acceptodds
Under review as a conference paper at ICLR 2027

RAOA: Alternating-Operator Neural Computation with Programmable Radio Propagation

Abstract

Can programmable radio propagation serve as computational depth rather than only as a communication channel or one-shot analog transform? We introduce the Radio Alternating Operator Ansatz (RAOA), a recurrent computing architecture that alternates an energy-derived problem update with a mixing update over a persistent latent state. Recomputing the problem field after each mix makes repeated passes compositional even when the same learned controls are reused across depth. We evaluate this idea through exact discrete optimization, constrained programmable-propagation simulation, and pretrained-model adaptation. On discrete objectives, repeated execution can improve solution quality without increasing the learned-control count, and the same formulation handles higher-order interactions directly. A passive phase-only free-space model further shows that the required operators can be approximated by programmable propagation while retaining useful downstream behavior despite realization error. When inserted as a zero-initialized residual adapter, RAOA adapts pretrained language models with WikiText performance close to a matched shallow MLP across three model families, while reasoning-task transfer remains model-dependent. Together, these results connect alternating-operator computation, programmable radio propagation, and neural adaptation within one recurrent framework. The RF realization evidence is simulation-based rather than a hardware demonstration.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Neural operator, Radio wave, Differentiable programming, Analog computing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_xt2XOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes RAOA, a recurrent update over a latent state that alternates a "problem" step defined as an energy minimization step, with a "mixer" step that adds a weighted feedback connection to the system. The motivation is that each pass could conceptually be executed by programmable radio propagation through a reconfigurable intelligent surface (RIS) with receiver feedback, so that repeated passes provide "computational depth" without extra hardware or parameters. The framework is tested on exact discrete optimization tasks, on a simplified 9x9 grid simulation and digitally as an LLM adaptation.

Critical strengths
  1. The paper repeatedly states the specific contributions compared to existing literature as well as being transparent on negative results.
  2. Each experiment performed seems well performed and is sensible in itself (Exact optimization, Simulation of small 9x9 grid, LLM adaptation on digital 64x64 grid)
Critical weaknesses
  1. The paper is very difficult to read and requires substantial rewriting.
  • 1.1 Many abbreviations are not introduced (HUBO/QUBO, Ancilla, QAOA, Max-Cut, 3-Sat,...), even key concepts such as "problem pass" and "mixture pass" are not clearly defined.
  • 1.2 The general positioning of the paper is not clear, clearer explanations on why such a design is desired in the first place would have been much appreciated.
  • 1.3 The text in the results section (5) contains no analysis of results but merely other results that are written out in numbers. The results section should contain at least some interpretation of the shown results.
  • 1.4 Figures 1/2 are not helpful and contain bad lay-out (Fig 1. arrows) and (at that point in the paper) irrelevant results (Fig 2. cubic MSE, 49.3% linear,..). In fact, Figure 5 is the clearest, but unfortunately pushed to the appendix
  1. The connection between theory/experiments and the radio-wave domain are somewhat unclear. Only results 5.4 actually rely on the "radio-interpretation", and from what I understand it performs quite poorly (36-40% Normalized Frobenius error: table 3)? The results on the LLM essentially ablate a single bottleneck layer (L=1, as more doesn't help), which is not so interesting as such. The interpretation through RIS/SIMs makes the paper substantially more difficult to parse and, in my opinion, the performed experiments and/or theory currently do not justify this.

  2. I feel like the comparison to LoRA is a bit of a stretch, LoRA shows plenty of advantages that additional layers do not (for instance that the fine-tuned weights can simply be added to the pretrained ones thanks to linearity). Differences should be described in more detail and RAOAs existence justified more solidly.

  3. This made me think a lot about other energy based frameworks, such as Hopfield Networks (Ising-like models). I felt that the connection to that particular research domain was lacking and could have been explored more, although seeing the current abundance of ML papers I believe it is natural to miss certain connections. Still I'd like to hear the authors opinion on connections between their method and Ising-like approaches.

