{"id":"9b6c12c2-0766-40c4-ac57-6ef41a2543f3","arxiv_id":"2607.06841","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Functional tensor trains plus BSDE regression solve the HJB score PDE, yielding a fast low-rank sampler that outperforms neural diffusion methods on multimodal targets.","lead":"The paper introduces Tensor Train Diffusion (TTD), which solves the high-dimensional HJB PDE for score-based sampling by representing the log-density with functional tensor trains and regressing via BSDEs. It offers a faster, more stable alternative to neural PDE solvers for sampling from unnormalized multimodal densities.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The paper's central claim is a methods claim: FTT + BSDE regression yields a practical, high-fidelity sampler for multimodal unnormalized densities that is faster and more accurate than neural baselines on the examined suite. The derivation (Lem. 2.1, discrete BSDE loss (10), ALS micro-steps (16)/(95), domain-extension and rank/degree adaptivity) is non-circular and correctly specialises known theory. Empirical support (Figs. 3–6, 14–15) is strong and includes wall-clock comparisons. The only modelling assumption that could invalidate the claim—exponential rank growth—is already flagged by the authors and by the reader; it does not undermine correctness for the targets where ranks stay modest. Consequently the ACCEPT verdict stands; no adjustment is warranted.","tokens_in":39757,"tokens_out":453,"duration_ms":6392,"concrete_test":"Re-run the d=50 Multiwell experiment of Fig. 3 with the rank-adaptation threshold δ raised from 10^{-4} to 10^{-2} (forcing systematically lower ranks) and check whether ESS remains >0.7 and relative log-Z error <10^{-3}; if both metrics collapse, the low-rank hypothesis is confirmed as necessary for the reported performance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption (low-rank FTT representability of V(·,t) on the trajectory-induced domains) is already the paper's own stated limitation (§B.1 and conclusion). It is not a hidden inconsistency: the method is correctly derived from HJB–BSDE theory, the ALS/BSDE losses are well-posed under that modelling hypothesis, and the Multiwell/Ginzburg-Landau experiments (including explicit rank-adaptivity diagnostics in Fig. 5/11 and Gaussian rank bounds in Thm. A.5–A.6) supply concrete evidence that the hypothesis holds for the claimed target class. When the hypothesis fails the method simply becomes inapplicable, which the authors acknowledge. No further load-bearing flaw in the central claim is visible.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Tensor Train Diffusion (TTD), a method for sampling from unnormalized high-dimensional densities by solving the HJB PDE that governs the log-density of a noising process. The value function is approximated in the functional/extended tensor-train format and learned via a sequence of BSDE-derived empirical regression losses (Eqs. 9–10) that are minimized by alternating least squares with adaptive rank, basis degree, and regularization. An outer iteration starting from annealed Langevin refines the control so that training trajectories concentrate on the relevant support. Numerical experiments on Multiwell (d=10/50), Ginzburg–Landau, and a Kitagawa state-space model report high ESS, low log-variance, and accurate log-Z estimates, with runtimes competitive with or better than neural diffusion samplers (DIS, PIS and several other baselines).","tokens_in":39909,"tokens_out":752,"duration_ms":8720,"significance":"If the low-rank hypothesis holds for the target class, TTD supplies a concrete, non-SGD alternative to PINN- and neural-SDE-based diffusion samplers that is both faster and more accurate on the reported multimodal benchmarks. The combination of HJB–BSDE theory with carefully stabilized FTT regression (domain extension, adaptive τ, rank and degree selection) is technically solid and the appendix supplies explicit Gaussian rank bounds (Thms. A.5–A.6) together with reproducible algorithmic detail. The public code link further strengthens the contribution. The work therefore advances the practical toolkit for high-dimensional sampling from unnormalized densities while remaining honest about the modelling assumption that limits its scope.","major_comments":[{"comment":"The central modelling hypothesis—that V(·,t) admits a sufficiently low-rank FTT representation on the trajectory-induced domains—is correctly identified as a limitation (§B.1 and conclusion) and is supported by the Gaussian analysis and the rank-adaptivity diagnostics (Figs. 5, 11). No load-bearing inconsistency is present; the method