# arXiv paper proposes PINC-Flow for generating fMRI dynamics during unseen tasks [Canonical reading](https://escapemeridian.com/readings/arxiv-paper-proposes-pinc-flow-for-generating-fmri-dynamics-310831dd) Sam Gijsen, Michał Łukomski, Marc-André Schulz and Kerstin Ritter propose PINC-Flow, a per-timestep conditioned diffusion transformer for generating fMRI brain dynamics during cognitive tasks not seen in training. It uses compositional language and can also use spatial priors—information about brain regions—to guide generation. The authors argue that this addresses a limitation of earlier generative models of neural time series, which they say were largely restricted to categorical conditioning and therefore precluded compositional and zero-shot generalization. They report evaluating the model across held-out task conditions. From language alone, it recovers region-specific recruitment across tasks and held-out spatial activation patterns, according to the abstract. The authors say spatial priors complement language when it degrades, while retaining compositional task specification. They present the approach as a possible way to design and evaluate novel cognitive experiments in silico before empirical validation. The abstract does not give quantitative performance results, so the extent of the reported capabilities is not clear from this account. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv q-bio.NC · Article Original source date: Jun 10, 2026 (2026-06-10). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-09T06:39:02.602Z. First observed: 2026-11-05T05:00:21.124Z. Retrieved: 2026-11-09T06:32:30.183Z. - [Flow Matching with In-Context Priors for Out-of-Distribution Brain Dynamics](https://arxiv.org/abs/2606.11833) ## Limits The supplied evidence is the arXiv abstract, not the full paper; it does not provide quantitative performance results or details of the evaluation. The authors describe counterfactual neuroscience and experimental design as potential applications; the abstract does not establish that these uses have been validated in practice. ## Open question How does predictive performance vary across held-out task conditions, and where does language alone degrade enough for spatial priors to help? ## Source-stated The authors propose PINC-Flow, a per-timestep conditioned diffusion transformer, to generate fMRI brain dynamics for cognitive tasks not seen in training, using compositional language and optional spatial priors. > we propose PINC-Flow, a per-timestep conditioned diffusion transformer for generating realistic fMRI brain dynamics during unseen cognitive tasks by injecting both compositional language and optional spatial priors in-context. - [Supporting source](https://arxiv.org/abs/2606.11833) ## Source-stated The authors say that earlier generative models of neural time series were largely limited to categorical conditioning, which prevented compositional and zero-shot generalization. > generative models of neural time series have largely remained restricted to categorical conditioning, precluding compositional and zero-shot generalization. - [Supporting source](https://arxiv.org/abs/2606.11833) ## Source-stated The authors report that language alone recovers region-specific recruitment and held-out spatial activation patterns, and that spatial priors complement the text pathway where language alone degrades. > From language alone, the model recovers region-specific recruitment across tasks and held-out spatial activation patterns. Spatial priors, when available, complement the text pathway by anchoring generation in regions of task space where language alone degrades - [Supporting source](https://arxiv.org/abs/2606.11833) ## Source-stated The authors propose that zero-shot generation could support in-silico design and evaluation of novel cognitive experiments before empirical validation. > Such zero-shot generation could enable counterfactual neuroscience by supporting in-silico design and evaluation of novel cognitive experiments before empirical validation. - [Supporting source](https://arxiv.org/abs/2606.11833) ## Source and contributor context Gijsen, Sam, Łukomski, Michał, Schulz, Marc-André, Ritter, Kerstin · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Time-series Observations arranged over time, such as monthly sales; forecasting uses their patterns to estimate later values. - [Definition reference](https://developers.google.com/machine-learning/glossary#time-series-analysis) ## Zero-shot Doing a task without worked examples (zero-shot) or with only a few examples given in the prompt (few-shot). - [Definition reference](https://developers.google.com/machine-learning/glossary#zero-shot-prompting) ## Counterfactual An argument or scenario that considers what would have resulted if events had gone differently from how they actually did, such as if a policy had never been adopted; as an adjective, relating to this kind of thinking about what did not happen but could have. - [Definition reference](https://dictionary.cambridge.org/dictionary/english/counterfactual) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: c0bc7cb5-38d3-4545-8ba3-a73b1a98732b. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-09T06:39:02.602Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 7639ff6236db46528f3f2f68337d76c3f839fbbe3bc9e5affc9adbb4462da108. Source text hash: b803f6c1fde9bf7d6bcb76f631caf6212906370de88f289ef90b345168dcc62b. ## Revision and source-check history ### Version 1 — 2026-11-09T06:39:02.602Z Sam Gijsen, Michał Łukomski, Marc-André Schulz and Kerstin Ritter propose PINC-Flow, a per-timestep conditioned diffusion transformer for generating fMRI brain dynamics during cognitive tasks not seen in training. It uses compositional language and can also use spatial priors—information about brain regions—to guide generation. The authors argue that this addresses a limitation of earlier generative models of neural time series, which they say were largely restricted to categorical conditioning and therefore precluded compositional and zero-shot generalization. They report evaluating the model across held-out task conditions. From language alone, it recovers region-specific recruitment across tasks and held-out spatial activation patterns, according to the abstract. The authors say spatial priors complement language when it degrades, while retaining compositional task specification. They present the approach as a possible way to design and evaluate novel cognitive experiments in silico before empirical validation. The abstract does not give quantitative performance results, so the extent of the reported capabilities is not clear from this account. ## Original source arXiv q-bio.NC · Article Original source date: Jun 10, 2026 (2026-06-10). