CS PhD Roundup — August 3, 2026: 8 active leads, no new ad cleared the author filter

What is actionable now

No newly posted opening in the July 27–August 3, 2026 window cleared both parts of this channel's filter: a named advisor and a verifiable NeurIPS, ICML, ICLR, CVPR, or ACL publication from 2022–2025. The useful result is a tighter carry-forward queue: eight active leads, with Oxford closing first and one Inria role kept watch-only because its posting date is not public.

Several public adverts dated this week were left out rather than presented as matches. The Oxford senior postdoc in statistical machine learning names the research topic and deadline but no advisor; the Cambridge LLM-safety studentship gives a research plan and funding but no supervisor on the public page; the Earlham biological-foundation-model advert names Ke Li but did not yield a target-venue publication record in the retrieved sources; and the ESA agentic-AI fellowship names an office rather than an individual supervisor. 1 2 3 4

Quick triage

Direction	Advisor and institution	Position	Deadline or status	First action
Safe, interpretable ML and LLMs	Philip Torr, University of Oxford	Postdoc	Aug 6, 06:00 in this channel's time zone (noon UK)	Submit the statement, CV, and two referees before the portal closes
LLM training and reasoning	Aurelien Lucchi and Ilija Bogunovic, University of Basel	PhD or postdoc	Page shows Aug 18, 16:59 in this channel's time zone, but its practical July 31 target has passed and the text also says rolling	Email the department to confirm which deadline is live
Mechanism-informed world models	Efstratios Gavves, University of Amsterdam	PhD	Aug 20	Prepare a month-by-month CV and two references
Physics-informed world models	Efstratios Gavves, University of Amsterdam	Postdoc	Aug 20	Lead with physics, simulation, or robot-learning evidence
Biological foundation models	Charlotte Bunne, EPFL AIMM	Two postdocs	Rolling; Fall 2026 preferred	Send the one-page research statement and two references
Neuromorphic reinforcement learning	Melih Kandemir, University of Southern Denmark	Postdoc	Aug 31, 16:59 in this channel's time zone (23:59 Denmark)	Assemble the publication list and PhD proof
Trustworthy and neuro-symbolic ML	Christoph Lampert, ISTA	Multiple postdocs	Rolling until filled	Send one PDF directly to the PI
Text-to-video generation	Stéphane Lathuilière, Inria Grenoble	Postdoc, watch-only	Aug 20; posting date not exposed	Confirm the date caveat before relying on it

Safe ML, LLMs, and optimization

University of Oxford: Torr Vision postdoc

Fit. Philip Torr's publication page lists Prompting a Pretrained Transformer Can Be a Universal Approximator at ICML 2024, When Do Prompting and Prefix-Tuning Work? A Theory of Capabilities and Limitations at ICLR 2024, and Universal In-Context Approximation By Prompting Fully Recurrent Models at NeurIPS 2024. The strongest match is an applicant working on LLMs, retrieval-augmented generation, agentic systems, continual learning, interpretability, or AI safety. 5

Role and application. The one-year postdoc sits in the Torr Vision Group and focuses on interpretability, safety, and alignment for LLMs and related generative architectures. The advert also names long-term-history retrieval, noisy multimodal observations, and empirical work with policymakers on AI regulation and governance. It was placed July 23 and closes August 6 at noon UK time, which is 06:00 in this channel's UTC-05:00 display. The advert's header says "permanent," while its body says fixed-term for one year; clarify the contract before accepting an offer. Salary is £39,424–£47,779. Applicants need a PhD or near-completed PhD in machine learning, a supporting statement with research interests, a CV, and two referees. 6

Lab signal. The policy-facing collaboration and the Department of Engineering's Athena Swan Bronze award are explicit signals in the advert. No lab-size figure or independent student feedback is published. Apply through the Oxford recruitment advert.

University of Basel: LLM training, reasoning, and adaptation

Fit. The opening is co-led by Aurelien Lucchi and Ilija Bogunovic. Lucchi's OMLS group lists Adaptive Methods through the Lens of SDEs at ICLR 2025 and Enhancing Optimizer Stability: Momentum Adaptation of The NGN Step-size at NeurIPS 2025. Bogunovic's page lists Group Robust Preference Optimization in Reward-free RLHF at NeurIPS 2024 and corruption-tolerant Gaussian-process bandit work at NeurIPS 2022. That combination points to LLM post-training, optimization, reasoning, reinforcement learning, and robust decision-making. 7 8

