REVIEW 3 major objections 5 minor 83 references
Expanding Flow Maps
T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash
Pith's one-line read Expanding flow maps make output size a learned, controllable degree of freedom rather than a fixed canvas, and the paper shows the same few-step flow-map recipe works for continuous and discrete data.
desk verdict A genuinely new variable-dimension flow framework with mostly careful theory, but the discrete consistency proof leans on Assumption A5, which is not well-formed as written—still worth serious refereeing. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the expanding flow map Phi_{s,t} = X_{s,t} composed with E_{s,t}: an expand operator E_{s,t} that lifts a state from dimension d(s) to d(t) by inserting noise coordinates (via concatenation, positional insertion, or child expansion), followed by a transport map X_{s,t} that denoises the augmented state along a local-time coordinate t_i = (t - t_ins^i)/(1 - t_ins^i). The local-time coordinate is what lets coordinates inserted at different global times share a common denoising clock. The paper packages the two-stage dynamics as a piecewise-deterministic Markov process whose generator is a transport term plus a jump kernel, and enforces consistency through the Lagrang
What would settle it
Take a trained discrete EFM, sample inputs x_s and times s,t, and compare the denoiser's clean prediction on the expanded sequence with the expansion of its clean prediction on the shorter sequence. If these differ substantially on held-out data, the discrete Eulerian identity and the objective built on it are not satisfied, so the claimed consistency reduction does not hold in practice.
Extended reading notes
Core claim
The paper introduces Expanding Generative Flows (EFlows) and Expanding Flow Maps (EFMs). EFlows are defined by an expanding interpolant that grows the dimension d(t) over time, with new coordinates inserted as conditional noise; the process is characterized as a piecewise-deterministic Markov process with a transport term and a jump kernel. An EFM is the composition Phi_{s,t} = X_{s,t} composed with E_{s,t}, where E_{s,t} lifts a state to a higher dimension and X_{s,t} moves it forward along the interpolant. The authors show this composed map satisfies the three flow-map consistency identities, making it a valid stochastic flow map that expands and denoises jointly, and they derive discrete
Load-bearing premise
The discrete consistency identities rely on Assumption A5: the model's clean prediction on an already-expanded sequence must equal the expansion of its clean prediction on the shorter sequence; the velocity mask does not enforce this, and the paper does not empirically verify it, so if it fails, the discrete training objective is not mathematically guaranteed.
Editorial extensions
If this is right
- Fixed-canvas flow maps are a special case: when the expand operator is the identity, an EFM reduces to a standard flow map.
- The same consistency objectives used for latent flow maps can be applied to variable-size output, and the paper demonstrates one- and two-step generation of molecular graphs and text.
- Because the map is a composition of learned operators, output size becomes a controllable degree of freedom rather than a fixed hyperparameter.
- Discrete EFM extends flow-map distillation to the simplex, enabling few-step generation without committing to a sequence length in advance.
- The PDMP characterization gives a single generative process for the entire expanding trajectory, not just a sequence of independent jumps.
Reading between the lines
- We infer the expand-transport split could be applied symmetrically to deletion, letting the same PDMP construction handle outputs that shrink as well as grow during inference.
- A concrete testable extension is to make the inserted noise distribution learnable, which the paper notes as a future direction and which may address the single-step mode collapse observed for language.
- The insertion head's predicted per-gap count could be repurposed as a controllable length or size knob at inference time, without retraining.
- The same consistency-identity proof strategy should transfer to other interpolants, not just linear ones, as long as the expand operator and velocity mask satisfy the stated structural assumptions.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces Expanding Generative Flows (EFlows) and Expanding Flow Maps (EFMs), a framework for flow-based generative modeling on state spaces whose dimension grows over time. It defines an expanding interpolant, an expand operator that augments the state by inserting noise, and a transport map that denoises the expanded state. The continuous construction is formalized as a piecewise-deterministic Markov process, and analogous discrete constructions are given for variable-length sequences and graphs. The authors derive consistency identities (Lagrangian, Eulerian, semigroup) for the expanding flow map in both continuous and discrete settings, and evaluate the method on molecular conformer generation, molecular graph generation, and language modeling, reporting competitive few-step results against fixed-length flow-map baselines.
