acceptodds
Under review as a conference paper at ICLR 2027

Resonance Net: Persistent Structured Dynamics for Post-Input and Multimodal Computation

Abstract

We introduce Resonance Net, a recurrent architecture in which persistent complex-valued signals evolve on the edges of a sparse directed graph. Motivated by graph-based signal propagation and filtering, the architecture treats each recurrent tick as a native internal computation step over a persistent structured substrate. Each tick pools incoming state, applies learned complex transformations and edge-selective routing, and updates the substrate, allowing computation to continue after external input ends. We evaluate this property on CLRS-style graph problems by holding the encoded instance fixed while varying the number of post-input recurrent ticks. On average, deeper dependencies favor larger computation budgets: at the deepest evaluated settings, 48 post-input ticks improve mean pointer accuracy across five seeds from 69.5% to 83.9% on BFS and from 69.8% to 84.7% on Bellman–Ford. We next evaluate heterogeneous modality-specific interfaces through a shared core. Across five seeds, increasing the core edge budget from 0.5× to 4× raises joint multimodal utility from 0.862 to 0.869, while fixed, permuted, and shared ingress assignments perform similarly. Eight-seed mechanism ablations reveal task- and seed-dependent effects rather than uniformly necessary components: no individual removal produces a consistent cross-seed degradation in the primary dependency-by-tick surface AUC, while edge-selective routing increases the BFS benefit obtained from post-input computation across all eight paired seeds. Together, these results show that persistent structured dynamics can support extended internal computation and heterogeneous task interfaces, while also exhibiting substantial variation across learned recurrent solutions.

Primary area:
learning on time series and dynamical systems
Keywords:
recurrent neural networks, recurrent computation, algorithmic reasoning, graph neural networks, complex-valued neural networks, multimodal learning, persistent state

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_tiH8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ResonanceNet, a novel recurrent architecture built around a sparse directed graphs and persistent complex-valued state on graph edges. These states are updated through recurrent message-passing dynamics. Standard input encoders and task-specific decoders interface with this shared recurrent reservoir.

The paper evaluates the architecture in several settings, including CLRS-style algorithmic tasks (notably BFS and Bellman-Ford), multimodal learning scenarios, numerous ablation studies and the analyses of post-input computation. A central claim is that allowing recurrent computation to continue after external inputs disappear improves performance on tasks involving deeper dependencies.

Overall, the paper's contribution is an original and carefully documented architecture. The authors provide extensive experimental evaluations and are very transparent about negative results, limitations. However, while the study demonstrates the feasibility of the proposed architecture, it lacks evidence about when, why, and under what conditions the architecture is preferable to simpler alternatives.

Critical strengths
  1. Original architecture

The proposed architecture is novel although it combines ideas from several research directions, including recurrent neural networks, state-space models, graph filtering, complex-valued neural networks, and dynamical systems. The design is clearly described.

  1. Honest and rigorous experimental reporting

One of the strongest aspects of the paper is its scientific honesty (or clarity):

  • Negative results are clearly reported
  • Results report confidence intervals and statistical tests
  • Seed sensitivity is explicitly discussed
  • Claims are measured and results are never overstated
  1. Interesting findings on post-input computation

The most interesting empirical result in my opinion concerns algorithmic tasks such as BFS and Bellman-Ford. The paper shows that increasing the number of post-input recurrent ticks improves perfs on deeper instances. Ex: large gains are observed at the deepest tested settings for both BFS and Bellman-Ford. This points that that recurrent computation can continue performing useful work even after the input sequence has ended. 

  1. Extensive empirical effort

The amount of experimental work is substantial. The paper includes ablations, control experiments, and also seed analyses (there is a seizable supplementary materials file). The effort invested in understanding the behavior of the architecture should be rewarded.

Critical weaknesses
  1. The paper demonstrates viability more than usefulness

My main concern is that the manuscript establishes that the architecture works, but does not convincingly establish why one should use it over alternatives.