Role and application. The University of Basel accepts either a PhD candidate or a postdoc for work on LLM pre-training and post-training, mathematical reasoning, decision-making, domain adaptation, fine-tuning, and reinforcement learning. The PhD is funded for four years; the postdoc starts at two years with a possible extension to four. The expected start is around October 2026. The application uses two PDFs: a motivation statement of no more than 150 words, a CV and publications where applicable, degree transcripts, recommender contact details, and a thesis or publication where applicable. 9

Deadline caveat. The page was published July 20. Its body encouraged applications before July 31, while the job metadata shows August 18 at 23:59 Europe/Zurich, equivalent to August 18 at 16:59 in this channel's display, and also says there is no fixed deadline. Treat July 31 as the missed practical target, not as proof that the role is closed; confirm the live status with Basel before investing in the application. The advert promises close supervision, collaboration with EPFL and ETH, conference participation, and a diverse working environment. Open the Basel application page.

World models and embodied AI

University of Amsterdam: mechanism-informed world-model PhD

Fit. Efstratios Gavves's public profile lists Dream to Manipulate: Compositional World Models Empowering Robot Imitation Learning with Imagination and CaPo: Cooperative Plan Optimization for Efficient Embodied Multi-Agent Cooperation at ICLR 2025, plus Mechanistic Neural Networks for Scientific Machine Learning at ICML 2024. The match is strongest for applicants working on causal or representation learning, world models, generative modeling, robotics, or dynamical systems. 10

Role and application. This four-year PhD asks how models can represent mechanisms, events, causes, and temporal relationships rather than only predict the next frame. The vacancy closes August 20 and pays €3,059–€3,881 gross per month, plus an 8% holiday allowance and an 8.3% year-end allowance. The start is as soon as possible or by discussion. Applicants need an MSc in AI, computer science, electrical engineering, physics, mathematics, or a related field; solid ML and programming skills; a detailed CV with months for education and work; a motivation letter; a publication or project list; and two referees. 11

Lab signal. CyPhAI is part of VISLab and the UvA Informatics Institute. The advert names substantial GPU compute, real robots, biomedical and physical-world data, and collaborations with Toyota Research, Qualcomm, Elekta, and the Netherlands Cancer Institute. It also expects publication at venues including CVPR, ICLR, ICML, and NeurIPS, plus teaching support. Apply through the UvA PhD vacancy.

University of Amsterdam: physics-informed world-model postdoc

Fit. This role is also with Gavves and CyPhAI. It suits a researcher with a PhD or near-completed PhD and a record in machine learning, computer vision, robotics, or scientific ML, especially with physics-informed learning, simulation, dynamical systems, robot learning, or world models. Gavves's ICLR 2025, ICML 2024, ICML 2023, NeurIPS 2023, and ICML/ICLR 2022 work provides the relevant advisor signal. 10

Role and application. The 21-month postdoc studies physics-informed and physics-grounded world models, including partial observation, faster simulation, cross-embodiment transfer, and agents acting inside learned models. It closes August 20 and pays €3,546–€5,538 gross per month, excluding the same holiday and year-end allowances. The preferred start is as soon as possible or by discussion. Submit a detailed CV, motivation letter, publication or project list, and two references. The role also expects the postdoc to guide PhD and MSc students. 12

Lab signal. The advert again names GPU compute, real robots, biomedical and physical-world data, Toyota Research, Qualcomm's QUVA 2.0, Elekta, the Netherlands Cancer Institute, and the ELLIS network. A knowledge-security check may be part of selection. Open the UvA postdoc vacancy.

Biology and reinforcement learning

EPFL AIMM: two biological foundation-model postdocs

Fit. Charlotte Bunne is a tenure-track assistant professor at EPFL and leads the Artificial Intelligence in Molecular Medicine group. Her CV lists Modeling Complex System Dynamics with Flow Matching Across Time and Conditions at ICLR 2025 and Supervised Training of Conditional Monge Maps at NeurIPS 2022. This is a good match for applicants who can combine large-scale ML with biological data, multimodal learning, generative models, or AI for science. 13

Role and application. EPFL is recruiting two postdocs for biological foundation-model work spanning model architecture, multimodal learning, agentic interfaces, and translation toward drug discovery and biomedicine. A PhD in ML, computer science, computational biology, or a related field is expected, along with large-scale pretraining or fine-tuning, biological experience, and a strong publication record. Review is rolling; Fall 2026 is preferred, with later starts possible. Send a CV, a one-page research-interest statement, and two references through the EPFL careers page. The page gives a competitive EPFL salary and full social benefits rather than a figure. 14

Lab signal. AIMM sits between computer science and life sciences, with pharmaceutical collaborations, an international and collaborative environment, mentorship toward independent research, and dedicated compute through CSCS and the Swiss AI Initiative. These are statements from the vacancy, not independent student feedback. 14

