acceptodds
Under review as a conference paper at ICLR 2027

Neural Bayesian Sequential Routing

Abstract

We introduce Neural Bayesian Sequential Routing (NBSR), a framework that turns neural inference into a sequential Bayesian decision process. NBSR combines four key ideas: a persistent global knowledge oracle, a hierarchical graph of routers and experts, an explicit Dirichlet belief state updated by accumulated neural evidence, and hard belief-conditioned routing trained end-to-end with Gumbel-Softmax. Together, these components unify conditional computation, uncertainty estimation, and sequential decision-making in one architecture. The evolving belief state enables early exit, OOD abstention, cost-aware evidence acquisition, and transparent path-wise audit trails. We prove monotone growth of Dirichlet precision and bounded predictive variance under positive evidence, formalizing sequential hypothesis sharpening, and show Bayes-optimal recovery under idealized assumptions. Extensions with recurrent memory and autoregressive routing support partially observable control and Bayesian experimental design. Across vision, diagnosis, language modeling, control, and active triage, NBSR achieves competitive predictive performance while providing uncertainty-aware, interpretable, and resource-rational inference.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Bayesian deep learning, evidential deep learning, dynamic routing, mixture of experts, uncertainty quantification

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_EkagOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

NBSR routes inputs through experts that add positive evidence to a Dirichlet state. The state conditions later routing and supports early exit and abstention. The paper derives concentration properties and evaluates variants on classification, synthetic language modeling, partially observable control, and diagnostic test selection.

Critical strengths
  1. Accumulating expert contributions provides inspectable intermediate predictions and a mechanism for adaptive computation.
  2. The paper reports both positive and negative results, including poorer diagnostic calibration and reduced language-modeling throughput. These results help expose the method's limitations.
Critical weaknesses
  1. Equation 3 has an unbounded concentration incentive.For α=c1\alpha = c\mathbf{1}, the predictive mean remains uniform and the classification loss remains log⁡K\log{K}, while the Dirichlet differential entropy tends to −∞-\infty as c→∞c \to \infty. Thus, for λ>0\lambda>0, the objective can improve without improving predictions or acquiring information. Theorem 1 does not validate concentration as epistemic certainty. A likelihood interpretation for adding learned outputs from the same input is also missing.
  2. The empirical evidence is insufficient for the uncertainty and efficiency claims. The CIFAR-10 accuracy gain is 0.14% in one seed. Diagnostic ECE is worse than the MLP's, and language modeling is slower with higher perplexity. The control experiment compares NBSR's abstention with baselines required to choose movement actions, which does not isolate the benefit of its uncertainty representation. An evidential baseline is also missing.
Questions for authors
  1. What prevents the concentration scaling degeneracy in Equation 3? Can you specify the constrained or regularized objective and show that its uncertainty scores remain meaningful?
  2. How do early exit and OOD detection compare across seeds against evidential and softmax baselines with abstention tuned on a validation set?
3 comments

Thank you for the thoughtful review. We clarify guarantees, architectural contribution, and propose revisions.

  1. core contributions and novelty. NBSR makes three contributions: (a) Belief-conditioned hard routing: a persistent oracle hxh_x, an evolving Dirichlet state αt\alpha_t, and routers selecting one expert based on [hx,αt][h_x,\alpha_t], with additive evidence updates αt+1=αt+et\alpha_{t+1}=\alpha_t+e_t (Secs. 3–4). Straight-through Gumbel-Softmax enables end-to-end training of discrete routing. (b) Explicit intermediate beliefs: each visited expert contributes model-generated evidence, yielding inspectable prediction trajectories, path-wise attribution, and uncertainty-based stopping/abstention (Figs. 1–3). (c) Sequential decision extensions: NBSR-Mem incorporates recurrent memory for partially observed control, while AR-NBSR performs sequential, cost-aware diagnostic panel selection (Secs. 5.4–5.5). The novelty lies in integrating these mechanisms into a shared inference framework, not in Dirichlet conjugacy or Gumbel-Softmax individually.

