acceptodds
Under review as a conference paper at ICLR 2027

Blackboard Intelligence Can Surpass Autoregressive on Globally Constrained Problems

Abstract

Next-token prediction has driven remarkable progress in large language models, yet a growing body of evidence suggests that they can struggle on problems governed by complex global constraints. In this work, we focus on this regime and ask whether some of these limitations arise from the inference interface induced by next-token prediction itself. We study this question through **blackboard intelligence**: an inference-time perspective in which a model works on a fixed, revisable canvas and searches over candidate solution states rather than committing to a causal, left-to-right trajectory. We instantiate this idea with diffusion language models, whose any-order prediction interface naturally exposes predictions over partially filled solution states. Our key observation is that mean confidence, a simple model-internal quantity available from the standard masked diffusion objective, provides a useful proxy for global coherence and can guide inference-time search and revision. Empirically, across ZebraLogic, Nurse Rostering, and Job-Shop Scheduling, Blackboard consistently improves inference while holding the fine-tuned LLaDA-8B-Instruct checkpoint fixed and substantially outperforms same-scale autoregressive baselines, reaching accuracy on ZebraLogic-Hard, exact feasibility on Nurse Rostering, and optimality on JSSP. Stronger autoregressive search and refinement also fail to close the gap on ZebraLogic-Hard, while Blackboard surpasses tested frontier LLMs there and on JSSP despite their substantially greater scale and strong test-time reasoning. These results suggest that, for globally constrained problems, diffusion language models can offer a distinct problem-solving interface.

Primary area:
generative models
Keywords:
Diffusion Language Models, Any-Order Generation, Globally Constrained Problems, Constraint Satisfaction, Combinatorial Optimization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_wEKXOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies inference-time control for fine-tuned masked diffusion language models (LLaDA-8B) on globally constrained problems, including ZebraLogic, Nurse Rostering, and JSSP. The authors use mean confidence over masked tokens to trigger corrective inference, guide lookahead search and backtracking, and selectively invoke Best-of-N sampling. The approach improves over standard greedy decoding on the same LLaDA checkpoint and outperforms matched-scale autoregressive baselines.

Critical strengths
  • The paper is clear and easy to follow.

  • Selectively triggering additional inference using late-phase confidence is a simple and practical idea. It substantially reduces NFE compared to unconditional intervention while leaving most already-correct greedy trajectories untouched.

  • The paper evaluates its approach on well-chosen tasks spanning constraint satisfaction and combinatorial optimization.

Critical weaknesses

1. Separating the known benefits of adaptive-order decoding from Blackboard's contribution

A central theme of the paper is that dLLMs can outperform AR models on globally constrained problems. However, prior work has already demonstrated the benefits of adaptive-order unmasking on puzzles such as Sudoku and ZebraLogic ([1, 2]). In this paper, most of the gap between LLaMA-8B and LLaDA-8B is already present under standard confidence-greedy decoding (+43.8 pp on ZebraLogic-Hard and +59.2 pp on Nurse Rostering), compared to Blackboard's additional gains of +12.0 pp and +3.4 pp. These additional gains are meaningful, but the framing should more clearly distinguish Blackboard's contribution from the established benefits of adaptive-order decoding.

[1] Ye et al., Beyond Autoregression: Discrete Diffusion for Complex Reasoning and Planning, in ICLR 2025.

[2] Kim et al., Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions, in ICML 2025.

2. The Gibbs bound does not justify mean confidence as a search criterion

In Section 3.2, the paper motivates mean confidence C(z)\mathcal{C}(z) as a proxy for global coherence through the Gibbs bound

