Research Scientist - Science of Evaluations

About the AI Security Institute

The AI Security Institute is the largest team in a government dedicated to understanding AI capabilities and risks in the world. 

Our mission is to equip governments with an empirical understanding of the safety of advanced AI systems. We conduct research to understand the capabilities and impacts of advanced AI and develop and test risk mitigations. We focus on risks with security implications, including the potential of AI to assist with the development of chemical and biological weapons, how it can be used to carry out cyber-attacks, enable crimes such as fraud, and the possibility of loss of control. 

The risks from AI are not sci-fi, they are urgent. By combining the agility of a tech start-up with the expertise and mission-driven focus of government, we’re building a unique and innovative organisation to prevent AI’s harms from impeding its potential. 

The Science of Evaluations Team 

AISI’s Science of Evaluations team will conduct applied and foundational research focused on two areas at the core of our mission: (i) measuring existing frontier AI system capabilities and (ii) predicting the capabilities of a system before running an evaluation. 

Measurement of Capabilities: the goal is to develop and apply rigorous scientific techniques for the measurement of frontier AI system capabilities, so they are accurate, robust, and useful in decision making. This is a nascent area of research which supports one of AISIs core products: conducting tests of frontier AI systems  and feeding back results, insights, and recommendations to model developers and policy makers.  

The team will be an independent voice on the quality of our testing reports and the limitations of our evaluations. You will collaborate closely with researchers and engineers from the workstreams who develop and run our evaluations, getting into the details of their key strengths and weaknesses, proposing improvements, and developing techniques to get the most out of our results. 

 The key challenge is increasing the confidence in our claims about system capabilities, based on solid evidence and analysis. Directions we are exploring include: 

• Running internal red teaming of testing exercises and adversarial collaborations with the evaluations teams, and developing “sanity checks” to ensure the claims made in our reports are as strong as possible. Example checks could include: performance as a function of context length, auditing areas with surprising model performance, checking for soft refusal performance issues, and efficient comparisons of system performance between pre-deployment and post-deployment testing. 
• Running In-depth analyses of evaluations results to understand successes and failures and using these insights to create best practices for testing exercises. 
• Developing our approach to uncertainty quantification and significance testing, increasing statistical power (given time and token constraints).  
• Developing methods for inferring model capabilities across given domains from task or benchmark success rates, and assigning confidence levels to claims about capabilities. 

Predictive Evaluations: the goal is to develop approaches to estimate the capabilities of frontier AI systems on tasks or benchmarks, before they are run. Ideally, we would be able to do this at some point early in the training process of a new model, using information about the architecture, dataset, or training compute. This research aims to provide us with advance warning of models reaching a particular level of capability, where additional safety mitigations may need to be put in place. This work is complementary to both safety cases—an AISI foundational research effort—and AISI’s general evaluations work.  

This topic is currently an area of active research (e.g., Ruan et al. 2024), and we believe it is poised to develop rapidly. We are particularly interested in developing predictive evaluations for complex, long-horizon agent tasks, since we believe this will be the most important type of evaluation as AI capabilities advance. You will help develop this field of research, both by direct technical work and via collaborations with external experts, partner organizations, and policy makers.  

Across both focus areas, there will be significant scope to contribute to the overall vision and strategy of the science of evaluations team as an early hire. You’ll receive coaching from your manager and mentorship from the research directors at AISI (including Geoffrey Irving and Yarin Gal), and work closely with talented Policy / Strategy leads and Research Engineers and Research Scientists.  

Responsibilities 

This role offers the opportunity to progress deep technical work at the frontier of AI safety and governance. Your work will include: 

• Running internal red teaming of testing exercises and adversarial collaborations with the evaluations teams, and developing “sanity checks” to ensure the claims made in our reports are as strong as possible. 
• Conducting in-depth analysis of evaluations methodology and results, diagnosing possible sources of uncertainty or bias, to improve our confidence in estimates of AI system capabilities. 
• Improving the statistical analysis of evaluations results (e.g. model selection, hypothesis testing, significance testing, uncertainty quantification). 
• Developing and implementing internal best-practices and protocols for evaluations and testing exercises. 
• Staying well informed of the details and strengths and weaknesses of evaluations across domains in AISI and the state of the art in frontier AI evaluations research more broadly. 
• Conducting research on predictive evaluations using the latest techniques from the published literature on AISI’s internal evaluations, as well as conducting novel research to improve these techniques.  
• Writing and editing scientific reports and other materials aimed at diverse audiences, focusing on synthesising empirical results and recommendations to key decision-makers, ensuring high standards in clarity, precision, and style. 