Significance. If the framework is correct, it unifies fixed-canvas flow maps as a special case (identity expansion) and provides a principled way to do few-step generation where output size is itself sampled. The continuous theory is carefully set up with explicit assumptions A1–A4, several of which are enforced by construction or by the velocity mask. The paper also makes good use of publicly released checkpoints (FLM, EFlow, DeFoG) for teacher-based distillation and provides a code release. The main theoretical risk is in the discrete setting, where one of the three claimed consistency identities rests on an ill-defined assumption. The empirical claims are promising but are presented without error bars, so their statistical strength is currently unverified.
major comments (3)
- [App. B.8 / Assumption A5 / Eq. (31b)] Assumption A5 as stated is not well-formed. E_{s,t} is defined in Eq. (27) on latent states in R^{d(s)×V} by inserting Gaussian noise tokens, but ψ_{s,s}(x_s) is a simplex-valued vector in (Δ^{V−1})^{d(s)}. The expression E_{s,t}(ψ_{s,s}(x_s)) therefore has no definition; a natural extension (e.g., inserting uniform or Gaussian tokens) would not lie in (Δ^{V−1})^{d(t)}, contrary to the claim in A5. The discrete Eulerian identity (31b) uses A5 to replace ψ_{s,s}(x^ε_s) with E_{s,t}(ψ_{s,s}(x_s)), so the derivation of (31b) is incomplete. Since Proposition 5.1 explicitly claims all three discrete identities, this is a load-bearing gap. The semigroup identity used in the training objective (32) does not appear to rely on A5, so the method might still be salvageable, but the proposition as stated is not proven. Please either provide a correct definition of E_{s,t} on denoiser outputs and enf
- [Sec. 6.3, App. C.5, Table 4] The one-step EFM result in Table 4 appears to use a map Φ_{0,1}(x_0)=ψ_{tdenoise-min,1}(E_{0,1}(x_0)) rather than the trained Φ_{0,1} defined in Eq. (30). App. C.5 states that true one-step sampling is degenerate and that a modified denoising interval [tdenoise-min, 1] is used. Yet Table 4 reports a one-step generative perplexity of 98.35 with no caveat in the main text (the text later admits mode collapse). This is inconsistent: the modification is a form of task-specific tuning, and the reported number does not correspond to a single application of the learned expanding flow map. Please clarify exactly what the 1-step entry measures, report the unmodified one-step performance, and discuss whether tdenoise-min is tuned per task.
- [Tables 1-4] No error bars, confidence intervals, or multiple seeds are reported in any of the main empirical tables. Key claims such as 'EFM's FCD remains strong at 0.44 vs. 0.49' and 'EFM's FCD is nearly 5× lower than CFM's' rest on single runs. Without variance estimates, the practical significance of these margins cannot be assessed. Please add standard deviations over at least 3 seeds (or bootstrap intervals) for the main headline metrics.
minor comments (5)
- [Eq. (27), App. D.5] The symbol E is used both for the edge matrix in the discrete graph construction and for the expand operator E_{s,t}; this creates confusion in equations like (98). Please use a distinct symbol for one of them.
- [Eq. (32)] The loss in (32) sums over d(t) positions, but the target \(\bar\psi_{s,t}\) is defined on the expanded state x^ε_s with positions inserted after u left empty. The exact masking of those positions in the cross-entropy sum should be stated explicitly.
- [Eq. (13)] The per-coordinate integral notation \(\int_{s_i}^{t_i} b_{\tau_i}(x_{\tau_i}[i]) d\tau_i\) is unclear: is \(b_{\tau_i}\) the same velocity field evaluated at a different time or a coordinate-specific field? Please define it precisely.
- [Figure 4 vs. Table 4] The 1-step example in Figure 4 reports a single-sample perplexity of 47.8, while the table reports 98.35. If the example is cherry-picked, say so; otherwise explain the discrepancy.
- [Title page] Typo: 'Correspondance' should be 'Correspondence'.
