
Fall
2017,
Fall
2018)
Instructors: Samir Khuller, Jessica Chang, Aravind Srinivasan - Discrete
Structures
(Fall
2014
and
Spring
2015)
Instructors: Clyde Kruskal, Fawzi Emad - Introduction
to
Programming
(Fall
2015
and
Spring
2016)
Instructors: Nelson Padua-Perez, Fawzi Emad - Paradigms
of
Programming(IITM)
Instructors: Narayanaswamy NS
I have served as a Teaching Assistant for the following classes during my time at UMD
As an undergrad I have served as a Teaching Assistant at IITM.
Service
Community Service
Program leadership and reviewing across top AI/ML conferences and journals, plus mentorship for new researchers.
Connect