Barnab\'as P\'oczos

Research profile

Research summary

Barnabás Póczos is a machine learning researcher affiliated with Carnegie Mellon University whose bibliometric record, resolved from OpenAlex, shows an h-index of 51 and an i10-index of 134. His indexed output rose steeply through the mid-2010s, peaking at 41 publications in 2018, and remains active with 26 works in the last three years and 14 already indexed for 2026. Topical clustering places stochastic gradient optimization (43 works), Gaussian processes and Bayesian inference (40), and neural network applications (39) at the core of his portfolio, with machine learning algorithms (34) and statistical methods (31) close behind. His current field distribution splits evenly across machine learning, machine learning theory (statistical distribution learning), and machine learning (multimodal AI agents). The three Evolve-reviewed 2026 preprints illustrate this spread: LLM-elicited priors for Bayesian optimization, minimax Bayes-risk bounds for sparse-prior distribution learning, and process-reward reinforcement learning for agentic vision-language models. Evolve's verdicts are cautiously positive, citing promising results alongside unreported implementation details, strong theoretical assumptions without experiments, and a thin margin over a reproduced baseline. His recent collaboration footprint records 16 distinct coauthors, with frequent collaborators including Xingming Long, Yu Liu, and Zhiwei Yang.

Research keywords

Highlights

• LLM-elicited priors improve Bayesian optimization

  Evolve's review finds that priors and auxiliary observations elicited from large language models improve Bayesian optimization on HPOBench and fusion tasks. The verdict is promising but flags that the LLM identity, hyperparameters, and full-grid results are unreported; reproducibility readiness is rated medium.

• Sparse priors yield provable distribution-learning risk bounds

  The reviewed theory paper establishes Omega(sqrt(k/n)) lower and O(sqrt(k)·log n/sqrt(n)) upper Bayes-risk bounds for distribution learning with sparse priors. Evolve judges the theory sound but notes strong assumptions and no experiments, with low reproducibility readiness.

• Process rewards for tool-using vision-language agents

  The NTEP-R method adds necessary tool-evidence path process rewards to reinforcement learning for agentic vision-language models, with clear efficiency gains. Evolve notes a thin 2.03-point margin over one reproduced baseline and reliance on judge-based rewards; reproducibility readiness is low.

Timeline

1. 2002First indexed works appear (2 publications), the start of a record now spanning more than two decades in the OpenAlex index.
2. 2016Output accelerates sharply: annual indexed works climb from 24 in 2015 to 33 in 2016.
3. 2018Publication count peaks at 41 indexed works, the highest annual total in the record.
4. 2022Indexed output falls to 3 works, the lowest annual count since the mid-2000s.
5. 2026Three arXiv preprints enter the Evolve catalog — LLM-augmented Bayesian optimization, sparse-prior distribution learning theory, and reward design for agentic vision-language models — within a year that already shows 14 indexed works.

Papers in the Evolve catalog

• 2026-09-18

  Bayesian Optimization with Rich Auxiliary Information via LLMs

• 2026-09-16

  Sparse Priors for Efficient Distribution Learning

• 2026-09-03

  Making Every Tool Call Count: Necessary Tool-Evidence Path Rewards for Agentic Vision-Language Models

Frequent co-authors