\leq \frac{|M|}{\beta}\psi(\mathcal{C}(z)).$$ However, this bound only controls the expected excess potential above $V^\star(z)$, the minimum potential attainable from the current state, rather than $V^\star(z)$ itself. When comparing search branches, what matters is whether a partial state can still lead to a valid or optimal solution. A wrong branch ($V^\star(z)>0$) can still have $\mathcal{C}(z)\approx 1$ if the remaining completions concentrate on a single least-bad solution. Conversely, a valid branch ($V^\star(z)=0$) can have lower confidence if probability is split across multiple optimal completions. Mean confidence can therefore favor a dead-end branch over a valid one. These limitations may also help explain some of the empirical results. On Nurse Rostering, unconditional confidence-guided correction overwrites already-correct rosters and reduces accuracy from 73.0% to 57.4% (Table 16), raising concerns about the reliability of local search decisions. On JSSP, where multiple optimal solutions exist, the pre-test trigger precision is only 21.4% (Table 21). These results raise questions about mean confidence as both a branch-ranking score and a failure-detection signal, neither of which is justified by the Gibbs bound. >**3. Lack of direct validation of confidence-guided reasoning trajectories** The paper presents BACKTRACK and SEARCH as ways to maintain global coherence, but does not directly evaluate the quality of their local decisions. For CSPs like ZebraLogic and Nurse Rostering, a useful reference would be the deduction trace of an efficient solver that derives later assignments from earlier ones through explicit, deterministic rules. Comparing the model's commitment order with this reference would help determine whether its search follows the problem's deductive dependencies rather than simply selecting high-confidence cells. Since both tasks also have a unique ground-truth solution $x^\star$, the authors could directly check whether BACKTRACK rejects incorrect fills and whether SEARCH selects branches consistent with $x^\star$. Without these analyses, it remains unclear whether confidence-guided search actually captures the logical structure of the problem. >**4. Overstated "revisable canvas" framing** The paper emphasizes a "persistent, revisable canvas" where earlier decisions can be revised, but Algorithm 1 implements a much more limited form of revision. When the trigger fires, the model discards the greedy output and restarts with a blank answer table. During the second pass, previously accepted cells are never revisited; BACKTRACK only cancels the immediately proposed fill. On JSSP and ATSP, mean confidence is not used to rank partial search states at all. It only triggers Best-of-10 sampling, where the best candidate is selected using the exact task objective (makespan or tour length). These experiments demonstrate confidence-guided inference and selective computation, but not the general canvas revision implied by the framing.
Questions for authors
  1. On ZebraLogic-Hard and Nurse Rostering, how closely does the model's unmasking order align with the deduction trace of an efficient reference solver? When the current partial state is consistent with the unique solution x⋆x^\star, how often does BACKTRACK reject an incorrect versus a correct fill? Similarly, how often does SEARCH select a branch consistent with x⋆x^\star when such a branch exists among the candidates?

  2. In SEARCH, how does mean confidence compare with alternative scoring rules, such as cumulative rollout log-probability or mean predictive entropy, under the same candidate pool and search budget?

  3. Could you clarify how NFE is counted in SEARCH? Section 4.1 describes branching over the top-kk candidates at each masked cell and running a depth-dd rollout per branch. With repeated SEARCH calls, a naive cost of ∣M∣kd|M|kd candidate evaluations per call could amount to thousands per puzzle, making the reported averages of 129 NFE for Blackboard and 280 for always-on (Table 1) surprisingly small. Could you provide the exact SEARCH pseudocode and clarify the branching scope and NFE accounting?

5 comments
AuthorsOpenReview ↗

Response to Reviewer wEKX - Part1

We appreciate the reviewer's careful technical scrutiny of our work and the detailed questions concerning Blackboard's foundations and implementation. These comments motivated us to further examine the proposed inference mechanisms through additional controlled experiments, direct decision-level validation, and implementation clarifications.

Our responses address five key concerns:

  • Contribution beyond adaptive-order decoding (W1): Distinguishing established adaptive-order gains from Blackboard's additional state-dependent inference capabilities.
  • Gibbs analysis and mean confidence (W2 / Q2): Clarifying the theoretical scope and directly evaluating confidence-based search and selective intervention.
  • Inference mechanism (W3 / Q1): Validating deductive dependencies and the decisions made by SEARCH and BACKTRACK.
  • Revisable canvas (W4): Clarifying the scope of revision and its correspondence to the implemented procedures.
  • SEARCH implementation and cost (Q3): Providing pseudocode and detailed NFE accounting.

W1. Separating Adaptive-Order Decoding from Blackboard

We agree that prior work [Ye et al., 2025; Kim et al., 2025] has established the benefits of adaptive-order decoding. Importantly, adaptive-order commitment is a foundational component of Blackboard, rather than a separate or competing mechanism. As stated in our original submission (Lines 67–73), our central claim is that the standard masked diffusion objective provides a state-level inference interface for problem solving under complex global constraints. Building on confidence-based adaptive decoding, which we already recognize as a standard dLLM inference strategy (Lines 131–139), Blackboard exploits any-order prediction and partial-state conditioning to enable state-dependent search, commitment deferral, and correction. This is particularly valuable when individual assignments are coupled through global constraints, where locally plausible decisions can compromise the feasibility of the entire solution. Our paper already quantifies these additional gains over adaptive greedy decoding using the same fine-tuned checkpoint (Lines 380–391). To further clarify these contributions, we conducted a forced left-to-right control using the same LLaDA weights:

InferenceZLNRJSSP (% opt.)
Forced left-to-right64.444.864.0
Adaptive greedy78.473.070.0
Blackboard90.476.480.2

The first improvement captures the established benefits of adaptive-order commitment (+14.0 / +28.2 / +6.0 pp), while the second demonstrates the additional gains from Blackboard's state-dependent inference procedure (+12.0 / +3.4 / +10.2 pp). These gains are empirically distinguishable but mechanistically connected. Together, they demonstrate how the dLLM's any-order inference interface and state-level confidence signal enable richer inference-time problem solving beyond greedy decoding on globally constrained problems.


