acceptodds
Under review as a conference paper at ICLR 2027

Grab a Coffee: Future-Aware Guidance for Discrete Diffusion with Compiled Objectives

12 views

Abstract

Discrete diffusion models generate sequences by iteratively resolving multiple tokens in parallel, offering a flexible alternative to left-to-right generation. However, guiding this process with a sequence-level objective is difficult because the value of one unresolved token depends on the other tokens with which it can form a high-reward sequence. Enumerating all such completions makes the whole guidance computation grow exponentially with the number of unresolved positions. We introduce COFFEE, a plug-and-play framework that avoids this enumeration by separating sequence dependence from the objective. At each diffusion step, a target-free carrier absorbs the marginal token distributions predicted by the denoiser to construct a joint model over the unresolved tokens, while a compiled finite-state model records how their combinations affect the sequence-level preference. Pairing their states allows COFFEE to transfer global preferences to unresolved positions and sample a clean reconstruction without retraining the diffusion model. The same framework supports explicit hard constraints and learned soft objectives. We evaluate COFFEE across multiple symbolic, language, and biological benchmarks, where it achieves strong control results with task-dependent quality and diversity trade-offs. By making objectives available to inference rather than only evaluation, COFFEE brings joint conditioning, completion-weighted guidance, and optimization-based constraints into pretrained neural generation, showing the potentials of neural-symbolic methods in diffusion guidance.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Discrete Diffusion Models, Probabilistic Circuits, Graphical Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_2fPmOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents COFFEE, this is a method for plug-and-play inference time guidance for discrete diffusion models which works with denoiser token predictions and a target independent probabilistic carrier to model dependencies between tokens. This method works by compiling sequence level constraints and rewards into finite-state representations which allows efficient dynamic programming to account for future completions and then samples jointly guided sequences without retraining the diffusion model. This method is evaluated with symbolic reasoning, language generation, and biological sequence design, against a comprehensive set of baselines.

Critical strengths
  1. COFFEE demonstrates strong empirical results across a diverse range of domains and tasks, including symbolic reasoning, language generation, and biological sequence design. This is important for demonstrating broad applicability.
  2. The paper provides extensive comparisons against a broad set of existing guidance methods, including recent approaches. These comparisons strengthen the empirical evidence supporting COFFEE's effectiveness relative to prior work.
  3. The evaluation on multiple pretrained diffusion backbones, including LLaDA and Dream, provides evidence of the method's transferability demonstrating its not limited to a particular model architecture.
  4. COFFEE can be integrated with pretrained diffusion models without modifying or retraining their parameters. This makes the approach attractive for applying new constraints and objectives to existing models.
Critical weaknesses
  1. A of COFFEE is the trade-off between objective optimization and generation diversity. Specifically, on DeepSTARR, the method produces only a single unique sequence, while on K562, only 32 distinct sequences are generated among 1,000 samples. This could limit the practical usefulness of the method for biological sequence design. It is worth noting that this limitation is explicitly acknowledged by the authors and does not directly negate the otherwise convincing empirical evidence presented in the paper, but would benefit from additional experimentation (See questions for authors).

  2. COFFEE presents a useful framework, several of its core ideas are very directly built upon existing approaches. For example:

  • Ctrl-G: Combined probabilistic sequence models with finite-state automata and dynamic programming for constrained generation, closely resembling COFFEE's use of paired states and backward inference to guide future completions.

  • CDD: Introduced inference-time constraint enforcement through projection during discrete diffusion denoising, establishing an earlier approach to steering diffusion trajectories toward satisfying constraints.

  • CoDD: Introduced structured probabilistic modeling of token dependencies in discrete diffusion, which COFFEE directly builds upon through its target-independent dependence carrier.

COFFEE does well in effectively integrates joint dependency modeling with compiled objectives, and demonstrates strong empirical results. However, its methodological novelty appears somewhat incremental, as the core techniques are established in prior work.

Questions for authors
  1. It would be helpful for the authors further clarify the main methodological advances of COFFEE relative to existing approaches such as CoDD and Ctrl-G. For example, would be helpful for the authors to make more clear what capabilities does COFFEE introduce more than combining established dependency modeling and finite-state inference techniques.

  2. Given the significant diversity collapse observed on DeepSTARR and K562, the authors would benefit from much deeper investigation of the trade-off between objective optimization and generation diversity. These results would be very helpful.

  3. A few of the experiments, such as Sudoku and Dyck repair, involve substantial task-specific construction of feasible solution spaces. Could the authors further isolate the benefits of COFFEE's joint guidance mechanism from those of the underlying feasibility construction. An additional ablation of this would be informative.

6 comments

We thank the reviewer for recognizing our empirical results across diverse domains, broad baseline comparisons, and plug-and-play transferability across pretrained diffusion backbones (LLaDA, Dream, D3LM, and EvoDiff). Your questions directly shaped Sections 3.3 and 3.4 and our clarified methodological positioning in Section 3.5 and Appendix J (highlighted in sapphire blue in the revised PDF).

