Pith. sign in

REVIEW 3 major objections 5 minor 37 references

A single function evaluation builds a full stochastic search tree for budget-constrained online discovery.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-12 06:08 UTC pith:4IOQAEAK

load-bearing objection Solid engineering of single-NFE tree paths for budgeted online search; the BSS construction and dynamic schedule are real advances over DTS, theory is standard but clean, main soft spot is the unquantified distillation residual. the 3 major comments →

arxiv 2607.02915 v1 pith:4IOQAEAK submitted 2026-07-03 cs.LG cs.AI

Bootstrap Flow-Map Tree Sampling Enables Online Feedback Driven Search

Binglin Ji , Anindya Sarkar , Hengchang Lu , Jens Sj\"olund , Yevgeniy Vorobeychik This is my paper
classification cs.LG cs.AI
keywords flow mapstree searchonline feedbackbudget-constrained samplingdiffusion alignmentbootstrap sufficient statisticinference-time scaling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

When the target you care about is unknown at the start and revealed only through costly sequential feedback, ordinary generative samplers either collapse to local modes or burn too many model evaluations to explore broadly. This paper claims that a distilled flow map, combined with a bootstrap sufficient-statistic construction, can synthesize an entire DDPM-like stochastic trajectory from any intermediate state with one network call. That single-call path becomes the edge of a search tree whose depth and step sizes can be chosen freely, so the sampler can start with small exploratory steps near the root and later enlarge steps for local refinement once high-utility modes appear. A budget-aware selection rule further shifts probability mass from exploration to exploitation as the remaining query budget shrinks. Empirically the resulting Bootstrap Flow-Map Tree outperforms existing tree, particle, and meta-flow baselines on both open-ended ImageNet search and compositional/quantitative alignment tasks while using far fewer function evaluations.

Core claim

BFMT shows that a teacher-distilled flow map together with a bootstrap sufficient-statistic loop yields exact DDPM-style stochastic transitions from any depth after a single network evaluation, and that the resulting tree, equipped with non-uniform transition schedules and a remaining-budget selection rule, converges in total variation to the soft optimal policy at rate O(β T² M^{-1/4}) while delivering higher reward under fixed NFE and feedback budgets than competing samplers.

What carries the argument

Bootstrap Sufficient Statistic (BSS) loop: after one flow-map evaluation that maps noise and the current state to a clean sample x0, intermediate nodes are obtained by a closed-form interpolant that re-uses the same pair (x0, ε); the construction is proved to match the true DDPM transition kernel at every step, enabling dynamic step sizes and full tree expansion for the cost of a single NFE.

Load-bearing premise

The distilled student flow map must recover the true conditional posterior accurately enough that the subsequent bootstrap interpolant still produces a valid DDPM trajectory; large distillation residual breaks both the theory and the single-evaluation claim.