W2 / Q2. The Gibbs Bound and Empirical Utility of Mean Confidence

We acknowledge that our Gibbs analysis does not guarantee global feasibility or optimal branch selection. Rather, it provides theoretical motivation for mean confidence as a state-level inference signal. As defined in Section 3.2 (Lines 218–226), the bound concerns excess potential relative to the best completion compatible with the current partial state; Appendix B (Lines 785–788) explicitly notes its idealized nature and the distinction from model-derived confidence.

Building on this theoretical motivation, the practical question is whether model-derived mean confidence provides a useful signal for guiding inference. Although individual confidence judgments may favor incorrect branches, Blackboard uses this signal across evolving partial states and multiple inference stages to guide commitment, deferral, search, and selective intervention. We examine its empirical utility through the following additional experiments.

AuthorsOpenReview ↗

Response to Reviewer wEKX - Part2

W2 / Q2. The Gibbs Bound and Empirical Utility of Mean Confidence (Continued)

(1) Alternative SEARCH scoring (Q2). We compared scoring rules using identical candidate pools, rollout depth (d = 3), and per-call evaluation budgets, changing only the SEARCH score.

Triggered puzzles solved (NFE)ZL (n = 92)NR (n = 98)
Mean confidence (ours)76.1% (558)21.4% (522)
Negative mean entropy72.8% (560)19.4% (532)
Cumulative rollout log-probability18.5% (1466)15.3% (1343)
Log-probability + confidence-gated commitment23.9% (1677)12.2% (1696)

Mean confidence performs comparably to negative mean entropy but substantially outperforms cumulative rollout log-probability. Both mean confidence and negative mean entropy evaluate the predictive state induced by a candidate continuation, aggregating confidence or uncertainty across the remaining masked positions rather than scoring the likelihood of the sampled path. In contrast, log-probability-selected paths fail to increase remaining-state confidence in 58% / 78% of SEARCH calls (ZL / NR), leading to repeated search and substantially higher NFE without comparable accuracy gains. Applying the same confidence-based commitment rule does not close this gap, highlighting the value of state-level evaluation under globally coupled constraints.

(2) Selective intervention (NR and JSSP). Beyond individual commitments, Blackboard uses late-phase trajectory-level mean confidence to decide whether additional inference is warranted, avoiding unnecessary intervention on reliable greedy trajectories.

On NR, unconditional correction reduces accuracy from 73.0% to 57.4%, whereas selective triggering achieves 95.9% precision for greedy failures and improves accuracy to 76.4%, with 19 recoveries and only 2 regressions. On JSSP, confidence instead triggers additional sampling, with the exact task objective selecting the final solution. Despite imperfect trigger precision (21.4% on training traces; 41.2% on test), retaining the greedy schedule prevents regressions, while 41 triggered instances improve to optimality.

Together, these results show that mean confidence need not perfectly evaluate every state to be useful. By exploiting this signal across evolving partial states and inference stages, Blackboard coordinates search and selective correction to improve end-to-end performance despite imperfect individual judgments.


W3 / Q1. Validating State-Dependent Inference Decisions

We directly evaluated commitment and search decisions against the unique ground-truth solution x∗x^* on ZebraLogic-Hard (ZL) and Nurse Rostering (NR).

(1) Deductive structure and error propagation.

We ran exact constraint propagation from the currently committed cells to examine whether the model's commitment order reflects logical dependencies.

ZLNR
Commits a propagation-forced cell when available98.8%89.2%
Commit accuracy on cells not yet resolved by propagation80.3%94.7%

The model frequently commits deductively determined cells when available, though not in the solver's exact sequence. Incorrect commitments have cascading consequences: after the first wrong fill, greedy's next-commit error rate rises from 1.6% to 63.5% (ZL) and from 1.5% to 50.9% (NR), consistent with the constraint propagation observed in Appendix D.3. This motivates evaluating alternative partial states rather than relying on a single irreversible greedy trajectory.

(2) Coordinated commitment and correction.

On triggered puzzles, the α-gate coordinates BACKTRACK and SEARCH using mean confidence over evolving partial states. BACKTRACK commits or defers proposed fills according to their effect on predictions across remaining masked positions, while SEARCH evaluates alternative 3-step continuations. Together, they adapt commitment order and guide corrective inference.

BACKTRACK regulates commitment rather than classifying correctness. It defers 29.6% / 73.0% of correct proposals (ZL / NR), alongside 0/1 and 24/45 incorrect proposals, respectively. Importantly, 100% / 98.6% of deferred correct fills are eventually completed correctly, illustrating how postponed commitments can be resolved as the partial state evolves.

SEARCH selects ground-truth-consistent branches in 85.7% / 34.1% of cases (random: 19.0% / 19.3%). On ZL, all 13 observed potential incorrect greedy commitments routed to SEARCH were corrected.