is simply inapplicable when the hypothesis fails. The experimental evidence on Multiwell and Ginzburg–Landau is therefore sufficient for the claimed target class, and no further major technical revision is required on this point.","section":null}],"minor_comments":[{"comment":"Notation for time inversion (⃗f) and the discrete loss (10) is dense; a short glossary or expanded first appearance would help readers less familiar with the BSDE literature.","section":null},{"comment":"Figures 3–4 and 14–15 would benefit from explicit wall-clock units and a brief statement of hardware used, so that the runtime comparisons are fully reproducible.","section":null},{"comment":"The hyper-parameter study (Figs. 12–13) is useful; a short table summarizing the default values of γ, q, p, δ and m_i would make the experimental protocol easier to replicate.","section":null},{"comment":"A few typographical inconsistencies appear (e.g., “HJB type” vs. “HJB-type”, occasional missing spaces around citations); a light copy-edit would polish the manuscript.","section":null}],"recommendation":"accept","confidential_remarks":"The manuscript is a clean, technically careful contribution that sits comfortably within the scope of a machine-learning / computational-statistics venue. The low-rank assumption is openly stated and does not undermine the claims that are actually made. I see no reason to request major revision or rejection."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean engineering paper that turns the authors’ earlier TT-BSDE machinery into a practical score-based sampler. The new piece is the full TTD pipeline: functional tensor trains for the HJB value function, BSDE-derived backward regression (loss (10)), outer-loop trajectory refinement, and a suite of adaptivity tricks (rank, degree, regularization, moving domains with linear extension). That combination is not in the prior TT-PDE or neural-diffusion literature.\n\nWhat they do well is concrete. On Multiwell (d=10 and 50) and Ginzburg-Landau they report higher ESS, lower log-variance, and better log-Z estimates than DIS, PIS and a broad set of neural baselines, at substantially lower wall-clock time. Rank-adaptivity diagnostics (Figs. 5/11) and the Gaussian rank bounds (Thms. A.5–A.6) give evidence that the low-rank hypothesis is not pure hope for these targets. Code is public, metrics come from path-space importance weights against the known unnormalized density, and the HJB–BSDE derivation is standard and correctly specialized. The free parameters (γ, p/q, δ, m_i, N) are ordinary numerical knobs, not free parameters that absorb the claim.\n\nThe soft spot is exactly the one they flag in B.1 and the conclusion: if V(·,t) does not stay low-rank on the trajectory domains, storage and ALS cost explode and the method is simply inapplicable. That is a modelling limitation, not a hidden inconsistency; the stress-test is right that nothing further is broken. Theory of FTT ranks for general HJB solutions remains open, which they also state.\n\nThis is for people who already care about high-dimensional sampling from unnormalized densities and who are willing to live with a low-rank modelling assumption. It deserves a serious referee. I would cite it when I need a non-neural baseline that is fast on multimodal problems of this type, and I would bring it to reading group if the group is methods-oriented.","headline":"Solid methods paper: FTT+BSDE for the diffusion HJB gives a real, faster sampler on multimodal targets when the low-rank assumption holds.","tokens_in":40574,"tokens_out":528,"would_cite":true,"duration_ms":7615,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Tensor trains plus backward SDEs solve the high-dimensional HJB equation that drives reverse diffusion sampling, yielding faster and more accurate samples from multimodal unnormalized densities than neural methods.","keywords":["tensor train","functional tensor train","score-based sampling","Hamilton-Jacobi-Bellman","backward SDE","diffusion models","high-dimensional sampling","low-rank approximation"],"falsifier":"On a multi-well target whose modes are known exactly, replace the learned TT control by the true score (available from a high-resolution 1-d finite-difference solve) and check whether the effective sample size and log-variance of the TT sampler match the true-score baseline within sampling error; a systematic gap that cannot be closed by increasing TT rank or number of time steps would falsify the claim that the FTT approximation is sufficiently accurate for high-fidelity sampling.","tokens_in":40583,"feed_emoji":"🔢","tokens_out":775,"duration_ms":14634,"temperature":0.7,"pith_summary":"Sampling