What would settle it

Replace the distilled map with a deliberately under-trained student whose instantaneous and consistency losses remain high, then measure whether the generated intermediate states still match the analytic DDPM kernel and whether the reported reward curves collapse relative to a multi-NFE baseline under identical budgets.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Any setting that must discover high-utility modes under a hard query budget (drug design, interactive recommendation, rare-disease treatment search) can replace multi-step rollouts with single-NFE tree edges.
  • Dynamic non-uniform transition schedules become a free design knob for controlling the exploration-exploitation trade-off without increasing network calls.
  • Budget-aware node selection (BASE) systematically outperforms classical UCT when the total number of feedback queries is known and finite.
  • Convergence rate improves quadratically with tree horizon T because a long horizon costs only one NFE, giving a concrete incentive to deepen trees rather than widen them.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same single-NFE path construction could be grafted onto other flow or consistency models that already expose an analytic SNR schedule, potentially turning any one-step denoiser into a tree sampler.
  • If the distillation residual can be bounded, the TV rate could be tightened further by replacing progressive widening with adaptive branching that depends on residual size rather than visit count alone.
  • The hierarchical early-explore / late-exploit pattern observed in the qualitative figures suggests a natural curriculum for online reward-model training: early feedback can be used only for coarse mode discovery, later feedback for fine alignment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces Bootstrap Flow-Map Tree (BFMT), a training-free tree sampler for online feedback-driven search and alignment under strict sampling budgets. It distills a flow map that maps any intermediate state xt to a one-step sample from the conditional posterior p(x0|xt), then uses a bootstrap sufficient-statistic (BSS) loop (Eq. 9) to synthesize a full DDPM-like stochastic trajectory from that single NFE. Flow-map flexibility further permits non-uniform transition schedules that shift from small early steps (global exploration) to large late steps (local exploitation), while a budget-aware selection rule (BASE, Eq. 13) replaces standard UCT. Soft Bellman backups propagate terminal rewards (from black-box feedback models) up the tree. Proposition 4.1 claims the BSS trajectory matches DDPM kernels; Proposition 4.3 bounds TV distance to the optimal soft policy by O(β T² M^{-1/4}). Experiments on ImageNet class search and compositional/quantity alignment report higher mean and max rewards than DTS, FKS, DAS and MFM at lower NFE and feedback budgets, with ablations isolating BSS, dynamic schedules and BASE.

Significance. Online feedback-driven discovery under evaluation budgets is a genuine bottleneck in scientific and engineering applications. Combining amortized one-step conditional sampling with tree search that needs only a single NFE per full path is a concrete efficiency advance over existing tree and SMC samplers. The dynamic-transition and budget-aware selection ideas are well-motivated and the soft-Bellman / progressive-widening analysis (Prop. 4.3) is standard and carefully written. Empirical ablations (Figs. 5–8) isolate each component and the qualitative exploration-to-exploitation visualizations (Figs. 9, 12) are informative. If the distillation residual is shown to be small and the statistical claims are tightened, the method would be a useful practical tool for inference-time alignment and search.

major comments (3)
  1. [§2, Prop. 4.1, Eqs. 7–9] Prop. 4.1 and the single-NFE claim rest on the assumption that the distilled student ˆv produces exact (or sufficiently accurate) draws from p(x0|xt). The teacher-distillation losses L_inst + L_cons are stated, and Fig. 5 contrasts BSS with an ISI baseline, yet the manuscript never reports a quantitative residual (conditional log-likelihood, Wasserstein distance, or multi-step ODE vs. one-step student error) as a function of noise level t. Without this, both the theoretical guarantee and the claimed NFE reduction remain conditional on an unmeasured approximation quality.
  2. [§3, Figs. 4, 7] Search experiments use only five randomly chosen ImageNet classes (Table 1 / Appendix). Mean and max reward curves (Figs. 4, 7) lack error bars, confidence intervals, or multi-seed statistics. Given the stochastic nature of tree expansion and BASE selection, the claim that BFMT “substantially outperforms” all baselines cannot be assessed for statistical reliability from the reported plots alone.
  3. [§2, Eq. 13, Eq. 15; §3 dynamic-schedule ablation] The free parameters of the method (β, progressive-widening C and ζ, the concrete sequence of transition times t′, distillation loss weights) are listed but never subjected to a sensitivity study. In particular, the dynamic schedule that is credited for the exploration–exploitation transition is described only qualitatively; an ablation that freezes the schedule to a few fixed non-uniform templates (or reports the exact schedule used) would make the contribution of “dynamic transition time steps scheduling” more reproducible and falsifiable.
minor comments (5)
  1. [§2 / Appendix] Proposition numbering jumps (Prop. 4.1, then 4.3); Prop. 4.2 appears to be missing or renumbered inconsistently between main text and appendix.
  2. [Fig. 1] Figure 1 caption and the surrounding text refer to “N Diffusion Depth / N BFMT depth” without defining N; a short legend would help.
  3. [Appendix §12] The reparameterization of the SANA/TrigFlow noise schedule (Appendix §12) is useful but dense; a short pseudocode block mapping (t_cm, α, σ) into the BSS loop of Eq. 9 would improve reproducibility.
  4. [Figs. 4, 11] Several figure panels (e.g., Fig. 4, Fig. 11) use overlapping markers without a clear legend for every series; increasing marker size or adding a table of final numbers would aid comparison.
  5. [Abstract, §2] Typographical inconsistencies: “Bootstrapped” vs. “Bootstrap” in the title/abstract; “L_distill_const” vs. “L_distill_cons”; occasional missing spaces before citations.