Circularity Check
No significant circularity; minor self-citations in related work are not load-bearing, and Assumption A5 is a soundness gap rather than a circular step.
full rationale
The derivation is self-contained rather than circular. The EFM Phi_{s,t} = X_{s,t} ∘ E_{s,t} is a compositional definition; the Lagrangian/Eulerian/semigroup identities in Prop. 4.1 and Prop. 5.1 are derived from the fixed-dimensional flow-map identities (external: Boffi et al. 2025; Lee et al. 2026; Potaptchik et al. 2026b) under structural assumptions A1–A4 that are enforced by the velocity mask, not by fitting the desired output. The consistency objectives (22) and (32) are self-consistency training losses, not fitted constants renamed as predictions. Teacher distillation uses the independently released FLM checkpoint (Lee et al. 2026) and external baselines; self-distillation from data is standard and does not smuggle the result into the input. Self-citations (Tang et al. 2025; Tang et al. 2026) appear only in related-work surveys and are not load-bearing. The main deductive weakness is Assumption A5 in App. B.8: ψ^{(t)}_{s,s}(E_{s,t}(x)) = E_{s,t}(ψ_{s,s}(x)) is asserted without empirical verification and is arguably ill-defined because E_{s,t} is defined on latent states while ψ_{s,s} outputs simplex vectors; this threatens the proof of the discrete Eulerian identity (31b). That is a soundness/correctness gap, not a circular reduction—the identity does not reduce to A5 by construction, and the practical semigroup target (33) does not depend on A5. No quoted step exhibits a fitted value being called a prediction or a load-bearing conclusion identical to its premise.
Assumptions & free parameters
free parameters (4)
- Insertion schedule shape and cutoff per task =
Conformer: cosine schedule on [0, 0.6]; graph: polynomial r=0.5; language: cosine with tins-end=0.5
- Refiner noise sigma_r =
0.4 for QM9, 1.0 for Drugs
- Noise/prior scales =
sigma_z=1.0, sigma_h=0.1, prior sigma=1.0
- Training stabilization hyperparameters =
c=1e-6, r=0.5, lambda_E=5.0, diagonal fraction=0.75
assumptions (5)
- domain assumption A1-A4: expand operator is differentiable with embedding Jacobian, composable, leaves inactive coordinates unchanged, and admits an active-inactive velocity split.
- ad hoc to paper A5: diagonal-expand consistency, psi^{(t)}_{s,s}(E_{s,t}(x)) = E_{s,t}(psi_{s,s}(x)).
- standard math Piecewise-deterministic Markov process generator and finite-jump/no-explosion conditions.
- standard math Stochastic interpolant background: velocity as conditional expectation and the conditional flow-matching objective.
- standard math Disintegration theorem for conditional measures.
invented entities (1)
-
Local time coordinates t_local
Cite this review
Pith. "Pith review of Expanding Flow Maps." pith.science (2026). https://pith.science/paper/5B7466YW
@misc{pith2026260721585,
author = {Pith},
title = {Pith review of: Expanding Flow Maps},
year = {2026},
howpublished = {\url{https://pith.science/paper/5B7466YW}},
note = {Machine review of arXiv:2607.21585}
}
read the original abstract
Flow-based generative models have enabled remarkable progress in fast and controllable generation across continuous and discrete state spaces, yet existing parameterizations are constrained to fixed dimensions or fixed sequence lengths. Here, we introduce Expanding Generative Flows (EFlows), which define flows between distributions of increasing dimensionality along an expanding interpolant that grows the state by augmenting it with conditional noise. Building on this construction, we propose Expanding Flow Maps (EFMs), a new class of flow maps that distill the expanding interpolant into efficient few-step generative models. Each EFM factors the map between any two timesteps into two learnable operations: an expand operator, which augments the state space with new coordinates or tokens conditioned on the current state, and a transport map, which pushes the expanded state forward along the interpolant. Composing these operators yields a single map that jointly expands and denoises the state, recovering existing fixed-canvas flows and flow maps as the special case in which the expand operator is the identity. We further extend the framework to the discrete simplex, enabling variable-size graph generation and variable-length sequence generation. Across both continuous and discrete modalities, we establish EFlows and EFMs as a principled framework for settings in which output size is itself a learned, controllable degree of freedom.