Several of the authors' own honest analyses point in this direction:

  • many effects are not statistically significant
  • strong seed sensitivity is observed
  • alternative architectures can also benefit from additional computation after input termination
  • some ablations fail to reveal clear contributions from individual components
  • the BABILong experiments do not strongly support the claimed long-term memory advantages

2. Limited understanding of the graph core topology

A substantial part of the model's identity lies in the graph reservoir, yet its topology is discussed only quickly.

I would like to better understand:

  • How are the graph structures defined beyond maximum depth?
  • How many input, hidden, and output nodes are used?
  • How are initial edges chosen?
  • How are skip probabilities determined?
  • Why were the specific topologies selected?
  • How sensitive is performance to topology design?

More controlled experiments varying this topology would be valuable

3. Insufficient justification of several design choices

The architecture incorporates many sophisticated elements, but it remains unclear which of them are necessary and which contribute meaningfully.

Edge-based state representation

The model stores persistent state on edges rather than nodes.

What is the advantage of edge memory over node memory? Acomparison against node-state variants would be highly informative.

Complex-valued dynamics

The motivation for complex-valued state is understandable, but I am not sure about empirical evidence. Only relatively small phase-rotation effects are observed. This naturally raises the question of how a real-valued recurrent graph model with comparable capacity would perform ?

Post-input computation

The paper emphasizes the value of continuing computation after external inputs cease.

However, this raises a broader question: how much of the observed benefit comes from the graph reservoir itself versus recurrent computation alone?

A useful control would be to evaluate recurrent architectures without the graph core but with comparable post-input computation budgets.

4. Multimodal section feels disconnected from the rest of the paper

The multimodal experiments are the least convincing aspect of the manuscript.

The study combines highly heterogeneous datasets (CMU-MOSEI and Vision & Touch) that:

  • do not contain paired samples,
  • do not share modalities,
  • do not share prediction targets,
  • do not obviously benefit from shared representations.

Consequently, I struggled to understand the scientific motivation.

The paper explicitly states that it is not attempting to build a multimodal foundation model. If so, it remains unclear why sharing a reservoir across such unrelated tasks is a meaningful idea ?

I believe stronger evidence could be obtained by focusing on a single multimodal dataset where genuine cross-modal interactions can occur within the recurrent substrate.

Questions for authors

Rather than expanding into many directions, I believe the paper would become considerably stronger by focusing on a smaller number of key questions and investigating them more deeply.

In particular, I would like to see the following points clarified : (see also "Weaknesses" for more details)

  1. Understanding the effect of post-input computation.
  2. Understanding the role of the graph substrate itself.
  3. Characterizing the importance of graph topology.
  4. Comparing edge-state and node-state formulations.
  5. Evaluating whether complex-valued dynamics are necessary.
  6. Providing deeper theoretical motivation for the architecture and its design choices.
3 comments

We thank the reviewer for recognizing the architecture's originality and the value of the post-input computation results. We agree that a stronger explanation of the substrate's design choices and practical usefulness would improve the paper.

We have added several controlled experiments and expanded the architectural discussion.

1. Usefulness of post-input computation and comparison with simpler models

We expanded the graph evaluation to eight seeds per task, with computation budgets selected on independent validation probes.

The selected budgets improve held-out accuracy by +9.20 points on BFS and +17.06 points on Bellman–Ford. Improvements occur in 14 of 16 task–seed cases.

We also compare Resonance with GRU and MPNN under the same graph-probe protocol. On Bellman–Ford, Resonance achieves 94.23%, compared with 72.58% for GRU and 94.52% for MPNN.

B200 profiling provides an additional comparison. At batch size 64 and K=48, Resonance takes 43.73 ms per batch, compared with 91.27 ms for the tested MPNN implementation.

These results give a more concrete picture of the substrate's predictive and computational characteristics. See Sections 5.1, 5.2, and 5.4.