  2. Concentration degeneracy and mathematical correction. We agree with the counterexample: for α=c1\alpha=c\mathbf{1}, the predictive mean remains uniform, the NLL is log⁡K\log K, but H(Dir(α))→−∞H(\mathrm{Dir}(\alpha))\to-\infty as c→∞c\to\infty. Thus Eq. (3), with λ>0\lambda>0 and unrestricted evidence, is unbounded below. Thm. 1 guarantees monotone concentration growth and a decreasing upper bound on variance, not calibrated epistemic uncertainty or information gain. We propose replacing the entropy penalty with intermediate proper log scores and bounded evidence: et,k=Emax⁡sigmoid⁡(gt,k)e_{t,k}=E_{\max}\operatorname{sigmoid}(g_{t,k}), mt=αt/∑kαt,km_t=\alpha_t/\sum_k\alpha_{t,k}, and L=−log⁡mT,y+λ∑t<T[−log⁡mt,y]+γ∑tc(vt)\mathcal L=-\log m_{T,y}+\lambda\sum_{t<T}[-\log m_{t,y}]+\gamma\sum_t c(v_t). For finite depth, total concentration is bounded by K+tKEmax⁡K+tKE_{\max}, while intermediate log scores reward early correct prediction rather than concentration inflation. Notably, medical NBSR uses λ=0\lambda=0, language omits the entropy penalty, and NBSR-Mem employs target-masked evidence regularization after the original penalty caused optimization difficulties (Sec. 5.4). We will update the objective and prioritize comparisons against λ=0\lambda=0, reporting evidence scales, calibration, and early-exit behavior under the revised formulation.

  3. Bayesian interpretation and uncertainty guarantees. The update in Eq. (2) is exact Dirichlet pseudo-count algebra: expert outputs can be represented by generalized-likelihood factors ∏kpket,k(x)\prod_k p_k^{e_{t,k}(x)}. However, successive experts often query the same oracle, so their evidence is not automatically equivalent to conditionally independent observations or a calibrated Bayesian posterior. Thm 2 establishes recovery of the Bayes-optimal predictive mean only under idealized capacity, optimization, and λ=0\lambda=0 assumptions. We will distinguish the mathematical belief-update mechanism from probabilistic calibration, and avoid interpreting growing concentration alone as proof of genuine epistemic certainty.

  4. Existing empirical evidence and limitations. On CIFAR-10, NBSR achieves 96.74% accuracy / 0.015 ECE, compared with 96.60% / 0.026 for flat ResNet-18 and 96.43% / 0.027 for sparse MoE (Table 1). Fig. 2(c) also shows separation in terminal entropy between CIFAR-10 and SVHN. These support the architecture, but do not establish calibrated epistemic uncertainty or reliable OOD detection. The current CIFAR comparison uses one seed, without OOD AUROC or matched selective-risk evaluation. Medical ECE is worse than the MLP (0.166 vs. 0.016; Table 2), and language modeling is slower with slightly higher perplexity. We do not claim improvements across every task. We will prioritize multi-seed evaluation, confidence intervals, and calibration/OOD comparisons after correcting the entropy objective.

  5. Early exit, OOD abstention, and fair baselines. CIFAR Fig. 2(b) demonstrates that nearly 90% of samples exit after one expert without reducing accuracy, supporting adaptive expert computation. However, total inference time remains comparable to the flat model (5.1 s vs. 5.0 s, Table 1), so this is not yet an end-to-end speedup. In POMDP navigation, NBSR-Mem and CNN-GRU both achieve 100% success, while NBSR-Mem abstains on all evaluated OOD hazards (Table 4). We agree that the baselines were not given an equivalent abstention option, making their 0% halt rate an unfair safety comparison. We will prioritize validation-tuned reject options for flat softmax, one-pass evidential, and MoE baselines, evaluating risk–coverage, OOD AUROC/FPR, ID false-rejection, and accuracy–latency tradeoffs across seeds. These tests will establish whether NBSR's evolving belief offers benefits beyond ordinary prediction confidence.

