{"id":"1a4d846f-7cef-4174-8d78-5dccb9d8529d","arxiv_id":"2504.14889","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"NF-BO is a latent Bayesian optimization framework that uses normalizing flows for a one-to-one input-latent mapping and token-level adaptive sampling, outperforming prior LBO methods on Guacamol and PMO benchmarks.","lead":"The paper presents NF-BO, a latent Bayesian optimization method that uses autoregressive normalizing flows to guarantee exact reconstruction between molecular sequences and their latent codes. It reports state-of-the-art results on molecule optimization benchmarks by also adding a token-level adaptive candidate sampling strategy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Exact-reconstruction guarantee rests on an unenforced distinctness condition on learned embeddings; Proposition 2's premise is an assumption, not an invariant of training.","rationale":"The paper's central value proposition is that NF-BO removes the reconstruction gap by composing an injective discrete map h with an invertible normalizing flow g. Proposition 2 is the only result that supplies h's injectivity and its left inverse; without it, Proposition 1 is vacuous for this application. The condition it imposes (distinct, L2-normalized embeddings) is not an invariant of the optimization defined by Eqs. (13)-(14): the contrastive loss encourages separation but cannot rule out exact or directional collapse, and no projection step is described. This is a gap between theorem and algorithm rather than an observed failure: Table 5's 0.000 discrepancy on the top-1000 initial points is encouraging, but the guarantee is claimed for the method in general and the embeddings are retrained during the BO loop (Algorithm 1), so that table does not fully settle the assumption. I agree with the reader that this is the weakest load-bearing point. The issue is addressable by explicit enforcement or by reporting minimum embedding separation and full reconstruction accuracy across the entire run; if those checks pass, the central claim stands. I therefore keep the CONDITIONAL verdict and would not adjust the reader's assessment.","tokens_in":18844,"tokens_out":20411,"duration_ms":208814,"concrete_test":"Run the released code (github.com/mlvlab/NFBO) on one Guacamol task (e.g., amlodipine MPO) for the full (10,000 initial, 10,000 oracle) budget. After every SeqFlow retraining step in Algorithm 1, record d_min = min_{i != j} (1 - cos(e_i, e_j)) over all vocabulary embeddings and compute the reconstruction accuracy dec(enc(x)) == x on all molecules in the accumulated dataset D, not just the top 1,000 used in Table 5. If d_min stays positive and reconstruction accuracy is 1.0 at every iteration, Proposition 2's premise holds empirically for this setting. If d_min reaches 0 or accuracy drops below 1, the exact-reconstruction guarantee is violated, and the paper should either enforce distinctness explicitly (e.g., renormalization plus a separation projection) or weaken the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central theoretical claim is that SeqFlow guarantees perfect reconstruction (Section 4.2, Propositions 1-2), which is what eliminates the value discrepancy problem. Proposition 2 holds only if the learned embedding set E consists of distinct, L2-normalized vectors (ei != ej, ||ei||_2 = 1). The training loss in Eqs. (13)-(14) encourages separation through a contrastive term but does not enforce it, and no renormalization or pairwise-separation projection after gradient updates is stated. If two embedding directions coincide, h is no longer injective; the argmax left-inverse h^{-1}(v) = [argmax_j sim(v_i, e_j)] in Eq. (8) is ambiguous for the collapsed pair, and the guarantee h^{-1}(h(x)) = x fails. The surrogate can then receive conflicting objective values at the same latent code, re-introducing the label inconsistency the method claims to remove. Table 5 reports a y != y_hat ratio of 0.000 on the top-1000 initial data, which is good empirical evidence for that subset, but it is not a guarantee and does not cover the embeddings after the iterative retraining in Algorithm 1, where E is updated jointly with g. The proof is also stated for x in N^L while E is finite; the intended statement should quantify over the finite token vocabulary.