# Applied Research Intern at NextToken

> Onsite · United States · Intern · Other · open. Source: https://alderhq.app/jobs/ashby%3Anexttoken%3Af202a966-912e-40b8-b55a-8cca851eef5e (read 2026-11-20).

## Facts (observed)

- Company: [NextToken](https://alderhq.app/companies/nexttoken)
- Location text: Mountain View; San Francisco
- Where it can be done from (inferred): Onsite · United States (evidence: "Location: Mountain View; San Francisco")
- Posted: 2026-10-24 (27 days ago)
- First seen by alder: 2026-11-02 · last verified: 2026-11-09
- Reposted: 0 times · edited: 0 times
- Apply on the employer's site: https://jobs.ashbyhq.com/nexttoken/f202a966-912e-40b8-b55a-8cca851eef5e
- Skills listed (inferred): LLMs

## Pay and sponsorship

- Posted pay: not posted
- Market band for similar posted roles (SF Bay Area, n 196): p25 $60k, median $88k, p75 $128k

## Description (the employer's text)

Sit at the edge of research and production. You'll take a real workflow from a real team, make its failure modes measurable, construct the reward system, and post-train models that beat the frontier baseline on the work that matters. We hire interns who want to be researchers who ship.

### What you'll own

Evaluation suites built from production traces: deciding what to measure, building the graders, and proving they track real quality.

Verifiers and reward models. Turning human corrections, outcomes, and standards of good work into training signal.

Post-training experiments end-to-end: hypothesis, data, SFT/DPO/RL runs, ablations, analysis, and the write-up that defends the win with numbers.

RL environments and agentic harnesses for the workflows we're specializing models for.

Reproducing recent post-training, reasoning, and agent papers, and translating the ones that work into our stack.

### About you

Currently a student (BS / MS / PhD) in CS, ML, or an adjacent field, or recently graduated.

Strong empirical ML background. You've run post-training or RL experiments on LLMs and can explain what worked, what didn't, and why.

Rigorous about evaluation. You can spot a benchmark that lies and design one that doesn't.

Comfortable in messy production data. Real traces, not curated datasets, are your raw material.

Bias to ship. Your experiments end in a deployed model or a killed hypothesis, not just a doc.

Available for at least 12 weeks at meaningful intensity.

### Bonus

Publications or OSS in post-training, RL, evals, or agents.

Experience with verifier models, synthetic data, or RL environments at scale.

Shipped research into a production system before.

Strong taste for what the field is over- vs. under-investing in.

---
Data: alder (https://alderhq.app). Cite https://alderhq.app/jobs/ashby%3Anexttoken%3Af202a966-912e-40b8-b55a-8cca851eef5e and the date you read it. Observed facts come from the employer's own listing; location reach, level, family and skills are inferences.