Overall, the mathematical issue concerns the unrestricted entropy regularizer, not the feasibility of belief-conditioned sequential routing itself. We will revise the objective and its theoretical interpretation.

Thank you for the acknowledgment of the concentration degeneracy and the clarification of the Bayesian interpretation. The log-score formulation addresses the unbounded-loss issue in principle. However, the revised manuscript states that this formulation has not been evaluated, and the reported CIFAR results retain the original objective. Proper scoring of predictive means also does not imply meaningful concentration-based uncertainty. The requested evidential baseline, and validation-tuned abstention/OOD comparisons remain future work. I therefore appreciate the corrections but remain unconvinced that the current evidence establishes the proposed benefits.

Thank you for the follow-up, and for acknowledging the corrections to the objective and Bayesian interpretation. We agree that the proposed bounded-evidence formulation has not yet been experimentally evaluated, and that proper predictive scoring does not by itself guarantee calibrated concentration-based uncertainty.

Regarding the objective, the revised manuscript explicitly identifies the unbounded direction in the original entropy-regularized loss and introduces a mathematically well-posed alternative using bounded evidence and proper intermediate log scores (Sec. 3). Importantly, this issue concerns the training objective rather than the underlying sequential routing and Dirichlet evidence-update mechanism, which remains applicable with the revised formulation. We have also clarified that the existing CIFAR results use the original objective and therefore do not establish the effectiveness of the replacement.

Regarding empirical evidence, we agree that matched evidential baselines and validation-tuned abstention/OOD evaluations are needed to establish the individual benefits of the proposed mechanisms. Nevertheless, the CIFAR-10 results (96.74% accuracy, 0.015 ECE) and adaptive-depth evaluation (nearly 90% first-stage exits at unchanged observed accuracy) provide encouraging initial evidence for the feasibility of sequential belief-conditioned computation, although they do not establish superior uncertainty calibration or computational efficiency.

We believe the methodological contribution remains useful despite these limitations. NBSR integrates hard expert routing, evolving predictive beliefs, additive evidence accumulation, and adaptive computation into a unified sequential inference framework. The revised mathematical formulation addresses the identified objective defect in principle, while the existing experiments demonstrate the feasibility of the architecture across several settings. Although further controlled evaluations are needed to establish its full benefits, we believe these contributions offer a valuable direction for neural inference, and outweigh the limitations of the current experimental validation.

Reviewer_xkfJOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a method that modifies existing model architectures to make them into a sequential Bayesian decision process. This is done through a hierarchical graph of routers and experts. This allows for uncertainty estimation and early exit. The experiments compare this approach with baselines on:

  • Visual Categorization. On CIFAR-10 the accuracy gains are relatively small with better predicted confidence over the baselines. The inference time is not faster.

  • Language modeling. The proposed approach has worse performance and throughout then the baselines with slightly better predicted confidence.

  • Medical diagnosis. The proposed method has the same accuracy as MLP but worse predicted confidence than MLP. The results show the cost performance frontier and the proposed method reaches the peak performance at lower cost, but it’s not uniformly better across different costs.

  • Navigation. There are some performance improvements in experiments without memory and no performance improvements in experiments with memory.

Critical strengths

The proposed new architecture incorporates uncertainty through sequential Bayesian updates and outputs potentially useful prediction traces. It allows for early stopping whic hcan be computationally useful.

Critical weaknesses

Weak Experiment Results.

The experiment results across the five experiments are very weak. The overall results do not demonstrate the method improves either performance or decreases the compute cost. There are some sporadic gains in performance, cost, and predicted uncertainty here and there, but nothing consistent.

The results do not establish strong accuracy gains in any of the experiments. The only experiments showing some small accuracy gains is CIFAR-10 with 0.14% improvement over the baselines and navigation experiment without memory with 0.6% improvement. The other four experiments show comparable or decreased performance compared to the baselines.

The hypothesis that this approach is improving predicted confidence is questionable due to predicted confidence being worse in some experiments, e.g., medical diagnosis so no consistent improvements there. It is unclear if the uncertainty estimates of the proposed Bayes approach are modeled well.

