Senior Lead Research Scientist, Agentic AI

We’re seeking a Senior Lead Research Scientist (Agentic AI) to push the frontier of autonomous, tool‑using AI and ensure that innovations make it into production. You’ll split your time between novel research (benchmarks, learning algorithms, publications, and thought leadership) and building the tools, datasets, and systems required to run rigorous experiments and ship results into our agentic platform. You will partner closely with ML engineers, product, platform, and safety teams to translate research into reliable, scalable capabilities for customers and developers on Upwork.

Responsibilities

50/50 Split between research and engineering/productionalization.
Advance agentic benchmarking. Define and maintain a rigorous evaluation suite for agents (task success, reliability, recovery, safety, latency, and cost). Establish protocols, datasets, and reproducible metrics aligned to best practices in agentic evaluation; continuously harden benchmarks against loopholes and overfitting.
Invent and publish. Lead novel studies on agent planning, tool use, reflection/memory, safety, and multi‑agent coordination. Publish at top venues (e.g., NeurIPS/ICML/ICLR/ACL) and present learnings internally and externally.
Explore RLEF for agents. Develop Reinforcement Learning from Execution Feedback (RLEF) approaches that ground agent behavior in environment/run‑time signals (e.g., execution traces, tool results, test outcomes), comparing to RLHF/RLAIF on agent tasks.
Continuous/online learning. Design safe, measurable loops for continual improvement (data selection, drift detection, reward model updates, policy refresh), with guardrails that protect quality and cost.
Human‑in‑the‑loop systems. Partner on data strategy, labeling protocols, and reviewer tooling for RLHF and workflow‑level judgment; instrument quality controls and reviewer calibration.
Build research tooling. Stand up agents‑at‑scale experiment infrastructure: simulators, sandboxes, and orchestration for long‑horizon tasks; evaluation harnesses; offline/online A/B; and dashboards for longitudinal tracking.
Train & align models. Implement high‑quality pipelines for SFT, DPO, RLHF/RLAIF/RLEF; manage data provenance, safety filters, and automated red‑teaming; integrate eval signals into CI/CD.
Ship to production. Collaborate with platform teams to graduate prototypes into reliable services (APIs/SDKs, auth, observability, rate limiting) and to integrate agents with developer protocols (e.g., MCP) and runtime services.

What it takes to catch our eye

PhD or equivalent research track record with peer‑reviewed publications in relevant venues; strong empirical methodology and scientific writing/presentation skills.
Demonstrated contributions to agentic evaluation/benchmarks or long‑horizon reasoning (e.g., designing tasks, metrics, robust protocols).
Hands‑on experience adapting LLMs for tool use and multi‑step plans; fluent in prompting, function/tool calling, and memory/critique patterns.
Practical mastery of alignment methods (SFT, DPO, RLHF, RLAIF, and RLEF) and reward‑modeling; you know when to prefer each and how to evaluate them.
Proficiency in Python and one or more of PyTorch/JAX; experience with distributed training (e.g., DDP/Ray), dataset curation, experiment tracking, and reproducibility.
Ability to build research‑grade tools that evolve into production‑grade services (APIs/SDKs, data stores, streaming/messaging, tracing/metrics).
Comfortable building end‑to‑end eval pipelines (offline + online), defining pass/fail gates, and quantifying trade‑offs (quality, safety, latency, cost).
Experience with safety testing and red‑teaming for agents; familiarity with risk taxonomies for autonomous systems.
Proven success mentoring senior ICs, leading cross‑functional initiatives, and educating internal/external audiences (talks, tutorials, blog posts, open‑source).

Come change how the world works.

Upwork is establishing its first international operational hub in Lisbon, Portugal. The new office is expected to be fully operational by Q4 2026.

This position will initially be employed through a partner to ensure a seamless hiring process while we establish the hub. Once the hub is established, there may be opportunities to transition to employment with Upwork depending on business needs and other requirements. While employed by the partner, you’ll work as part of Upwork’s team, with access to our resources, culture, and growth opportunities.