W2 and Q1. Why COFFEE is defined by the scattered-mask query and advances both communities (Revised Secs. 2.3 and 3.5; App. J, Table 23)

COFFEE is defined by the scattered-mask inference query rather than an additive pairing of existing parts. Combining an efficient density model and an efficient constraint scorer does not automatically make their expectation tractable under arbitrary query patterns (Khosravi et al., 2019), yet parallel discrete diffusion requires evaluating E_q₀[W_C(X₀) | x_t, X₀ⁱ = u] across scattered masked positions M_t where multiple non-adjacent tokens commit simultaneously. Answering this query in a single O(L · A · |V| · H²) sweep requires aligning carrier transitions and objective states on a shared product graph sᵢ = (hᵢ, aᵢ) (Eq. 13), advancing both communities in distinct ways.

1. Neurosymbolic and probabilistic-circuit advances (Ctrl-G, TRACE, CoDD; App. J, Table 23). Autoregressive neurosymbolic controllers (Ctrl-G, TRACE) never face scattered parallel masks because unresolved tokens form a contiguous suffix with one known boundary; additionally, Ctrl-G handles only hard deterministic text constraints, while TRACE uses local token-linear weights that miss multi-token biological motifs (nZVs Q2). COFFEE unifies both via weighted regular-language Aho-Corasick automata supporting hard and multi-token soft constraints over parallel diffusion masks. Conversely, whereas CoDD and Discrete Copula Diffusion model unconditional token dependencies without a control objective, COFFEE stacks a modular objective graph over a CoDD-style dependency carrier (App. C.1, Table 11), where the carrier acts as a data prior curbing reward exploitation during parallel commits.

2. Discrete-diffusion guidance advances (CDD, DG-TAG). Continuous simplex projections (CDD) and predictor gradients (DG-TAG) update relaxed per-position marginals without tracking discrete automaton states, which leads to hard grammar violations in repair (Table 3), enhancer mode collapse (Table 2), and homopolymer reward hacking on 3IXT protein scaffolding (Table 35). By running exact backward-forward DP over discrete states, COFFEE coordinates parallel commits and composes finite-state diversity rules (HR16/HR24, B_t) in one pass without per-step backpropagation.

Method familyDecoding topologySupported objectivesParallel coordination & diversity invariants
Ctrl-G / TRACELeft-to-right AR suffixHard DFA only (Ctrl-G) or token-linear (TRACE)Sequential single-token commits; no parallel diffusion
CoDD / Discrete CopulaParallel discrete diffusionNone (unconditional dependency modeling only)Captures joint over M_t; no reward or constraint layer
CDD / DG-TAGParallel discrete diffusionContinuous simplex projection or predictor gradientsRelaxed marginals; breaks hard repair & repeat bounds
COFFEE (Ours)Parallel scattered masks M_tUnified hard DFA + multi-token soft regular weightsExact product-state DP + composable HR16/HR24 & B_t

Table note. Detailed capability matrix across all baseline families appears in Appendix J (Table 23).

Q2 and W1. Optimization-diversity trade-off and inference-time remedies (New Sec. 3.3, Table 2; App. K, Tables 24 to 28)

Evaluating Diversity = N_unique / N over official 249-bp (DeepSTARR) and 200-bp (K562) windows shows that COFFEE maintains near-ceiling uniqueness across ten of twelve main settings without extra penalties (App. M.2, Table 36), confirming that diversity loss is not a general side effect of DP lookahead. Instead, concentration arises strictly in unconstrained DeepSTARR and K562 from two separable factors, namely neural enhancer predictors assigning extreme scores to poly-T repeats and cooled replay (τₐ = 0.3) sharpening onto those peaks. Because Eq. 13 acts over discrete states, two inference-time automata target distinct layers of diversity without retraining. HR16 and HR24 enforce within-sequence homopolymer limits, while the exclusion bank (B_t) prevents exact duplication of previously delivered evaluator windows.

1. Separating replay temperature from oracle reward bias (App. M.2, Table 36). Restoring neutral temperature (τₐ = 1) leaves our ten diverse tasks steady while lifting decoded uniqueness markedly on K562; by contrast, DeepSTARR remains concentrated because its oracle peaks on poly-T tracts even at unit temperature.

2. Composable homopolymer bounds (HR16, HR24) and batch deduplication (B_t, Sec. 3.3, Table 2; App. K). Composing 16-bp or 24-bp repeat bounds with six-mer enhancer weights (GATA-1 and HNF4A motifs in Table 41, using a validation-calibrated soft scale on K562) prevents poly-T shortcuts during backward message passing and redirects score mass onto diverse regulatory motifs above competing baselines at natural genomic perplexity. Independently, freezing a prefix-tree exclusion bank (B_t) over delivered evaluator windows avoids exact cross-sample duplicates in N=32 and growing N=128 batches (accelerated cleanly via vectorized state projection), and the two automata compose directly when both repeat caps and zero batch duplicates are desired.