The language model example shows the proposed method achieves worse performance and worse throughput than the standard transformer, so it’s not clear there are benefits to having the readable reasoning trace.

Methodological Limitations.

The paper does not present an explicit probabilistic model.

Questions for authors

See the weakness plus these additional questions.

Can the authors include experiments across different seeds and statistical significance?

Do NBSR heads and the original model heads use the same number of parameters? If not, can the authors provide additional experiments keeping the number of parameters the same across modified and unmodified models?

Have the authors tried modifying the prediction head through some other ways, e.g., a standard MoE instead of the proposed sequential Bayesian decision process?

Can the authors add additional ablations isolating the effect of belief-conditioned hard routing and Dirichlet update independently.

1 comment

Thank you for the thoughtful review. We clarify the contribution, reported results, and the controls needed to isolate its benefits.

  1. core contribution and scope. NBSR is not proposed as an accuracy-only replacement for a dense head. Its novelty is the integration of (a) hard routing that uses a shared oracle hxh_x and the evolving Dirichlet belief αt\alpha_t, with one selected expert contributing et>0e_t>0 (Eqs. 1–2); (b) intermediate predictive beliefs supporting selective computation, abstention, and an additive, path-wise model-evidence record (Figs. 1–3); and (c) extensions of this interface to recurrent, partially observed control (NBSR-Mem) and sequential cost-aware panel selection (AR-NBSR; Secs. 5.4–5.5). These are architectural capabilities, not claims that every task improves accuracy, uncertainty, and latency simultaneously. We will make this distinction central, and separate the results validating each component.

  2. Explicit probabilistic model. The belief is p∼Dir(αinit)p\sim\mathrm{Dir}(\alpha_{\rm init}); for a visited expert, αt+1=αt+et(x)\alpha_{t+1}=\alpha_t+e_t(x), and the terminal predictive probability is qθ(y=k∣x,T)=αT,k/∑jαT,jq_\theta(y=k\mid x,\mathcal T)=\alpha_{T,k}/\sum_j\alpha_{T,j} (Sec. 3). The update is algebraically equivalent to multiplying the Dirichlet density by a generalized pseudo-likelihood Lt(p;x)∝∏kpket,k(x)L_t(p;x)\propto\prod_k p_k^{e_{t,k}(x)}. This is not a fitted generative likelihood p(x∣p)p(x\mid p) or proof that repeated neural evaluations of the same input are independent Bayesian observations. Theorem 1 guarantees increasing concentration and a decreasing variance bound, not calibration. We will explicitly state the pseudo-count interpretation and its assumptions rather than equating neural evidence with independent measurements.

  3. What the empirical results establish. on CIFAR-10 (Table 1), NBSR obtains 96.74% accuracy / 0.015 ECE, versus 96.60% / 0.026 for flat ResNet-18 and 96.43% / 0.027 for top-2 MoE. Accuracy gains are modest, but the ECE difference in this run and the nearly 90% first-stage exits at unchanged accuracy (Fig. 2b) are concrete evidence for the combined design. They do not show lower wall-clock inference time: NBSR is 5.1 s versus 5.0 s flat. We will report conditional expert computation separately from end-to-end latency.

  4. Other tasks: tradeoffs and contributions. Medical NBSR matches flat MLP accuracy (97.62%) but has poorer ECE (0.166 vs. 0.016; Table 2). On controlled language modeling it has worse PPL/throughput (13.81 / 37,376 tokens/s vs. Transformer 13.43 / 62,004; Table 3); its value in this experiment is the intermediate evidence trace, not speed. For navigation, NBSR-Mem and CNN-GRU both reach 100% success (Table 4), confirming the need for memory rather than an NBSR accuracy advantage. NBSR's OOD-halting result is promising, but baselines were not given a matched reject option. In active triage (Fig. 7), AR-NBSR reaches about 90.5% at roughly USD 45–50; the greedy comparator reportedly examines unpurchased outcomes, so it cannot substantiate superiority over a feasible EIG policy. We will correct that comparator description and qualify the safety/BOED conclusions.

