iiinova — where your digital self keeps evolving

Naveen Raman

@naveenraman.bsky.social

40 followers 82 following 7 posts

Machine Learning PhD at Carnegie Mellon University. AI + Decision Making for social good. naveenraman.com

Posts Replies Media

Naveen Raman @naveenraman.bsky.social · 08/04/2026

7/ It was great working on this project with @stephmilani.bsky.social and my advisor Fei Fang!

010 1

Naveen Raman @naveenraman.bsky.social · 08/04/2026

6/ We finally investigate whether automatic selection matches manual human selection. On the CUB Birds dataset, DRS selects the same concepts as manual curation. Moreover, it identifies subsets that achieve within 0.6% of manual performance while using only 70% of concepts.

110 1

Naveen Raman @naveenraman.bsky.social · 08/04/2026

5/ Well-selected concepts also improve the effectiveness of human intervention by up to 87%. The intuition is that better-selected concepts allow for more impactful interventions.

110 1

Naveen Raman @naveenraman.bsky.social · 08/04/2026

4/ Using DRS to select concepts outperforms baseline heuristics on both standard RL environments and a real-world Glucose simulator.

120 2

Naveen Raman @naveenraman.bsky.social · 08/04/2026

3/ The DRS algorithm selects concepts which minimize state abstraction error, thereby separating states with different optimal actions. We prove upper bounds on the gap between the optimal and concept-based policy when using DRS.

110 1

Naveen Raman @naveenraman.bsky.social · 08/04/2026

2/ Our key insight is that concept selection is really a state abstraction problem. A concept is decision-relevant if its removal would cause agents to confuse states that require different actions. This insight allows us to borrow results from the state abstraction literature.

120 2

Naveen Raman @naveenraman.bsky.social · 08/04/2026

Training concept-based models relies on concept selection which is labor-intensive and slow. We introduce Decision-Relevant Selection (DRS), a principled algorithm for automatic concept selection in RL. Paper: arxiv.org/abs/2604.04808 Website: naveenraman.com/projects/con... 🧵 1/n

141 4

Reposted by Naveen Raman

Russ Salakhutdinov @rsalakhu.bsky.social · 07/12/2024

blog.ml.cmu.edu/2024/12/06/s...

blog.ml.cmu.edu

ScribeAgent: Fine-Tuning Open-Source LLMs for Enhanced Web Navigation

TL;DR: LLM web agents are designed to predict a sequence of actions to complete a user-specified task. Most existing agents are built on top of general-purpose, proprietary models like GPT-4 and rely ...

021 2