Diagnostic or N=1,000 methodDS Target%DS Valid HitsDS LOOK562 Target%K562 Valid HitsK562 LOO
Other 10 main settings (Table 36)N/AN/AN/AN/A97.5% to 100% uniqueN/A
CDD / D-CBG (HR16 valid hits)100.0% / 5.1%159 / 511.20 / 1799.742.5% / 100.0%423 / 170.1 / 1.20
DG-TAG / MDM-VGB (HR16 valid hits)99.2% / 60.8%441 / 5972.60 / 56.697.2% / 84.7%818 / 8353.40 / 55.5
COFFEE soft only (τₐ = 0.3, HR16)100.0% (all-T 5.31)0 (1 uniq)1.0599.5%5 (28 uniq)1.08
COFFEE soft only (τₐ = 1, uniq /1000)100.0% (score 5.28)20 uniq1.0685.1%638 uniq1.84
HR16 hard DFA only (W_soft ≡ 1)13.9%139 (174 uniq)1.3243.3%357 (914 uniq)2.99
COFFEE + HR16 (Table 2)96.5%819 (841 uniq)2.7399.9%846 (847 uniq)3.93
COFFEE + HR24 (App. K)99.3%946 (950 uniq)2.9199.9%833 (MDM-VGB 845)4.00
Duplicate-exclusion bank (B_t)DS UniqueDS Target%DS LOOK562 UniqueK562 Target%K562 Speedup
Empty (N=32) to Growing (N=128)1/32 to 128/128100% / 99.2%1.05 / 1.222/32 to 128/128100% / 96.1%8.58× (LOO 1.30)

Table note. DS denotes DeepSTARR (249-bp window, target 1.17) and K562 uses 200-bp views; full 7-baseline HR16/HR24 and bank ablations appear in Sec. 3.3 (Table 2) and App. K (Tables 24 to 28).

Reviewer_knxyOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper introduces COFFEE, a training-free framework for guiding discrete diffusion models toward sequence-level objectives without computationally infeasible enumeration. The proposed procedure combines neural token predictions with finite-state models to support both hard constraints and soft preferences. The result is validated in data examples across symbolic, language, and biological tasks.

Critical strengths

COFEEE avoids exponentially expensive enumeration of possible sequence completions and is computationally efficient.

The procedure can supports both hard constraints and learned soft objectives.

Critical weaknesses

Notation of this paper wasn't clear. It is a bit hard to follow. Line 304 "the sampling conditional is proportional to" this seems not a natural or standard noun or concept.

The performance seems wasn't that as impressive compared to benchmark. More comprehensive study might be helpful.

"Due to the Markov property, any two partial paths ending in this state have the same weighted completions." It is unclear to me how this conclusion is reached.

Questions for authors

Questions would be how we could further clarify and strengthen the weaknesses.

4 comments

We thank the reviewer for highlighting the computational efficiency of COFFEE and its unified support for both hard constraints and learned soft objectives, and for candidly flagging where our notation, exposition, and empirical summary needed sharper presentation. We tightened imprecise phrasing and added explicit mathematical links throughout Sections 1 and 2, presented exact normalized sampling conditionals at Eq. 19, and added two focused experimental subsections (Sections 3.3 and 3.4), with all revisions highlighted in sapphire blue in the uploaded manuscript.

Q1. Clarified notation, explicit CoDD objective, and normalized conditionals at line 304 (Revised Secs. 1 and 2, Eqs. 19 and 20)

In Sections 1 and 2, we replaced informal phrasing (including the Figure 1 caption and Introduction transitions), stated our target-free joint-construction objective explicitly under the CoDD masked-reconstruction loss L_CoDD(ω) (Sec. 2.2), and replaced the unnormalized proportional form at preprint line 304 with normalized closed-form conditionals (Sec. 2.4, Eqs. 19 and 20).

1. Composite product-state setup. Let hᵢ₋₁=h denote the carrier state and aᵢ₋₁=a denote the compiled automaton state before emitting position i, forming the composite boundary state (hᵢ₋₁=h, aᵢ₋₁=a). Choosing candidate token u ∈ V and next carrier state h′ deterministically updates the objective state to a′ = δ_{C,i}(a, u), with composite edge weight K̄_{C,i}(h, a, u, h′) = Kᵢ(h, u, h′) · ψ_{C,i}(a, u) and denoiser marginal e_{t,i}(u).

2. Algebraic cancellation of Bᵢ(h′) and direct evaluation via M. In the preprint's tilted factor form, the base continuation term Bᵢ(h′) appeared in both numerator and denominator and cancels algebraically, while the backward message Mᵢ₋₁(h, a) from Eq. 15 already equals the exact normalizing partition sum over all (u, h′). Dividing by Mᵢ₋₁(h, a) gives the normalized joint token-state conditional (revised Eq. 19),

q_C^joint(x₀ⁱ=u, hᵢ=h′ | hᵢ₋₁=h, aᵢ₋₁=a, x_t) = [ e_{t,i}(u) · Kᵢ(h, u, h′) · ψ_{C,i}(a, u) · Mᵢ(h′, a′) ] / Mᵢ₋₁(h, a),

whose sum over h′ yields the normalized next-token marginal conditional q_C^joint(x₀ⁱ=u | h, a, x_t). Revised Eq. 20 then multiplies and divides by Bᵢ(h′) solely to clarify its interpretation as expected future objective weight; practically, forward sampling evaluates Eq. 19 directly from M without a separate B pass. If any remaining passages, symbols, or line numbers still feel unclear, we would appreciate the reviewer pointing them out so we can clarify them further.

Q2. Overall empirical performance, benchmark breadth, and structural capabilities (Table 1; New Secs. 3.3 and 3.4, Tables 2 and 3; App. B.2)

We appreciate this concern and clarify why both our benchmark evidence and the capabilities unlocked by product-state inference extend beyond incremental scalar gains. Evaluating one unchanged sampling mechanism across eight benchmarks, four serious backbones including billion-parameter masked diffusion language models (Dream 7B and LLaDA 8B), and seven ported baseline families reveals strong performance alongside qualitative capabilities that relaxed continuous guidance and unguided solvers cannot provide individually.