University of Southern Denmark: neuromorphic reinforcement learning postdoc

Fit. Melih Kandemir's publication page lists Deterministic Uncertainty Propagation for Improved Model-Based Offline Reinforcement Learning at NeurIPS 2024, Learning Interacting Dynamical Systems with Latent Gaussian Process ODEs at NeurIPS 2022, and Evidential Turing Processes at ICLR 2022. The role fits an RL researcher who can also handle probabilistic modeling, uncertainty, continuous control, or spiking neural networks. 15

Role and application. The two-year ADIN Lab postdoc develops sample- and energy-efficient reinforcement learning for high-dimensional dynamical systems, combining deep RL, probabilistic modeling, and spiking neural networks. It starts no earlier than October 1. The local deadline is August 31 at 23:59 in Denmark, or August 31 at 16:59 in this channel's display. Applicants need a PhD in computer science, mathematics, statistics, or theoretical physics, at least two first-author papers at flagship ML venues, strong RL foundations, and Python with PyTorch or JAX. Submit a CV, full publication list, and PhD diploma or supervisor statement as PDFs through the SDU application page. Salary follows the Danish collective agreement; no figure is published. 16

Lab signal. ADIN sits in IMADA's Data Science and Statistics section, where mathematicians and computer scientists work together. The advert expects collaboration toward top-tier publications and standard teaching-assistant duties. 16

Trustworthy ML and vision generation

ISTA: Lampert group postdocs

Fit. Christoph Lampert's group lists recent work at the target venues, including Federated Learning with Unlabeled Clients: Personalization Can Happen in Low Dimensions and ASIDE: Architectural Separation of Instructions and Data in Language Models at ICLR 2025, Banded Square Root Matrix Factorization for Differentially Private Model Training at NeurIPS 2024, and 1-Lipschitz Layers Compared: Memory, Speed, and Certifiable Robustness at CVPR 2024. 17

Role and application. Multiple one- or two-year postdocs are available with flexible starts and no fixed deadline. Topics include privacy, fairness, robustness, federated and continual learning, LLM security, transfer learning, and the interface between symbolic and statistical ML. Applicants should hold a PhD in ML or a related area, with strong mathematics, programming, analytical reasoning, and first-tier publications. Send one PDF containing a CV, research statement, publication list, transcripts, and three references to Prof. Christoph Lampert at chl(at)ist.ac.at. The page says applications remain open until the positions are filled. 18

Lab signal. This is explicitly 100% research, with no mandatory teaching or administrative duties. The page also names state-of-the-art compute, equipment and conference-travel funding, an English-language institute, and optional supported teaching at PhD level. 18

Inria Grenoble: text-to-video postdoc, watch-only

Fit. Stéphane Lathuilière is a research scientist in Inria Grenoble's ROBOTLEARN team. DBLP lists his Collaborating Foundation Models for Domain Generalized Semantic Segmentation at CVPR 2024, An eye for an ear at NeurIPS 2024, and Zero-shot spatial layout conditioning for text-to-image diffusion models at CVPR 2023. 19

Role and application. The two-year postdoc focuses on editing and conditional generation with text-to-video models, including diffusion and transformer architectures, video editing, temporal consistency, semantic fidelity, efficient training, and reinforcement-learning post-training. It starts October 1, closes August 20, and pays €2,788 gross per month. A PhD or equivalent in speech processing, computer vision, or machine learning is required, with solid programming skills; LLMs, SpeechLMs, RL, or robotics are useful but not mandatory. Applications must go through the Inria online route. 20

Why watch-only. The official page does not expose a posting date. The role is actionable on its own terms, but it cannot be counted as a newly posted opening for this week's date-bounded roundup. Confirm the vacancy is still live before applying.

Deadline watch

August 6: Oxford Torr Vision postdoc, 06:00 in this channel's UTC-05:00 display time.
August 18: Basel page metadata shows 16:59 in this channel's display time, but the advert also carries a missed July 31 practical target and rolling language. Confirm before applying.
August 20: Both UvA world-model roles. The Inria text-to-video role also closes then, but remains watch-only.
August 31: SDU neuromorphic-RL postdoc, 16:59 in this channel's display time for the stated 23:59 Danish deadline.
Rolling: EPFL AIMM biological foundation-model postdocs and ISTA Lampert group postdocs.

The practical order is straightforward: finish Oxford first, ask Basel which deadline is operative, then choose between the two UvA roles by whether your evidence is closer to a PhD proposal or an independent postdoc record. The remaining options are rolling or leave more time, but SDU's two-first-author-paper requirement is a hard screening condition rather than a detail to postpone.