Questions for authors
  1. Related to W2: Why is 35% error good enough? Why is the connection to radio waves so essential to this work, what would change if the term is dropped from theory and experiments?

  2. Related to W2: Isn't there also some additional cost (as in energy, transmission inaccuracy etc) that should be included in the framework for it to really be a physically plausible model? For the moment it seems very far from true RIS/SIMs, if the goal is really to have a physically plausible AI method as the title suggests, discussions on where limitations lie are crucial.

  3. Related to W3: How does this framework relate to Hopfield Networks?

2 comments

Thank you for the direct presentation and positioning criticism. We agree that important definitions and the physical/digital boundary are currently too difficult to find, and that the results section needs interpretation rather than a sequence of tables.

1. Presentation and figures. We will introduce QUBO/HUBO, ancillas, QAOA, Max-Cut, 3-SAT, the problem step, mixer step, and modeled pass depth before using them. We will redraw Fig. 1 with a clearer state-feedback path and visible separation between the programmed radio operator and digital normalization/feedback.

We will move the small cubic-MSE/delayed-product annotations out of the main operator-ladder figure, and promote a clean vector version of the more intuitive Appendix Fig. 5 concept. Each results subsection will lead with the scientific question and end with what the evidence supports and what it does not.

2. Why radio and what does the 35%-40% error mean? The physical experiment is a constrained operator-compilation study, not a claim that 35% error is intrinsically good.

The original nine-mode model had about 0.35-0.41 normalized operator error at four panels; our new 13-mode, 12-variable Max-Cut experiment reports about 0.495 problem-operator error at four panels and downstream success 0.0878 versus 0.1604 for the simpler problem-only recurrence at equal nominal propagation passes. More panels improve average approximation but individual geometries still fail.

Simulated four-panel full RAOA preserves 0.0856 success at an assumed normalized 20-dB readout SNR versus 0.0878 without noise, and four-bit feedback gives 0.0856 versus 0.0878 with unquantized feedback. These are not received-power measurements.

Significant modeled phase drift degrades success, and hypothetical phase-offset calibration can restore much of it; calibration overhead remains open.

The radio component supplies a concrete constrained substrate for reconfigurable linear operators, while all tanh, normalization, controller updates, and feedback are digital. We will state explicitly that a measured closed-loop RF implementation and energy/link-budget accounting are future work.

3. LoRA distinction. We agree that LoRA has advantages RAOA does not share: a low-rank linear weight update can often be merged into frozen weights for inference, whereas RAOA's nonlinear recurrent residual path generally cannot. Our trainable-parameter matching does not imply equal inference operations or hardware costs.

We use LoRA and MLP as reference adaptation baselines rather than claiming RAOA dominates them. Indeed the matched digital recurrent Qwen baseline is better than RAOA on WikiText in our study (NLL 2.6727 vs 2.6851), so the correct claim is feasible structured adapter transfer, not unique digital superiority.

4. Hopfield/Ising connection. RAOA shares energy-derived local-field updates with Hopfield/Ising-type networks and with unrolled optimization. It differs in its explicit finite-depth tied-control recurrence, optional cross-coordinate mixer, and proposed correspondence between the linear portions of the update and reprogrammable propagation.

Unlike standard energy-descent interpretations, the complete tanh/mixer stage has no guaranteed monotonic decrease of the energy. Its relationship to QAOA is architectural (alternating roles), not quantum equivalence.

We will make these connections explicit and, in light of our additional ablations, identify dynamic local-field recomputation rather than universal Hadamard-mixer benefit as the consistent mechanism.

I thank the authors for their answers to our questions. Even though the responses are meaningful, the required and announced adaptations to the manuscript are considerable. Since the manuscript seems not to have been reworked, it is difficult to evaluate whether my concerns regarding clarity and writing would be properly addressed. The answers to the other questions also remain rather vague, particularly regarding W2 and W3. Overall my original concerns remain, and for this reason, I maintain my score.