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NF-BO, a latent Bayesian optimization method that uses an autoregressive normalizing flow (SeqFlow) to map discrete sequences to a continuous latent space. The key claimed advantage is a one-to-one encoding function from the input space to the latent space with a left-inverse decoding function, which is said to eliminate the value discrepancy problem that arises from imperfect reconstruction in VAE-based LBO. The method also introduces Token-level Adaptive Candidate Sampling (TACS) to focus trust-region search on important tokens. The authors evaluate NF-BO on Guacamol and PMO molecular-design benchmarks, reporting strong performance against existing LBO and generative-model baselines.","tokens_in":19102,"tokens_out":8592,"duration_ms":78739,"significance":"If the central claims hold, the paper would make a useful contribution to latent Bayesian optimization by providing a principled alternative to VAE-based encoders and by proposing a concrete way to address the value discrepancy problem. The empirical evaluation is extensive, covering Guacamol and PMO benchmarks with multiple oracle-budget settings, and the manuscript provides code, pseudocode, and proofs in the appendix. However, the theoretical guarantee of perfect reconstruction rests on an unenforced distinctness assumption on the learned embeddings, and the construction does not fully address the consistency between arbitrary latent-space candidates and re-encoded evaluated points. The TACS temperature is also tuned on a benchmark task whose results are then reported. These issues are load-bearing for the paper's main claims, so the manuscript needs substantial revision before the claims can be considered fully established.","major_comments":[{"comment":"The injectivity of the mapping h is proven under the assumption that the embedding set E consists of distinct, L2-normalized vectors. This assumption is not enforced during training: the contrastive loss in Eq. (13) encourages separation but does not guarantee it, and Algorithm 1 has no projection or pairwise-distinctness step after gradient updates. If two embeddings collapse to the same vector, the argmax in Eq. (8) is not unique, h^{-1} is not a well-defined function, and the claimed guarantee h^{-1}(h(x)) = x fails. Table 5 reports a zero y != y_hat ratio on the top-1000 initial data, but because Algorithm 1 updates E jointly with the flow, this measurement does not cover the embeddings used later in optimization. The proof of Proposition 2 also quantifies over x in N^L while E is finite; the intended statement should be over the finite token vocabulary.","section":"Section 4.2, Proposition 2 and Eqs. (13)-(14)"},{"comment":"The construction gives a left-inverse for the encoding h, but it does not give a two-sided inverse between the discrete input space and the latent space. For an arbitrary latent candidate z sampled by TACS (e.g., within a trust region), the decoded input x = h^{-1}(g^{-1}(z)) may encode to a different latent vector z' = g(h(x)) that is not equal to z. The surrogate is trained on the latent encodings z^{(i)} = g(e_{x^{(i)}}) of the evaluated points, while acquisition is performed at the candidate z. Thus the latent coordinate at which the surrogate is queried can differ from the latent coordinate at which the resulting evaluation is stored, potentially reintroducing a form of value inconsistency that the paper claims to remove. Propositions 1-2 only establish h^{-1}(h(x)) = x, not that decoding an arbitrary z and re-encoding the result returns the same z; the paper should either restrict the candidate set to the image of the discrete embeddings under g, or analyze/control this discrepancy.","section":"Section 4.2 and Section 4.4 / Algorithm 1"},{"comment":"The TACS temperature tau is tuned on a single Guacamol task (searching {400, 200, 100}) and then fixed at tau=400 for all benchmarks, including the same seven Guacamol tasks whose results are reported in the paper. This constitutes hyperparameter selection on the evaluation tasks, which can bias the comparison with baselines. The sensitivity analysis in Table 7 shows that tau=200 achieves a higher summed score (5.565) than the selected tau=400 (5.544), so the chosen value is not clearly the best even within the same benchmark. The authors should either tune hyperparameters on a separate validation set or report results for all settings without selecting on the test tasks.","section":"Section 5.3 and Appendix K"},{"comment":"For the valt task, the authors state that 'we also added two nonzero data points in the initial dataset of 100 for all methods.' This is a modification to the standard Guacamol benchmark setup. Although applied uniformly to all methods, it should be reported as a limitation and its impact on the ranking should be discussed, especially because the y-axis is on a log scale and without these points the task may be degenerate for all methods.","section":"Section 5.1 and Figure 8"}],"minor_comments":[{"comment":"In Eq. (9), the notation \\hat{x}_i is used both for the argmax