Circularity Check

0 steps flagged

No significant circularity; BFMT derivations are self-contained from independent target definition, flow-matching distillation, and standard soft-Bellman error propagation.

full rationale

The target policy π* is defined independently of the sampler via the standard soft-optimal form (Eq. 1–2) using an external pretrained prior and online black-box reward r(x). The single-NFE claim and Prop. 4.1 follow from an explicit construction (Eq. 9) whose law is proved to match the DDPM kernel by matching conditional means/covariances under the GLASS sufficient statistic (Appendix proof of Prop. 4.1); this is a derivation of equivalence, not a reduction of a prediction to its own input. Soft values and the optimal policy (Eq. 11–12) are the ordinary soft Bellman recursion. The TV bound of Prop. 4.3 is ordinary progressive-widening + Hoeffding + Lipschitz error propagation under stated boundedness assumptions; no free parameter is fitted to data and then re-presented as a prediction. Empirical rewards come from external models (CLIP-FlanT5 VQA, ImageReward, DINO+SAM counts). Citations to flow matching, consistency distillation, DTS, and progressive widening are external and non-load-bearing for uniqueness. No self-definitional loop, fitted-input-as-prediction, or self-citation chain appears. The distillation residual is an approximation assumption (already flagged by the reader), not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 3 invented entities

The central claims rest on standard generative-model machinery (flow matching, DDPM kernels, soft value functions) plus three paper-specific constructions (distilled flow map, bootstrap interpolant, BASE). Free parameters are ordinary algorithmic hyper-parameters; no new physical entities are postulated.

free parameters (4)
  • inverse temperature β
    Controls the softness of the target policy π*; chosen by the user and appears in both theory and experiments.
  • progressive-widening constants C, ζ
    Determine maximum branching factor B(xt)=⌈C N(xt)^ζ⌉; set by hand following prior continuous UCT literature.
  • distillation loss weights / training schedule for ˆv
    Student flow map is trained with L_inst + L_cons; exact coefficients and number of distillation steps are free choices that affect how closely Prop. 4.1 holds.
  • dynamic transition schedule (sequence of t′)
    Non-uniform time steps are chosen by the practitioner; the paper demonstrates one effective schedule but does not derive an optimal one.
axioms (4)
  • domain assumption Existence of an ODE that transports the Gaussian prior to the conditional posterior p(x0|xt) whose drift is given by the conditional flow-matching field (Eq. 4-5).
    Standard result from flow-matching literature; invoked to justify the teacher field that is distilled.
  • domain assumption Reward r(x0) is bounded and L-Lipschitz; transition kernels are Lipschitz.
    Required for the Monte-Carlo error bounds and the TV convergence rate of Prop. 4.3.
  • ad hoc to paper The distilled student ˆv approximates the analytic teacher field closely enough that one-step samples are exact draws from p(x0|xt).
    Assumed after distillation; residual error is not quantified in the main theorems.
  • standard math Soft Bellman optimality yields the optimal policy expressed via the BSS transition (Eq. 12).
    Standard soft-value recursion; derivation supplied in appendix.
invented entities (3)
  • Bootstrap Sufficient Statistic (BSS) loop (Eq. 9) independent evidence
    purpose: Synthesize a full DDPM-like stochastic trajectory from a single flow-map evaluation while preserving arbitrary time steps.
    Core algorithmic novelty; independent evidence is the matching of moments shown in Prop. 4.1.
  • Budget-Aware Stochastic Exploration (BASE) selection rule (Eq. 13) no independent evidence
    purpose: Shift node-selection temperature with remaining budget fraction so the tree explores early and exploits late.
    Paper-specific alternative to UCT; ablated against UCT in Fig. 8.
  • Bootstrap Flow-Map Tree (BFMT) sampler no independent evidence
    purpose: End-to-end framework combining the above for online feedback-driven search.
    The overall method being introduced.