1. Breadth across large backbones and standardized speed (Sec. 3.1, Table 1; App. B.2 and F.1). Whereas prior studies mostly evaluated smaller text backbones or narrower constraint sets, COFFEE finishes first or second in every main control setting (ten firsts and two seconds) under matched budgets and disjoint splits, while achieving the best mean steady-state guided runtime rank across thirteen RTX 4090 settings at unguided Base VRAM.

2. Capabilities beyond scalar benchmarks (New Secs. 3.3 and 3.4, Tables 2 and 3). Rather than relying on task-specific decoding heuristics, discrete product-state compilation enables two practical capabilities within the same backward-forward pass. On regulatory DNA (Sec. 3.3), finite-state diversity rules separate within-sequence repeat caps (HR16/HR24 homopolymer bounds) from cross-sample exact deduplication (B_t exclusion bank) to curb poly-T collapse without retraining. On Python bracket repair (315 HumanEval instances across k ∈ {1, ..., 5} masked slots) and unobserved all-position repair (Sec. 3.4), pairing compiled syntax graphs with denoiser likelihoods maintains full validity where gradient methods break grammar rules and more than doubles known-slot unit-test passes over unguided solvers.

Main benchmark rank (Table 1, 12 settings)First placeSole firstTied firstSecond place
Target control ↑10 / 12642 / 12 (Top-2 in 12/12)
Reference quality ↓6 / 12422 / 12 (Top-2 in 8/12)
Benchmark or rebuttal suiteStrongest baseline or solverCOFFEE (Ours)Revised PDF
Matched RTX 4090 speed (13 settings)MDM-VGB 2nd (rank 2.77)Rank 2.38 (1st); 2.98× steady, 2.43× coldApp. B.2, Tables 6 to 8
DeepSTARR / K562 active + HR16 (/1000)MDM-VGB 597 / 835; DG-TAG 441 / 818819 / 846 (LOO 2.73 / 3.93)Sec. 3.3, Table 2; App. K
Exclusion bank (N=32 / growing N=128)Empty bank (N=32) 1/32 and 2/32 unique32/32 and 128/128 unique (93.8% to 100% hit)Sec. 3.3; App. K, Table 27
Known-slot Python unit tests (/315)Uniform 71 (22.5%); CPU solver 66 (21.0%)179 (56.8%) (all 100% syntax valid)Sec. 3.4, Table 3; App. L
All-pos Python pass (/128, Dream / LLaDA)Solver 30 / 30; DG-TAG 54% / 47% valid33 / 33 (100% rule valid on both)Sec. 3.4, Table 3; App. L

Table note. Full HR24 DNA diversity, exclusion-bank speedups, and all-position language repair breakdowns appear in Appendices K and L.

Q3. The missing step from the Markov property to path sharing (Revised Secs. 2.2 and 2.3, Fig. 2; App. A.4, Table 5)

You are right that a logical step was skipped between the two halves of that sentence. The core insight is that linear-time dynamic programming depends on state sufficiency between prefix and suffix; neither raw token sequences under nonlocal rules nor factorized denoiser outputs are Markovian on their own, but lifting the state space to the paired product state sᵢ = (hᵢ, aᵢ) forms a minimal Markov blanket between prefix (1 ... i) and suffix (i+1 ... L) so exponentially many distinct histories share one cached continuation value.

1. Explicit prefix-suffix independence step (Secs. 2.2 and 2.3, Eqs. 8 and 13, Fig. 2(B)). In Eq. 6, every carrier factor after position i depends on the prefix x₀^(1:i) solely through hᵢ; partial paths ending in the same carrier state hᵢ therefore face identical suffix completions with identical weights and reuse Bᵢ(hᵢ) (Eq. 8). Once compiled objective transitions are included (Eq. 13), all downstream carrier, objective, and terminal weights depend on the prefix solely through sᵢ = (hᵢ, aᵢ), allowing distinct prefixes landing in (hᵢ, aᵢ) to share Mᵢ(hᵢ, aᵢ) (Fig. 2(B) shows four prefixes merging at M₄(2, a_init)).

2. Concrete five-token worked example (App. A.4, Table 5). Consider length L=5 over alphabet {A, B} with positions 1 through 4 masked and x₅ = A observed, paired with a two-state HMM carrier (h ∈ {0, 1}) and a soft bigram automaton (a ∈ {∅, A}) multiplying path weight by 3 whenever AB occurs. At i=3, prefix ABA along carrier trajectory (0, 1, 0) (already earning one AB bonus) and prefix BBA along (1, 0, 0) (earning none) differ in tokens, carrier history, and prefix weight, yet both land in s₃ = (h=0, a=A). Both therefore share M₃(h=0, a=A) = 11,879 / 24,000 and induce the identical next-token distribution at position 4. Crucially, dropping aᵢ conflates prefixes ending in A versus B and flips the preferred next token, whereas dropping hᵢ conflates distinct future carrier dynamics, showing why both coordinates of (hᵢ, aᵢ) are essential for exact path sharing.

