Open-Ended AI PhD | World Models for Emergent Multi-Agent Behaviour

Open-Ended AI PhD | World Models for Emergent Multi-Agent Behaviour

Review date: 6 November 2026
Recommendation: Promising with material uncertainties

Quick Decision

This is a four-year UCL collaborative studentship asking how generative world models can keep producing varied, coherent multi-agent environments in which AI characters learn, cooperate and compete. The advertised doctoral work is unusually cross-functional: develop open-ended generation methods; design metrics and curricula for emergent complexity; compare against frontier models; and test the results with designers and players. The official UELA catalogue identifies the opportunity as project 2531cc1056, hosted by UCL Electronic and Electrical Engineering (EEE), with Laura Toni and Jack Parker-Holder as academic supervisors and Iconic Interactive as industry partner. Project catalogue

The funding details are stronger than the short listing alone indicates: the project-specific collaborative catalogue says Home and International fee-status candidates may apply and lists a £3,900 annual partner stipend enhancement. UELA still caps international offers at 15 across the whole cohort. The UCL award page also gives two different year labels for £26,022, calling it a 2027/28 estimate in one section and a 2026/27 amount in another. Confirm the 2027/28 base stipend, international fee arrangement, partner payment and award conditions in writing. Collaborative catalogue UELA award page

The central research question is timely, but the listing leaves the model family, training data and rights, compute, evaluation protocol and success measures open. First clarify project availability and applicant eligibility, then ask about model/data/compute and partner terms at interview. The application deadline is 8 January 2027 at 12:00 UK time; UCL recommends submitting the application and references by 16 December 2026. UELA award page Applicant guidance

Research Problem and Doctoral Work

What Is Unsolved and Why It Matters

A world model can generate interactive scenes, but a visually plausible scene or a long video does not establish that agents can act in it usefully, that different agents will continue to produce diverse behaviour, or that the world remains coherent over a long interaction. The project asks how to prevent generated multi-agent environments from becoming repetitive while keeping characters understandable and engaging. Its application setting is interactive entertainment, but the research question sits at the intersection of generative modelling, reinforcement learning, multi-agent systems, quality-diversity search and evaluation. Project catalogue

Recent work illustrates why several kinds of evaluation may be needed. Genie demonstrated an 11-billion-parameter interactive world model trained from unlabelled internet video, with learned dynamics and latent actions; it is a useful frontier reference, not a model the UCL project promises to use. CASCADE, coauthored by Parker-Holder, studies population-driven, reward-free data collection for learning general world models. QDAC, coauthored by Iconic’s Borja González León, learns high-performing diverse behaviours in continuous control. MADRID applies diversity search to uncover weaknesses in multi-agent policies. These methods address different problems and should not be treated as interchangeable. Genie CASCADE QDAC MADRID

A plausible research challenge is to measure more than novelty: whether environments remain action-controllable, internally consistent, learnable, robust to different agents and valuable to people. This is an evidence-based interpretation of the advertised tasks, not a confirmed work package. The project description does not define “emergent complexity” quantitatively or identify the specific frontier models, baselines, game or environment. A closed-loop world-model benchmark argues that visual quality in open-loop generation is not enough to establish value for acting agents. Social-agent evaluation work also cautions that language-model judges can overestimate performance after targeted interaction training. World-in-World SOTOPIA-π

Work, Methods, and Potential Outputs

The catalogue describes four connected activities:

1. Develop open-ended methods using world models to generate a continuing stream of multi-agent environments.
2. Design metrics and curricula that measure and encourage emergent complexity.
3. Benchmark the methods against frontier models.
4. Test the work with real designers and players, then publish openly and present to technical and creative audiences.

This implies a full research loop from model/environment generation through quantitative assessment and human-facing evaluation. It may require systems programming, reinforcement-learning experiments, benchmark design, data governance and participant research. None of the following is fixed in the public project text: world-model architecture or size; training corpus, licence or consent; action representation; evaluation metric definitions; baseline models; compute/storage; game platform; player-study design or ethics review; or partner IP and publication-review terms. These should be treated as interview and offer questions, not silently filled in from papers on Genie or another system.