Person Specification 

To set you up for success, we are looking for some of the following skills, experience and attitudes, but we are flexible in shaping the role to your background and expertise. 

• Experience working within a world-leading team in machine learning or a related field (e.g. multiple first author publications at top-tier conferences). 
• Strong track-record of academic excellence (e.g. PhD in a technical field and/or spotlight papers at top-tier conferences). 
• Comprehensive understanding of large language models (e.g. GPT-4). This includes both a broad understanding of the literature, hands-on experience with designing and running evaluations, scaling laws, fine-tuning, scaffolding, prompting. 
• Broad experience in empirical research methodologies, potentially in fields outside of machine learning, and statistical analysis (T-shaped: some deep knowledge, lots of shallow knowledge, in e.g. experimental design, A/B testing, Bayesian inference, model selection, hypothesis testing, significance testing). 
• Deeply care about methodological and statistical rigor, balanced with pragmatism, and willingness to get into the weeds. 
• Experience with data visualization and presentation. 
• Proven track record of excellent scientific writing and communication, with ability to understand and communicate complex technical concepts for non-technical stakeholders and synthesize scientific results into compelling narratives.  
• Motivated to conduct technical research with an emphasis on direct policy impact rather than exploring novel ideas.  
• Have a sense of mission, urgency, and responsibility for success, demonstrating problem-solving abilities and preparedness to acquire any missing knowledge necessary to get the job done. 
• Ability to work autonomously and in a self-directed way with high agency, thriving in a constantly changing environment and a steadily growing team. 
• Bring your own voice and experience but also an eagerness to support your colleagues together with a willingness to do whatever is necessary for the team’s success. 

Salary & Benefits

We are hiring individuals at all ranges of seniority and experience within this research unit, and this advert allows you to apply for any of the roles within this range. Your dedicated talent partner will work with you as you move through our assessment process to explain our internal benchmarking process. The full range of salaries are available below, salaries comprise of a base salary, technical allowance plus additional benefits as detailed on this page.

• Level 3 - Total Package £65,000 - £75,000 inclusive of a base salary £35,720 plus additional technical talent allowance of between £29,280 - £39,280
• Level 4 - Total Package £85,000 - £95,000 inclusive of a base salary £42,495 plus additional technical talent allowance of between £42,505 - £52,505
• Level 5 - Total Package £105,000 - £115,000 inclusive of a base salary £55,805 plus additional technical talent allowance of between £49,195 - £59,195
• Level 6 - Total Package £125,000 - £135,000 inclusive of a base salary £68,770 plus additional technical talent allowance of between £56,230 - £66,230
• Level 7 - Total Package £145,000 inclusive of a base salary £68,770 plus additional technical talent allowance of £76,230

 

This role sits outside of the DDaT pay framework given the scope of this role requires in depth technical expertise in frontier AI safety, robustness and advanced AI architectures.

 

There are a range of pension options available which can be found through the Civil Service website. 

The Department for Science, Innovation and Technology offers a competitive mix of benefits including: 

• A culture of flexible working, such as job sharing, homeworking and compressed hours. 
• A minimum of 25 days of paid annual leave, increasing by 1 day per year up to a maximum of 30. 
• An extensive range of learning & professional development opportunities, which all staff are actively encouraged to pursue. 
• Access to a range of retail, travel and lifestyle employee discounts. 
• The Department operates a discretionary hybrid working policy, which provides for a combination of working hours from your place of work and from your home in the UK. The current expectation for staff is to attend the office or non-home based location for 40-60% of the time over the accounting period. 

Selection Process 

In accordance with the Civil Service Commission rules, the following list contains all selection criteria for the interview process. 

The interview process may vary candidate to candidate, however, you should expect a typical process to include some technical proficiency tests, discussions with a cross-section of our team at AISI (including non-technical staff), conversations with your workstream lead. The process will culminate in a conversation with members of the senior team here at AISI. 

Candidates should expect to go through some or all of the following stages once an application has been submitted: 

• Initial interview 
• Technical take home test 
• Second interview and review of take home test 
• Third interview 
• Final interview with members of the senior team 

Required Experience 

We select based on skills and experience regarding the following areas: 

• Empirical research and statistical analysis
• Frontier AI model architecture, training, evaluation knowledge     
• AI safety research knowledge
• Written communication     
• Verbal communication     
• Research problem selection 

 

---

Additional Information