Reference graph
Works this paper leans on
-
[1]
arXiv preprint arXiv:2602.16813 , year=
Flow Map Language Models: One-step Language Modeling via Continuous Denoising , author=. arXiv preprint arXiv:2602.16813 , year=
-
[2]
International Conference on Learning Representations , year=
Any-order flexible length masked diffusion , author=. International Conference on Learning Representations , year=
-
[3]
arXiv preprint arXiv:2511.01464 , year=
Split-Flows: Measure Transport and Information Loss Across Molecular Resolutions , author=. arXiv preprint arXiv:2511.01464 , year=
-
[4]
International Conference on Machine Learning , year=
Categorical flow maps , author=. International Conference on Machine Learning , year=
-
[5]
arXiv preprint arXiv:2604.09784 , year=
Discrete Flow Maps , author=. arXiv preprint arXiv:2604.09784 , year=
-
[6]
International Conference on Machine Learning , year=
Consistency models , author=. International Conference on Machine Learning , year=
-
[7]
International Conference on Learning Representations , year=
Improved techniques for training consistency models , author=. International Conference on Learning Representations , year=
-
[8]
Advances in Neural Information Processing Systems , year=
How to build a consistency model: Learning flow maps via self-distillation , author=. Advances in Neural Information Processing Systems , year=
Show all 83 references
-
[9]
Advances in Neural Information Processing Systems , year=
Align your flow: Scaling continuous-time flow map distillation , author=. Advances in Neural Information Processing Systems , year=
-
[10]
Advances in Neural Information Processing Systems , year=
Mean flows for one-step generative modeling , author=. Advances in Neural Information Processing Systems , year=
-
[11]
International Conference on Learning Representations , year=
One step diffusion via shortcut models , author=. International Conference on Learning Representations , year=
-
[12]
arXiv preprint arXiv:2507.16884 , year=
Splitmeanflow: Interval splitting consistency in few-step generative modeling , author=. arXiv preprint arXiv:2507.16884 , year=
-
[13]
Journal of Machine Learning Research , volume=
Stochastic interpolants: A unifying framework for flows and diffusions , author=. Journal of Machine Learning Research , volume=
-
[14]
Advances in Neural Information Processing Systems , volume=
Denoising diffusion probabilistic models , author=. Advances in Neural Information Processing Systems , volume=
-
[15]
arXiv preprint arXiv:2210.02747 , year=
Flow matching for generative modeling , author=. arXiv preprint arXiv:2210.02747 , year=
-
[16]
arXiv preprint arXiv:2302.00482 , year=
Improving and generalizing flow-based generative models with minibatch optimal transport , author=. arXiv preprint arXiv:2302.00482 , year=
-
[17]
International Conference on Learning Representations , year=
Denoising diffusion implicit models , author=. International Conference on Learning Representations , year=
-
[18]
Transactions in Machine Learning Research , year=
Flow map matching with stochastic interpolants: A mathematical framework for consistency models , author=. Transactions in Machine Learning Research , year=
-
[19]
International Conference on Machine Learning , year=
Inductive moment matching , author=. International Conference on Machine Learning , year=
-
[20]
Advances in Neural Information Processing Systems , year=
Multistep distillation of diffusion models via moment matching , author=. Advances in Neural Information Processing Systems , year=
-
[21]
International Conference on Learning Representations , year=
Consistency trajectory models: Learning probability flow ode trajectory of diffusion , author=. International Conference on Learning Representations , year=
-
[22]
International Conference on Learning Representations , year=
Score-based generative modeling through stochastic differential equations , author=. International Conference on Learning Representations , year=
-
[23]
International Conference on Learning Representations , year=
Building normalizing flows with stochastic interpolants , author=. International Conference on Learning Representations , year=
-
[24]
Score-based generative modeling of graphs via the system of stochastic differential equations , author=
-
[25]
International Conference on Machine Learning , year=
Graph generation with diffusion mixture , author=. International Conference on Machine Learning , year=
-
[26]
Advances in Neural Information Processing Systems , year=
Variational flow matching for graph generation , author=. Advances in Neural Information Processing Systems , year=
-
[27]
International Conference on Machine Learning , year=
Defog: Discrete flow matching for graph generation , author=. International Conference on Machine Learning , year=
-
[28]
International Conference on Learning Representations , year=
PairFlow: Closed-Form Source-Target Coupling for Few-Step Generation in Discrete Flow Models , author=. International Conference on Learning Representations , year=
-
[29]
Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining , pages=
