---
id: "20261014-0229-bi-fork-zimmel-et"
title: "Bi-FORK (Zimmel et al., arXiv:2610.12449) has hit the single-source concentration cap — what corroborates its Beam3D runtime and rejection-rate findings?"
type: "capture"
status: "promoted"
promoted_to: ["30-notes/claim-bi-fork-stflow-baseline-shares-author-reduces-independence.md","40-entities/entity-vlado-menkovski.md","40-entities/entity-johannes-brandstetter.md","50-questions/question-corroborate-zimmel-2026-bi-fork-2610-12449-single-source-cap.md (progress log appended, not closed)"]
not_promoted: ["Claim 1 (no independent corroboration found yet) — the capture's own research finding IS the open question's unresolved status; routed into the question's progress log instead of a 4th/5th claim-note resting on the capped Bi-FORK preprint, per Cali's 2026-10-14 review flag (5).","Claim 3 (GeoTDM's paper corroborates the mechanisms Bi-FORK cites for GeoTDM's cost) — dropped. Premise is false per Cali's 2026-10-14 review (1): Bi-FORK's own text (sha 4bd190de…, l. 474) attributes GeoTDM's cost to being 'GNN-based, were already slow to train and sample from,' not 'diffusion-style multi-step sampling.' GeoTDM corroborating a mechanism Bi-FORK never cites is not corroboration of Bi-FORK. Correction folded into the question's progress log instead of a claim-note.","Further leads (github.com/ml-jku/Bi-FORK repo check, the metamaterials dataset paper, the predecessor flow-matching paper, deeplearn.org trailhead, the Eindhoven author-cluster mapping) — not claims; the repo check folded into the question's progress log (no sha, informal), the rest left as unread leads.","Entity candidates not promoted to hub (co-authorship alone is a mention, per the 2026-10-01 ruling): Stefano Ermon, Fleur Hendriks (see 00-meta/seek-flags.md 2026-10-15 [entity] — candidate once her dataset paper is read directly), Anna Zimmel, Jiaqi Han, Minkai Xu, Aaron Lou, Haotian Ye, Kiet Bennema ten Brinke, Koen Minartz — left as mentions. 'Single-source concentration cap' (concept) — already documented in sources.md, not a domain entity; no page created. Particle guidance, deflation (concept) — already covered by existing claim-notes, no new work needed."]
origin: "batch"
writer_model: "claude-sonnet-5"
date_created: "2026-10-14T00:00:00.000Z"
provenance: "Batch research run, 2026-10-14, targeting question-corroborate-zimmel-2026-bi-fork-2610-12449-single-source-cap"
derived_from: []
verifies: "question-corroborate-zimmel-2026-bi-fork-2610-12449-single-source-cap"
tags: ["single-source-cap","corroboration","physics-ml","bifurcation-theory","generative-modeling","unverified-quant","arxiv"]
seek_code_commit: "d40ee39"
---


Research task: find independent corroboration for two held findings from Zimmel, Hendriks,
Holzleitner, Sestak, Weichselbaumer, Menkovski & Brandstetter, "Bi-FORK: Generative Modeling
of High-Dimensional Bifurcating Systems" (arXiv:2610.12449, submitted 2026-10-08) — the
Beam3D runtime comparison (Bi-FORK ~18× faster than [[claim-bi-fork-particle-guidance-sampler-is-parallelized-deflation-analog|STFlow]] and ~1,966× faster than GeoTDM) and the Beam3D rejection-rate comparison (Bi-FORK 0% vs. GeoTDM 99%). Three claim-notes already rest on this single preprint ([[claim-bi-fork-whole-trajectory-latent-flow-matching-for-global-branch-consistency]], [[claim-bi-fork-260k-point-scale-is-first-demonstration-of-one-to-many-bifurcation-recovery]], [[claim-bi-fork-particle-guidance-sampler-is-parallelized-deflation-analog]]), triggering the single-source concentration cap; see [[question-corroborate-zimmel-2026-bi-fork-2610-12449-single-source-cap]].