pith-pipeline@v1.1.0-grok45 · 27834 in / 3119 out tokens · 33061 ms · 2026-07-12T06:08:23.166264+00:00 · methodology

0 comments
read the original abstract

In many scientific and engineering domains, maximizing discovery within a limited sampling budget demands strategic, observation-guided exploration. While generative models have enabled training-free reward alignment, current methods typically excel in local searches within narrow regions of the underlying distribution. These approaches struggle when preferences are unknown a priori and only revealed through sequential feedback-a scenario demanding broad exploration to uncover high-utility regions. To address this, we introduce Bootstrap Flow-Map-Tree (a.k.a BFMT), a novel computationally efficient sampling framework designed for history-aware global search and alignment under sampling budget constraints. BFMT enables full tree-path construction from any tree depth using a single function evaluation, drastically reducing computational overhead while providing critical foresight for sequential sampling. By enabling dynamic transition time steps scheduling, BFMT efficiently allocates its sampling budget, smoothly transitioning from broad global exploration to fine-grained local refinement of high-utility modes discovered through exploration. Extensive experiments and ablations across diverse search and alignment tasks demonstrate that BFMT substantially outperforms baseline approaches.

Figures

Figures reproduced from arXiv: 2607.02915 by Anindya Sarkar, Binglin Ji, Hengchang Lu, Jens Sj\"olund, Yevgeniy Vorobeychik.