2. Graph construction and topology sensitivity

We expanded Section 3.5 to explain the graph construction.

The CLRS configuration uses 16 input, 64 hidden, and 8 output nodes. It has a maximum forward depth of 12, distance-dependent skip connections, and additional recurrent/feedback connections. The reference configuration contains 509 directed edges at seed 0.

We also added a five-seed topology study. It varies skip length and recurrent connectivity while approximately preserving the edge budget.

The local- and long-skip variants retain similar BFS accuracy to the original topology. Long skips reach 98.36% on Bellman–Ford, compared with 93.14% for the reference. Feedback variations produce task-dependent performance differences.

Post-input computation improves performance in 34 of 40 additional task–seed evaluations.

These findings characterize the role of connectivity and show useful computation across different sparse graph configurations. See Section 5.3 and Appendix G.

3. Edge-local versus node-local state

Our motivation is to maintain persistent information separately for each directed relationship. Nodes perform transient pooling, while edges retain their own evolving states.

We added both equal-parameter and memory-matched node-state controls.

The memory-matched node implementation uses approximately 1.95M parameters, compared with 1.00M for complex-edge Resonance. On Bellman–Ford, complex edge achieves 94.23%, versus 88.34% for the memory-matched node model, and leads in seven of eight paired seeds.

The comparisons clarify the capacity and parameterization trade-offs behind edge-local storage. See Section 3.1 and Table 1.

4. Complex-valued dynamics

We added a real-valued persistent-edge alternative with matched parameter count and state-memory capacity.

The complex and real versions achieve 94.23% and 93.97% on Bellman–Ford, respectively. Their BFS results are also close.

We view complex arithmetic as a structured formulation for learned attenuation, rotation, and phase-sensitive dynamics. The real-valued implementation provides another viable choice within the same persistent-edge organization.

5. Multimodal motivation

Our multimodal experiment examines whether heterogeneous peripheral systems can be trained through a common persistent computational interface.

We added a jointly trained no-core bypass with the same task schedule and readout. It obtains joint utility 0.8693, compared with 0.8663 for the shared-core setting.

We also retain the original positive capacity trend within Resonance, where increasing the core budget from 0.5× to 4× improves utility by 0.00652.

Together, these results describe the shared interface and the contribution of its peripheral and central components.

We believe the revised manuscript now provides a substantially clearer architectural motivation, supported by direct comparisons and a controlled topology assessment.

Thank you very much for these updates, I really do appreciate the efforts you put in the revision. I remain unconvinced by your answer to the complex-valued dynamics - you seem to say there is hardly any difference with the real-valued case ? If yes, what is the point ? The reservoir is greatly clarified, thanks a lot. I believe the multimodal aspect is also better explained.

Thank you for clarifying this concern. I do agree that similar predictive accuracy raises an important question about what the complex formulation offers.

I carried out a small follow-up study tracing edge activity during recurrent computation. Across four seeds on BFS and Bellman–Ford, complex-edge recurrence maintained more localized activity during post-input computation. Its effective active-edge fraction was 64–68%, compared with 90–91% for the real-valued model. Both started with similar activity concentration, but the real-valued model became more distributed over subsequent ticks.

This relates to my original design goal. Complex arithmetic provides an explicit formulation of damped rotation and phase evolution. I want a persistent substrate where internal activity remains organized and controlled as computation continues, rather than spreading broadly across connections.

However, the exact relationship between this behavior and predictive performance remains unclear. Both models achieve similar accuracy, and the masking tests did not establish a consistent pruning advantage. Still, I think this difference in internal dynamics is a meaningful property of the complex formulation and may support future work on state organization and structural adaptation.

I kept this exploratory diagnostic outside the main manuscript to maintain its focus on the controlled performance comparisons. But I wanted to share it here because it directly addresses the motivation behind my design choice.

