CS PhD Roundup — September 28, 2026: deep-learning theory at TU Darmstadt, an egocentric-4D postdoc at Bocconi, and two calls closing September 30

The 21–28 September 2026 window produced two openings, both from labs with a recent top-conference record: a PhD in the mathematical foundations of deep learning at TU Darmstadt, posted on 24 September, and a two-year postdoc at Bocconi University in Milan that turns first-person video into training data for a humanoid robot, posted on 23 September. A funded PhD in Leipzig is open with a 15 October deadline, a VU Amsterdam postdoc advertised at the end of August stays open until 10 October, and two postdoc calls close on 30 September.

A role reaches the queue below only when its named supervisor has a recent paper at NeurIPS, ICML, ICLR, CVPR or ACL. Four leads that failed that test, or fell outside the window, sit in their own section at the end.

Quick triage

Institution	Position & unit	Key adviser(s)	Direction	Deadline	Direct application
TU Darmstadt (Germany)	PhD, 3 years, full time, Signal Theory and Statistical Learning / new Emmy Noether group	Mariia Seleznova; chair led by Felix Krahmer	Mathematical foundations of deep learning: random matrix theory, stability, generalisation, scaling laws	22 Oct 2026	TU Darmstadt job T416-2026
Bocconi University, Milan (Italy)	Post-doctoral research contract, 2 years, Department of Computing Sciences (project E4RL)	Chiara Plizzari, with Francis Engelmann and Antonio Loquercio	Egocentric 4D scene representations from first-person video, used to train humanoid robot control	23 Nov 2026	Bocconi job market
Leipzig University (Germany)	PhD, 3 years, TV-L E13, LION Lab	Leonie Weissweiler, with Clara Schoeder	Mechanistic interpretability of protein language models	15 Oct 2026, then open until filled	LION Lab posting
VU Amsterdam (Netherlands)	Postdoc, 2 years, 1 FTE, Learning and Reasoning	Filip Ilievski, Emile van Krieken, Cees Snoek	Neurosymbolic reasoning and abstraction in multimodal world models	10 Oct 2026	VU vacancy

New this week

Mariia Seleznova, TU Darmstadt

Fit. The position sits in the Signal Theory and Statistical Learning group and joins a newly founded Emmy Noether research group, funded by the German Research Foundation's project RMT4DL, Random Matrix Theory for Deep Learning. 1 Two lines of work are named: methods from probability theory and random matrix theory that analyse the stability, generalisation and feature learning of networks with very many parameters, and empirically informed theory that formalises observed phenomena such as scaling laws. 1

Advisor evidence. Mariia Seleznova leads the new group. She wrote her PhD at LMU Munich under Gitta Kutyniok and stayed in the same chair as a postdoc, and she holds a guest position in the Data Science group at the Max Planck Institute of Quantum Optics. 2 Her NeurIPS 2023 paper Neural (Tangent Kernel) Collapse, written with Dana Weitzner, Raja Giryes, Gitta Kutyniok and Hung-Hsu Chou, is the venue check for this position. 3

Role and application. The contract runs three years at full time, paid at EG 13 of the TU Darmstadt collective agreement, and starts as soon as possible. The reference is T416-2026 and applications close on 22 October 2026. 1 Applicants need a master's degree in mathematics, computer science or a closely related field, with strong grounding in machine learning theory, probability theory, optimisation and linear algebra; familiarity with random matrix theory helps, and so does Python with PyTorch or JAX. 1 The dossier is a CV, a motivation letter, bachelor's and master's transcripts and diplomas, and the names and contact details of one or two academic referees. 1 Questions go to mariia.seleznova@tu-darmstadt.de. 1

Lab signal. The group opens this year, so the first doctoral researcher joins a team of a few people and works directly with its leader. The advert lists 30 days of annual leave on a full-time contract, a Hessen regional transport pass, and mobile working. 1

Chiara Plizzari, Bocconi University

Fit. The E4RL project turns recordings of human activity into material for training robots, which reduces the reliance on hand-built simulations and on data collected by driving robots around. 4 The project converts first-person video into egocentric 4D representations: 3D reconstructions of a scene and their evolution over time. It recovers the structure of the environment, the objects in it, and the person's pose and motion, so the representation records where things are, how they are used, and what part they play in finishing a task. 4 Those representations then train control policies that run on a physical humanoid robot using only its onboard sensors and computing. 4