Figure 1
Figure 1. Figure 1: Overview of BFMT. that synergizes history-aware tree search with highly efficient Flow Map dynamics. While standard tree search requires computationally expensive rollouts to estimate intermediate node values, BFMT utilizes Flow Maps to collapse this evaluation into a single NFE. Although the deterministic ODE nature of Flow Maps typically bottlenecks the stochastic, DDPM-like transitions required for chil… view at source ↗
Figure 2
Figure 2. Figure 2: Different Key Components of BFMT Where, w1, w2, and w3 represent scalar weights. S serves as a linear sufficient statistic, and t ∗ denotes a reparametrized time parameter: Ss,t(xs, xt) = αsσ 2 t xs + αtσ 2 s xt σ 2 t α2 s + α 2 tσ2 s , t∗ (s, t) = g −1  σ 2 t σ 2 s σ 2 t α2 s + α 2 tσ2 s  , g(t) = σ 2 t α 2 t (6) Although this reparameterization ensures the accessibility of bs, the process of generating… view at source ↗
Figure 3
Figure 3. Figure 3: Search Visualizations. Search Result [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Search Performance Analysis of Competitive Approaches on Imagenet Target Classes. bootstrap sufficient statistic mechanism, which enables the construction of the entire sampling trajectory using a single NFE. Crucially, this gain is mathematically grounded: as established in Proposition 4.4, convergence to the target distribution scales quadratically with the diffusion horizon. By simulating extensive diff… view at source ↗
Figure 5
Figure 5. Figure 5: Impact of BSS. The Impact of Bootstrap Sufficient Statistic (BSS): To evaluate the impact of the BSS mechanism within BFMT, we compare our approach against BFMT-ISI, a variant that replaces BSS with an alternative ODE-based, single-NFE trajectory generation scheme: mapping xt to x1 via a pretrained flow-map, followed by an Iterative Stochastic Interpolant (ISI) between x1 and ϵ ∼ N (0, I) to con￾struct int… view at source ↗
Figure 6
Figure 6. Figure 6: Impact of Dynamic Transition step on Search. denoising phase, BFMT achieves broader exploration. Conversely, larger steps in the final phases promote localized exploitation. Given a fixed NFE budget, this non-uniform transition scheduling enables BFMT to conduct a more efficient search compared to its uniform-step counterpart, BFMT￾US. This enhanced exploratory capacity directly yields a more diverse set o… view at source ↗
Figure 7
Figure 7. Figure 7: Search Performance Analysis on Imagenet Classes. Search Performance Evaluation with Best Reward In this section, we evaluate the search performance of BFMT against several baseline methods, including MFM Best-Of￾N (BoN), which is specifically op￾timized for best-reward evaluation. Detailed comparative results are pre￾sented in the figure 7. Consistent with our earlier analysis, we use Im￾ageReward for eval… view at source ↗
Figure 9
Figure 9. Figure 9: Exploration Strategy of BFMT in Online Feedback-Driven Search Settings. Alignment Experimental Setting: To assess the effectiveness of BFMT on online, feedback-driven alignment problems, we design two challenging experimental settings. First, for the compositional alignment task, we utilize 50 randomly selected prompts from GenAI-Bench [15] that necessitate at least three advanced compositional skills. Sec… view at source ↗
Figure 10
Figure 10. Figure 10: Alignment Visualization. (left) Quantity Prompt. (right) Compositional Prompt. ◦ DAS □ FKS △ MFM (BoN) ⋄ DTS ▽ BFMT (Our) Mean Reward Best Reward NFE NFE Feedback Budgets Feedback Budgets NFE Feedback Budgets [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Analysis of Competitive Approaches on Compositional Alignment. Inheriting the properties of tree search, BFMT uses Bellman value backups after each observation to aggregate historical statistics and drive informed global exploration. By prioritizing branches based on total posterior mass rather than pointwise reward, this mechanism mitigates reward over-optimization. Moreover, unlike SMC, BFMT’s gradient-… view at source ↗
Figure 12
Figure 12. Figure 12: BFMT’s Hierarchical Search Strategy. shown in the figure 12, smaller transition steps in the early stages allow BFMT to explore the space more broadly (e.g., BFMT samples a car with a different color and a different background), while larger transition steps in the later stages enable it to focus on and sample from high-utility modes (e.g., BFMT samples a red sports car). Additional visualizations are pro… view at source ↗
Figure 13
Figure 13. Figure 13: Performance Comparisons on Quantity-Aware Alignment Task. [PITH_FULL_IMAGE:figures/full_fig_p025_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Alignment Performance Analysis on Imagenet Classes. Additional Quantitative Com￾parisons with Baselines on the Quantity-aware Alignment Task We also evaluate the alignment per￾formance of BFMT against several baselines in [PITH_FULL_IMAGE:figures/full_fig_p025_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Additional Visualizations on the impact of BSS. 10 Additional Visual Illustration of BFMT’s Exploratory Search Strategy Figures 16 and 17 illustrate how BFMT’s search strategy evolves across feedback stages. Initially, BFMT explores broadly, generating semantically diverse samples within the target category, though not yet precisely aligned with the fine-grained target prompt. As feedback accumulates, it … view at source ↗
Figure 16
Figure 16. Figure 16: Exploration Strategy of BFMT on Online Feedback-Driven Search Settings. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Exploration Strategy of BFMT on Online Feedback-Driven Search Settings. 11 Additional Visualizations of Hierarchical Exploration Strategy of BFMT Figures 18 and 19 present additional visualizations of BFMT’s hierarchical exploration strategy, further reinforcing its effectiveness in enabling efficient online feedback-driven search [PITH_FULL_IMAGE:figures/full_fig_p027_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: BFMT’s Hierarchical Search Strategy. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: BFMT’s Hierarchical Search Strategy. 12 Adapting the TrigFlow Noise Schedule to the BFMT Framework In this section, we describe how we adapt the trigonometric noise schedule from TrigFlow [19] and the consistency model (SANA-Sprint) [19] to our BFMT framework, which operates under a different noise schedule. We first review the TrigFlow parameterization, then describe the consistency model reparameterizat… view at source ↗
Figure 20
Figure 20. Figure 20: Additional Search Visualizations. 17 Additional Competitive Alignment Visualization Figures 21, 22, 23 presents additional qualitative comparisons in the online feedback-driven alignment setting, further validating the effectiveness of BFMT over competitive baselines. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Additional Alignment Visualization. 32 [PITH_FULL_IMAGE:figures/full_fig_p032_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Additional Alignment Visualization. 33 [PITH_FULL_IMAGE:figures/full_fig_p033_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Additional Alignment Visualization. 34 [PITH_FULL_IMAGE:figures/full_fig_p034_23.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 19 linked inside Pith