State configuration at i=3 (App. A.4, Table 5)Cached continuation M₃Pr(x₄ = A) at s₃Pr(x₄ = B) at s₃
Exact product state (h=0, a=A) (shared by ABA and BBA)11,879/24,000 (0.4950)34.3%65.7%
Alternate automaton state (h=0, a=∅)6,677/24,000 (0.2782)61.0%39.0%
Alternate carrier state (h=1, a=A)13,659/24,000 (0.5691)29.8%70.2%

Table note. Exact rational and decimal values from the five-token worked walkthrough in Appendix A.4 (Table 5).

Reviewer_nZVsOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper adds sequence control to frozen discrete diffusion models by combining denoiser evidence with a dependence carrier and a objective. The carrier supplies a joint distribution over unresolved tokens, and the objective represents hard acceptance conditions or learned additive token and pattern rewards. Backward dynamic programming on their graph computes continuation weights, then joint ancestral sampling of a complete reconstruction for the host diffusion update. The paper evaluates symbolic constraints, lexical coverage, detoxification, DNA, and protein motif scaffolding. Its practical results also reveal important limitations, including reliance on precomputed feasible support in symbolic tasks, quality trade-offs in language, and severe concentration or complete collapse in some DNA outputs.

Critical strengths
  1. The method addresses a failure of independently sampling guided token marginals. Results directly supports the need for coherent parallel commitments.

  2. The fixed step inference seems mathematically well specified. The implementation on local temperature is explicitly distinguished from globally tempering path weights, and resource failures are separated from infeasibility. These details improve the credibility of the limited formal claim.

Critical weaknesses
  1. Exact fixed-step inference does not imply exact guided diffusion or preservation of the original model's joint law. The carrier constructs a surrogate q0 from denoiser evidence, and the implementation truncates token candidates and state support. Main language selection also searches ancestral temperatures in {0.3, 0.6, 0.9}, whereas the exact replay theorem requires 1. The objective-approximation proposition assumes uniform log-weight error on a fixed support and does not bound support errors or error accumulated across diffusion updates. These are substantial limitations for interpreting exact reasoning. Practical objective construction is an additional modeling problem. The uniform log-error condition in Proposition 2.1 is not empirically established, especially on the optimized distribution where the DNA collapse occurs.

  2. Multiple claims cannot be fully supported by existing experiments. The DNA evidence contains severe failures of useful design. Appendix F shows that DeepSTARR produces one distinct all-T sequence across 1,000 outputs, while K562 produces only 32 distinct token sequences. These outcomes weaken the biological contribution. Broad speed claims are also stronger than what the table supports. The rows mix GPU environments and native batches, which the authors themselves warn are descriptive. A matched hardware/batch/precision benchmark is needed before claiming generally superior inference speed. Main paper statistical support and biological generality are also limited. The stronger bootstrap analysis is on reused CommonGen validation records, while Table 1 mainly gives single point estimates. Some final panels were previously evaluated during the study and are explicitly not fresh prospective holdouts. Protein evidence uses one motif 1YCR and the same OmegaFold evaluation family for both structural metrics.

Questions for authors
  1. How much do the main results change with unit ancestral temperature and progressively expanded retained support?

  2. What benefit remains over a CoDD-style carrier with a minimal hard-automaton controller or a tractable soft-lookahead controller?

  3. Can authors compare against solver-only and uniformly weighted feasible-support samplers with complete construction costs included?

  4. Do DeepSTARR and K562 improvements survive homopolymer restrictions and a requirement to produce multiple successful designs?

9 comments

We thank the reviewer for recognizing our problem motivation, coherent parallel-commitment evidence, and fixed-step mathematical formulation, and for prompting us to state the exactness scope and empirical boundaries of our framework more precisely. All new experiments and clarifications are integrated into the revised manuscript (highlighted in sapphire blue) and detailed below.

W1. Exact fixed-step inference, multi-step diffusion, and the original joint (Revised Secs. 2.2 to 2.4, 3.5; App. A.7, C.1)

Our exactness result applies to a fixed reconstruction query under the constructed carrier and compiled objective, with unit ancestral temperature and full support. Within that step-level query (Proposition A.2), backward-forward message passing samples q_C(x₀ | x_t) ∝ q₀(x₀ | x_t) · W_C(x₀) directly. When W_C encodes an exact indicator constraint (as in Sudoku, Dyck, and CommonGen) and the parameter-free identity carrier (H=1, used in APARENT, HepG2, K562, and Protein) sets q₀ equal to the denoiser product, fixed-step inference matches the conditioned denoiser law at that step. Across diffusion steps, Eq. 21 performs iterative guided reconstructions under q_host rather than MCMC stationary to a trajectory tilt. Revised Section 2.4 distinguishes three sources of departure between intentional modeling choices and algorithmic approximations.

1. Constructing a data-closer joint carrier q₀ (Sec. 2.2; App. C.1, Tables 10 and 11). Pretrained masked diffusion denoisers emit conditionally independent marginals ∏_{i ∈ M_t} p_θ(x₀ⁱ | x_t) across simultaneously unmasked tokens, omitting inter-token coupling that can invite degenerate joint assignments. Rather than mimicking that factorized product, our canonical CoDD formulation defines L_CoDD(ω) = E[ −log q₀(x₀ | x_t; ω) ], while deployed carriers follow distinct asset-specific recipes (App. C.1, Table 10). Specifically, LLaDA tracks denoiser-product joint NLL and DeepSTARR fits a clean-sequence order-4 Markov prior, whereas Dream reports standalone sequence-prior HMM NLL with an unpinned historical runtime package and higher normalized composed-query NLL under a fresh fixed-corruption audit. Because standalone prior fit does not establish composed masked-reconstruction improvement, downstream benefit rests on our paired ablation (LLaDA · RTP, Table 11), where the learned HMM improves safety and fluency by curbing reward exploitation during parallel commits.