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

Resonance Net is a recurrent architecture with persistent complex-valued state on the edges of a sparse directed graph. Its updates combine local pooling, complex routing, geometric attenuation, and phase rotation, with optional structural plasticity. The main experiment holds the encoded input fixed and varies the number of subsequent recurrent ticks on custom CLRS-style BFS and Bellman–Ford probes. Deeper dependencies generally favor larger budgets in the aggregate, although the gains vary substantially across training seeds. The paper also studies heterogeneous peripherals around a shared multimodal core, mechanism ablations, and BABILong retention. Its supported contribution is a feasibility demonstration and controlled study of post-input computation.

Critical strengths
  1. Controlled evaluation of additional computation. Holding the encoded instance fixed during the tick sweep separates additional processing from additional evidence. The paper reports training-seed variation rather than treating probe examples as independent model replications, and it discloses deteriorating runs and confidence intervals that include zero (§5.1, Tables A2–A4).

  2. A concrete architecture and a positive capacity result. The persistent state and synchronous update are specified clearly enough to inspect. Increasing the multimodal core budget from 0.5× to 4× improves joint utility by 0.00652, with a paired 95% interval of [0.00462, 0.00843] (Appendix D). This supports a capacity trend within the architecture, with the limitations of the fixed evaluation splits stated explicitly.

Critical weaknesses
  1. The main computation result is not yet reliable across seeds. At the deepest conditions, the mean gains are 14.4 percentage points for BFS and 14.8 for Bellman–Ford, but their five-seed paired intervals are [−45.8, 74.6] and [−21.3, 51.0] points (§5.1). One BFS run falls from 85.9% to 30.5% at K=48, and one Bellman–Ford run also deteriorates (Table A3). The budget frontier is selected on the displayed evaluation sweep. Eq. (5) bounds state magnitude, but this does not protect accuracy as recurrence continues. The evidence supports a descriptive pattern, rather than a reliable way to allocate additional computation.

  2. The distinctive mechanisms and the joint core's added value remain uncertain. The MPNN gains more on Bellman–Ford (+32.7 points), while its BFS accuracy deteriorates and Resonance improves (Table A4). Thus extra-compute benefits are not unique to Resonance, but the comparison also does not establish that its state representation is irrelevant. All primary mechanism-ablation intervals include zero (Table 1), which is inconclusive rather than evidence of equivalence. The BFS selectivity result is positive in all eight seeds on the secondary post-input-gain metric, with interval [0.038, 0.173] and unadjusted p=0.0078. This is encouraging, but the intervention combines pooling, complex modulation, and CReLU, and the analysis spans 40 comparisons. It does not isolate one mechanism or establish a confirmed general effect.

    The multimodal controls also leave the core's added value unresolved. V&T validation data also inform learning-rate and checkpoint selection. The architecture comparisons match parameters but use different post-input transition counts (§4). The peripheral-only bypass achieves strong isolated-task results, but it is not a matched joint-training control (§5.2, Table A6). Similarly, unsuccessful frozen transfer limits a reusable-core interpretation without establishing that a jointly trained core is unused. These qualifications matter when interpreting the positive capacity trend.

Questions for authors
  1. Does the depth-dependent budget pattern survive when budgets are selected on separate validation probes and evaluated with more training seeds? Can a stopping rule avoid the observed accuracy deterioration?
  2. Can you compare equal-capacity real versus complex state, and node versus edge state, while reporting FLOPs, runtime, and persistent state size? The phase ablation alone does not answer these representation questions.
  3. What happens with a jointly trained peripheral-only bypass using the same task schedule, readout, parameter budget, and model selection procedure? This would complement the existing isolated bypass.
  4. At QA5/128k, erasing all persistent state gives mean accuracy 0.340 versus 0.307 intact (Table A8). This does not demonstrate a retention benefit, but aggregate accuracy does not establish that predictions are unchanged. Can you report paired accuracy uncertainty, prediction changes, and probability changes under erasure? Analogous interventions before the CLRS readout, accounting for any retained readout trace, would help identify how the edge state contributes at large K.
1 comment

We thank the reviewer for the detailed experimental suggestions. They were especially helpful in identifying where additional controls could strengthen our conclusions.

