New opportunity

Researcher, Post Training

Cartesia · San Francisco, United States

About this role

Cartesia seeks a Researcher on its Post-Training team to develop methods and systems that improve the capabilities and alignment of multimodal foundation models. The role combines machine learning research, evaluation, and infrastructure, with a focus on using feedback to improve model reasoning across modalities.

Responsibilities

Lead research initiatives in multimodal model alignment and capabilities, including preference optimization and feedback-driven learning.
Develop post-training techniques, evaluation frameworks, and metrics to measure model improvement.
Work with research, product, and platform teams on best practices for creating specialized models.
Implement, debug, and scale experimental systems for reliable, reproducible training runs; turn research findings into production-ready systems that improve reasoning, consistency, and human alignment.

Required qualifications

Deep knowledge of preference optimization and alignment methods, including RLHF and related approaches.
Experience designing evaluations and metrics for generative or multimodal models.
Strong engineering and debugging skills, experience building or scaling complex machine learning systems, and the ability to diagnose model behavior throughout the training and evaluation pipeline.

Preferred qualifications

Experience training multimodal models, including text, audio, or vision-language models.
Contributions to alignment research or open-source model evaluation or fine-tuning projects.
Experience designing or implementing human-in-the-loop evaluation systems.

Location and compensation:

Full-time, on-site role at Cartesia’s San Francisco headquarters. The posted compensation range is $180,000–$350,000 USD per year; the posting also describes base salary and equity.
Visa sponsorship support is assessed case by case and is not guaranteed. US-employee benefits include fully covered medical, dental, and vision insurance for employees and families; 9 weeks of paternity leave or 12 weeks of maternity leave; a 401(k); a monthly commuter allowance; flexible PTO; and daily meals and snacks.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

Preference optimizationModel alignmentRLHFModel evaluationEvaluation metricsFeedback-driven learningMultimodal modelsPost-trainingMachine learning systemsDebuggingFine-tuningHuman-in-the-loop evaluation

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Own multimodal alignment research
The role owns initiatives to improve model capabilities and alignment, with emphasis on preference optimization, RLHF-related methods, and feedback-driven learning across modalities.
Prove model improvement
Develop post-training methods alongside evaluation frameworks and metrics that measure changes in reasoning, consistency, and alignment rather than relying on training results alone.
Make experiments reliable
Implement, debug, and scale experimental systems for reproducible training runs; trace unexpected model behavior through the training and evaluation pipeline.
Turn findings into usable systems
Translate research into production-ready systems and partner with research, product, and platform teams on best practices for specialized models.

Likely challenges to think through

Interpreting feedback across modalities
Likely preparation hypothesis: a feedback signal that helps one type of reasoning may not transfer cleanly to another. Prepare to explain how you would compare signals and inspect behavior by modality before claiming an alignment gain.
Separating metric gains from real gains
Likely preparation hypothesis: optimizing against an evaluation can improve its score without improving the intended behavior. Prepare a validation approach using complementary metrics and qualitative error analysis.
Balancing research speed with reproducibility
Likely preparation hypothesis: rapid iteration can make training runs harder to compare or findings harder to deploy. Prepare to discuss experiment tracking, debugging checkpoints, and criteria for moving a result into a reliable system.

Practice your answers

Technical practice

How would you compare RLHF with other preference-optimization approaches for a post-training objective?

What to coverThe role calls for deep knowledge of preference optimization and alignment, including RLHF. Suggested preparation: Explain how you would choose an approach based on the available feedback, training stability, and the behavior you need to improve.

How would you turn feedback into a learning signal for reasoning across modalities?

What to coverThe posting explicitly focuses on feedback-driven learning and reasoning across modalities. Suggested preparation: Discuss how you would define the desired behavior, assess feedback quality, and check whether improvement in one modality comes at a cost in another.

How would you design an evaluation framework to establish whether post-training improved a multimodal model?

What to coverDeveloping evaluation frameworks to measure model improvement is a stated responsibility. Suggested preparation: Cover baselines, representative tasks, held-out evaluation, and how you would distinguish a genuine gain from adaptation to a narrow test set.

What metrics would you use to assess reasoning, consistency, and human alignment?

What to coverThe role aims to improve all three behaviors and calls for experience designing generative or multimodal model metrics. Suggested preparation: Explain what each metric captures, where automated measures may be misleading, and when human judgments would add useful evidence.

What post-training challenges arise when a model handles text, audio, or vision-language inputs?

What to coverMultimodal training experience in those modalities is listed as a nice-to-have, not a requirement. Suggested preparation: Discuss how you would detect uneven improvements across modalities and test whether feedback reflects the input information the model actually used.

How would you trace a model-behavior regression through the training and evaluation pipeline?

What to coverThe posting calls for diagnosing complex behavior across that pipeline. Suggested preparation: Walk through how you would isolate changes in data, feedback, training, and evaluation before attributing the regression to the model itself.

How would you make experimental training runs reliable and reproducible?

What to coverImplementing and debugging systems for reliable, reproducible runs is an explicit responsibility. Suggested preparation: Address controlled configurations, data and model versioning, run records, and checks that reveal failures or unintended differences between runs.

What trade-offs would you make when scaling a post-training experiment?

