xAI · Post-training

Joseph Jennings

Member of Technical Staff

I work on post-training for the Grok model family — graders, evals, and reward design that shape agentic coding behavior. Most recently the Grok 4.5 release.

Before xAI: agentic trading systems at Robinhood, diagnostic agents at Apple, and ML at Sprout Social and JPMorgan — plus a GNN soup paper and gossip-based KV cache research along the way.

Research interests

RL data mixtures · post-training · reward & grader design · evaluation of agentic models

Google Scholar ORCID CV

News

Experience

xAI

Member of Technical Staff

May 2026 — Present
  • Design reward signals, graders, and evaluations for RL post-training of agentic coding — shipped in the Grok 4.5 release
  • Study how grader and eval design shifts model behavior, with a focus on reward signals that resist Goodharting
Robinhood

Software Engineer

Aug 2025 — May 2026
  • Designed the evaluation methodology for agentic stock trading — curated eval suites that gated launch quality — and built the MCP layer that lets agents execute trades autonomously — robinhood.com/us/en/agentic-trading
  • Reduced credit card update API latency by 45% (41s to 342ms p95) by redesigning Kafka partitioning and resolving PostgreSQL transaction contention to support 35k daily updates — robinhood.com/creditcard
Apple

Graduate Software Engineer Intern

May 2025 — Aug 2025
  • Investigated whether LLM agents can autonomously root-cause server failures at fleet scale (430K nodes): built an agent that correlates Splunk trace logs against system-architecture guidelines to generate diagnostic hypotheses — 79% diagnostic accuracy, 62% fewer ops tickets
  • Shipped the agent's diagnoses into the internal server-management UI, so engineers see AI-generated failure explanations instead of manually triaging failed servers
Sprout Social

Software Engineer II

Jun 2024 — May 2025
  • Architected and deployed a feature flagging engine handling 1,200+ requests/sec with 99.98% uptime for 85k daily users; enabled decoupled, real-time deployments and rollbacks
  • Implemented backend using Spring Boot, AWS S3 (persistence), Redis (caching), and PostgreSQL (customer metadata); optimized read/write paths for high QPS workloads
JPMorgan Chase

Machine Learning Engineer Intern

Jun 2023 — Aug 2023
  • Tested whether latent graph structure beats tabular baselines: designed a Graph Neural Network over 1.1M nodes that raised F1 from 0.22 (XGBoost) to 0.61
  • Ran feature-ablation studies with Recursive Feature Elimination, distilling 1.1M records into 16 high-impact features that improved model signal and downstream prediction quality
NASA

Software Engineer Intern

Aug 2022 — Dec 2022
  • Parallelized motor telemetry collection for a flight control system via multi-threading in C++; persisted data into structured logs for real-time monitoring

Publications

Enhanced Soups for Graph Neural Networks

Joseph Zuber, Aishwarya Sarkar, Joseph Jennings, Ali Jannesari

GrAPL @ IPDPS 2025

A gradient-based "souping" method that merges independently trained GNNs into a single model, reducing training memory by 76% and achieving a 24.5× speedup with a 1.2% accuracy improvement on OGB benchmarks.

arXiv PDF

Learned Souping Several independently trained GNNs are merged into one model. Interpolation weights alpha are learned by gradient descent, producing a single souped GNN. θ₁ θ₂ ⋮ θₖ α₁ α₂ αₖ Σ αᵢθᵢ souped GNN trained independently α fit by grad descent ensemble accuracy
+1.2% accuracy (up to) 24.5× souping speedup −76% souping memory

Mixing weights α are learned on validation data instead of grid-searched; the partitioned variant soups parameter slices to scale to large graphs (ogbn-products, GraphSAGE).

More on Google Scholar →

Projects

Gossip-Based Distributed KV Cache for LLM Inference

Graduate research project — University of Illinois Urbana-Champaign

A prototype inference system that gossips compressed KV-cache entries across nodes, so previously seen prompts hot-start on any replica. In our test cluster, this cut P99 latency by 40% and preserved availability through 80% node failure.

What I'm Reading

Research

  • On-Policy Distillation — Thinking Machines Lab

    Dense per-token teacher grading of the student's own rollouts — RL's on-policy relevance with distillation's reward density.

  • DeepSeekMath — arXiv 2402.03300

    The GRPO paper: group-relative advantage estimation that drops the critic entirely — still the backbone of much of reasoning RL.

  • TVCache: A Stateful Tool-Value Cache for Post-Training LLM Agents — arXiv 2602.10986

    Longest-prefix matching over tool-call sequences so parallel rollouts stop re-executing slow tool calls — GPUs shouldn't idle during agentic post-training.

  • Scaling Data-Constrained Language Models — arXiv 2305.16264

    Scaling laws for when unique tokens run out: how much repetition still helps, and how to think about data mixtures under constraint.

Life

Education

University of Illinois at Urbana-Champaign

Master's in Computer Science

GPA: 3.88 / 4.00

Specialization in Distributed Systems and Cloud Computing

Iowa State University

Bachelor of Science in Computer Science

Teaching Assistant for Data Structures, Object-Oriented Programming, and Introduction to Python

Technical Focus

Research — RL post-training for LLMs, reward and penalty design, grader and eval construction, Goodhart-resistance

Engineering — Python, PyTorch, Go, Java, C++; distributed systems (Kafka, PostgreSQL, Redis, Kubernetes, AWS)