Together, these results show how Blackboard exploits evolving partial states to coordinate commitment, deferral, and corrective search, rather than merely selecting locally high-confidence cells.

AuthorsOpenReview ↗

Response to Reviewer wEKX - Part3

W4. Scope of the Revisable Canvas

As discussed in W1–W3, Blackboard refers to a state-level inference interface enabled by the joint use of any-order prediction, partial-state conditioning, and model-internal mean confidence, rather than unrestricted revision of previously committed tokens. Its central capability is to evaluate and adapt decisions over evolving partial solutions under globally coupled constraints. For ZebraLogic and Nurse Rostering, Algorithm 1 implements this perspective through selective restarting, local backtracking, and search over alternative partial-state continuations. BACKTRACK tentatively fills a cell and reverts (remasks) it if mean confidence over the remaining masked positions decreases. While previously accepted commitments are not arbitrarily revisited, this local revision, together with SEARCH, allows Blackboard to adapt the commitment order and redirect inference before potentially unfavorable decisions become irreversible, as supported by our decision-level analysis in W3/Q1. For JSSP and ATSP, confidence serves a different role: it determines when additional complete-solution sampling is warranted, while the exact task objective selects among candidates. These experiments instantiate confidence-based selective computation rather than partial-state revision.

Thus, Blackboard is an inference-time perspective that supports different forms of state-dependent computation, with task-specific implementations rather than a single universal revision operator.


Q3. SEARCH Implementation and NFE Accounting

The corrective procedure can be summarized as follows:

Correction (restart from blank answer region):
  while masked cells remain:
    evaluate current partial state          # 1 NFE

    if mean confidence < α:                 # SEARCH
      for each masked cell i and top-k value v:
        evaluate depth-d continuation       # ~d NFE
      accept best improving continuation,
      or fall back to greedy commitment

    else:                                   # BACKTRACK
      evaluate candidate commitments
      accept a non-decreasing-confidence fill,
      or invoke SEARCH if none is accepted

One NFE corresponds to one model forward pass. As the reviewer correctly notes, a SEARCH call evaluates the top-k values at each remaining masked position through depth-d rollouts, requiring approximately |M|kd model forward passes, all of which are included in the reported NFE. The total inference cost is:

NFE = NFE_greedy + I(trigger) × NFE_correction

Indeed, corrective inference can be computationally expensive: on ZebraLogic-Hard, triggered puzzles incur an average of 558 NFE. However, Blackboard's confidence-guided inference strategy selectively allocates this additional computation. Adaptive greedy decoding already constructs solutions through state-dependent commitments, while mean confidence determines when further correction is warranted.

On ZebraLogic-Hard, 81.6% of puzzles retain their greedy solutions, with only 18.4% triggering correction. Among these, SEARCH is invoked approximately 2.1 times on average, with its cost decreasing as fewer masked positions remain. Consequently, Blackboard achieves an overall average of 129 NFE, substantially below the cost of applying corrective search to every puzzle.


Taken together, these additional analyses clarify Blackboard's contribution beyond adaptive-order decoding, provide direct evidence for the practical utility of state-level mean confidence, and validate its role in coordinating commitment and corrective search.*

We hope this evidence, together with the clarified theoretical scope and implementation details, addresses the reviewer's soundness concerns and supports reconsideration of the overall assessment.

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

This paper proposes an inference-time framework for solving globally constrained problems with masked diffusion language models. It treats a partially completed structured answer as an editable state and uses mean confidence over the remaining masked positions to decide when to apply additional search or correction. The implementation combines selective lookahead and rollback for ZebraLogic and nurse rostering, and confidence-triggered sampling with makespan-based selection for job-shop scheduling. Experiments show improvements over the same task-fine-tuned diffusion checkpoint. I find the idea simple and effective, but the current evaluation does not yet establish how broadly it applies or its advantage over a sufficiently comprehensive set of autoregressive decoding approaches.