The supervisors’ recent work offers useful conceptual anchors. Toni’s 2025 study compares intrinsic rewards that promote diversity at different levels of RL exploration; it is relevant to curriculum and measurement design but does not test generated worlds. Parker-Holder’s CASCADE addresses population-generated task-agnostic trajectories; his 2024 Rainbow Teaming paper uses open-ended quality-diversity search to generate varied adversarial prompts, not game worlds. González León’s QDAC balances return with diverse behaviours in continuous control. Toni et al. 2025 Rainbow Teaming QDAC

School, Programme and Field Context

UCL is the degree-awarding institution and EEE is the project’s host department. The catalogue names Computer Science co-supervisor Jack Parker-Holder and industry collaborator Iconic Interactive, but it does not advertise a joint or industry-awarded degree. UELA projects are registered and funded for four years, with thesis submission expected within that period. EEE’s general MPhil/PhD information says that students initially register for an MPhil and transfer after a first-year examination; it describes the usual PhD period as three to four years. The generic course page lists three calendar years full-time. These are general course descriptions; the project-specific four-year UELA registration and funding terms should govern this studentship. Project catalogue UELA award page EEE research degree page EEE course page

For the EEE course, UCL requires a UK Master’s or at least an upper-second-class UK bachelor’s degree, or an overseas equivalent, in a relevant subject; the listed English-language requirement is Level 1. The course page gives a 2027/28 international fee of £36,100 per year outside a funded award. The studentship’s Home-rate fee coverage and UCL international differential award are separate funding terms and should be confirmed for the candidate’s assessed fee status. EEE course page UELA award page

UCL reports EEE ranked second in the UK in The Times UK University Rankings 2026 subject table. UCL also reports Computer Science first in England and second in the UK for research power in REF 2021. The two signals concern different subjects and measures; neither is a ranking of this project or a prediction of supervision or doctoral outcomes. EEE subject ranking Computer Science research and REF signal

Career-market context is broad rather than specialty-specific. The UK government’s 2026 plan estimates around 86,000 people were employed in UK AI companies in 2024 and says demand for AI-skilled roles rose in 2025 even as overall job-posting growth slowed. UKIE estimates 26,000 direct employees across 2,000 UK games businesses. These figures suggest adjacent AI and games ecosystems, but do not measure vacancies for open-ended world-modelling specialists, establish an Iconic placement, or predict a graduate’s outcome. UK Digital and Technologies Jobs Plan UKIE games-industry facts

Public group pages also show an organizational transition: the former Learning and Signal Processing group and UCL DARK page both state that they have been sunsetted and merged with the British Open-Ended Learning & Discovery Lab. This is not evidence of a problem with the project, but it makes it sensible to ask which group will host the student, how often the supervisors meet, and what compute, research staff and facilities are assigned. No project-specific access is published on those pages. LASP group page DARK Lab page

Supervisor and Lab Assessment

Supervisor Research Directions

The official project catalogue identifies Dr Laura Toni (UCL EEE) and Dr Jack Parker-Holder (UCL Computer Science) as academic supervisors, with Dr Borja González León, Iconic Interactive’s Chief Scientist, providing industrial supervision. The student is expected to spend time at Iconic’s central London office with company researchers, engineers and game designers. No schedule, employment relationship or minimum time allocation is published. Project catalogue Iconic Interactive

Toni’s profile identifies her as an Associate Professor in EEE. Her recent work on intrinsic rewards and RL exploration fits the advertised interest in encouraging and measuring behavioural diversity. Parker-Holder’s researcher profile describes interest in world models trained from large-scale data and open-endedness; the project catalogue, not the profile, is the primary confirmation of his role on this project. His and colleagues’ CASCADE, Rainbow Teaming and MADRID papers are relevant methodological precedents. González León’s profile lists him as Iconic’s co-founder and Chief Scientist, and his QDAC paper connects to quality-diversity methods. Laura Toni profile Jack Parker-Holder profile Borja González León profile Toni et al. 2025 CASCADE Rainbow Teaming MADRID QDAC