Reviewer_X2JAOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces the Radio Alternating Operator Ansatz (RAOA), a recurrent computing architecture that alternates an energy-derived problem update with a mixing update over a persistent latent state. The problem field is recomputed after each mix, and the same learned controls are reused across depth.

  1. On exact QUBO/HUBO objectives with n=8,12n=8,12, each instance learns only two shared scalar controls. Extending the same two-control program from two to 4, 8, and 16 passes increases mean success for all four tested objective families. The same formulation handles higher-order interactions directly.
  2. A passive phase-only free-space model shows that the required operators can be approximated by programmable propagation. Increasing panel count reduces operator error and improves downstream Max-Cut success.
  3. Single-stage RAOA adapters recover 93.9–99.2% of the shallow MLP WikiText NLL improvement across three Transformer families, while GSM8K transfer remains model-dependent. All foundation-model experiments execute the differentiable operator in software, and all radio-propagation results are simulated.
Critical strengths
  1. Using native higher-order fields is a useful design choice. It avoids the ancilla burden of quadratization, keeping the optimization problem smaller instead of introducing extra variables to obtain a pairwise representation.
Critical weaknesses
  1. The advantage of the problem/mixer alternation is not well established. The ablations mainly support recomputing the local field. In Table 16, removing the mixer barely affects 3-SAT and hypergraph Max-Cut and improves partition success from 0.0356 to 0.105. The mixer helps Max-Cut, but this benefit is not tested across the independent-instance depth study. The observed depth gains therefore do not establish what the alternation contributes beyond repeated local-field updates. All instances have only eight or twelve variables, and there is no comparison with standard solvers under matched computational budgets. The results therefore provide limited evidence for adopting this particular alternating structure.

  2. The language-model experiments do not establish what RAOA's particular architecture contributes beyond existing adapters. The MLP achieves lower WikiText NLL on all three model families, and the matched recurrent baseline also outperforms RAOA on Qwen (Table 2 and lines 362–365). Increasing RAOA depth gives no consistent improvement. On GSM8K, RAOA shows no clear improvement over frozen Qwen and Gemma. On Phi, the RAOA adapter achieves 79.45% accuracy, compared with 83.73% for the MLP adapter.

  3. The propagation simulation does not establish the practical value of the proposed radio architecture. It only tests small pairwise operators. Even with 512 programmable phases for a nine-mode target, normalized operator errors remain approximately 0.35–0.41 (Table 3). Neither the native higher-order fields nor the 64-dimensional language-model adapters are realized. No end-to-end latency or energy advantage is established. The paper therefore leaves a substantial gap between its application results and a useful radio implementation.

The paper combines several existing ideas into one architecture. The small-scale and software experiments do not sufficiently establish why this combination should be adopted.

Questions for authors
  1. Can you test larger discrete optimization instances and compare RAOA with the β=0\beta=0 ablation and standard solvers under the same computational budget, including the cost of training the controls?

  2. What are the target matrices TT for the problem and mixer passes in Eq. (37), and how are the compiled operators combined with the normalization and tanh⁡\tanh operations in Eq. (1)?

2 comments

Thank you for identifying the two central gaps: the evidence for the chosen alternation beyond field recurrence, and the distance between software objectives and physical propagation. We have extended both analyses and now support a more restricted claim.

1. Larger instances, no-mixer controls, and classical solvers. We extended Max-Cut and 3-SAT from 8/12 variables to 16 and 20 variables, including a harder 3-SAT distribution with approximately 4.25 clauses per variable. Ground-truth optima are checked exactly.

The table uses 12 independent instances and three optimizer seeds, controls trained at one stage and evaluated at eight stages. The simulated-annealing (SA) column receives the same one-thread CPU time as full RAOA's 120-step, 64-replica instance-specific training plus 512-trajectory evaluation; it is not the same time budget as problem-only RAOA.