  1. [1]

    Feedback efficient online fine-tuning of diffusion models.arXiv preprint arXiv:2402.16359, 2024

    Masatoshi Uehara, Yulai Zhao, Kevin Black, Ehsan Hajiramezanali, Gabriele Scalia, Nathaniel Lee Diamant, Alex M Tseng, Sergey Levine, and Tommaso Biancalani. Feedback efficient online fine-tuning of diffusion models.arXiv preprint arXiv:2402.16359, 2024

    Pith/arXiv arXiv 2024

  2. [2]

    A general framework for inference-time scaling and steering of diffusion models.arXiv preprint arXiv:2501.06848, 2025

    Raghav Singhal, Zachary Horvitz, Ryan Teehan, Mengye Ren, Zhou Yu, Kathleen McKeown, and Rajesh Ranganath. A general framework for inference-time scaling and steering of diffusion models.arXiv preprint arXiv:2501.06848, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    Feynman- kac correctors in diffusion: Annealing, guidance, and product of experts.arXiv preprint arXiv:2503.02819, 2025

    Marta Skreta, Tara Akhound-Sadegh, Viktor Ohanesian, Roberto Bondesan, Alán Aspuru- Guzik, Arnaud Doucet, Rob Brekelmans, Alexander Tong, and Kirill Neklyudov. Feynman- kac correctors in diffusion: Annealing, guidance, and product of experts.arXiv preprint arXiv:2503.02819, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Test-time alignment of diffusion models without reward over-optimization.arXiv preprint arXiv:2501.05803, 2025

    Sunwoo Kim, Minkyu Kim, and Dongmin Park. Test-time alignment of diffusion models without reward over-optimization.arXiv preprint arXiv:2501.05803, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Debiasing guidance for discrete diffusion with sequential monte carlo.arXiv preprint arXiv:2502.06079, 2025

    Cheuk Kit Lee, Paul Jeha, Jes Frellsen, Pietro Lio, Michael Samuel Albergo, and Francisco Vargas. Debiasing guidance for discrete diffusion with sequential monte carlo.arXiv preprint arXiv:2502.06079, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Training-free guidance beyond differentiability: Scalable path steering with tree search in diffusion and flow models.arXiv preprint arXiv:2502.11420, 2025

    Yingqing Guo, Yukang Yang, Hui Yuan, and Mengdi Wang. Training-free guidance beyond differentiability: Scalable path steering with tree search in diffusion and flow models.arXiv preprint arXiv:2502.11420, 2025

    Pith/arXiv arXiv 2025

  7. [7]

    Diffusion tree sampling: Scalable inference-time alignment of diffusion models.arXiv preprint arXiv:2506.20701, 2025

    Vineet Jain, Kusha Sareen, Mohammad Pedramfar, and Siamak Ravanbakhsh. Diffusion tree sampling: Scalable inference-time alignment of diffusion models.arXiv preprint arXiv:2506.20701, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Pith/arXiv arXiv 2022

  9. [9]

    Glass flows: Transition sampling for alignment of flow and diffusion models.arXiv preprint arXiv:2509.25170, 2025

    Peter Holderrieth, Uriel Singer, Tommi Jaakkola, Ricky TQ Chen, Yaron Lipman, and Brian Karrer. Glass flows: Transition sampling for alignment of flow and diffusion models.arXiv preprint arXiv:2509.25170, 2025

    arXiv 2025

  10. [10]