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-09T06:39:02.602Z. First observed: 2026-11-05T05:00:21.124Z. Retrieved: 2026-11-09T06:32:30.183Z. - [Flow Matching with In-Context Priors for Out-of-Distribution Brain Dynamics](https://arxiv.org/abs/2606.11833) ## Limits The supplied evidence is the arXiv abstract, not the full paper; it does not provide quantitative performance results or details of the evaluation. The authors describe counterfactual neuroscience and experimental design as potential applications; the abstract does not establish that these uses have been validated in practice. ## Open question How does predictive performance vary across held-out task conditions, and where does language alone degrade enough for spatial priors to help? ## Source-stated The authors propose PINC-Flow, a per-timestep conditioned diffusion transformer, to generate fMRI brain dynamics for cognitive tasks not seen in training, using compositional language and optional spatial priors. > we propose PINC-Flow, a per-timestep conditioned diffusion transformer for generating realistic fMRI brain dynamics during unseen cognitive tasks by injecting both compositional language and optional spatial priors in-context. - [Supporting source](https://arxiv.org/abs/2606.11833) ## Source-stated The authors say that earlier generative models of neural time series were largely limited to categorical conditioning, which prevented compositional and zero-shot generalization. > generative models of neural time series have largely remained restricted to categorical conditioning, precluding compositional and zero-shot generalization. - [Supporting source](https://arxiv.org/abs/2606.11833) ## Source-stated The authors report that language alone recovers region-specific recruitment and held-out spatial activation patterns, and that spatial priors complement the text pathway where language alone degrades. > From language alone, the model recovers region-specific recruitment across tasks and held-out spatial activation patterns. Spatial priors, when available, complement the text pathway by anchoring generation in regions of task space where language alone degrades - [Supporting source](https://arxiv.org/abs/2606.11833) ## Source-stated The authors propose that zero-shot generation could support in-silico design and evaluation of novel cognitive experiments before empirical validation. > Such zero-shot generation could enable counterfactual neuroscience by supporting in-silico design and evaluation of novel cognitive experiments before empirical validation. - [Supporting source](https://arxiv.org/abs/2606.11833) ## Source and contributor context Gijsen, Sam, Łukomski, Michał, Schulz, Marc-André, Ritter, Kerstin · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Time-series Observations arranged over time, such as monthly sales; forecasting uses their patterns to estimate later values. - [Definition reference](https://developers.google.com/machine-learning/glossary#time-series-analysis) ## Zero-shot Doing a task without worked examples (zero-shot) or with only a few examples given in the prompt (few-shot). - [Definition reference](https://developers.google.com/machine-learning/glossary#zero-shot-prompting) ## Counterfactual An argument or scenario that considers what would have resulted if events had gone differently from how they actually did, such as if a policy had never been adopted; as an adjective, relating to this kind of thinking about what did not happen but could have. - [Definition reference](https://dictionary.cambridge.org/dictionary/english/counterfactual) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: c0bc7cb5-38d3-4545-8ba3-a73b1a98732b. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-09T06:39:02.602Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 7639ff6236db46528f3f2f68337d76c3f839fbbe3bc9e5affc9adbb4462da108. Source text hash: b803f6c1fde9bf7d6bcb76f631caf6212906370de88f289ef90b345168dcc62b. ### passed — 2026-11-09T06:39:02.602Z — anthropic/claude-opus-5.5 Publication: c0bc7cb5-38d3-4545-8ba3-a73b1a98732b attribution — pass: The reading correctly attributes PINC-Flow to Sam Gijsen, Michał Łukomski, Marc-André Schulz and Kerstin Ritter. It frames the claims about prior models, results and applications as the authors' statements or the abstract's report, not as established fact. > Authors:Sam Gijsen, Michał Łukomski, Marc-André Schulz, Kerstin Ritter summary — pass: The title, summary and question faithfully convey the proposal: a model that generates fMRI brain dynamics for unseen tasks from compositional language plus optional spatial priors. They also cover the claimed limitation of categorical conditioning, the reported findings and the hedged experiment-design use. Phrases such as 'tasks not seen in training' and 'before empirical validation' carry the core meaning. Terms such as fMRI, categorical conditioning and in silico would benefit from term-library tooltips, but they do not block understanding. Omitting the 'first generative model' and 'hundreds of held-out task conditions' framing is acceptable condensation. date — pass: The originalPublished value of 2026-06-10 matches the v1 submission date. The 4 Nov 2026 revision and the November feed date are correctly not used as the original date. > \[Submitted on 10 Jun 2026 (v1), last revised 4 Nov 2026 (this version, v2)\] support — pass: Each claim matches its verbatim excerpt. The application claim keeps the source's 'could' hedge. The limitations correctly note that the abstract has no quantitative results and that counterfactual uses are not validated. The question follows from the abstract's own statement that language alone degrades in some regions of task space. The reading adds no unsupported benefits or costs. > Spatial priors, when available, complement the text pathway by anchoring generation in regions of task space where language alone degrades media — pass: The selected media is the verified direct link to the original arXiv abstract page, which is the source itself and the most direct way for readers to access it. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/310831dd841cbba9191f323a5e97bed84b60835010999745e466ea8a58dbc332/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/arxiv-paper-proposes-pinc-flow-for-generating-fmri-dynamics-310831dd.json)