PhD in Language Models: Post-Training Diffusion Models for Reasoning with Reinforcement Learning

PhD in Language Models: Post-Training Diffusion Models for Reasoning with Reinforcement Learning

Quick Decision

Strong but verify fit. This is a full-time, four-year PhD in reinforcement learning and reasoning for diffusion-based large language models at the University of Basel, funded through a Swiss National Science Foundation project. The advertised work includes post-training methods, adaptive unmasking, credit assignment and cross-domain reasoning benchmarks, including mathematics. The position is available from January 2027 or later. The official employer page asks candidates to apply preferably before 20 November 2026, yet also says there is no fixed deadline and the position remains open until a suitable candidate is found. A separate opportunity metadata field shows 2 December 2026, 23:59 Europe/Zurich. The employer’s current page is the primary timing source; treat 20 November as the prudent date and ask the employer to confirm the discrepancy in writing. The salary amount is not disclosed. Before accepting, obtain the gross salary, progression and contract details, and confirm compute resources and doctoral registration.

Research Problem and Doctoral Work

Diffusion language models generate text by repeatedly refining masked or noisy token sequences. Unlike conventional autoregressive systems that generate strictly left-to-right, they can update multiple token positions in parallel. That offers new options for generation speed and modeling dependencies, but creates a different post-training problem: policy likelihoods can be difficult to evaluate, making standard reinforcement-learning objectives expensive or statistically unreliable.

The official vacancy asks the researcher to develop reinforcement-learning methods for post-training large diffusion LLMs, explore reasoning-aware generation strategies such as adaptive unmasking and credit assignment, and build benchmarks for reasoning across domains, including mathematical problem solving. The likely research loop is to choose a model and benchmark, define a training or reward objective, run large-scale experiments, diagnose speed and accuracy tradeoffs, and publish results. The position requires large-scale experimentation, but does not specify the base model, training corpus, reward signals, hardware allocation, or exact evaluation suite.

Recent work by Professor Bogunovic and collaborators helps map the technical frontier. The 2025 wd1 preprint proposes a ratio-free policy-optimization objective to reduce repeated likelihood approximations. The 2026 GDSD preprint uses advantage-guided denoiser self-distillation to avoid relying on an ELBO likelihood surrogate, while the 2026 Temporal Self-Distillation preprint aligns predictions across denoising times to improve the speed-quality tradeoff of parallel decoding. These are active preprints, not established project methods; their reported results should be read as author-reported evidence. They point to possible dissertation questions around credit assignment across partially denoised states, choosing which tokens to reveal or commit, combining reward with self-distillation, and testing whether math or planning gains transfer to other domains.

A central risk is evaluation validity. Reasoning benchmarks can be contaminated or overfit, reward functions can be gamed, and decoding choices can change results independently of training. Large-scale experiments also depend on model licenses, GPU time and reproducibility infrastructure. A useful project will need careful baselines that separate improvements from the base model, sampler, unmasking schedule and post-training objective.

School, Programme and Field Context

The University of Basel is the named employer and apparent degree-awarding institution; the position is in its Department of Mathematics and Computer Science. The Rhine AI group also describes a connection to University College London, but the vacancy does not state joint registration or a joint degree. Basel's department provides computer laboratories, seminar and study spaces, and a library. The Spring 2026 computer-science course offer includes Professor Bogunovic's “Modern Reinforcement Learning: From Agents to LLMs” and a separate course in generative modeling and diffusion models. This indicates relevant departmental teaching, not a promise that a doctoral candidate will take those exact courses. The vacancy promises close supervision and mentoring, and names national and international research interactions, including access to collaborators at EPFL and ETH.

The University's general doctoral rules require a master's degree from a recognized institution, faculty approval, matriculation at the next possible date and continuous enrollment. A doctoral agreement with the first supervisor is due within the first six months. The precise Faculty of Science regulations, course-credit requirements and committee structure should be confirmed for this appointment. The project is a salaried research position rather than a named structured doctoral school in the advertisement.

The University of Basel's 2026 institutional ranking page lists ARWU 91, QS 158 and THE 120. These broad university rankings do not establish a computer-science subject position or the quality of this individual research group. The research area is specialized and fast-moving: it combines machine learning theory, reinforcement learning, generative modeling, benchmarking and large-scale systems work. The applicant should assess the group's compute and engineering capacity as carefully as its publication direction.

Supervisor and Lab Assessment

Professor Ilija Bogunovic joined Basel's DMI as professor for AI and foundation models in 2025. Rhine AI describes its program as research on reasoning, human-aligned intelligence, exploration, reinforcement learning, generative AI and agentic systems. Its recent papers show direct activity in diffusion LLM post-training. The 2025 wd1 paper addresses repeated likelihood estimation in policy optimization; GDSD studies likelihood-free guided self-distillation for RL; Temporal Self-Distillation investigates a speed-quality tradeoff across denoising timesteps. Together, these methods connect closely to the advertised topics of adaptive unmasking and credit assignment.

The main limitation is that the professor's Basel group is recent and the vacancy does not identify the day-to-day supervisor structure, group size, compute budget, or which post-training approaches are within the funded grant. The job says close supervision and mentoring are offered and mentions collaborations with national and international teams; ask for specifics about meetings, co-supervision, research-engineering support, access to models and GPUs, conference travel, and how the project will balance principled analysis with experiments.

