Wenbo Chen

Wenbo ChenLLM agents: post-training, evaluation & infra Making LLM agents better, and proving they really are.

I'm a researcher at Amazon, post-training the LLMs behind Amazon Ads Agent. I work on LLM agents end to end: post-training the model, building the benchmarks that measure it, and the infrastructure that makes training fast and grading honest.

ML PhD from Georgia Tech / NSF AI4OPT with Pascal Van Hentenryck, where my work on scalable data-driven decision making saw major industrial adoption.

See the work Email

What I work on

Agent, eval, infra

A better agent needs a benchmark that can tell it is better, and infrastructure fast enough to try again. I work on all three, and every piece ships with a check.

01 · Agent

Better training, cheaper reasoning.

Amazon Ads Agent: a conversational AI experience for advertisers

Amazon Ads Agent post-training its LLMs

Amazon's AI agent for advertisers: describe a goal in plain language, and it plans, launches, optimizes, and measures the campaigns.

  • In production
Product
LSFlow: spherical flow policy with a combinatorial optimization solver

LSFlow

RL where every move combines many choices, with a solver that keeps every move valid: +20.6% over the prior best.

  • ICML 2026 Spotlight
Paper
02 · Eval

Measuring what agents can really do.

SkillsBench pipeline: benchmark construction, quality filtering, and paired evaluation

SkillsBench co-first author

Do skills actually help AI agents? 87 expert tasks with exact checkers: curated skills add 16.6 points.

PaperWebsitePress
03 · Infra

Fast to train, honest to grade.

ATTUNE architecture: a coding agent's diagnosis-driven search loop, deterministic correctness gates, an autonomous harness, durable state, and a self-evolving skill

ATTUNE

A coding agent tunes LLM training on its own, and a correctness gate checks every change: 1.7× throughput over expert recipes.

PaperCode
BenchShield catching a Terminal-Bench 3 reward hack where submitted source disables Lean kernel checking

BenchShield

Catching agents that game their tests: hacking paths found before a run, 96% detection across 31k+ runs.

Paper
  • LLM Post-Training
  • Reinforcement Learning
  • Test-Time Scaling
  • Distillation
  • Agent Evaluation
  • Benchmarks
  • RL Environments
  • Hallucination Detection
  • Training Efficiency
  • Distributed Training
  • Reward Hacking
More on post-training, test-time scaling, evaluation, and data-driven decision making →

Latest

News