What to coverThe role combines research with scaling complex ML systems and measuring behavior at scale. Suggested preparation: Explain how you would balance iteration speed, compute cost, observability, and confidence that results remain comparable as the system grows.

How would you design a human-in-the-loop evaluation for model alignment?

What to coverHuman-in-the-loop evaluation is a listed nice-to-have. Suggested preparation: Describe how you would frame judgments, assess disagreement or inconsistent feedback, and use human results alongside automated evaluations.

How would you decide whether a post-training finding is ready for a specialized production model?

What to coverThe role involves creating specialized models with other teams and translating research into production-ready systems. Suggested preparation: Discuss evidence of improved reasoning and consistency, reliability checks, and trade-offs between a promising experimental result and dependable deployment.

Behavioural practice

Tell me about a research initiative you led from an open question to a defensible finding.

What to coverOwning research initiatives is part of the advertised role. Suggested preparation: Use a real situation, explain how you chose the question and evaluated the evidence, and state the result, including any limits to the finding.

Describe a time you worked across research, product, and platform perspectives on a model decision.

What to coverThe role calls for partnering with those teams on best practices for specialized models. Suggested preparation: Give a real situation, your actions in resolving differing priorities, and the resulting decision or outcome.

Tell me about a difficult training or evaluation failure you diagnosed.

What to coverDebugging experimental systems and tracing model behavior are stated responsibilities. Suggested preparation: Describe the situation, the steps you personally took to isolate the cause, and the result, including how you verified the fix.

Describe a time you moved a research result toward a dependable system without sacrificing quality.

What to coverThe posting emphasizes production-ready research and fast execution without sacrificing quality. Suggested preparation: Use a real situation, explain the trade-offs and safeguards you chose, and report the outcome rather than assuming the transition was successful.

Sharpen your knowledge

Topics to prepare

Preference optimization and alignment
Required: deep knowledge of preference optimization, RLHF and related alignment approaches. Revision exercise: compare two approaches in terms of the feedback they use, what they optimize and how each could fail to reflect human preferences.
Feedback-driven multimodal reasoning
The role explores how feedback can improve reasoning across modalities; text, audio and vision-language training experience is a nice-to-have. Revision exercise: sketch how a feedback signal could assess a text-and-audio response, including one ambiguity the signal might miss.
Evaluation validity and metrics
The posting requires experience designing evaluations and metrics for generative or multimodal models. Revision exercise: define separate measures for reasoning, consistency and human alignment, then identify a case where one measure improves without the underlying behavior improving.
Research-to-production trade-offs
The role calls for turning research findings into production-ready systems and partnering with research, product and platform teams on specialized models. Revision exercise: outline what evidence would justify advancing a post-training method beyond an experiment, and what reliability questions each team might raise. The questions are a preparation suggestion, not a stated company process.

Skills to revise

Build a comparative evaluation
The role develops evaluation frameworks to measure model improvement. Practice designing a small before-and-after evaluation: specify examples, scoring criteria, failure categories and how results would be compared. This is a suggested exercise, not a prescribed Cartesia framework.
Trace model-behavior failures
The posting requires diagnosing behavior across the training and evaluation pipeline. Practice working through a hypothetical drop in alignment scores: check the feedback data, training changes, evaluation inputs and scoring, and state what observation would distinguish each cause.
Make training runs reproducible
The role involves implementing, debugging and scaling experimental systems for reliable, reproducible runs. Practice drafting a run record that captures data version, configuration, changes, results and failures; explain how to investigate two runs that disagree.
Plan a post-training experiment
Researchers in this role own initiatives to improve multimodal model alignment and capabilities. As preparation, write a one-page experiment plan with a hypothesis, feedback source, comparison condition, evaluation measures and a decision rule. The format is a suggestion, not an employer requirement.

Your preparation checklist

Prepare a post-training research walkthrough
The role involves owning research on multimodal alignment and capabilities. If you have a relevant project, prepare a brief account of its hypothesis, method, your decisions, results, limitations, and next experiment.
Explain your preference-optimization choices
The posting calls for deep knowledge of RLHF and related alignment methods. Prepare a comparison of approaches, the feedback each uses, and how you would test whether an apparent improvement is real.
Draft a model-evaluation exercise
As interview preparation, design a hypothetical evaluation for reasoning and consistency across two modalities. Specify metrics, comparison baselines, failure cases, and how you would interpret conflicting results; this is a practice exercise, not a stated interview task.
Map a training-to-evaluation debugging case
The role requires diagnosing model behavior and scaling reproducible experiments. Prepare an example, if available, showing how you traced a failure through data, training, and evaluation, plus the controls and records used to reproduce the run.
Outline a path from finding to usable system
Cartesia describes turning research findings into production-ready systems with research, product, and platform teams. Sketch the validation, reliability checks, and cross-team decisions you would propose before deploying a post-training change.
Gather evidence for relevant nice-to-haves
If applicable, select concrete artifacts from multimodal training, alignment research, open-source evaluation or fine-tuning, or human-in-the-loop evaluation. Be ready to state your contribution and what the evidence does—and does not—show.
Write questions for the team
Prepare questions about which feedback signals and modalities are priorities, how model improvement is evaluated, where experiment reproducibility is hardest, and how research, product, and platform teams decide a method is ready to use.

Availability last checked: Nov 4, 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in United States. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…