2. Data-driven objective compilation and theoretical boundary (Secs. 2.3 and 3.5). Translating downstream specifications into tractable surrogates is standard in guided generation; COFFEE uses inspectable, composable weighted automata fitted from data. We acknowledge the tension between Proposition 2.1's uniform ℓ_∞ log-weight error assumption (which is uncertified on fitted surrogates and homopolymer tails, and is never treated as an empirical certificate) and our strong empirical performance with lightweight automata (Q2). Developing tighter distribution-dependent bounds remains an important theoretical direction, while our claims rest on independent held-out evaluators and unseen ESMFold.

3. Finite support (K, S) and ancestral temperature τₐ (Sec. 2.4; App. M.2). Retaining K token candidates and S active states bounds memory and latency on large vocabularies, while τₐ ∈ (0, 1] locally tempers replay transitions. Omitting τₐ = 1 from {0.3, 0.6, 0.9} was an oversight for which we apologize; we evaluate τₐ = 1 and expanded (K, S) in Q1.

Carrier audit (App. C.1, Tables 10 and 11)Baseline or initialFitted carrier
Dream 7B standalone prior / LLaDA 8B joint NLL ↓381.9 / 94.2196.9 / 89.5
Dream 7B composed-query audit / DeepSTARR clean NLL ↓131.6 / 1.386154.9 (audit) / 1.280
LLaDA · RTP non-toxic% (3,072 draws) ↑91.3% (H=1 identity)93.1% (+1.82 pp, CI
+0.88,+2.74+0.88, +2.74

) | | LLaDA · RTP external PPL ↓ / native LOO ↓ | 24.8 / 7.49 (H=1 identity) | 22.7 / 5.80 (H=64 HMM) |

Table note. The Dream composed-query audit (131.6 to 154.9 vs. teacher 131.4 nats/window) uses freshly reconstructed initialization and fixed corruptions on identical windows; APARENT, HepG2, K562, and EvoDiff protein scaffolding use an exact parameter-free identity carrier (H=1) with zero fitting error.

W2 (Part 1). Diversity scope, matched runtime, and bootstrap intervals (App. B.2, F.1)

1. Biological diversity scope (Sec. 3.3; Q1 and Q4). Uniqueness already stays high across ten of twelve main settings without extra penalties, showing that collapse is not inherent to DP guidance. Instead, concentration occurs strictly on unconstrained DeepSTARR and K562 due to poly-T enhancer-oracle optima amplified by cooled replay, and is mitigated at inference time in Q4.

2. Standardized latency and memory (App. B.2, Tables 6 to 8). Re-evaluating Base and all seven guided families across thirteen settings on identical RTX 4090 GPUs under matched window and minibatch caps separates true algorithmic cost from codebase differences. The underlying efficiency insight is that compiling constraints once and executing one vectorized GPU DP pass scales with automaton state transitions rather than per-step predictor backpropagation (DG-TAG, CDD), candidate rollouts (D-CBG, GILC-DB), or particle trajectory replication (CDM). Consequently, COFFEE ranks first in steady-state latency and ties for first on cold requests at unguided Base peak memory (outpacing Python decoding loops in Base DNA, while paying modest subword state-tracking time on open-ended 7B/8B LLMs in exchange for full concept satisfaction).

Metric (13 settings)COFFEEMDM-VGBCDMDG-TAGGILC-DBD-CBGCDD
Steady / cold rank ↓2.38 / 2.692.77 / 2.694.08 / 4.084.54 / 4.544.54 / 4.464.54 / 4.385.15 / 5.15
COFFEE faster (steady)Ref. (60/78)7/1312/1310/1310/1310/1311/13
Steady / cold ratioRef. (2.98× / 2.43× faster)1.57× / 1.31×2.61× / 2.11×2.65× / 2.21×3.18× / 2.57×4.45× / 3.68×4.57× / 3.57×
Max VRAM Bio / LLM (GiB)0.15 / 16.20.27 / 16.23.92 / 21.60.15 / 16.98.04 / 16.90.93 / 17.10.53 / 16.9

Table note. Ratios divide comparator latency by COFFEE latency (Base max VRAM is 0.15 / 16.2 GiB; per-task latencies appear in App. B.2, Tables 6 and 7).

3. Disjoint splits and percentile bootstrap intervals (App. F.1, Table 20). Carrier fitting, surrogate regression, and DEV selection use partitions strictly disjoint from evaluation panels (retaining Appendix D's panel-history disclosure). Percentile bootstrap intervals across all twelve settings indicate exact ties on formal puzzles and consistent positive gains on language, enhancer DNA, and protein scaffolding. The lone negative interval occurs on APARENT under compact default support (K=16), reflecting early candidate pruning over a large 4⁶ = 4,096 six-mer alphabet that reverses past DG-TAG once K widens in Q1.