Recent Important Publications

Work	Contribution and project relevance	Limitation for this project
Toni et al. (2025), intrinsic rewards and exploration	Compares state-, policy- and skill-level diversity incentives; useful for the project’s metrics and curriculum question.	Does not study world-model generation or multiplayer environments.
Xu et al. (2022), CASCADE	Uses a population of agents to collect diverse task-agnostic data for world-model learning.	Simulated benchmark setting; predates current video foundation models and is not a player-facing world.
Grillotti et al. (2024), QDAC	Learns diverse high-performing behaviours in continuous control, including adaptation to perturbed environments.	Does not generate worlds or evaluate character coherence and human preference.
Samvelyan et al. (2024), Rainbow Teaming	Applies open-ended quality-diversity search to generate varied adversarial prompts.	Safety-prompt domain rather than interactive worlds.
Samvelyan et al. (2024), MADRID	Generates diverse adversarial settings to find vulnerabilities in trained multi-agent policies.	Evaluated in a football simulator; not a general measure of novelty, long-horizon coherence or player enjoyment.
Bruce et al. (2024), Genie	Demonstrates large-scale interactive world generation with latent actions from video.	Not stated as the project model; its data rights and compute implications cannot be transferred to this award.
Hughes et al. (2024), open-endedness perspective	Defines open-endedness around novelty and learnability and discusses safety.	A position paper, not an implementation recipe.
Zhang et al. (2025), World-in-World	Argues that world models should be evaluated in closed-loop agent interactions rather than by open-loop appearance alone.	Not a multi-agent entertainment benchmark.
Wang et al. (2024), SOTOPIA-π	Finds that LLM judges can overestimate social-agent ability after targeted interaction training.	Focuses on language-based social interaction, not world generation.

Lab, Resources, and Collaboration Environment

The project catalogue gives the clearest project-specific resource statement: time at Iconic’s central London offices and collaboration with its research, engineering and game-design staff. Iconic’s public site describes real-time AI for gameplay and interactive storytelling, which supports the industrial relevance of the setting but does not establish access to a particular product, dataset, proprietary model or compute cluster. Ask what is actually allocated, whether access continues throughout the doctorate, and how work can be published openly while respecting partner IP and confidentiality. Project catalogue Iconic Interactive

Funding, Application, and Living Cost

The UELA 2027/28 call is expected to provide around 50 fully funded four-year studentships, with up to 15 International offers. The scheme covers four years of fees at the Home rate, a rising enhanced stipend and a Research Training Support Grant. International awardees receive a UCL award for the Home/International fee difference. This project is specifically marked collaborative and eligible to Home and International fee-status candidates, and the collaborative catalogue lists an additional £3,900 stipend enhancement per year from Iconic. International eligibility remains competitive, and applicants must self-assess fee status; the formal assessment is made by UCL. UELA award page Collaborative catalogue

There is a material stipend-label conflict on the official UELA page: its key-information section estimates £26,022/year for 2027/28, while its detailed funding section labels £26,022 as the 2026/27 total, £1,500 above the cited UKRI London-weighted rate of £24,522. The specific project catalogue adds £3,900/year from the partner. A simple £26,022 + £3,900 = £29,922 annual figure is only a conditional illustration if both amounts apply concurrently to this cohort. It is not a confirmed 2027/28 award value. The project-level RTSG amount and whether any additional partner RTSG applies are also unstated. Obtain the exact year-by-year stipend and fee terms in the offer letter. UELA award page Collaborative catalogue

UCL’s indicative guide lists postgraduate-hall rent at £7,956–£27,508 for 52 weeks and £5,314 for listed essential costs over 52 weeks. Adding the endpoints gives a rough £13,270–£32,822 range before costs outside the guide, not a full personal budget. The range is wide because accommodation varies; private shared rent, utilities, travel choices, visa/health costs and personal spending can shift the total. The UCL course page lists a 2027/28 international EEE fee of £36,100/year outside the studentship, but UELA funding is intended to cover the Home/International differential for an eligible International awardee. UCL cost guide EEE course page

Apply through the official UELA 2027/28 portal by 8 January 2027 at noon UK time. Two references must be received by the deadline; UCL recommends a complete application and references by 16 December 2026 because support is unavailable during the Christmas closure. An applicant may apply for up to three projects. The guidance includes motivation, suitability and knowledge/skills responses (up to 300 words each), plus a research-experience answer split into two sections (up to 500 words each). Shortlisting is expected to finish at the end of January 2027, interviews in February and offers around mid-March. Applicant guidance UELA award page