## Claim: No independent source has yet corroborated or contradicted Bi-FORK's Beam3D benchmark figures

verifies: question-corroborate-zimmel-2026-bi-fork-2610-12449-single-source-cap

A targeted search (web search for citations of arXiv:2610.12449 and of "Bi-FORK" by name; search for a reachable GitHub reproduction) found no independent author group engaging with, replicating, or citing Bi-FORK's reported Beam3D numbers as of 2026-10-14. The preprint was submitted 2026-10-08 — six days before this search — which is too recent to expect citing papers or peer review to have landed. The paper's own reproducibility statement points to a public code repository, https://github.com/ml-jku/Bi-FORK, confirmed to exist and to contain a README describing training/evaluation scripts for Beam3D among other datasets, but as of this check it shows only a single commit and no evidence of outside reproduction (no forks beyond the repository's own history, no independent stars pattern suggesting external use). This is a lead for a later check, not corroboration.

This claim directly answers the core question in the negative-but-honest sense the question anticipated: **the specific runtime multipliers (18×, 1,966×) and the specific rejection-rate figures (0% vs. 99%) remain [unverified-quant — needs primary] beyond Bi-FORK's own preprint.** [source: arXiv:2610.12449, Tier 1, absence of corroborating literature established by search rather than a document to quote]

## Claim: One of Bi-FORK's two Beam3D baselines, STFlow, is not an independent comparison — it shares an author and a department with Bi-FORK itself

verifies: question-corroborate-zimmel-2026-bi-fork-2610-12449-single-source-cap

Bi-FORK's Beam3D benchmark compares against two baselines: GeoTDM and STFlow. STFlow is "Kiet Bennema ten Brinke, Koen Minartz, and Vlado Menkovski. Stflow: Data-coupled flow matching for geometric trajectory simulation. In International conference on machine learning. PMLR, 2026" (Bi-FORK's own reference [8]), published as Bennema ten Brinke, Minartz & Menkovski, "STFlow: Data-Coupled Flow Matching for Geometric Trajectory Simulation," arXiv:2505.18647 / PMLR vol. 306 (ICML 2026). Its author line reads "Kiet Bennema ten Brinke 1 Koen Minartz 1 Vlado Menkovski 1" with affiliation "Machine Learning for Physical Sciences (ML4Sci/e) Group, Department of Mathematics & Computer Science, Eindhoven University of Technology."

Bi-FORK's own author list includes "Vlado Menkovski† 4 Johannes Brandstetter† 1 5" with affiliation 4 given as "Department of Mathematics and Computer Science, Eindhoven University of Technology" — the same person, the same university department, as a shared senior/corresponding author across the paper under evaluation and one of its two comparison baselines.

This does not mean the Beam3D numbers are fabricated or even biased — STFlow is itself a peer-reviewed, independently published method the Bi-FORK authors did not design — but it means the "independent baseline" framing is only half-independent: GeoTDM (Han, Xu, Lou, Ye & Ermon, Stanford) shares no authors with Bi-FORK, while STFlow was co-authored by one of Bi-FORK's own senior authors. A reader treating "beats two independent baselines" as two equally arm's-length data points would be overstating the independence of one of them. [source_url: https://arxiv.org/pdf/2505.18647, source_title: "STFlow: Data-Coupled Flow Matching for Geometric Trajectory Simulation", source_author: "Kiet Bennema ten Brinke, Koen Minartz, Vlado Menkovski", source_date: 2026-05-27 (v3), source_venue: "arXiv:2505.18647 [cs.LG] / PMLR vol. 306, ICML 2026", source_tier: 1, source_quote: "Kiet Bennema ten Brinke 1 Koen Minartz 1 Vlado Menkovski 1", source_quote_2: "Machine Learning for Physical Sciences (ML4Sci/e) Group, Department of Mathematics & Computer Science, Eindhoven University of Technology", source_sha: 9b75706f09235a08e0e8b489d58db20fa540a6f2a89b23d91925b6d8a7578539; cross-checked against arXiv:2610.12449, source_quote: "Vlado Menkovski† 4 Johannes Brandstetter† 1 5", source_quote_2: "Department of Mathematics and Computer Science, Eindhoven University of Technology", source_sha: 4bd190dea95f4aaefd0d4832bd08b7d812926e48d40f8a7176a551de9c79df2e]

> [!note] Seek's commentary:
> This is the shape corroboration-hunting should have: not "did anyone confirm the number" (nobody has, yet) but "how much is the number already leaning on itself." A within-lab baseline isn't disqualifying — labs reasonably compare against their own prior work, and STFlow went through its own separate ICML review — but it quietly halves the independence of "beats two baselines" into "beats one external baseline and one internal one." Worth flagging precisely because it's the kind of thing a reader skimming for "18× faster than STFlow, 1,966× faster than GeoTDM" would never think to ask about.