TasknFull RAOA successProblem-only successSA at full-RAOA CPU-time budget
Max-Cut160.08440.11430.6791
Max-Cut200.02930.04490.5686
Hard 3-SAT160.36940.40870.7405
Hard 3-SAT200.28650.27220.6761

We agree that these results do not establish a competitive general-purpose solver or superiority to simulated annealing. They show a repeatable mechanism: under additional ablations at 8/12 variables, freezing the initial energy field sharply reduces deep optimum-hit rates; removing the mixer often matches or exceeds full RAOA. CPU time is an implementation-specific proxy, not a hardware-energy or matched-FLOP study. We will make the training budget and the harder instance distribution explicit.

Separately, a fresh independently replayed control-transfer study (M9) showed that a pair of controls trained on 16-variable source instances can be evaluated on disjoint 20-variable instances without target-instance retraining; each new instance still supplies its own energy coefficients. This is control transfer, not an unchanged programmable radio matrix.

2. What does Eq. (37) compile? For a pairwise Ising energy with local field u = Jz + b, the problem target is an (n+1) x (n+1) augmented local-field matrix acting on [z;1]; the extra fixed +1 coordinate supplies the affine bias. The mixer target is the Hadamard-derived matrix M in the top-left n x n block of an otherwise zero augmented matrix.

The two separate surface phase programs approximate these real target matrices by the in-phase component of the normalized complex transfer.

The surface does not implement all of Eq. (1): after reading the problem field, software electronics evaluate the state-dependent normalization and tanh, and apply scalar sin(gamma)/cos(beta)/sin(beta) combinations and feedback. The nonlinear objective is recomputed on each updated state. These hybrid digital steps and gain normalization are essential to state explicitly.

To test a less trivial propagation target, we extended the phase-only compiler to 13 modes (12-variable Max-Cut) using three problem instances, three RF geometries, and 1/2/4 passive 128-element panels:

PanelsReal problem-operator errorFull RAOA success at 8 nominal passesProblem-only success at 8 nominal passes
10.7640.01780.0197
20.6340.05570.0917
40.4950.08780.1604

Approximation errors remain high, and the no-mixer variant is better in this tested ensemble. Full RAOA and problem-only are matched in nominal pass count, not field recomputations, configuration switches, electronic processing, energy, or latency.

The phase-only study only targets pairwise operations; native 3-/4-body HUBO fields are implemented in software, not synthesized by these radio panels. We will narrow the physical-realization and efficiency claims accordingly and add a transparent diagram of which computations are propagated and which remain electronic.

Thank you for the detailed response and the additional experiments. I appreciate that the authors took my questions seriously, especially by testing larger instances and comparing against standard solvers.

However, some of the new results make me question the motivation for this particular architecture even more. The no-mixer variant performs better in most of the new optimization experiments, and the standard baseline performs substantially better under the reported CPU-time budget. If the main benefit comes from repeatedly recomputing the local field rather than the alternating structure, then I am not sure what we actually gain from RAOA specifically.

The new radio simulations are useful, but the approximation errors remain large, and there is still no evidence of a practical efficiency advantage. I understand that the authors have clarified and narrowed their claims, but I don't think this fully addresses the gap between showing that the proposed computation is possible and demonstrating why this particular approach is worth using.

Overall, I appreciate the additional effort and think the rebuttal has clarified several technical questions. However, my main concern about the significance and motivation of the proposed method remains unchanged. Therefore, I will maintain my original score.

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

This paper studies recurrent architectures for programmable radio propagation. They specifically propose RAOA, which stands for Radio Alternating Operator Ansatz. RAOA is characterized by a 2-stage recurrent architecture. In the first stage, the programmable path applies a problem-dependent computation to the recurrent state. In the second stage (which is processed on top of the feedback that is given by the receiver to the transmitter), a mixer is used instead. Therefore, the primary contribution seems to be this alternation between problem-dependent and mixer computations.