Critical strengths
  1. The approach is intuitive and makes practical use of an existing model signal. Using confidence to decide whether a solution needs further work is appealing, especially because unnecessary correction can damage an already correct answer. The method does not require an additional learned feasibility evaluator.
  2. The same-checkpoint improvements provide meaningful evidence that the controller is useful in the evaluated settings. The selective-intervention ablations are particularly informative: they suggest that allocating computation carefully can be more effective than correcting every instance. This supports the practical value of the core idea within the tested tasks.
Critical weaknesses
  1. The task coverage remains concentrated in a relatively narrow setting. The main experiments use task-specific fine-tuning and highly structured, largely synthetic problems with compact answer representations. The appendix broadens the task list, but the additional tasks remain closely related forms of structured constraint solving. This leaves uncertainty about robustness across different constraint structures, less templated representations, and larger problem sizes, even within the paper's intended domain. The results demonstrate that the method can work well on selected benchmarks; they provide less evidence that the same design transfers reliably across globally constrained problems. The interventions differ across tasks, making general applicability harder to assess. ZebraLogic and rostering use local search and correction, whereas JSSP relies on additional complete samples ranked by a task objective. Trigger statistics and thresholds are also task dependent. These choices may be reasonable, but it is unclear how much adaptation is needed for a new problem or which characteristics determine the appropriate intervention. The appendix's observation that local lookahead is ineffective on Latin-square scheduling further suggests that the applicable regime needs clearer empirical characterization.
  2. The autoregressive comparisons are useful but do not sufficiently isolate the proposed framework's comparative advantage. Matching downstream training settings does not remove differences between the pretrained LLaDA and LLaMA models. The revision experiments also ask a model trained for direct answer generation to operate through a critique-and-rewrite interface without revision-specific training. Consequently, a gap could reflect the starting models or training/interface mismatch as well as the decoding framework. A stronger comparison would give autoregressive approaches competitive search or editing procedures, suitable training where needed, and comparable inference budgets. The existing positive autoregressive search results on JSSP and ATSP make this distinction especially important. The accuracy–compute comparison is not yet comprehensive enough to support the broader decoding comparison. NFE is informative within similar diffusion implementations, but a cached autoregressive step and a full-state diffusion evaluation have different costs. The existing timing and FLOP measurements are helpful; a common-budget comparison across the principal diffusion and autoregressive alternatives would make it clearer whether the proposed controller offers a favorable practical trade-off.
Questions for authors
  1. Clarification: Which existing results most strongly support transfer to different constraint structures, representations, and problem sizes with limited controller retuning? In particular, what explains when local correction works and when whole-solution sampling is required?
  2. Clarification: How should the reported autoregressive comparisons be interpreted after accounting for pretrained-checkpoint differences, revision-specific training, and actual inference cost? A controlled comparison that isolates these factors would materially affect my assessment of the claimed decoding advantage.
5 comments
AuthorsOpenReview ↗

Response to Reviewer AwNg - Part1

We appreciate the reviewer's recognition of the practical value of selective intervention and the meaningful improvements achieved within a fixed model. We also recognize the concerns regarding the scope of applicability and the interpretation of autoregressive comparisons, which we address through further clarification and additional controlled experiments. Our responses focus on two main concerns:

  • Generality and task adaptation (W1 / Q1): Clarifying the target problem regime, the task characteristics guiding corrective operators, and the extent of required adaptation.
  • Autoregressive comparisons and inference cost (W2 / Q2): Examining checkpoint differences, expanded AR search budgets, and within-checkpoint accuracy–latency trade-offs.

W1 / Q1. Generality and Task Adaptation

Our study focuses on structured globally constrained problems, where complex, interacting constraints require progressively evaluating and refining partial solutions. As motivated in our paper (Lines 55–66, 176–189), autoregressive models struggle in this regime, motivating Blackboard's any-order, state-level inference approach.

Two task characteristics guide the choice of corrective operator, depending on where mean confidence is most informative:

  1. Feasibility problems. In densely coupled constraint settings such as ZebraLogic and Nurse Rostering, local commitments substantially affect predictions over remaining variables, making remaining-state mean confidence useful for guiding partial-state search and correction. In rotating-shift rostering, however, incorrect local assignments can remain plausible under Latin-square constraints, making local lookahead less effective. Here, confidence-based ranking of complete candidate solutions is more useful (Appendix I.1).

  2. Optimization problems. In JSSP and ATSP, multiple feasible solutions can differ in quality according to an exact objective. We therefore use late-phase mean confidence to determine when additional complete-solution sampling is warranted, while the exact objective selects the best candidate.

The required adaptation is modest within our evaluated settings. We use a common validation-based pipeline to select task-specific parameters, with settings fixed before test evaluation. In practice, the main inference hyperparameters are largely shared across tasks (k = 5, d = 3 for partial-state search; N = 10 for complete-solution sampling). We then evaluate generalization on held-out test sets, without test-time retuning.

Thus, Blackboard's generality lies in exploiting the same state-level inference interface across structured globally constrained problems, while adapting the corrective operator to where mean confidence is most informative under each task's constraint structure.

AuthorsOpenReview ↗

Response to Reviewer AwNg - Part2

W2 / Q2. Autoregressive Comparisons: Checkpoint Differences and Inference Cost

We appreciate the reviewer's concerns regarding pretrained-checkpoint differences, revision training, and inference cost. Beyond these cross-model considerations, we would like to emphasize a central aspect of our evidence: how much inference-time control can improve problem-solving performance while holding the underlying model fixed. This within-checkpoint perspective is particularly relevant to our interface-level hypothesis.