Moflow: an invertible flow model for generating molecular graphs , author=. Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining , pages=
-
[30]
International Conference on Learning Representations , year=
Top-n: Equivariant set and graph generation without exchangeability , author=. International Conference on Learning Representations , year=
-
[31]
International Conference on Machine Learning , year=
Diamond Maps: Efficient Reward Alignment via Stochastic Flow Maps , author=. International Conference on Machine Learning , year=
-
[32]
International Conference on Machine Learning , year=
Meta Flow Maps enable scalable reward alignment , author=. International Conference on Machine Learning , year=
-
[33]
Advances in Neural Information Processing Systems , year=
Edit Flows: Flow Matching with Edit Operations , author=. Advances in Neural Information Processing Systems , year=
-
[34]
arXiv preprint arXiv:2510.03506 , year=
OneFlow: Concurrent Mixed-Modal and Interleaved Generation with Edit Flows , author=. arXiv preprint arXiv:2510.03506 , year=
-
[35]
International Conference on Learning Representations , year=
GLASS Flows: Transition Sampling for Alignment of Flow and Diffusion Models , author=. International Conference on Learning Representations , year=
-
[36]
International Conference on Learning Representations , year=
Beyond fixed: Training-free variable-length denoising for diffusion large language models , author=. International Conference on Learning Representations , year=
-
[37]
International Conference on Learning Representations , year=
Block diffusion: Interpolating between autoregressive and diffusion language models , author=. International Conference on Learning Representations , year=
-
[38]
International Conference on Learning Representations , year=
DreamOn: Diffusion Language Models For Code Infilling Beyond Fixed-size Canvas , author=. International Conference on Learning Representations , year=
-
[39]
International Conference on Learning Representations , year=
Beyond Masks: Efficient, Flexible Diffusion Language Models via Deletion-Insertion Processes , author=. International Conference on Learning Representations , year=
-
[40]
arXiv preprint arXiv:2601.22527 , year=
-EOS: Training-free Bidirectional Variable-Length Control for Masked Diffusion LLMs , author=. arXiv preprint arXiv:2601.22527 , year=
-
[41]
Chemical science , volume=
MoleculeNet: a benchmark for molecular machine learning , author=. Chemical science , volume=. 2018 , publisher=
2018
-
[42]
1997 , publisher=
Foundations of modern probability , author=. 1997 , publisher=
1997
-
[43]
Advances in Neural Information Processing Systems , year=
Simplified and generalized masked diffusion for discrete data , author=. Advances in Neural Information Processing Systems , year=
-
[44]
Advances in Neural Information Processing Systems , year=
Simple and effective masked diffusion language models , author=. Advances in Neural Information Processing Systems , year=
-
[45]
International Conference on Learning Representations , year=
Your absorbing discrete diffusion secretly models the conditional distributions of clean data , author=. International Conference on Learning Representations , year=
-
[46]
International Conference on Learning Representations , year=
Masked diffusion models are secretly time-agnostic masked models and exploit inaccurate categorical sampling , author=. International Conference on Learning Representations , year=
-
[47]
International Conference on Machine Learning , year=
The diffusion duality , author=. International Conference on Machine Learning , year=
-
[48]
OpenAI blog , volume=
Language models are unsupervised multitask learners , author=. OpenAI blog , volume=
-
[49]
arXiv preprint arXiv:1312.3005 , year=
One billion word benchmark for measuring progress in statistical language modeling , author=. arXiv preprint arXiv:1312.3005 , year=
-
[50]
Preuer, Kristina and Renz, Philipp and Unterthiner, Thomas and Hochreiter, Sepp and Klambauer, Gunter , journal=. Fr. 2018 , publisher=
2018
-
[51]
International Conference on Machine Learning , year=
Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design , author=. International Conference on Machine Learning , year=
-
[52]
Nucleic acids research , volume=
The protein data bank , author=. Nucleic acids research , volume=. 2000 , publisher=
2000
-
[53]
Journal of the Royal Statistical Society: Series B (Methodological) , volume=
Piecewise-deterministic Markov processes: A general class of non-diffusion stochastic models , author=. Journal of the Royal Statistical Society: Series B (Methodological) , volume=. 1984 , publisher=
1984
-
[54]
arXiv preprint arXiv:2605.10938 , year=
Elf: Embedded language flows , author=. arXiv preprint arXiv:2605.10938 , year=
-
[55]
International Conference on Learning Representations , year=
Beyond autoregression: Fast llms via self-distillation through time , author=. International Conference on Learning Representations , year=
-
[56]
International Conference on Machine Learning , year=