We have carried out new experiments addressing the main questions about computation reliability, state representation, and joint multimodal training.

1. Independent selection of post-input computation budgets

We agree that the original five-seed tick sweep was primarily descriptive. We therefore conducted a new evaluation with eight training seeds for each task.

Each model selects one computation budget from the original K grid using 128 independent validation examples per dependency depth. We then evaluate that choice on 256 separate test examples per depth.

The results are:

TaskK=0Selected KMean gain
BFS74.39%83.59%+9.20 pp
Bellman–Ford77.17%94.23%+17.06 pp

The paired 95% confidence intervals are [+0.10, +18.30] points for BFS and [+5.98, +28.14] for Bellman–Ford.

Selected computation improves 14 of 16 task–seed cases. One case is unchanged, and one decreases by only 0.27 points. By comparison, fixed K=48 improves 10 of 16.

The selected budget also avoids several severe fixed-K failures. For example, BFS seed 4 selects K=0 instead of using a longer budget that reduces accuracy.

This establishes a practical validation-based method for allocating post-input computation. The method selects one fixed budget per trained model, rather than using an input-adaptive stopping rule.

These results appear in Section 5.1 and Figure 2.

2. Real versus complex state and node versus edge state

We added a real-valued edge-state control matching the original model's trainable parameters and persistent-state capacity.

The two implementations achieve comparable performance: 83.59% versus 85.09% on BFS and 94.23% versus 93.97% on Bellman–Ford, for complex and real edge states respectively.

We also added two complex-node controls. One matches trainable parameters; the other approximately matches persistent-state storage.

The memory-matched node model achieves 82.22% on BFS and 88.34% on Bellman–Ford. It uses 1.95M parameters, compared with 1.00M for complex-edge Resonance.

These comparisons allow us to discuss complex dynamics as one structured implementation and edge-local persistence as an explicit choice of state organization.

Full paired comparisons are given in Section 5.2 and Appendix G.

3. Computation and memory costs

We added B200 resource profiling.

At batch size 64 and K=48, Bellman–Ford inference takes 43.73 ms for complex-edge Resonance, 38.93 ms for real edge, 31.12 ms for GRU, and 91.27 ms for MPNN.

A separate K=8 training-step measurement reports peak allocated memory of 1.22 GB for Resonance and 5.70 GB for MPNN.

We additionally measure state-storage scaling as graph size and connectivity change. Runtime, persistent-state size, and parameter counts are reported separately in Section 5.4 and Appendix H.

4. Jointly trained peripheral-only bypass

We implemented the requested joint no-core control.

It retains the original peripheral architectures, task heads, D2 readout, and 6k MOSEI / 12k Vision & Touch training schedule. Its learning rate is calibrated separately.

Across five paired seeds, joint utility is 0.8663 for Resonance and 0.8693 for the bypass. The paired Resonance-minus-bypass difference is −0.00303, with a 95% interval of [−0.01038, +0.00432].

This provides the joint-training comparison that was missing from the submitted experiments. See Section 5.5 and Appendix I.

5. Mechanism and long-context interpretation

We retained the eight-seed mechanism ablations and BABILong stress tests in the appendix.

The BFS edge-selectivity ablation shows a +0.106 full-minus-ablation difference in post-input gain, with all eight paired seeds positive. The intervention jointly changes pooling, complex modulation, and CReLU, as specified in Appendix E.

Appendix F also retains the QA5/128k state-erasure comparison. Our revised central claims rest on the independent held-out computation experiments rather than on the long-context retention results.

We believe these additions directly strengthen the reliability and fairness of the evaluation while making the architecture's computational trade-offs clearer.