Funding, Application, and Living Cost

The official vacancy specifies a 100% position funded by an SNSF project, a four-year PhD duration and a start from January 2027 or later. The advertisement describes salary, benefits and working conditions as excellent but gives no amount. The SNSF has salary-range rules for doctoral staff, but they do not determine the exact gross pay, percentage-specific salary, annual progression or net income for this post. Do not infer those from another Swiss university or from a generic funding regulation. Ask the University for the actual salary scale, level, yearly progression, social-insurance deductions, pension and contract term.

Doctoral candidates pay CHF 350 per semester in university fees. The University of Basel's 2025 student sample budget indicates about CHF 1,300–2,150 per month, depending on housing and personal circumstances. It is a planning estimate, not an employee budget; the stated insurance assumption includes under-25 students, and premiums vary. Without the salary figure and individual tax, insurance and housing assumptions, there is no reliable way to calculate take-home buying power.

The deadline field also requires care. The official employer page encourages applications before 20 November 2026 and states there is no fixed closing date, with recruitment continuing until a suitable candidate is found. A separate listing metadata record shows 2 December 2026 at 23:59 Europe/Zurich. Do not rely on the later date without confirmation. Apply promptly through the official employer link and ask whether 20 November is an internal review cutoff.

The employer asks candidates to submit two PDF files containing: a motivation statement of no more than 150 words; CV with publication list if applicable; bachelor's and master's transcripts; contact details for referees (not direct letters); and a recent thesis or publication if applicable. English fluency is required. The role does not state a nationality restriction, but international candidates should check recognized-degree equivalence and residence/work-permit steps. The University notes that immigration procedures for third-country employees can take weeks or months.

Applicant Fit and Preparation

The formal profile requires a master's degree in computer science, mathematics, physics or a related discipline, exceptional academic performance, strong mathematical foundations, coding and implementation skills, and the ability to conduct large-scale experiments. LLM, RL or generative-model experience is an advantage, but is not mandatory; previous research experience is also beneficial but not required.

A persuasive application should show more than general interest in AI. Useful evidence includes mathematical reasoning, well-documented research code, experience designing controlled experiments, model training or optimization, and a clear understanding of why diffusion-model likelihoods and token-generation schedules differ from autoregressive training. Applicants from adjacent disciplines should explain the equivalence of their master's training and include quantitative work that demonstrates implementation ability. Candidates with only prompt-engineering experience and little mathematics or research coding may find the advertised work less aligned with their background.

The position could prepare researchers for academic work in machine learning and generative modeling, research engineering, applied AI, or industrial model development. Swiss federal labor-market reporting describes AI vacancy growth across research, industry, healthcare and other areas; roughly 40% of reported AI vacancies were in data engineering and machine learning. The same source says permanent junior vacancies were around 5% and that Basel-Stadt advertisements increased 30% in the measured period. These are broad, time-bounded vacancy indicators, not a forecast for PhD graduates or specifically for reasoning-model researchers. The practical career value will depend on research quality, systems experience, publication record and the ability to move between theory and implementation.

Next Actions and Questions to Verify

Before Applying

• Prepare the two-PDF package, keeping the motivation statement within 150 words.
• Ask the employer whether 20 November is an internal review date and whether the separate 2 December metadata date is valid.
• Request the gross salary by year, pay grade, social-insurance/pension treatment, benefits, contract end date and confirmation that funding covers the full four years.
• Ask what models, benchmark suites, compute budget and engineering support are available.
• Confirm master's-degree equivalence, faculty admission steps, matriculation date and any work-permit timing.

At Interview

• Ask how the project will balance new RL objective design with adaptive unmasking, credit assignment and benchmark construction.
• Clarify the planned baselines and safeguards against benchmark contamination, reward hacking and sampler confounds.
• Ask about publication expectations, mentoring cadence, group size and opportunities to collaborate with EPFL/ETH or other partners.

Before Accepting an Offer

• Review the written employment contract and formal doctoral-admission decision separately.
• Confirm the salary, annual progression, employment percentage, tuition/semester fees and permitted teaching duties.
• Build a Basel budget using current housing, insurance and tax assumptions rather than relying only on the University's student sample budget.
• Confirm the start date and immigration/residence process for the candidate's nationality.

Evidence and Sources

Core Institutional and Programme Sources

• Official University of Basel PhD vacancy
• Rhine AI Lab
• Department of Mathematics and Computer Science
• Ilija Bogunovic profile
• Departmental announcement: Professor Ilija Bogunovic
• Computer Science Spring 2026 course offer
• Doctoral Studies: admission and matriculation
• Doctoral supervision and agreements
• University of Basel rankings
• University of Basel contact and directions

Supervisor and Research Sources

• wd1: weighted policy optimization
• GDSD: guided denoiser self-distillation
• Temporal Self-Distillation for diffusion language models
• Large Language Diffusion Models

Funding, Cost, International and Career Sources

• Doctoral tuition fees
• University of Basel sample living-cost budget
• SNSF doctoral salary rules
• Entry and work permits for international employees
• Swiss AI labor-market data