  5. Seeds, parameters, and routing baselines. The current CIFAR result uses one seed (111), with the same pretrained, unfrozen ResNet-18, data pipeline, and 150-epoch schedule (Appendix F); this does not establish statistical significance. The heads have different structures and are not parameter-matched. A top-2 sparse MoE on CIFAR and a Transformer-MoE for language are already included (Tables 1, 3), but these do not isolate the Bayesian update. We will prioritize multiple seeds with confidence intervals, head/active-FLOP parameter matching, and validation-tuned stopping comparisons; claims will remain single-run where repetitions are unavailable.

  6. Ablations and corrected objective. To isolate the mechanism we propose matched comparisons of (i) flat evidential head, (ii) hard feature-only router, (iii) hard belief-conditioned router, and (iv) soft MoE; with and without sequential accumulation, keeping the backbone and budget fixed. We will also remove the unbounded +λH(Dir(αt))+\lambda H(\mathrm{Dir}(\alpha_t)) incentive identified in the other reviews: bounded per-step evidence combined with intermediate proper log scores is a candidate replacement, not an already validated result. We will test evidence scales, calibration, early-exit risk–coverage and latency after this correction. This directly distinguishes the architectural contribution from effects of additional parameters or concentration inflation.

Reviewer_vwedOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary
  • The paper proposes Neural Bayesian Sequential Routing, a neural architecture that combines conditional expert selection with an explicit Dirichlet belief state. Experts produce nonnegative class evidence, which accumulates across stages; routers can use the input representation and current belief to select subsequent experts. The resulting intermediate predictions support early stopping, abstention, and inspection of evidence accumulation.

  • The paper evaluates variants on classification, synthetic language modeling, partially observable control, and cost-sensitive feature acquisition. The supplied code clarifies that these experiments use different objectives and routing mechanisms: CIFAR implements repeated belief-conditioned routing, while several other variants reuse an initial routing decision.

Critical strengths
  • A potentially useful integration of evidential prediction and conditional computation. Accumulating evidence provides intermediate predictions and an explicit state for selecting subsequent computation. Although evidential learning, expert routing, and early exits have substantial precedents, their particular combination here is a plausible contribution. The CIFAR implementation provides a concrete realization with straight-through routing during training and sparse expert execution during inference.

  • Encouraging classification and calibration results. On CIFAR-10, the paper reports ECE of 0.015 versus 0.026 for the flat baseline, with comparable accuracy of 96.74% versus 96.60%. These results motivate further investigation, although additional controls are needed to attribute the improvement to sequential belief-conditioned routing.

Critical weaknesses
  • The positive differential-entropy penalty makes the stated objective unbounded below. For an objective containing +λH(Dir(α))+\lambda H(\mathrm{Dir}(\alpha)), with λ>0\lambda>0, increasing concentration can reduce the loss without improving predictive probabilities. For example, α=c1\alpha=c\mathbf{1} retains uniform predictions while the Dirichlet differential entropy tends to −∞-\infty as cc increases. This undermines the interpretation of the penalty as encouraging useful evidence or reliable uncertainty. The supplied CIFAR and acquisition implementations retain this penalty, and gradient clipping does not bound concentration. This is an objective-level concern, not evidence that the reported runs diverged. The criticism does not apply directly to medical and language experiments, which disable or omit the penalty.

  • The experiments do not adequately establish the benefit of the central routing mechanism or adaptive stopping. The medical and language implementations reuse an initial routing decision rather than selecting later experts from the updated belief. Their “Fast” thresholds also exceed the maximum possible Dirichlet differential entropy for their respective class counts, forcing every example to stop after the first evidence stage. Thus, these configurations do not demonstrate input-dependent depth. Together with the absence of matched flat-evidential and feature-only-routing ablations, this leaves unclear whether the reported improvements arise from evidence parameterization, additional capacity, or sequential belief-conditioned routing.