Reviewer_x6oCOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Resonance Net, a recurrent architecture that maintains persistent complex-valued states on the edges of a sparse directed graph. Each recurrent tick combines signal aggregation, edge-specific routing, phase rotation, and bounded state updates, allowing computation to continue after external inputs stop. Experiments examine post-input computation on custom graph tasks, shared multimodal processing, mechanism ablations, and long-context retention. The results show average gains from additional computation and increased core capacity, but also substantial seed variability and limited evidence for the distinctive architectural mechanisms. The manuscript reports these limitations transparently, while the practical advantage of the proposed substrate remains unresolved.

Critical strengths

1, The architecture is clearly formulated, with explicit definitions of persistent edge states, synchronous updates, and amplitude constraints.

2, Holding the encoded instance fixed while varying recurrent ticks provides a useful test of additional computation without additional evidence.

3, The graph experiments include adapter-only, GRU, and MPNN controls with shared task interfaces and approximately matched parameter counts.

4, Multiple training seeds, mechanism ablations, randomized-core controls, and no-core bypasses provide a thorough examination of alternative explanations.

5, Figures 1 and A1 communicate the computation process clearly. The paper also distinguishes its custom graph probes from official CLRS benchmark results and explicitly acknowledges negative findings.

Critical weaknesses

1, The distinctive architectural contribution remains insufficiently established. The primary ablations do not demonstrate consistent benefits from phase dynamics, geometry, selectivity, or plasticity. A simpler real-valued model with persistent edge states would help establish the value of the proposed design.

2, Additional computation is not reliably beneficial across training seeds. At the deepest BFS condition, one seed drops from 85.9% to 30.5%, and the gain intervals include zero for both primary tasks. The preferred tick budget is also selected on the displayed evaluation grid, requiring independent validation.

3, The multimodal experiments do not clearly establish the contribution of the shared core. Isolated no-core controls perform comparably, while frozen trained cores do not reliably outperform randomized controls. A joint no-core baseline and tasks requiring cross-modal interaction would provide stronger evidence.

4, The evaluation protocols limit the comparative conclusions. V&T validation data also inform model selection, and Resonance uses a 12-tick flush plus two extra ticks while baseline wrappers use two post-input steps. Independent test evaluation and comparisons against measured computation or runtime are needed.

5, The main figures should present seed variability and the relevant negative controls alongside aggregate gains. The interpretation of complementary mechanisms also remains speculative because the ablations do not directly demonstrate compensation among components. A more focused presentation would clarify the contribution supported by the evidence.

Questions for authors

see weaknesses

3 comments

We thank the reviewer for the careful assessment. We understand the main concern to be whether the proposed substrate provides a useful architectural contribution beyond demonstrating that it can be trained.

We have revised the manuscript around this question and added direct comparisons with simpler recurrent alternatives, independent computation-budget selection, and measured resource costs.

1. Stronger evidence for post-input computation

The original five-seed results showed substantial variation. We expanded the evaluation to eight seeds per task and separated budget selection from final testing.

The new held-out experiment shows mean gains of +9.20 percentage points on BFS and +17.06 on Bellman–Ford.

Post-input processing improves 14 of 16 task–seed outcomes with validation-selected budgets, compared with 10 of 16 under fixed K=48.

Figure 2 now shows individual-seed results alongside the aggregate gains. This directly addresses the concern about reliability and test-set budget selection.

2. Direct comparison with simpler recurrent models

We expanded the architectural comparison to six models: complex-edge Resonance, real-edge recurrence, two node-state alternatives, GRU, and MPNN.

All use the same custom graph-probe inputs and independently selected computation budgets.

On BFS, complex-edge Resonance achieves 83.59%, compared with 81.32% for GRU and 81.11% for MPNN.

On Bellman–Ford, Resonance achieves 94.23%, compared with 72.58% for GRU and 94.52% for MPNN.