from an unnormalized high-dimensional density is hard because the normalizing constant is unknown and modes are often separated. Diffusion-style methods reverse a noising process whose score is the gradient of the log-density; that log-density obeys a Hamilton-Jacobi-Bellman PDE that is expensive to solve with neural networks. This paper shows that the same PDE can be solved by representing the value function in functional tensor-train format and marching backward in time with a sequence of cheap least-squares regressions that come from the associated backward SDE. The resulting Tensor Train Diffusion sampler needs far fewer density evaluations, converges in a few outer iterations, and produces higher-fidelity samples on multi-well and Ginzburg-Landau targets than existing neural diffusion samplers while running substantially faster. A sympathetic reader cares because the method turns a previously brittle, hyper-parameter-sensitive training problem into a stable, rank-adaptive linear-algebra procedure that scales with the intrinsic low-rank structure of the target.","feed_headline":"Tensor trains solve the reverse-diffusion HJB faster than nets","feed_subtitle":"Low-rank functional trains plus BSDE regression yield high-fidelity samples from multimodal densities with far fewer density calls.","key_machinery":"Functional tensor-train (FTT / extended TT) representation of the HJB value function, combined with a BSDE-derived empirical loss that is linear in the value function and its gradient; the loss is minimized by alternating least squares on the TT cores, with adaptive rank, basis degree, regularization, and domain-extension strategies that keep the approximation stable outside the training trajectories.","core_discovery":"The authors establish that the value function of the HJB equation governing the log-density of a noising process admits an accurate, low-rank functional tensor-train approximation along the trajectories of an iteratively refined reverse SDE. By discretizing the corresponding BSDE into a sequence of regularized least-squares problems and solving them with alternating least squares on the tensor-train cores, one obtains both the value function and its gradient (the score) at every time step. The resulting control produces high-quality samples from challenging multimodal targets after only a few outer iterations and with far lower wall-clock cost than neural competitors.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Tensor trains approximate reverse-diffusion HJB via low-rank FTT and BSDE","FTT cores plus ALS solve reverse HJB for high-fidelity multimodal samples","Low-rank functional trains cut reverse-SDE score cost versus neural nets","BSDE least-squares on tensor trains yields reverse-diffusion scores fast","Iterative FTT solves high-dim HJB log-density for score-based sampling"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"The value function of the HJB equation must stay low-rank enough in the functional tensor-train format on the moving domains visited by the sampling trajectories; if the ranks grow exponentially the storage and solve cost become prohibitive.","fun_headline_variants_meta":{"raw":{"variants":["Tensor trains approximate reverse-diffusion HJB via low-rank FTT and BSDE","FTT cores plus ALS solve reverse HJB for high-fidelity multimodal samples","Low-rank functional trains cut reverse-SDE score cost versus neural nets","BSDE least-squares on tensor trains yields reverse-diffusion scores fast","Iterative FTT solves high-dim HJB log-density for score-based sampling"]},"model":"grok-4.5","effort":"low","cost_usd":0.006214,"raw_usage":{"total_tokens":1637,"prompt_tokens":806,"num_sources_used":0,"completion_tokens":108,"cost_in_usd_ticks":62140000,"prompt_tokens_details":{"text_tokens":806,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":723,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":806,"tokens_out":108,"duration_ms":8088,"temperature":1.0,"reasoning_tokens":723,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T20:11:56.541151+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a multi-well target whose modes are known exactly, replace the learned TT control by the true score (available from a high-resolution 1-d finite-difference solve) and check whether the effective sample size and log-variance of the TT sampler match the true-score baseline within sampling error; a systematic gap that cannot be closed by increasing TT rank or number of time steps would falsify the claim that the FTT approximation is sufficiently accurate for high-fidelity sampling.","supporting_citations":[],"review_version":1}