Advisor evidence. Chiara Plizzari is an assistant professor in Bocconi's Department of Computing Sciences. She was a postdoc at the Smart Eyewear Lab run by EssilorLuxottica and the Politecnico di Milano, and a student researcher in Federico Tombari's team at Google Research in Zurich. 5 Her E2(GO)MOTION paper on motion-augmented event streams for egocentric action recognition appeared at CVPR 2022, and work she did at Google was accepted to CVPR 2025. 5

Role and application. The call is a post-doctoral research contract in information processing systems, group 09/IINF-05, at the Department of Computing Sciences, posted on 23 September 2026 with a closing date of 23 November 2026. 4 Plizzari's own announcement describes one fully funded two-year position, with the deadline at 23:59 on 23 November. 6 Applications go through the university's job market portal, which carries the call text; the listing publishes no salary figure. 4

Lab signal. Not published on the call page.

Open now, no publication date published

The Leipzig advert below carries a deadline and no publication date, so it runs here as a live call rather than as this week's news.

Leonie Weissweiler and Clara Schoeder, Leipzig University

Fit. The position belongs to a German Research Foundation priority programme called Artificial Intelligence for Protein Function. It adapts mechanistic interpretability methods built for language models to protein language models, which are trained on billions of protein sequences and acquire knowledge of protein structure without supervision. 7 The computational half of the project tests what a protein language model knows about toxicity and immunogenicity, turns those findings into a classifier, and then uses the classifier to steer protein generation towards sequences that provoke a smaller immune response. A second PhD student in the Schoeder Lab handles the wet-lab side and establishes immunogenic profiles for the generated proteins. 7

Advisor evidence. Leonie Weissweiler leads the LION Lab as an assistant professor of natural language processing at Leipzig University. Her papers at the venue this channel checks include A Crosslingual Investigation of Conceptualization in 1335 Languages at ACL 2023 and CaMEL: Case Marker Extraction without Labels at ACL 2022. 8

Role and application. The post is fully funded for three years at TV-L E13, full time, with a preferred start of 1 January 2027 and a later start possible. Applications close on 15 October 2026, after which the position stays open until it is filled. 7 Candidates need a master's degree in computational linguistics, bioinformatics, computer science, data science or a related field, strong mathematics and programming, and strong written and spoken English. 7 Applications arrive by email to leonie.weissweiler@uni-leipzig.de with the header [PROTEIN-PHD], and the dossier is a CV, transcripts, a copy or draft of the master's thesis, and a one-page research proposal that sets out the applicant's own idea inside the project rather than a plan for the whole project. 7

Lab signal. There are no formal teaching obligations, and teaching is possible through tutorials in the lab's undergraduate NLP and machine learning lectures. The project funds a research stay abroad of several months for each of the two PhD students, and the lab is a new group inside a university network that includes Max Planck, Helmholtz, Fraunhofer and Leibniz institutes. 7

Advertised earlier, still open

Filip Ilievski, Emile van Krieken and Cees Snoek, VU Amsterdam

The advert was published on 31 August 2026 and closes on 10 October, so it stays live while the two roles above are newer. 9

Fit. Multimodal world models have sharpened their predictions without capturing the commonsense and causal structure of the physical and social world. The postdoc learns explicit, effective abstractions from limited human-labelled data and treats them as dynamic neurosymbolic world models, evaluated on tasks that need commonsense reasoning over space, time and causality. 9

Advisor evidence. The project is funded by the NWO AiNed talent programme project Human-centric AI with common sense, led by Filip Ilievski in the Learning and Reasoning group at VU Amsterdam. The postdoc is co-supervised by Emile van Krieken in the same group and by Cees Snoek, who directs the Video & Image Sense Lab at the University of Amsterdam. 9 Snoek's university profile lists current CVPR papers from 2025 and 2023, which is the venue check for this role. 10

Role and application. The appointment is 1 FTE for two years and preferably starts in the first quarter of 2027. Salary runs from €3,706 to €5,760 gross per month on scale 10, with an 8% holiday allowance and an 8.3% end-of-year bonus, and the advert also points international staff to relocation support. Applications close on 10 October 2026 and consist of a CV and a cover letter through the VU portal; questions go to Filip Ilievski. 9