Applicant Fit and Preparation

Required Background and Skills

The project specifically asks for a strong degree in computer science, engineering, mathematics or a related area and solid programming ability, with Python and a deep-learning framework as examples. The separate EEE course requires a relevant UK Master’s, upper-second UK bachelor’s or equivalent and English Level 1. The individual’s degree equivalence and English evidence are assessed through UCL’s formal admissions process. Project catalogue EEE course page

Preferred Experience and Suggested Preparation

Experience in reinforcement learning, multi-agent systems, generative models or game development is welcome, not essential. A competitive application should show concrete evidence of programming ownership, careful experiment design, reproducible work, and the ability to explain methodological choices and limitations. A previous simulation, agent, generative-model or interactive prototype can make the fit legible; these are preparation ideas, not additional formal requirements. Since the project asks for testing with designers and players, examples of communicating across disciplines would also be relevant. UELA’s application is anonymised at initial review, so evidence should be explicit in the written responses rather than relying on personal details. Project catalogue Applicant guidance

Next Actions and Questions to Verify

Before Applying

• Confirm the project remains open and ask whether the supervisors recommend a short project-fit conversation.
• Confirm fee-status eligibility for project 2531cc1056. The official catalogue permits International fee candidates, but the overall award is limited to 15 International offers and the applicant must self-assess status.
• Obtain two references early. The final deadline is 8 January 2027 noon UK time; the recommended working deadline is 16 December 2026.
• Ask which applicant experience the supervisors most value among RL, generative modelling, multi-agent systems and game development.

At Interview

• Which world-model family, action interface and training regime are being considered? Which frontier systems and baseline models will be compared?
• What training data may be used, under which licences/consents, and is any partner product or dataset access guaranteed?
• How will “emergent complexity” be operationalised alongside novelty, coherence, learnability, robustness and player/design value?
• What compute, storage and engineering support is budgeted, and what scale is realistic over four years?
• What participant numbers, designer/player tasks, ethics review, compensation and data-handling plan are envisaged?
• How often will the student work at Iconic, who provides day-to-day supervision, and which UCL research group currently hosts the project?

Before Accepting an Offer

• Obtain the exact year-by-year stipend, the £3,900 partner enhancement terms, Home/International fee-difference award, RTSG and any exclusions in writing. Resolve the £26,022 year-label conflict.
• Confirm IP ownership, confidentiality, publication-review timelines and open-source expectations.
• Confirm project-specific compute/data access and the group/supervisory arrangements described at interview.

Evidence and Sources

Project, Programme and Application

1. UELA 2027-28 Advertised Projects Catalogue — project 2531cc1056
2. UELA 2027-28 Collaborative Projects Catalogue
3. UCL EPSRC Doctoral Landscape Award studentships
4. UELA 2027-28 applicant guidance
5. Electronic and Electrical Engineering MPhil/PhD course page
6. EEE postgraduate research degree information

School, Supervisors and Industry

1. UCL EEE subject ranking, 2026
2. UCL Computer Science research and REF 2021 signal
3. Laura Toni’s UCL profile
4. Learning and Signal Processing group page
5. Jack Parker-Holder’s research profile
6. UCL DARK Lab page
7. Iconic Interactive
8. Borja González León’s profile and research

Research Papers and Methods

1. Genie: Generative Interactive Environments
2. The impact of intrinsic rewards on exploration in Reinforcement Learning
3. Learning General World Models in a Handful of Reward-Free Deployments (CASCADE)
4. Quality-Diversity Actor-Critic (QDAC)
5. Rainbow Teaming: Open-Ended Generation of Diverse Adversarial Prompts
6. Open-Endedness is Essential for Artificial Superhuman Intelligence
7. World-in-World: World Models in a Closed-Loop World
8. SOTOPIA-π: Interactive Learning of Socially Intelligent Language Agents
9. SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents
10. Multi-Agent Diagnostics for Robustness via Illuminated Diversity (MADRID)

Costs and Career Context

1. UCL cost-of-study guide
2. UK Digital and Technologies Jobs Plan
3. UKIE: UK Video Games Industry in Numbers, 2025