index and for the reconstructed token; this is confusing and should be clarified (e.g., by writing the argmax index explicitly as j^*).","section":"Section 4.2, Eq. (9)"},{"comment":"The description of the constrained sampling distribution q_1(v_i|x_i) is terse; the relationship between the normalization constant Z and the acceptance probability should be spelled out so that the rejection step is unambiguous.","section":"Section 4.2, Eq. (11)"},{"comment":"The notation D[r - N_q:] is not defined; presumably r refers to the current iteration index, but this should be stated explicitly.","section":"Algorithm 1, line 2"},{"comment":"The sentence 'TACS temperatures above 200 consistently show better optimization results compared to not using TACS' is accurate according to the table, but the conclusion would be clearer if the authors also noted that temperatures 100, 50, and 20 perform worse than the no-TACS baseline, indicating that the method has a limited favorable range.","section":"Appendix K, Table 7"},{"comment":"Several PMO tasks have very large standard deviations (e.g., celecoxibrediscovery 0.799+-0.164, scaffoldhop 0.769+-0.172), which may weaken the statistical significance of some reported improvements; the authors should discuss this variation.","section":"Appendix A, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is already published at ICLR 2025, but this arXiv version has the same content. The main theoretical claim is overstrong relative to what is established, and the TACS tuning issue is a standard benchmarking concern. The authors have a good empirical record and the code is available, so the issues seem addressable with additional analysis, an enforced embedding-separation mechanism, and a fairer hyperparameter-selection protocol. I would not reject on the basis of the empirical results alone, but the central claim about eliminating the value discrepancy problem needs to be substantially qualified or strengthened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a solid, well-executed application of normalizing flows to latent BO. It won’t reorganize the field, but it solves the reconstruction gap in a direct way and supports it with strong empirical work. The main caveat is that the perfect-reconstruction guarantee depends on an unenforced distinctness condition on learned token embeddings.\n\nThe genuinely new part is using an invertible normalizing flow as the encoder-decoder inside LBO. SeqFlow adapts existing autoregressive flows (TextFlow) to this setting, and TACS is a sensible token-level candidate sampling scheme. Props 1 and 2 are correct: with distinct, L2-normalized embeddings, the argmax decoder is a left inverse, so reconstruction is exact. That is a real argument, not hand-waving. The empirical evidence is also good: the y-vs-yhat mismatch is 0.000 on the top-1000 initial points, and NF-BO ranks first on five of six PMO metrics and beats the VAE-based LBO baselines on Guacamol. The code is available.\n\nThe soft spots are real but fixable. The distinct-embedding assumption is load-bearing but not enforced. The contrastive loss encourages separation, but nothing guarantees ei != ej during or after the iterative retraining in Algorithm 1. So the exact-reconstruction claim is an empirical property, not a proven invariant. Table 5 only checks the top-1000 initial data, not the retraining loop. That’s not fatal, but it should be stated honestly. Second, the TACS temperature was tuned on one Guacamol task and then fixed for everything else; the sensitivity analysis shows the result is fairly robust, but it is still benchmark tuning. Third, the valt task has a disclosed modification (two nonzero initial points added), which slightly muddies that comparison. And Proposition 2 quantifies over N^L while the embedding set is finite; the intended statement should be over the token vocabulary.\n\nWho is it for? People doing latent BO, discrete sequence optimization, or molecular design. It’s a useful baseline and a clear proof of concept for NF-based latent spaces. I’d definitely give it a serious referee; the core idea holds up and the weaknesses are addressable. If I were the reviewer, I’d ask for an explicit separation constraint or post-training distinctness check, and a finite-vocabulary statement of the proposition. I’d also want the temperature selection reported as a minor limitation. But none of that undermines the central finding.","headline":"Solid NF-based latent BO with a real but unenforced reconstruction guarantee; strong experiments, minor tuning concerns.","tokens_in":19652,"tokens_out":3720,"would_cite":true,"duration_ms":33436,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Latent Bayesian