    How to build a consistency model: Learning flow maps via self-distillation.arXiv preprint arXiv:2505.18825, 2025

    Nicholas M Boffi, Michael S Albergo, and Eric Vanden-Eijnden. How to build a consistency model: Learning flow maps via self-distillation.arXiv preprint arXiv:2505.18825, 2025

    arXiv 2025

  11. [11]

    Continuous upper confidence trees

    Adrien Couëtoux, Jean-Baptiste Hoock, Nataliya Sokolovska, Olivier Teytaud, and Nicolas Bonnard. Continuous upper confidence trees. InInternational conference on learning and intelligent optimization, pages 433–445. Springer, 2011

    2011

  12. [12]

    Meta flow maps enable scalable reward alignment.arXiv preprint arXiv:2601.14430, 2026

    Peter Potaptchik, Adhi Saravanan, Abbas Mammadov, Alvaro Prat, Michael S Albergo, and Yee Whye Teh. Meta flow maps enable scalable reward alignment.arXiv preprint arXiv:2601.14430, 2026

    Pith/arXiv arXiv 2026

  13. [13]

    Evaluating text-to-visual generation with image-to-text generation

    Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2024

    2024

  14. [14]

    Imagereward: Learning and evaluating human preferences for text-to-image generation

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903–15935, 2023

    2023

  15. [15]

    Genai arena: An open evaluation platform for generative models.Advances in Neural Information Processing Systems, 37:79889–79908, 2024

    Dongfu Jiang, Max Ku, Tianle Li, Yuansheng Ni, Shizhuo Sun, Rongqi Fan, and Wenhu Chen. Genai arena: An open evaluation platform for generative models.Advances in Neural Information Processing Systems, 37:79889–79908, 2024

    2024

  16. [16]

    Kaiyi Huang, Chengqi Duan, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i- compbench++: An enhanced and comprehensive benchmark for compositional text-to-image generation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(5):3563–3579, 2025. 13

    2025

  17. [17]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. InEuropean conference on computer vision, pages 38–55. Springer, 2024

    2024

  18. [18]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023

    2023

  19. [19]

    Sana-sprint: One-step diffusion with continuous-time consistency distillation

    Junsong Chen, Shuchen Xue, Yuyang Zhao, Jincheng Yu, Sayak Paul, Junyu Chen, Han Cai, Song Han, and Enze Xie. Sana-sprint: One-step diffusion with continuous-time consistency distillation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 16185–16195, 2025

    2025

  20. [20]

    Sequential monte carlo samplers.Journal of the Royal Statistical Society Series B: Statistical Methodology, 68(3):411–436, 2006

    Pierre Del Moral, Arnaud Doucet, and Ajay Jasra. Sequential monte carlo samplers.Journal of the Royal Statistical Society Series B: Statistical Methodology, 68(3):411–436, 2006

    2006

  21. [21]

    Particle denoising diffusion sampler.arXiv preprint arXiv:2402.06320, 2024

    Angus Phillips, Hai-Dang Dau, Michael John Hutchinson, Valentin De Bortoli, George Deligiannidis, and Arnaud Doucet. Particle denoising diffusion sampler.arXiv preprint arXiv:2402.06320, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Monte carlo guided diffusion for bayesian linear inverse problems.arXiv preprint arXiv:2308.07983, 2023

    Gabriel Cardoso, Yazid Janati El Idrissi, Sylvain Le Corff, and Eric Moulines. Monte carlo guided diffusion for bayesian linear inverse problems.arXiv preprint arXiv:2308.07983, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Diffusion posterior sampling for linear inverse problem solving: A filtering perspective

    Zehao Dou and Yang Song. Diffusion posterior sampling for linear inverse problem solving: A filtering perspective. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  24. [24]

    Diffusion probabilistic modeling of protein backbones in 3d for the motif-scaffolding problem.arXiv preprint arXiv:2206.04119, 2022

    Brian L Trippe, Jason Yim, Doug Tischer, David Baker, Tamara Broderick, Regina Barzilay, and Tommi Jaakkola. Diffusion probabilistic modeling of protein backbones in 3d for the motif-scaffolding problem.arXiv preprint arXiv:2206.04119, 2022