The real-edge formulation also performs competitively, reaching 85.09% and 93.97% on the two tasks.

These experiments give a clearer comparison of persistent edge-state computation with conventional recurrent alternatives. See Table 1 and Figure 3.

3. Practical computational characteristics

We added direct B200 measurements of inference latency, training memory, and sparse-state scaling.

At batch size 64 and K=48, Resonance takes 43.73 ms per Bellman–Ford batch, compared with 91.27 ms for the tested MPNN implementation.

At K=8, peak training-step allocated memory is 1.22 GB for Resonance and 5.70 GB for MPNN.

The sparse edge-state representation also scales with the realized edge count. Our synthetic scaling experiment measures this behavior from 88 to 704 nodes.

These results provide a concrete resource perspective alongside predictive accuracy. See Section 5.4 and Appendix H.

4. Joint multimodal core and simpler controls

We added the requested jointly trained no-core bypass.

The shared Resonance core obtains joint utility 0.8663; the bypass obtains 0.8693 across five paired seeds.

The original within-Resonance capacity study also shows a consistent positive trend. Increasing the core budget from 0.5× to 4× improves utility by 0.00652, with a paired 95% interval of [0.00462, 0.00843].

We retain the original split definitions and computation schedules in the revised Evaluation section. The joint bypass provides a direct account of what can be achieved using the same peripheral and readout framework without the central core.

5. Clearer presentation of the contribution

The revised main figures now place seed variation, alternative architectures, and resource measurements alongside the main computation results.

We also moved the detailed mechanism ablations into the appendix and narrowed their interpretation. In particular, the combined edge-selectivity result is reported as a measured effect, without attributing it to an individually isolated component.

Our intended contribution is a persistent, sparse, relation-structured computational substrate. The new experiments establish useful post-input computation, competitive predictive performance, and concrete trade-offs in state organization and computational resources.

We hope these additions make the architecture's contribution and operating characteristics substantially clearer.

Thanks for the response. My concerns remain unresolved, so I will maintain the original score.

Thank you for taking another look at the response.

I would really appreciate some clarification on which concern remains the most important in your assessment, since several experiments have been added directly targeting the issues raised in the original review, including independent validation, architectural controls, and computational profiling.

This would help me better understand where the current evidence falls short and whether there is a specific point I can clarify during the remaining discussion period.

Thank you again for your time and feedback.

Comments on the paper

1 comment

We thank all three reviewers for their careful assessments and constructive suggestions. We appreciate their recognition of the architecture's originality, the controlled post-input experiments, and the extensive empirical evaluation.

We have substantially revised the manuscript to address the main concerns about reliability, architectural justification, and practical usefulness. The revision preserves the original Resonance Net architecture while adding several focused experiments.

First, we strengthened the post-input computation results. We expanded the evaluation to eight training seeds per task. Each model now selects its computation budget using separate validation probes and is evaluated on independent test examples. This yields mean gains of +9.20 percentage points on BFS and +17.06 on Bellman–Ford. Performance improves in 14 of 16 task–seed cases, with one unchanged and one small decrease.

Second, we added direct architecture comparisons. These include real-valued edge recurrence, equal-parameter and memory-matched node-state models, GRU, and MPNN. The results establish competitive performance across several formulations and clarify the trade-offs between persistent-state organization, model capacity, and computation.

Third, we examined topology and computational resources. A five-seed topology study evaluates different skip and feedback connections. B200 profiling measures inference latency, training memory, and sparse-state scaling. These results provide a more concrete account of the substrate's operating characteristics.

Finally, we added a jointly trained multimodal no-core control. This complements the original capacity experiments and allows a direct comparison with the peripheral/readout system.

The revised Results and Discussion now focus on these findings. Detailed protocols, seed-level analyses, and the original experimental controls remain in the appendix.

We hope this revision provides a clearer account of Resonance Net as a persistent, structured computational substrate, including its demonstrated capabilities and practical architectural choices.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.