optimization can be made lossless by an invertible flow that maps discrete inputs one-to-one into latent space, eliminating the reconstruction gap that corrupts VAE-based latent search.","keywords":["latent Bayesian optimization","normalizing flows","value discrepancy problem","autoregressive flow for sequences","molecular design","discrete sequence optimization","token-level candidate sampling","pointwise mutual information"],"falsifier":"Train SeqFlow on any sequence dataset, then compute the reconstruction failure rate: the fraction of training inputs $x$ for which $h^{-1}(g^{-1}(g(h(x))))$ differs from $x$, or for which two distinct tokens have cosine similarity equal to 1. Any non-zero failure rate, or any observed embedding collapse, would falsify the exact-reconstruction claim central to the method.","tokens_in":18637,"feed_emoji":"🧪","tokens_out":7227,"duration_ms":60565,"temperature":0.7,"pith_summary":"Latent Bayesian optimization embeds discrete inputs, such as molecules, into a continuous space, runs Bayesian optimization there, and decodes the result; when the generative model is a variational autoencoder, the reconstruction is imperfect and the objective value attached to a latent point can disagree with the value of the decoded input. The paper claims that this value discrepancy problem is avoidable: replace the VAE with an invertible normalizing flow, SeqFlow, whose autoregressive coupling layers transform discrete token embeddings into a latent code, with a nearest-embedding argmax as the left-inverse decoder. Under distinct, L2-normalized token embeddings, every input $x$ is reconstructed exactly from its latent code, so latent space and input space support the same objective values throughout optimization. On molecular design benchmarks, the resulting NF-BO, augmented with token-importance sampling, ranks first on five of six PMO metrics, outperforming both VAE-based latent optimization methods and strong generative baselines.","feed_headline":"Flow-based latent BO beats VAE-based LBO on molecules","feed_subtitle":"An invertible flow with exact decode closes the reconstruction gap, lifting PMO rank from 19th to 1st.","key_machinery":"The carrier of the argument is SeqFlow, an autoregressive normalizing flow for discrete sequences, together with its left-inverse decoding rule. Each token index $x_i$ is first represented by an L2-normalized embedding $e_{x_i}$ from a learned embedding set $E$; an invertible stack of $K$ autoregressive coupling blocks maps the embedding sequence $v$ into a latent code $z$, and decoding is the inverse flow followed by an argmax over cosine similarities to $E$. Propositions 1 and 2 establish that when the embeddings are pairwise distinct, $h$ is injective with $h^{-1}(v) = [\\arg\\max_j \\mathrm{sim}(v_i, e_j)]$ as its left inverse, so $f = g \\circ h$ is left-invertible and reconstruction is exact. The second component, token-level adaptive candidate sampling, uses the pointwise mutual information between each latent coordinate and the decoded sequence to define a sampling probability over which tokens to perturb around an anchor point, biasing search toward positions with high influence on the objective.","core_discovery":"The central discovery is that perfect reconstruction in latent Bayesian optimization follows from combining an invertible flow with an injective token lookup rather than from any approximation. Writing $h$ for the map that sends a discrete sequence $x$ to the sequence of learned token embeddings, and $g$ for a normalizing flow, the composition $f = g \\circ h$ is left-invertible because $h$ is injective (distinct L2-normalized embeddings) and $g$ is invertible by construction; decoding via $h^{-1} \\circ g^{-1}$ returns exactly $x$. SeqFlow instantiates this with an autoregressive flow over the sequence of embeddings, so the surrogate model in latent space always refers to a well-defined input and a well-defined objective value. Empirically, the paper reports that this exactness transfers to performance: NF-BO outperforms six latent Bayesian optimization baselines on Guacamol and takes first rank on five out of six PMO scoring metrics, improving the SELFIES VAE baseline's average rank from 19th to 1st.","pith_inferences":["If embedding distinctness degrades under training, the argmax decoder is still locally stable, so monitoring the minimum pairwise embedding distance during optimization could serve as a cheap early warning for reconstruction failures; this diagnostic is not reported in the paper.","Because the latent code is an invertible transformation of the input rather than a stochastic summary, interpolations and small perturbations in $z$ may be more semantically