Manifund foxManifund
Home
Login
About
People
Orgs
Categories
Newsletter
HomeAboutPeopleCategoriesLoginCreate
🌴
🌴
Pradeep

@pradeep02272002

0

Independent AI alignment researcher focusing on reward model reliability, multi-objective preference learning, and evaluator failure modes under optimization pressure.

https://www.linkedin.com/in/pradeep016/
$0total balance
$0charity balance
$0cash balance

$0 in pending offers

About Me

I am an independent AI safety researcher based in India working on empirical reward modeling and test-time evaluation reliability. My work investigates how preference evaluators fail under selection pressure and how multi-dimensional rubric supervision and counterfactual data curation prevent reward hacking and style bias.

Recent work:

  • Author of "Evaluator Reliability Under Selection Pressure: Implications for Agentic Test-Time Search" (accepted at the 3rd Workshop on Lifelong Agents @ AIAS+ 2026).

  • Developer of STELLAR-ARM, an open-source 6-dimensional associative reward model architecture (code and checkpoints on GitHub at pradeep-016/stellar-arm and Hugging Face at Stellar-lab/Stellar-Qwen3-8B-ARM).

Projects

Scaling Associative Reward Models (ARM)

unfunded

By Pradeep1Following2
Scaling Associative Reward Models (ARM)
🌴

Pradeep

about 21 hours ago

Quick update: I just pushed the ablation notebooks and evaluation logs to the GitHub repo (https://github.com/pradeep-016/stellar-arm).

The results break down how much comes from our data pipeline versus the model architecture. Our counterfactual data drives most of the raw accuracy gains, while the 6D ARM head prevents training instability on hard pairs and keeps reasoning-correctness consistency intact.