## Claim: GeoTDM's own paper independently confirms the two mechanisms Bi-FORK cites to explain GeoTDM's Beam3D behavior, without confirming the specific benchmark figures

verifies: question-corroborate-zimmel-2026-bi-fork-2610-12449-single-source-cap

Bi-FORK attributes GeoTDM's Beam3D results to two architectural properties: that GeoTDM is slow because it requires diffusion-style multi-step sampling, and that it achieves comparatively low JSD "due to its built-in equivariance." GeoTDM's own paper — Han, Xu, Lou, Ye & Ermon, "Geometric Trajectory Diffusion Models," arXiv:2410.13027, NeurIPS 2024, a fully independent author group (Stanford University, no shared authors with Bi-FORK or STFlow) — corroborates both properties as general mechanisms in its own words: "Akin to other diffusion models, GeoTDM resorts to multi-step sampling which may require more compute," stated in its own Limitations section, and that it "develop[s] a novel transition kernel leveraging SE(3)-equivariant spatial convolution and temporal attention," describing its EGTN denoising network.

This corroborates, at Tier 1 and from a genuinely independent venue, that the *mechanisms* behind Bi-FORK's characterization of GeoTDM are real and self-acknowledged by GeoTDM's own authors — not that Bi-FORK's specific Beam3D numbers (1,966× slower; 99% rejection rate) are independently confirmed, since GeoTDM's own paper does not evaluate the Beam3D benchmark or report a rejection-rate metric at all; Beam3D is Bi-FORK's own benchmark construction. The mechanism claim clears the Tier 1-2 floor; the magnitude remains [unverified-quant — needs primary] per the first claim above. [source_url: https://arxiv.org/pdf/2410.13027, source_title: "Geometric Trajectory Diffusion Models", source_author: "Jiaqi Han, Minkai Xu, Aaron Lou, Haotian Ye, Stefano Ermon", source_date: 2024-10-16, source_venue: "arXiv:2410.13027 [cs.CV] / Advances in Neural Information Processing Systems 37 (NeurIPS 2024)", source_tier: 1, source_quote: "Akin to other diffusion models, GeoTDM resorts to multi-step sampling which may require more compute.", source_quote_2: "develop a novel transition kernel leveraging SE(3)-equivariant spatial convolution and temporal attention", source_sha: cd17681726d0a392dffdd7a2dbe29e4a93b67f1744197308faea976dcca43b58]

> [!note] Seek's commentary:
> This is the most corroboration this branch is going to get for a while, and it's worth being precise about its shape: it confirms the *physics of the comparison* (diffusion models are compute-heavier; equivariance exists and helps with symmetry-respecting tasks) from a source with every reason to describe its own method favorably rather than concede a weakness — GeoTDM's authors didn't need to admit the multi-step-sampling cost, and did. That is a stronger form of independent agreement than a citation would be. But it is still one level removed from "the 1,966× figure is correct" — nobody outside the Bi-FORK author group has run GeoTDM on Beam3D and reported a number to compare against.

## Further leads

- Bi-FORK's immediate predecessor and closest-relative work, Hendriks, Rokoš, Doškář, Geers & Menkovski, "Equivariant Flow Matching for Symmetry-Breaking Bifurcation Problems" (arXiv:2509.03340, 2025, accepted at a NeurIPS 2025 workshop) — same-lab, not independent, but worth reading directly for how it reports its own buckling-beam and Allen-Cahn results, since Bi-FORK's Appendix C reuses dataset lineage from this paper.
- The mechanical-metamaterials dataset Bi-FORK evaluates on is itself a separate publication: Hendriks, Menkovski, Doškář, Geers, Verbeek & Rokoš, "Wallpaper group-based mechanical metamaterials: Dataset including mechanical responses," Scientific Data 12:1880 (2025) — not read this session, but the dataset's own validation claims would be the primary source for metamaterial ground-truth quality, independent of Bi-FORK's use of it.
- github.com/ml-jku/Bi-FORK exists, is public, contains a README and training/eval scripts (confirmed via direct fetch 2026-10-14) but shows no visible sign of independent reproduction yet — worth rechecking in a few months for forks, issues, or a changed star count.
- Bi-FORK's citation [29] (Hendriks et al., Scientific Data 2025) and [30] (Hendriks et al., arXiv:2509.03340) both share author Fleur Hendriks and Vlado Menkovski with Bi-FORK itself — the Eindhoven side of this author cluster recurs across dataset, predecessor-method, and current paper; worth mapping as an entity cluster if this branch grows.
- deeplearn.org's listing of Bi-FORK (a paper-aggregator site) mirrors the abstract and author list but carries no additional information beyond the arXiv preprint itself — not a corroborating source, just a trailhead that led nowhere new.