RAOA is validated in two main settings. First, in classical binary optimization problems (such as Max-Cut, 3-SAT and Partition). Then, they also evaluate whether the approach could be inserted within a pretrained LLM (such as Qwen3-4B) as a residual adapter. Finally, they also validate that RAOA could be actually be compiled to a programmable propagation framework.

Critical strengths

The paper shows the integration of a programmable propagation framework within Transformers, which is very interesting. It also shows that increasing the depth of the computation using this alternating-operator approach increases performance even without increasing the number of parameters. Finally, it demonstrates that their approach could be realizable through physical simulations.

Critical weaknesses

W 1 - Presentation In several cases, the papers uses terms without previously introducing or defining them, and includes information in sections that feel out of place. This makes it harder to understand the work and gather the claims and evidences that support them. I mention below a few of these situations below:

The introduction uses the term depth several times, without previously clarifying what it means in the context of the paper.

(~ l. 48) The terms "alternating-operator" and "QUBO/HUBO" are used without defining them.

(~ l. 48 - 49) "known optima let us vary modeled pass depth without ..." it is not clear why having known optima would imply that it is possible to vary the modeled pass depth without changing learned-control count.

(~ l. 49) "random instances". In this context, I think "instance" is not clear. It becomes clearer only later in the experiments sections.

(~ l. 50) "two-control program". At this point, I think mentioning that there are two controls feels unclear.

(~ l. 51 - 52) It could be benefitial to give some background on what is a "constrained passive phase-only free-space model" and on the physical realizability simulations.

(~ l. 53) "digital adapter". This probably refers to the MLP and LoRA, but to the best of my knowledge this is not a standard term to refer to these model adapters, specially considering that it was not previously explained that the text was talking about LoRAs and related architectures.

(~ l. 53) "Large-model gains are not depth evidence: the primary 4B comparison fixed depth.". At this point, this sentence is not clear. We do not know what is the "primary 4B comparison", and what exactly fixing depth means. Maybe it should not be in the Introduction, it seems to Experiment-specific.

(~ l. 97 - 101) These sentences feel like they do not belong to the list of contributions in the introduction.

In Section 3.2, several terms in the equations are not defined, including the indices ℓ\ell and ii in Equation 1. Then WinW_{in}, JJ, wsw_s and cc in Equation 2.

(l. 228 - 230) "all reported quantitative values are generated from frozen JSON/CSV artifacts". The reference to using JSON/CSV files seems somewhat out of place. This would not be expected in the main body of the paper.

(~ l. 255 - 257) What does "cap-hit" rates mean and what is the "preregistered audit"? The work preregistered also appears in line 316 where is it not clear again.

(~ l.262) it talks about a "synthetic perturbation study", which is not described.

W 2 - Apparent Lack of Motivation and Evidence for RAOA

W 2.1 - As described in Section 3.2, the RAOA method seems under-motivated. Why is a mixer used? Why this particular Hadarmard-derived orthogonal mixer? Why the specific equations involving sines, cosines, and hyperbolic tangents? Then, in Equation (2), why this specific formulation for the Transformer adapter?

W 2.2 - RAOA is not well supported by the results to the extent that the mixer was found to not be that important in an ablation. Apparently, it didn't produce significant changes in many of the binary optimization problems, and led to reduced performance in the partition problem. Those results are presented in Section 5.2 and in Appendix G.8. The mixer is an integral part of RAOA, and it's not clear to me, according to the paper, whether a RAOA without the mixer would be novel.

W 2.3 - In the experiments on LLMs (Section 5.3), lines 362 to 365 describe a parameter-matched Qwen depth-2 digital recurrent bottlnect that obtained similar or better performances than RAOA. As written in the paper, this seems to suggest that other recurrent methods would work besides RAOA: "Thus WikiText establishes effective RAOA transfer, not a unique advantage over generic recurrence.".