    Pith/arXiv arXiv 2022

  25. [25]

    Practi- cal and asymptotically exact conditional sampling in diffusion models.Advances in Neural Information Processing Systems, 36:31372–31403, 2023

    Luhuan Wu, Brian Trippe, Christian Naesseth, David Blei, and John P Cunningham. Practi- cal and asymptotically exact conditional sampling in diffusion models.Advances in Neural Information Processing Systems, 36:31372–31403, 2023

    2023

  26. [26]

    Learning to summarize with human feedback

    Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea V oss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. Advances in neural information processing systems, 33:3008–3021, 2020

    2020

  27. [27]

    Theoretical guarantees on the best-of-n alignment policy

    Ahmad Beirami, Alekh Agarwal, Jonathan Berant, Alex D’Amour, Jacob Eisenstein, Chirag Nagpal, and Ananda Theertha Suresh. Theoretical guarantees on the best-of-n alignment policy. 2024.URL https://api. semanticscholar. org/CorpusID, 266741736(3), 1879

    2024

  28. [28]

    Webgpt: Browser-assisted question-answering with human feedback, 2022.URL https://arxiv

    Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christo- pher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback, 2022.URL https://arxiv. org/abs/2112.09332, 35, 2022

    Pith/arXiv arXiv 2022

  29. [29]

    Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

    2022

  30. [30]

    Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288, 2023

    Pith/arXiv arXiv 2023

  31. [31]

    Aligning text-to-image models using human feedback.arXiv preprint arXiv:2302.12192, 2023

    Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback.arXiv preprint arXiv:2302.12192, 2023. 14

    Pith/arXiv arXiv 2023

  32. [32]

    Better aligning text-to- image models with human preference.arXiv preprint arXiv:2303.14420, 1(3), 2023

    Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Better aligning text-to- image models with human preference.arXiv preprint arXiv:2303.14420, 1(3), 2023

    Pith/arXiv arXiv 2023

  33. [33]

    Imagereward: Learning and evaluating human preferences for text-to-image generation

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024

    2024

  34. [34]

    Is conditional generative modeling all you need for decision-making?arXiv preprint arXiv:2211.15657, 2022

    Anurag Ajay, Yilun Du, Abhi Gupta, Joshua Tenenbaum, Tommi Jaakkola, and Pulkit Agrawal. Is conditional generative modeling all you need for decision-making?arXiv preprint arXiv:2211.15657, 2022

    Pith/arXiv arXiv 2022

  35. [35]

    Planning with diffusion for flexible behavior synthesis.arXiv preprint arXiv:2205.09991, 2022

    Michael Janner, Yilun Du, Joshua B Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis.arXiv preprint arXiv:2205.09991, 2022

    Pith/arXiv arXiv 2022

  36. [36]

    Training diffusion models with reinforcement learning.arXiv preprint arXiv:2305.13301, 2023

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning.arXiv preprint arXiv:2305.13301, 2023. 15 Bootstrap Flow-Map Tree Sampling Enables Online Feedback Driven Search (Appendix) Appendix Table of Contents Appendix: Theoretical Derivations Page 1 Derivation of BFMT Convergence Rate ....

    Pith/arXiv arXiv 2023

  37. [37]

    a red sports car

    Z min(ˆq(xt−1|xt), π∗(xt−1|xt))dx≥e −2βϵ Substitute this back into the TV distance definition: DT V (ˆq(xt−1|xt)∥π ∗(xt−1|xt))≤1−e −2βϵ For anyy≥0, the exponential inequality1−e −y ≤yholds true. Settingy= 2βϵ: 1−e −2βϵ ≤2βϵ Thus, we conclude that: DT V (ˆq(xt−1|xt)∥π ∗(xt−1|xt))≤2βϵ= 2βsup xt−1 | ˆVt−1(xt−1)−V t−1(xt−1)| Substituting our propagated value ...