Questions for authors
  • What prevents concentration inflation under the positive differential-entropy penalty? Can you provide either a bounded formulation or evidence-scale diagnostics and controlled ablations showing that the reported uncertainty and stopping benefits persist when this degeneracy is removed?

  • Can you reconcile the general description of repeated belief-conditioned routing with the medical and language implementations, which reuse the initial gate? What comparisons isolate the benefit of updated-belief routing from a flat evidential head and an otherwise matched router using only backbone features?

  • For KK classes, Dirichlet differential entropy is at most −log⁡Γ(K)-\log\Gamma(K). How were the medical threshold of −100-100 and language threshold of 11 selected, given that they force first-stage exit? Do thresholds producing genuinely variable depth yield favorable accuracy–calibration–latency tradeoffs?

  • The supplied greedy baseline evaluates each patient’s actual unpurchased panel values before choosing a panel. Is this intended as an oracle comparator? If so, please distinguish it from a feasible expected-information-gain policy and clarify what the comparison establishes.

3 comments

Thank you for the careful review and for recognizing the potential of combining evidential learning with conditional computation.

  1. Core contributions. NBSR contributes an integrated sequential inference architecture, not a new Dirichlet identity or Gumbel estimator. (a) Belief-conditioned hard routing: a shared oracle hx=fϕ(x)h_x=f_\phi(x) is computed once; at step tt, a router selects one expert using [hx,αt][h_x,\alpha_t], and its positive evidence updates αt+1=αt+et\alpha_{t+1}=\alpha_t+e_t (Secs. 3–4). Straight-through Gumbel-Softmax trains discrete, path-dependent execution. (b) Explicit intermediate beliefs: each step yields a predictive distribution, precision/entropy-based stopping signals, and an additive record of the selected experts' model-generated evidence (Figs. 1–3). On CIFAR-10, repeated belief-conditioned routing achieves 96.74% accuracy / 0.015 ECE, versus 96.60% / 0.026 (flat ResNet-18) and 96.43% / 0.027 (soft MoE; Table 1). (c) Sequential decision extensions: NBSR-Mem adds recurrent memory for partially observed navigation (Sec. 5.4); AR-NBSR selects costly diagnostic panels sequentially (Sec. 5.5). These demonstrate different uses of the same evolving-belief interface, though they do not individually establish optimal control or BOED. We will foreground these 3 core contributions and specify which implementation tests each.

  2. Unbounded entropy penalty and remedy. We agree that for α=c1\alpha=c\mathbf 1, the predictive mean remains uniform while H(Dir(α))→−∞H(\mathrm{Dir}(\alpha))\to-\infty. Thus +λH+\lambda H in Eq. (3) (also used in triage) can reward meaningless concentration inflation; Softplus and gradient clipping cannot rule it out. Theorem 1 guarantees increasing concentration and a shrinking variance upper bound, not calibrated epistemic certainty or genuine information gain. A mathematically bounded replacement is et,k=Emax⁡ sigmoid(gt,k)e_{t,k}=E_{\max}\,\mathrm{sigmoid}(g_{t,k}), with St=∑kαt,k≤K+tKEmax⁡S_t=\sum_k\alpha_{t,k}\le K+tKE_{\max}, and a loss L=−log⁡mT,y+λ∑t<T[−log⁡mt,y]+γ∑tc(vt)\mathcal L=-\log m_{T,y}+\lambda\sum_{t<T}[-\log m_{t,y}]+\gamma\sum_t c(v_t), where mt=αt/Stm_t=\alpha_t/S_t. Intermediate proper log scores reward correct predictions, not mere scaling of α\alpha. This is a proposed modification, not a submitted result; medical and language omit the positive entropy penalty, and NBSR-Mem uses target-masked evidence regularization (Sec. 5.4). We will replace/qualify the defective objective, compare λ=0\lambda=0 against the bounded alternative, and report evidence-scale distributions, ECE, stopping curves, and OOD/selective-risk metrics before attributing uncertainty benefits to concentration.