faithful, which could make NF-BO useful for property-guided molecular interpolation; this is an extension the paper does not explore.","The PMI-based token importance, computed with a trained flow, gives a per-position attribution of which parts of a molecule drive the objective, so TACS doubles as an interpretability tool for medicinal chemists beyond its role in candidate sampling.","The same construction transfers to other discrete structures with learned embeddings, such as DNA/RNA sequences or program tokens, where the left-inverse guarantee could replace approximate decoders; this is a conjecture, not a claim of the paper."],"forward_implications":["Latent Bayesian optimization no longer needs re-encoding or iterative inversion: because every latent code decodes to a unique input, an acquisition function can trust that the surrogate value it sees at $z$ is the objective value of the decoded molecule.","The one-to-one mapping turns the latent space into a faithful coordinate system for discrete structures, so trust-region local search and Thompson sampling behave as they do in continuous domains.","The exact-reconstruction property removes the need for repeated re-encoding of newly found data during optimization, which the paper notes previously caused additional value discrepancy.","The method is not tied to molecule-specific inductive biases, and the same SeqFlow/TACS pipeline is applicable to other discrete sequence optimization problems where exact reconstruction matters."],"supporting_citations":[{"why":"Defines normalizing flows as invertible transformations with tractable Jacobian determinants, the backbone of SeqFlow's one-to-one mapping.","marker":"Rezende & Mohamed, 2015"},{"why":"Introduces latent normalizing flows for discrete sequences, which SeqFlow adapts for its autoregressive flow and variational distribution over discrete inputs.","marker":"Ziegler & Rush, 2019"},{"why":"Establishes latent Bayesian optimization for molecular design by embedding discrete inputs with a VAE, the paradigm whose reconstruction gap the paper targets.","marker":"Gómez-Bombarelli et al., 2018"},{"why":"Formulates weighted-retraining LBO and states the complete-reconstruction assumption that the paper identifies as the source of the value discrepancy problem.","marker":"Tripp et al., 2020"},{"why":"Provides the trust-region latent BO baseline (LOL-BO) and the SELFIES VAE used in the paper's motivating demonstration of value discrepancy.","marker":"Maus et al., 2022"},{"why":"Supplies the CoBO baseline that refines the latent space during optimization, supporting the paper's claim that re-encoding amplifies value discrepancy.","marker":"Lee et al., 2023"},{"why":"Contributes the trust-region and Thompson-sampling machinery that NF-BO uses for local search and that TACS modifies.","marker":"Eriksson et al., 2019"},{"why":"Supplies the variational dequantization construction (the $q(v_i|x_i)$ distribution) used by SeqFlow's ELBO objective.","marker":"Ho et al., 2019"}],"fun_headline_variants":["Exact latent decoding with flows boosts molecule BO","Flow BO eliminates reconstruction gap, ranks 1st on PMO","Invertible flow encoding: latent BO becomes exact","SeqFlow: autoregressive flows for exact latent BO","NF-BO: normalizing flows close the latent BO gap"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee of perfect reconstruction rests on the learned token embeddings remaining pairwise distinct and L2-normalized; if any two embeddings collapse together, the argmax decoding fails and the exact-reconstruction claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Exact latent decoding with flows boosts molecule BO","Flow BO eliminates reconstruction gap, ranks 1st on PMO","Invertible flow encoding: latent BO becomes exact","SeqFlow: autoregressive flows for exact latent BO","NF-BO: normalizing flows close the latent BO gap"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000335,"raw_usage":{"total_tokens":1868,"prompt_tokens":968,"completion_tokens":900,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":820}},"tokens_in":584,"tokens_out":900,"duration_ms":7975,"temperature":1.0,"reasoning_tokens":820,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:37:23.431145+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SeqFlow on any sequence dataset, then compute the reconstruction failure rate: the fraction of training inputs $x$ for which $h^{-1}(g^{-1}(g(h(x))))$ differs from $x$, or for which two distinct tokens have cosine similarity equal to 1. Any non-zero failure rate, or any observed embedding collapse, would falsify the exact-reconstruction claim central to the method.","supporting_citations":[],"review_version":1}