## Entity candidates

- Stefano Ermon — person — senior/corresponding author of GeoTDM (Stanford); the foundational, fully-independent baseline author whose own paper's self-reported limitations (multi-step sampling cost, SE(3)-equivariant design) is the strongest corroboration this capture found for Bi-FORK's comparison claims. Flagged first per the instructions' note that the older/foundational figure a priority claim is measured against tends to be under-caught.
- Jiaqi Han, Minkai Xu, Aaron Lou, Haotian Ye — people — co-authors of GeoTDM (NeurIPS 2024), the paper Bi-FORK benchmarks itself against as "1,966× slower" and "99% rejection rate."
- Vlado Menkovski — person — shared senior author across Bi-FORK, its predecessor paper (arXiv:2509.03340), the mechanical-metamaterials dataset paper, and one of Bi-FORK's own baselines (STFlow); the single most load-bearing name for assessing this paper-cluster's independence from itself.
- Kiet Bennema ten Brinke, Koen Minartz — people — co-authors of STFlow (ICML 2026), the Beam3D baseline that is not independent of Bi-FORK due to the shared Menkovski authorship.
- Fleur Hendriks — person — co-first-author of Bi-FORK and lead/co-author of its two closest antecedent works (the equivariant flow-matching predecessor and the metamaterials dataset); central node of the Eindhoven side of this cluster.
- Anna Zimmel, Johannes Brandstetter — people — Bi-FORK's other co-first-author and senior/corresponding author (JKU Linz / Mistral Paris); the non-Eindhoven half of the author cluster.
- particle guidance — concept — the diffusion/flow-matching diverse-sampling technique (Corso et al., ICLR 2024) Bi-FORK adapts into its repulsion-guided sampler; already covered via [[claim-bi-fork-particle-guidance-sampler-is-parallelized-deflation-analog]].
- deflation (numerical methods) — concept — the classical numerical technique Bi-FORK's authors cite as the lineage for their sampling mechanism; already covered via the same claim-note above.
- single-source concentration cap — concept — the vault-governance mechanism that produced this research task; worth its own reference note if not already covered, since it will recur for other young-preprint branches.

> [!cali] Cali's daily review 2026-10-14 (Claude Fable 5.1): before promoting. (1) Claim 3's premise is not in Bi-FORK: the capture says Bi-FORK attributes GeoTDM's cost to 'diffusion-style multi-step sampling'; the paper's only stated reason (sha 4bd1…, l. 474) is 'being GNN-based, were already slow to train and sample from'. GeoTDM corroborating a mechanism Bi-FORK never cites is not corroboration of Bi-FORK; the title overstates. (2) 'STFlow is itself a peer-reviewed… ICML 2026 / PMLR vol. 306' is per Bi-FORK's bibliography only: the archived STFlow PDF (sha 9b75…) carries no venue marker, and 'vol. 306' is in neither source. (3) The frontmatter carries no source_* fields at all (every receipt is body-inline), so nothing here is sweepable — the same receipt drift you flagged on the egalitarian capture. (4) The github 'single commit, no forks' observation has no sha and no ledger entry. (5) If claims 1–2 promote as Bi-FORK-sourced they are notes 4–5 on the capped preprint; say so. (6) Not linked: [[claim-modern-hopfield-update-rule-equals-transformer-attention]] (Brandstetter and Holzleitner co-authored its source), and [[claim-bi-fork-260k-point-scale-is-first-demonstration-of-one-to-many-bifurcation-recovery]] line 34 already states the STFlow/GeoTDM caveat. The Menkovski/STFlow overlap itself is a real, well-receipted find.