  3. Which tasks use updated-belief routing? We agree that CIFAR implements repeated belief-conditioned routing, whereas medical and language reuse the initial gate. The latter provide multi-stage evidence accumulation and path records, not an experiment isolating re-routing from the updated belief. We will label those as fixed-gate variants and prioritize a matched-backbone/parameter comparison of (i) flat evidential head, (ii) feature-only hard router, (iii) belief-conditioned hard router, with the same evidence heads and training budget; we will separately compare a reused gate with a re-evaluated gate. The current MoE baseline alone does not isolate these effects.

  4. Early-exit thresholds. Since H(Dir(α))≤−log⁡Γ(K)H(\mathrm{Dir}(\alpha))\le-\log\Gamma(K), the maxima are approximately −110.32-110.32 for medical (K=41K=41) and −205.17-205.17 for language (K=65K=65). Hence η=−100\eta=-100 and η=1\eta=1 respectively force first-stage exit, rather than input-dependent depth. We will correct the Fast/Deep interpretation and evaluate thresholds within the attained entropy range, reporting depth histograms, accuracy, ECE, and wall-clock latency. CIFAR already provides a nontrivial threshold sweep (Fig. 2b): nearly 90% exit at depth 1 without an accuracy drop, but total inference time 5.1 s vs. 5.0 s for the flat network (Table 1); we will not call this demonstrated end-to-end acceleration.

  5. Triage comparator and information access. If, as indicated by the implementation review, the greedy baseline evaluates each patient's unobserved panel values before selection, it is a hindsight/outcome-aware comparator, not a feasible expected-information-gain (EIG) policy. We will audit and label it accordingly, and prioritize a feasible cost-matched greedy EIG baseline that integrates over predicted unobserved outcomes. We will report the full accuracy–cost frontier and restrict claims of superior BOED until such a comparison is available. A decrease in the model's Dirichlet differential entropy alone is not evidence of true diagnostic information gain.

Thank you to the authors for answering my questions. Their replies resolve several implementation questions, but two substantive concerns from my earlier reply remain:

  • The positive Dirichlet differential-entropy penalty admits an unbounded direction: increasing concentration can lower the loss without improving predictive probabilities.
  • The experimental evidence does not sufficiently isolate the benefit of sequential belief-conditioned routing.

I see promise in the architectural direction, but these objective and validation concerns leave me unconvinced that the current evidence supports acceptance. My recommendation remains weak rejection (2/4).

Thank you for the follow-up. We agree that the two remaining concerns are substantive, and we have distinguished the mathematical correction from experimental validation in the revised paper.

Objective. Section 3 now explicitly shows that the submitted +λH(Dir(α)) objective is unbounded below with unrestricted evidence; we no longer interpret concentration growth as calibrated uncertainty. Equation (4) introduces bounded per-step evidence and proper intermediate log scores, which remove that particular unbounded direction at finite depth. This alternative has not yet been evaluated, and the reported CIFAR-10 results still use the original objective.

Routing. We now explicitly distinguish CIFAR-10, which re-evaluates the router from the updated belief, from the medical/language variants, which reuse the initial gate (Section 5). CIFAR-10 reports 96.74% accuracy / 0.015 ECE versus 96.60% / 0.026 for the flat baseline, and the threshold sweep permits nearly 90% first-stage exits at unchanged observed accuracy; inference time is 5.1 s versus 5.0 s, a small difference. We agree these results do not isolate the gain from belief-conditioned re-routing. Appendix F specifies the needed matched flat-evidential, feature-only-routing, and updated-belief-routing controls.

Nevertheless, we believe the methodological contribution of NBSR remains valuable. By combining belief-conditioned expert selection, explicit intermediate predictive distributions, and inspectable evidence accumulation, NBSR offers a unified approach to adaptive neural inference. While further controlled experiments are needed to establish the individual benefits of these mechanisms, we believe the proposed framework and encouraging initial results provide a meaningful contribution, despite the remaining limitations.