We extended LLaMA beam search to B = 128 and evaluated Blackboard at different search widths on the same LLaDA checkpoint. The following results compare within-checkpoint improvements on ZebraLogic-Hard:

ModelInferenceAccuracyΔ vs. GreedyLatency
LLaMA-8B (NTP)Greedy34.6%–2.40 s
Beam-6435.0%+0.4 pp5.59 s
Beam-12835.0%+0.4 pp14.5 s
LLaDA-8B (masked diffusion)Adaptive greedy78.4%–1.5 s
Blackboard (k=5)90.4%+12.0 pp6.5 s
Blackboard (k=7)90.8%+12.4 pp8.2 s

Applying Blackboard itself yields substantial gains over adaptive greedy using identical LLaDA weights, whereas LLaMA accuracy plateaus despite substantially increased inference budgets. Temperature sampling up to N = 64 likewise remains at 34.6–35.4%. Nurse Rostering shows a similar pattern (LLaMA: 13.8% → at most 14.8%; LLaDA: 73.0% → 76.4%).

For optimization tasks such as JSSP and ATSP, AR search can also improve performance because complete candidates can be ranked by an exact objective. Here, Blackboard provides a complementary capability: state-level confidence selectively determines when additional sampling is warranted, avoiding unnecessary computation while benefiting from higher-quality candidate solutions (Appendix F.4, Table 20).

Together, these results support our interface-level hypothesis: the dLLM's any-order prediction and state-level internal confidence enable richer inference-time control, including corrective search and selective computation, beyond standard greedy decoding. The central contribution is not simply higher absolute accuracy, but the additional problem-solving capability enabled within a fixed model.


We hope these clarifications and additional experiments address the reviewer's main concerns. In particular, they demonstrate the applicability of Blackboard's state-level inference principle across structured constraint settings and its substantial within-checkpoint gains, contrasted with limited improvements from increased autoregressive inference budgets. We would appreciate reconsideration of the overall assessment in light of this new evidence.

AuthorsOpenReview ↗

Manuscript Revision Update

To address the concerns raised in your review, we have incorporated the clarifications and additional analyses from our rebuttal into the revised manuscript. The main revisions are as follows:

  • Generality and task adaptation (W1/Q1): We clarified the common state-level inference-control framework and its task-specific instantiations in §4.1 (L330–379), explaining how the choice between partial-state correction and complete-candidate exploration depends on task structure, with supplementary transfer studies in Appendix I.
  • Controlled inference comparisons (W2/Q2): We incorporated the forced left-to-right and always-on Blackboard controls in §4.3 (Table 1, L449–467), together with same-checkpoint comparisons against existing dLLM inference methods in Table 5 and Appendix D.6.1 (L1335–1403).
  • Autoregressive comparisons and inference cost (W2/Q2): We clarified the scope of autoregressive search and refinement comparisons in §4.2 (L384–436) and Appendix D.5 (Tables 9–10), including hardware-matched compute analysis, and added corrective-inference pseudocode and SEARCH cost accounting in Appendix D.6.2 (Algorithm 2, L1404–1426).

These revisions clarify how a shared state-level inference principle supports task-appropriate search, correction, and selective computation, while providing a more precise basis for interpreting the experimental comparisons.

We thank the reviewer again for the constructive feedback and welcome further questions or discussion regarding the revised manuscript and our responses.

Reviewer_4D5sOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether the inference interface of masked diffusion language models is useful for problems with strong global constraints. The main idea is to treat generation as working on a fixed and revisable canvas, called Blackboard Intelligence. The authors use the mean confidence over the remaining masked positions as a model-internal signal for deciding when to trust a trajectory, when to search, and when to revise. The method is evaluated on ZebraLogic, Nurse Rostering, and Job-Shop Scheduling. With the same fine-tuned LLaDA-8B checkpoint, Blackboard improves over standard diffusion inference on all three tasks and also performs better than the tested same-scale autoregressive baselines.

Critical strengths
  1. I like the paper's main perspective. Instead of treating diffusion language models mainly as a way to decode faster, the paper asks what kind of intermediate state the model exposes during problem solving. The fixed and revisable canvas is a natural fit for globally constrained tasks, and mean confidence is a simple signal that does not require an external verifier.

  2. The empirical study is fairly thorough. The most convincing comparison is within the same LLaDA checkpoint, where Blackboard improves ZebraLogic-Hard from 78.4% to 90.4%, Nurse Rostering from 73.0% to 76.4%, and JSSP optimality from 70.0% to 80.2%. The paper also includes useful ablations, difficulty breakdowns, uncertainty estimates, stronger autoregressive inference baselines, and additional transfer studies. I also appreciate that the authors openly discuss several limitations, including the idealized nature of the theory, the task-dependent corrective operators, and the fact that the frontier-model comparisons are not compute normalized.