Better & faster large language models via multi-token prediction , author=. International Conference on Machine Learning , year=
-
[57]
International Conference on Machine Learning , year=
Distillation of discrete diffusion through dimensional correlations , author=. International Conference on Machine Learning , year=
-
[58]
International Conference on Machine Learning , year=
Fast inference from transformers via speculative decoding , author=. International Conference on Machine Learning , year=
-
[59]
Advances in Neural Information Processing Systems , year=
Redi: Rectified discrete flow , author=. Advances in Neural Information Processing Systems , year=
-
[60]
Scientific data , volume=
GEOM, energy-annotated molecular conformations for property prediction and molecular generation , author=. Scientific data , volume=. 2022 , publisher=
2022
-
[61]
International Conference on Learning Representations , year=
Geodiff: A geometric diffusion model for molecular conformation generation , author=. International Conference on Learning Representations , year=
-
[62]
Advances in Neural Information Processing Systems , year=
Geomol: Torsional geometric generation of molecular 3d conformer ensembles , author=. Advances in Neural Information Processing Systems , year=
-
[63]
Greg Landrum , volume=
RDKit: A software suite for cheminformatics, computational chemistry, and predictive modeling , author=. Greg Landrum , volume=
-
[64]
International Conference on Machine Learning , year=
Learning gradient fields for molecular conformation generation , author=. International Conference on Machine Learning , year=
-
[65]
arXiv preprint arXiv:1909.11459 , year=
A generative model for molecular distance geometry , author=. arXiv preprint arXiv:1909.11459 , year=
1909 arXiv
-
[66]
Advances in Neural Information Processing Systems , year=
Et-flow: Equivariant flow-matching for molecular conformer generation , author=. Advances in Neural Information Processing Systems , year=
-
[67]
International Conference on Learning Representations , year=
Learning neural generative dynamics for molecular conformation generation , author=. International Conference on Learning Representations , year=
-
[68]
International Conference on Machine Learning , year=
An end-to-end framework for molecular conformation generation via bilevel programming , author=. International Conference on Machine Learning , year=
-
[69]
Advances in Neural Information Processing Systems , year=
Subgdiff: A subgraph diffusion model to improve molecular representation learning , author=. Advances in Neural Information Processing Systems , year=
-
[70]
Foundations of Crystallography , volume=
A solution for the best rotation to relate two sets of vectors , author=. Foundations of Crystallography , volume=. 1976 , publisher=
1976
-
[71]
CCP4 Newsl
Pymol: An open-source molecular graphics tool , author=. CCP4 Newsl. protein crystallogr , volume=
-
[72]
arXiv preprint arXiv:2604.11748 , year=
Langflow: Continuous diffusion rivals discrete in language modeling , author=. arXiv preprint arXiv:2604.11748 , year=
-
[73]
arXiv preprint arXiv:2504.10283 , year=
-Flow: A Unified Framework for Continuous-State Discrete Flow Matching Models , author=. arXiv preprint arXiv:2504.10283 , year=
-
[74]
arXiv preprint arXiv:2211.15089 , year=
Continuous diffusion for categorical data , author=. arXiv preprint arXiv:2211.15089 , year=
-
[75]
Advances in Neural Information Processing Systems , year=
Continuous diffusion model for language modeling , author=. Advances in Neural Information Processing Systems , year=
-
[76]
ArXiv , pages=
Gumbel-softmax flow matching with straight-through guidance for controllable biological sequence generation , author=. ArXiv , pages=
-
[77]
International Conference on Machine Learning , year=
Dirichlet flow matching with applications to dna sequence design , author=. International Conference on Machine Learning , year=
-
[78]
arXiv preprint arXiv:2605.11125 , year=
Language modeling with hyperspherical flows , author=. arXiv preprint arXiv:2605.11125 , year=
-
[79]
Advances in Neural Information Processing Systems , year=
Fisher flow matching for generative modeling over discrete data , author=. Advances in Neural Information Processing Systems , year=
-
[80]
arXiv preprint arXiv:2605.18530 , year=
Continuous diffusion scales competitively with discrete diffusion for language , author=. arXiv preprint arXiv:2605.18530 , year=
-
[81]
International Conference on Learning Representations , year=
Hierarchical Multi-Scale Molecular Conformer Generation , author=. International Conference on Learning Representations , year=
-
[82]
arXiv preprint arXiv:2606.13565 , year=
A2D2: Fine-Tuning Any-Length Discrete Diffusion for Adaptive Decoding , author=. arXiv preprint arXiv:2606.13565 , year=
-
[83]
The Journal of chemical physics , volume=
PSI4 1.4: Open-source software for high-throughput quantum chemistry , author=. The Journal of chemical physics , volume=. 2020 , publisher=
2020
Reviewed August 1, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.