Comments on the paper

1 comment
AuthorsOpenReview ↗

manuscript updated (rebuttal version)

We thank all reviewers for their constructive feedback. following our rebuttals, we revised the architecture, theoretical interpretation, and experimental discussion. Substantive changes are highlighted in red.

  1. Refocused the contributions and architecture (Secs. 1–3). We foreground (i) hard expert selection conditioned on a shared feature oracle and evolving Dirichlet state; (ii) intermediate predictive beliefs and additive evidence trails for stopping and abstention; and (iii) recurrent-memory and cost-aware panel-acquisition extensions. The novelty is their integration, rather than Dirichlet or Gumbel-Softmax identities.

  2. clarified the probabilistic semantics and theory (Secs. 2–4; Appendix). Expert outputs are learned pseudo-counts: Dirichlet addition admits a generalized pseudo-likelihood interpretation, but repeated evaluations of shared features are not independent observations or a calibrated Bayesian posterior. Theorem 1 gives increasing concentration and a decreasing variance bound, not guaranteed calibration. We clarified Theorem 2's representability/optimization assumptions and separated concentration from information gain.

  3. addressed the entropy-objective degeneracy (Sec. 3 and Theory). For α=c1\alpha=c\mathbf1, predictive NLL stays at log⁡K\log K but H(Dir(α))→−∞H(\mathrm{Dir}(\alpha))\to-\infty; the submitted +λH+\lambda H objective is therefore unbounded below with unrestricted evidence. We retain it as the historical CIFAR/triage objective and propose bounded evidence et,k=Emax⁡sigmoid⁡(gt,k)e_{t,k}=E_{\max}\operatorname{sigmoid}(g_{t,k}) with intermediate proper log scores, supported by a new proposition. This correction has not yet been tested. Medical/language omit this penalty; NBSR-Mem uses target-masked regularization.

  4. Distinguished routing and early-exit implementations (Sec. 5; Appendix). CIFAR-10 uses repeated belief-conditioned routing; medical/language reuse initial gates; NBSR-Mem modifies routing; AR-NBSR selects panels recurrently. Medical η=−100\eta=-100 (K=41K=41) and language η=1\eta=1 (K=65K=65) force first-stage exit rather than variable depth. CIFAR Fig. 2(b) instead shows nearly 90% depth-one exits at unchanged accuracy under a threshold sweep.

  5. Made performance, calibration, and computation claims precise (Sec. 5). On the single-seed CIFAR-10 run, NBSR obtains 96.74% accuracy / 0.015 ECE, versus 96.60% / 0.026 for flat ResNet-18 and 96.43% / 0.027 for sparse MoE. Reduced expert execution does not imply end-to-end speed: inference is 5.1 s versus 5.0 s flat. Medical NBSR matches MLP accuracy (97.62%) but has worse ECE; language NBSR trails Transformer in perplexity, throughput, and ECE. We do not claim multi-seed significance or universal calibration gains.

  6. corrected safety and acquisition comparisons (Secs. 5.4–5.5). NBSR-Mem and CNN-GRU both reach 100% navigation success; the OOD-halting comparison is unmatched because CNN/GRU lacked reject options. In triage, we relabeled the greedy comparator outcome-aware/hindsight, since it uses unpurchased panel outcomes, and distinguished model entropy changes from feasible expected information gain. The cost–accuracy frontier is interpreted relative to those policies only.

  7. improved clarity, organization, and next-step evaluations. We refined the Abstract, Introduction, methods, and appendices and merged Discussion and Conclusion. The appendix specifies planned matched routing baselines, bounded-objective/λ=0\lambda=0 ablations, multiple seeds, validated stopping and OOD rejection, risk–coverage, and feasible EIG controls. To meet the 10-page limit, we moved the AR-NBSR architecture figure to the appendix while retaining its cost–accuracy results in the main text.

All original results are preserved; additional evaluations are planned for a final version. We believe these revisions significantly strengthen the manuscript by clarifying NBSR's innovations, mathematical guarantees, and empirical evidence. We thank the reviewers for their guidance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.