Critical weaknesses
  1. The novelty is mostly in the overall framing, combination, and evaluation rather than in every individual component. Any-order decoding, remasking, uncertainty-guided lookahead, mean-confidence branch scoring, and search over partial diffusion trajectories all have close precedents. The paper acknowledges this, which I appreciate, but the incremental difference from accuracy-oriented methods such as LookUM, Order-Token Search, and POKE-SMC could still be clearer. It would help to compare against at least one such method using the same checkpoint and a similar inference budget, or explain why a direct adaptation is not meaningful for these structured outputs. I do not view this as a fatal issue, but it affects how I assess the methodological novelty.

  2. The experiments support the usefulness of Blackboard inference, but they do not fully isolate the broader claim that the inference interface itself explains the gap between autoregressive and diffusion models. LLaMA and LLaDA differ in more than their decoding interface, including their pretraining objectives and learned weights. The within-LLaDA greedy-to-Blackboard comparison cleanly validates the proposed inference method, while the stronger interface-level interpretation remains suggestive rather than causal. Similarly, the Gibbs analysis is useful motivation, but it is stated for confidence under the true conditional distribution and bounds excess potential relative to the best completion compatible with the current partial state. It therefore does not guarantee that the model-derived confidence identifies a globally feasible state. Could the authors clarify the intended strength of this claim and, if possible, provide a same-backbone control such as forced left-to-right filling or disabling revision?

Questions for authors

Please see the Critical Weaknesses section.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 4D5s - Part1

We sincerely thank the reviewer for the thoughtful and encouraging assessment, particularly for recognizing the value of the Blackboard perspective and our within-checkpoint improvements. We greatly appreciate the constructive suggestions, which motivated additional controlled experiments and empirical analyses to strengthen our claims.

Our responses address two main concerns:

  • Methodological novelty (W1): Comparing Blackboard with existing dLLM inference methods on the same checkpoint and clarifying its contribution beyond adaptive greedy decoding.
  • Interface-level interpretation and theoretical grounding (W2): Providing forced left-to-right controls, expanded AR search comparisons, and empirical validation of mean confidence while clarifying the scope of the Gibbs analysis.

W1. Comparison with Existing dLLM Inference Methods

To directly address the reviewer's suggestion, we reimplemented LookUM, Order-Token Search, and POKE-SMC on the same fine-tuned LLaDA checkpoint and evaluated them on the full ZebraLogic-Hard test set. Hyperparameters were selected using a separate training-split validation set.

InferenceAccuracyX-LargeNFELatency
Greedy78.4%51.2%351.5 s
POKE-SMC (M=4)76.8%48.8%862.2 s
LookUM (k=8)77.4%47.2%1413.4 s
LookUM (k=16)77.8%48.0%2826.7 s
Order-Token Search (K=8)79.4%53.6%41310.2 s
Blackboard90.4%75.2%1296.5 s

Under the evaluated same-checkpoint settings, existing methods provide little or no improvement over adaptive greedy decoding, despite comparable or greater inference computation. In contrast, Blackboard achieves a +12.0 pp gain, with particularly large improvements on the hardest puzzles.

While prior methods primarily target general diffusion decoding settings, Blackboard focuses on structured globally constrained problems, where autoregressive next-token generation struggles with tightly coupled decisions. Building on the dLLM's any-order interface, Blackboard uses mean confidence to evaluate evolving partial states at multiple stages of inference, coordinating adaptive commitments, corrective search, and selective intervention. This unified state-level inference procedure yields substantial end-to-end improvements over adaptive greedy decoding in our target domain.

AuthorsOpenReview ↗

Response to Reviewer 4D5s - Part2

W2. Interface-Level Interpretation, Same-Backbone Control, and Gibbs Analysis

We appreciate the reviewer's observation that LLaMA and LLaDA differ in pretrained objectives and weights, as well as their recognition of our within-checkpoint improvements as the strongest empirical evidence. We therefore emphasize what can be achieved through inference-time control within fixed model weights, rather than attributing the entire cross-model gap to the inference interface.

Following the reviewer's helpful suggestion, we conducted a forced left-to-right control using the same LLaDA checkpoint:

InferenceZLNRJSSP (% opt.)
Forced left-to-right64.444.864.0
Adaptive greedy78.473.070.0
Blackboard90.476.480.2

These results distinguish two complementary gains enabled by the same inference interface: adaptive any-order commitment improves over forced left-to-right filling (+14.0 / +28.2 / +6.0 pp), while Blackboard further improves over adaptive greedy (+12.0 / +3.4 / +10.2 pp). The latter improvement builds directly on the any-order prediction and partial-state conditioning capabilities underlying the former.

For comparison, we also extended LLaMA beam search to B = 128 on the same task-fine-tuned checkpoint. Despite increasing inference latency from 2.4 s to 14.5 s, ZebraLogic-Hard accuracy improves only from 34.6% to 35.0%, with a similar plateau on Nurse Rostering (13.8% → 14.6%). In contrast, Blackboard achieves substantial within-checkpoint gains by exploiting any-order predictions and mean confidence over the remaining partial state to evaluate the broader consequences of commitments, rather than relying on sequential next-token decisions.