Our partner will offer competitive benefits. When Upwork’s hub is established, we will be excited to offer employment and benefits directly as business needs require.

Upwork is committed to building a diverse, inclusive, and equitable workforce. Employment decisions are made without regard to race, color, religion, gender, sexual orientation, gender identity, national origin, disability, or any other status protected by applicable law.

To learn more about how Upwork processes and protects your personal information as part of the application process, please review our Global Job Applicant Privacy Notice

Responsibilities

50/50 Split between research and engineering/productionalization.

Advance agentic benchmarking. Define and maintain a rigorous evaluation suite for agents (task success, reliability, recovery, safety, latency, and cost). Establish protocols, datasets, and reproducible metrics aligned to best practices in agentic evaluation; continuously harden benchmarks against loopholes and overfitting.

Invent and publish. Lead novel studies on agent planning, tool use, reflection/memory, safety, and multi‑agent coordination. Publish at top venues (e.g., NeurIPS/ICML/ICLR/ACL) and present learnings internally and externally.

Explore RLEF for agents. Develop Reinforcement Learning from Execution Feedback (RLEF) approaches that ground agent behavior in environment/run‑time signals (e.g., execution traces, tool results, test outcomes), comparing to RLHF/RLAIF on agent tasks.

Continuous/online learning. Design safe, measurable loops for continual improvement (data selection, drift detection, reward model updates, policy refresh), with guardrails that protect quality and cost.

Human‑in‑the‑loop systems. Partner on data strategy, labeling protocols, and reviewer tooling for RLHF and workflow‑level judgment; instrument quality controls and reviewer calibration.

Build research tooling. Stand up agents‑at‑scale experiment infrastructure: simulators, sandboxes, and orchestration for long‑horizon tasks; evaluation harnesses; offline/online A/B; and dashboards for longitudinal tracking.

Train & align models. Implement high‑quality pipelines for SFT, DPO, RLHF/RLAIF/RLEF; manage data provenance, safety filters, and automated red‑teaming; integrate eval signals into CI/CD.

Ship to production. Collaborate with platform teams to graduate prototypes into reliable services (APIs/SDKs, auth, observability, rate limiting) and to integrate agents with developer protocols (e.g., MCP) and runtime services.

What it takes to catch our eye

PhD or equivalent research track record with peer‑reviewed publications in relevant venues; strong empirical methodology and scientific writing/presentation skills.

Demonstrated contributions to agentic evaluation/benchmarks or long‑horizon reasoning (e.g., designing tasks, metrics, robust protocols).

Hands‑on experience adapting LLMs for tool use and multi‑step plans; fluent in prompting, function/tool calling, and memory/critique patterns.

Practical mastery of alignment methods (SFT, DPO, RLHF, RLAIF, and RLEF) and reward‑modeling; you know when to prefer each and how to evaluate them.

Proficiency in Python and one or more of PyTorch/JAX; experience with distributed training (e.g., DDP/Ray), dataset curation, experiment tracking, and reproducibility.

Ability to build research‑grade tools that evolve into production‑grade services (APIs/SDKs, data stores, streaming/messaging, tracing/metrics).

Comfortable building end‑to‑end eval pipelines (offline + online), defining pass/fail gates, and quantifying trade‑offs (quality, safety, latency, cost).

Experience with safety testing and red‑teaming for agents; familiarity with risk taxonomies for autonomous systems.

Proven success mentoring senior ICs, leading cross‑functional initiatives, and educating internal/external audiences (talks, tutorials, blog posts, open‑source).

Come change how the world works.

Upwork is establishing its first international operational hub in Lisbon, Portugal. The new office is expected to be fully operational by Q4 2026.

Our partner will offer competitive benefits. When Upwork’s hub is established, we will be excited to offer employment and benefits directly as business needs require.

To learn more about how Upwork processes and protects your personal information as part of the application process, please review our Global Job Applicant Privacy Notice