Setting (App. F.1, Table 20)COFFEEStrongest full-panel baseline95% CI (COFFEE minus baseline)
Dyck / Sudoku (Dream & LLaDA)100.0%DG-TAG / CDM 100.0%Exact tie /
−0.22,+0.22-0.22, +0.22

pp | | CommonGen (Dream / LLaDA) | 100% / 100% | CDM 57.2% / GILC-DB 41.8% | **

+40.2,+45.2+40.2, +45.2

/

+55.7,+60.7+55.7, +60.7

pp** | | RTP (Dream, matched 1,024 stress) | 77.6% | MDM-VGB 62.4% (DG-TAG 54.6%) | **

+12.1,+18.4+12.1, +18.4

(

+19.6,+26.5+19.6, +26.5

) pp** | | RTP (LLaDA) / Protein-1YCR | 91.6% / 28.2% | CDD 81.7% / DG-TAG 14.5% | **

+7.23,+12.5+7.23, +12.5

/

+8.50,+19.0+8.50, +19.0

pp** | | DeepSTARR / K562 (vs. DG-TAG) | 5.31 / 99.8% | DG-TAG 3.64 / 97.2% (D-CBG 1 uniq) | **

+1.63,+1.71+1.63, +1.71

/

+1.60,+3.70+1.60, +3.70

pp** | | APARENT (K=16, corr. ×100) | 21.5 | DG-TAG 24.3 |

−3.62,−2.14-3.62, -2.14

(surpasses at K=64, Q1) |

Table note. Conditional on frozen systems and declared panels; D-CBG reaches 100.0% on K562 by collapsing onto a single unique sequence (Q4).

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Reply to All Reviewers

We thank all reviewers for their careful reading and constructive evaluations of our manuscript (ICLR 2027 Submission 3982). We appreciate that reviewers found fixed-step inference "mathematically well specified" (Reviewer nZVs), noted that COFFEE "avoids exponentially expensive enumeration of possible sequence completions and is computationally efficient" while supporting "both hard constraints and learned soft objectives" (Reviewer knxy), and recognized our "strong empirical results across a diverse range of domains and tasks" and plug-and-play transfer across "multiple pretrained diffusion backbones, including LLaDA and Dream" (Reviewer 2fPm).

Core Insights and Contributions Across Reviews

Our central insight is that guided parallel discrete diffusion hinges on coordinating simultaneous commitments across scattered masks with exact future lookahead. Evaluating one unified, training-free graphical-model formulation across domains yields five contributions supported by the manuscript and rebuttal studies (highlighted in sapphire blue in the revised PDF).

1. Cross-domain scope on billion-parameter backbones (Sec. 3.1, Table 1; App. F.1, Table 20). Prior discrete diffusion guidance papers evaluated isolated modalities, small text models, or only hard or only soft constraints, leaving cross-paradigm failure modes hard to diagnose. Porting seven guidance families onto symbolic puzzles, executable repair, open-ended text (Dream 7B and LLaDA 8B), regulatory DNA (D3LM), and multi-motif protein scaffolding (EvoDiff) under matched diffusion budgets and disjoint splits establishes a unified testbed where COFFEE ranks first or second in every main control setting while remaining competitive on reference quality. Placing study scopes side by side clarifies how this suite broadens empirical comparison for the community.

StudyConstraintsSymbolicText (backbone)DNAProtein
CDDHard onlyWord count/posDetox (≤169M)N/AN/A
CDMSoft onlyN/ADetox (15p), align (8B)1 enhancerFold co-design
D-CBG / D-CFGSoft onlyN/ALM only* (≤139M)1 enhancer*N/A
DG / TAGSoft onlyN/AN/A1 enhancerStability
GILCSoft onlyN/AN/A1 enhancerStability
MDM-VGBHard or SoftSudoku, Dyck1-char ban (0.6B)1 enhancer1 motif (1 eval)
COFFEE (Ours)Hard + SoftSudoku, Dyck, CodeCommonGen, RTP (7B/8B)3 tasks + diversity3 motifs (2 evals)

Table note. Entries marked with * evaluate unconditional perplexity or synthetic labels rather than downstream task evaluators; several prior works also evaluate small molecules or discrete image tokens.

2. Single-pass efficiency in a standardized hardware environment (App. B.2, Tables 6 to 8). Competing guidance methods differ sharply in test-time scaling because some backpropagate through predictors at every step (DG-TAG, CDD), expand candidate rollouts (D-CBG, GILC-DB), or replicate particle trajectories (CDM). Compiling constraints once and running a vectorized backward-forward sweep over denoiser marginals avoids external gradient and rollout loops, achieving the lowest mean steady-state guided runtime rank (and tied first on cold requests) across thirteen RTX 4090 benchmark settings while keeping peak GPU memory at the unguided backbone level.

3. Single-pass composition of hard rules and soft potentials without retraining (Secs. 2.3, 3.3, and 3.4). Compiling exact finite-state rules and learned soft pattern potentials into one product graph (Eq. 13) lets hard constraints and soft preferences interact inside a single joint distribution rather than competing through outer-loop heuristics, enabling diversity-constrained enhancer design (Sec. 3.3) and syntax-preserving code repair (Sec. 3.4) without modifying host denoisers.