Regarding the Gibbs analysis, we recognize that the bound does not guarantee global feasibility. We use it as theoretical motivation for mean confidence as a state-level inference signal. To further address the reviewer's concern about its practical reliability, we conducted additional analyses directly evaluating its roles in selective intervention and partial-state search.

On held-out test sets, late-phase mean confidence detects greedy failures with 89.1% precision / 75.9% recall on ZebraLogic-Hard and 95.9% precision / 69.6% recall on Nurse Rostering. As a partial-state search signal, SEARCH selects a ground-truth-consistent branch in 85.7% of cases on ZebraLogic-Hard, compared with 19.0% under random selection. These results provide empirical evidence that mean confidence serves as an informative signal at multiple levels of Blackboard inference, guiding both selective intervention and partial-state search despite imperfect individual judgments.


Together, these results strengthen our central claim that the dLLM's any-order inference interface and model-internal state-level confidence enable more effective inference-time problem solving under complex global constraints. We thank the reviewer again for the constructive suggestions, which helped us further clarify and substantiate our contribution.

AuthorsOpenReview ↗

Manuscript Revision Update

To address the concerns raised in your review, we have incorporated the clarifications and additional analyses from our rebuttal into the revised manuscript. The main revisions are as follows:

  • Methodological contribution (W1): We clarified Blackboard's contribution beyond established adaptive-order decoding in §1 (L67–74) and §4.1 (L330–336), emphasizing state-dependent commitment deferral, corrective search, and selective reconstruction of candidate solutions.
  • Comparison with existing dLLM methods (W1): We incorporated same-checkpoint comparisons against LookUM, POKE-SMC, and Order-Token Search in Table 5 (L1146–1161), with implementation details, hyperparameter selection, and inference-cost analysis in Appendix D.6.1 (L1335–1403).
  • Interface-level interpretation and theoretical grounding (W2): We added the forced left-to-right control in §4.3 (Table 1, L449–467) to distinguish adaptive-order gains from Blackboard's additional improvements, and clarified the intended scope of the Gibbs bound in §3.2 (L218–227).

Together, these revisions clarify how Blackboard builds on the dLLM's any-order interface to evaluate, search over, and reconstruct evolving solution states, providing a clearer basis for assessing its methodological contribution and controlled inference-time improvements.

We thank the reviewer again for the constructive feedback and welcome further questions or discussion regarding the revised manuscript and our responses.

We thank again for the reviewer for their constructive feedback, and hope our answers addressed the concerns appropriately. If so, we kindly ask the reviewer for potential reconsideration of the score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response — Manuscript Revision Update

We sincerely thank the reviewers and Area Chair for their thoughtful feedback and constructive suggestions. The discussion has helped us clarify the central contribution of Blackboard and strengthen the empirical and methodological evidence supporting our work.

Our work focuses on globally constrained problems, where individual decisions are tightly coupled and locally plausible commitments can compromise the coherence of an entire solution. In this setting, Blackboard explores how the dLLM's any-order inference interface can support problem solving through the evaluation, search, and revision of evolving partial solution states.

We have incorporated the key clarifications and additional analyses from our rebuttal into the revised manuscript. The main updates are summarized below:

  • Clarifying Blackboard's methodological contribution: We refined §1 and §4.1 to distinguish Blackboard from established adaptive-order decoding, emphasizing state-level confidence as a signal for coordinating commitment deferral, corrective search, and selective intervention beyond greedy decoding.

  • Strengthening controlled inference comparisons: We incorporated additional same-checkpoint dLLM baselines and forced left-to-right controls (§4.3; Appendices D.1 and D.6), while clarifying autoregressive search, refinement, and hardware-matched inference-cost comparisons (§4.2; Appendix D.5).

  • Clarifying theoretical motivation and validating inference decisions: We refined the interpretation of the Gibbs analysis (§3.2) and incorporated alternative SEARCH-scoring comparisons and direct decision-level validation (Appendix D.6.3), examining how mean confidence can guide inference over partially specified solutions under globally coupled constraints.

  • Clarifying revisable inference, task adaptation, and computational cost: We expanded the description of state-dependent exploration, local revision, and solution-level reconstruction, including corrective-inference pseudocode and SEARCH cost accounting (§4.1; Appendix D.6.2). We also clarified how the shared inference-control principle supports task-appropriate interventions across different constraint structures (§4.1; Appendix I).

Together, these revisions strengthen our central perspective: on problems governed by complex global constraints, the inference interface matters not only for how solutions are generated, but also for how evolving candidate states can be evaluated, searched, and revised. The revised manuscript provides a clearer methodological account and more direct empirical evidence for this perspective.

We appreciate the reviewers' careful engagement with our work and hope that the revised manuscript and our detailed rebuttal responses address the main concerns raised during the discussion. We welcome any further questions or feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.