Lab signal. The department has around 250 members, 50 tenured staff and 40 to 50 PhD students. The duties include supervising bachelor's and master's students and co-supervising PhD students on multimodal abstraction and world models, alongside teaching in lectures or lab courses. The advert states that where two candidates are equally qualified the university will select a female candidate. 9

Closing on 30 September

Two postdoc calls close on Wednesday. The University of Copenhagen's CopeNLU postdoc in the mechanistic understanding of AI reasoning, supervised by Isabelle Augenstein and Pepa Atanasova, runs two years from 1 February 2027 at DKK 39,900 a month plus an 18.07% pension contribution, and the deadline is 23:59 CET on 30 September. 11 Augenstein wrote on 28 September that the call closes in two days. 12 Its dossier is the longest in this list: a cover letter, a research plan, a CV, diplomas, a full publication list, three paper reprints and three referees. 11

Imperial College London closes the same day for a Research Associate on the AToM ERC grant, job number ENG04037, working with Edoardo M. Ponti on latent tokenization, sparse attention and permanent memory. The post runs two years at £50,733 to £59,484, or £45,399 to £48,876 if the PhD is not yet awarded. 13 The statement to write is two pages aimed at the project's directions rather than at the applicant's own thesis. 14

Screened out this week

KU Leuven: PhD in virtual knowledge graphs, edge computing and artificial intelligence. Posted on 28 September with a 30 October deadline, the position sits in the Stream Intelligence Lab under Pieter Bonte and builds cognitive digital twins and virtual knowledge graphs for edge and stream reasoning. 15 Bonte's record sits in the semantic web and stream-processing literature, in the Semantic Web journal and at ESWC, rather than at the five venues this channel checks.
University of Sheffield: research associate in computer vision and environmental modelling. Placed on 22 September with an 18 October deadline at £38,784 to £41,064, the role develops open-set recognition and uncertainty-aware methods for tree species mapping from UAV imagery in the Computer Vision Lab's PRISMA team. 16 The advert names no individual supervisor, so there is no advisor record to check.
ETH Zürich: an incoming robotics group, watch-only. Carlo Sferrazza announced on 24 September that he returns to ETH Zurich as an assistant professor of robotics and artificial intelligence next summer and will recruit PhD students for a lab on general-purpose embodied intelligence for humanoid robots and dexterous manipulation, with details to follow. 17 No advert, deadline or application route exists yet, and his publication record sits in robotics venues such as RSS rather than at the five this channel checks. 18
Université Laval: PhD in Gaussian-based uncertainty-aware modelling for multi-robot intelligence. The university's graduate research board lists the project as offer 492319, a four-year PhD in the Vision and Digital Systems Laboratory supervised by Hanqing Zhao and co-supervised by Jean-François Lalonde, whose CVPR 2024 and CVPR 2023 papers are on his Laval page. 19 20 The announcement of the position on LinkedIn is dated seven months ago, and the project board carries no publication date and no closing date, which puts the call outside this week's window with its current status unconfirmed. 21

Deadline watch

30 September — University of Copenhagen, CopeNLU postdoc. Ask referees and reprint permissions now: the dossier needs three papers and three referees, and interviews are expected in the week of 19 October. 11
30 September — Imperial College London, AToM research associate. Write the two-page statement against the AToM directions, and submit before the closing date because Imperial can close the advert early if applications run high. 13
10 October — VU Amsterdam, neurosymbolic reasoning postdoc. A CV and a cover letter through the portal, with questions to Filip Ilievski. 9
15 October — Leipzig University, LION Lab PhD. The one-page research proposal is the part that carries the application, and the call stays open until filled after that date. 7
22 October — TU Darmstadt, deep-learning theory PhD. The dossier is a CV, a motivation letter, transcripts and one or two referees, and the contract can start immediately. 1
1 November — Aarhus University, Physical AI PhD. Carried from the 21 September issue; the A3 Lab's robotics PhD under Behzad Bozorgtabar stays open, with a preferred start of 1 January 2027. 22
23 November — Bocconi University, E4RL postdoc. Two funded years on egocentric 4D representations for humanoid control. 4