4. Separate inference-time controls for within-sequence repeats and batch deduplication (New Sec. 3.3, Table 2; App. K, Tables 24 to 28). Because guidance runs over an explicit discrete state space rather than relaxed continuous marginals, practitioners can compile distinct counting-based diversity rules alongside soft biological potentials without retraining. Specifically, HR16 and HR24 enforce within-sequence homopolymer limits, while the exclusion bank (B_t) prevents exact duplication of previously delivered evaluator windows (Sec. 3.3).

5. Dual-community methodological bridge (Revised Secs. 2.2 to 2.4 and 3.5; App. J, Table 23). Rather than simply combining existing modules, COFFEE solves a different query (Eq. 13). For neurosymbolic circuits, it extends compiled regular-language lookahead (Ctrl-G, TRACE) from left-to-right decoding to scattered parallel masks under combined hard and weighted objectives; for discrete diffusion, it shows that modular graphical-model carriers (CoDD) act as a learned structural prior that curbs reward exploitation by factorized marginals.

Recurring Questions Across Reviews

1. Diagnosing and resolving biological mode collapse (nZVs W2, Q1, Q4; 2fPm W1, Q2; New Sec. 3.3, Table 2). Uniqueness already stays high across ten of twelve main settings without extra penalties (App. M.2, Table 36), showing that DP guidance does not intrinsically cause collapse. Mode collapse occurs specifically on unconstrained DeepSTARR and K562 because neural enhancer predictors assign peak scores to poly-T tracts, which cooled ancestral replay (τₐ = 0.3) further amplifies. Restoring unit temperature (τₐ = 1) separates replay sharpening from oracle bias, while compiling HR16/HR24 homopolymer caps limits within-sequence repeats and the exclusion bank (B_t) prevents exact duplication of previously delivered evaluator windows.

2. Distinguishing constraint feasibility from semantic disambiguation (nZVs Q3; 2fPm Q3; knxy Q2; New Sec. 3.4, Table 3). On purely formal puzzles (Sudoku, Dyck), any valid completion is equally acceptable, so CPU solvers match full validity rapidly (App. L.1, Table 29) and Table 1 strictly tests whether parallel diffusion preserves feasibility. Once syntactic legality no longer determines meaning (315 HumanEval Python repair instances across k ∈ {1, ..., 5} masked slots, plus unobserved code and text repair), solvers select arbitrary valid strings while gradient baselines frequently break hard grammar; weighting the shared feasibility graph by denoiser marginals recovers functional semantics without breaking syntax.

Overview of Sapphire-Blue Manuscript Revisions and Headline Outcomes

We uploaded a revised manuscript (10-page main text plus expanded appendices, all edits in sapphire blue) adding Section 3.3 (Table 2) on biological diversity, Section 3.4 (Table 3) on semantic repair, clarified derivations in Sections 1, 2.1 to 2.4 (Eqs. 8, 19, 20, Fig. 2), and 3.5, and Appendices A.4, B.2, C.1, F.1, J, K, L, and M (Tables 5 to 8, 10, 11, 20, and 23 to 41). The table below pairs each suite with its main numerical summary and exact PDF reference.

Evaluation suiteCOFFEE outcomeStrongest baseline or controlRevised PDF location
Main benchmark control (12 settings)Top-2 in 12/12 (10 firsts, 2 seconds)7 guided families ported (95% CIs in Table 20)Sec. 3.1, Table 1; App. F.1, Table 20
Main benchmark quality (12 settings)Top-2 in 8/12 (6 firsts, 2 seconds)Dream 7B, LLaDA 8B, D3LM, EvoDiffSec. 3.1, Table 1
Matched RTX 4090 speed (13 settings)Rank 2.38 (1st of 7); 2.98× steady, 2.43× cold60/78 steady wins; Base-level peak VRAMApp. B.2, Tables 6 to 8
DeepSTARR active + HR16 / HR24 (/1000)819 / 946 (LOO 2.73 / 2.91)MDM-VGB 597 / 604; DG-TAG 441 / 511; CDD 159Sec. 3.3, Table 2; App. K, Tables 24, 26
K562 active + HR16 / HR24 (/1000)846 / 833 (99.9% hit, LOO 3.93 / 4.00)MDM-VGB 835 / 845; DG-TAG 818; D-CBG 1Sec. 3.3, Table 2; App. K, Tables 25, 26
Exclusion bank (N=32 / growing N=128)32/32 and 128/128 unique (93.8% to 100% hit)Empty bank (N=32) 1/32 (DS) and 2/32 (K562)Sec. 3.3; App. K, Tables 27 and 28
Known-slot Python repair (/315 tests)179 (56.8%), 100% syntax validUniform feasible 71 (22.5%); Solver 66 (21.0%)Sec. 3.4, Table 3; App. L, Table 30
All-pos Python repair, Dream / LLaDA (/128)33 / 33 (100% rule valid)Solver 30 / 30; Uniform-DP 19 / 19; DG-TAG 54%/47% validSec. 3.4, Table 3; App. L, Table 31
3-motif protein scaffolding (/600)291 Ω (1st), 186 Ω ∧ ESM (2nd)CDD 222 Ω, 189 Ω ∧ ESM; DG-TAG 242 Ω, 165 Ω ∧ ESMApp. M.1, Tables 34 and 35

Table note. Ranks in Table 1 exclude archival report-only entries (Dream RTP CDM, marked with †) and include exact full-budget ties (6 sole firsts and 4 tied firsts on control, 4 sole firsts and 2 tied firsts on quality), with point-by-point responses below.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.