Questions for authors

Q 1 - In Section 5.3, the authors described a baseline with depth-2 digital recurrent bottleneck Qwen that performs on par or actually better than RAOA. Can't the recurrence relation of this baseline be used instead of RAOA on the programmable-propagation framework?

Q 2 - In Table 2, does "Frozen" correspond to the base model NLL loss on wikitext2? I'm surprised by the level of loss reduction by these adapters. Also, the tables gives a NLL that corresponds to a perplexity of about 27 for Qwen3-4b, whereas Qwen3-4b ppl is typically around 13.6 on WikiText-2. Is my interpretation correct?

3 comments
AuthorsOpenReview ↗

Q2: WikiText-2 NLL and Perplexity

Thank you for raising this question. Yes, the "Frozen" entry in Table 2 corresponds to the pretrained base model evaluated on WikiText-2 without any adaptation. We have conducted an additional controlled experiment to clarify the discrepancy in absolute perplexity.

Checkpoint and dataset. We use the following publicly available Hugging Face checkpoint and dataset:

  • Model: Qwen/Qwen3-4B (base model, not instruction-tuned).
  • Model revision: 1cfa9a7208912126459214e8b04321603b3df60c
  • Dataset: Salesforce/wikitext, configuration wikitext-2-raw-v1, full test split.
  • Original evaluation: Non-overlapping sequences of 128 tokens, with next-token negative log-likelihood (NLL) averaged over all scored positions.

The 128-token context was chosen for our controlled short-context adaptation experiments, partly motivated by the relatively short question inputs in GSM8K. This is one of many possible evaluation settings.

Additional context-length evaluation. We evaluated the impact of context lengths, without training or changing model parameters. We compared non-overlapping windows (used in our submission) and sliding windows with overlapping context.

Context lengthNon-overlap PPLSliding-window PPL
128 (original)27.130718.9157
25621.077116.4407
51217.348114.3451
102414.883512.7960
204813.199311.5948
409612.044711.0199
  1. Exact reproduction: The 128-token non-overlapping evaluation reproduces the submitted frozen-model NLL of 3.300667 (PPL 27.1307), scoring 296,545 next-token positions.
  2. Context-length dependence: Increasing the context length substantially reduces perplexity for the same frozen checkpoint. Under the original non-overlapping protocol, PPL decreases from 27.13 at 128 tokens to 12.04 at 4096 tokens.
  3. Windowing-protocol dependence: Sliding-window evaluation produces lower perplexity than non-overlapping evaluation at every tested context length. The sliding protocol evaluates the same 298,937 target positions at all context lengths, whereas non-overlapping evaluation discards boundary predictions.

These results establish that absolute WikiText-2 perplexity depends substantially on both context length and evaluation protocol. Consequently, perplexity values obtained under different protocols should not be compared directly.

More importantly, all methods in our submitted Table 2 (Frozen, LoRA, RAOA, and MLP) were evaluated using the same protocol, preserving a consistent comparison within that experiment. We do not claim that the reported absolute perplexities are directly comparable to other protocols.

We will clarify the exact checkpoint, dataset configuration, context length, and evaluation procedure in the revised manuscript and include the above sensitivity analysis. We thank the reviewer for identifying an ambiguity in our experimental reporting.

I appreciate the author's additional experiments clarifying the Qwen3-4b NLL results in WikiText-2. However, my primary concerns regarding motivation, empirical evidence and presentation remain unchanged.

Thank you for the detailed questions and response. We agree that key definitions and the motivation for our design choices were insufficiently explicit, and the new mechanism study motivates a more precise architectural claim.

W1: Presentation and definitions. We will define a stage as one problem update followed by one mixer update; depth L denotes stages, corresponding to 2L modeled, not measured, RF passes for full RAOA. A QUBO is a quadratic unconstrained binary objective; a HUBO includes higher-order binary products. An instance is one sampled objective with fixed coefficients. The two learned controls refer to gamma and beta per instance; the objective coefficients are supplied, not learned. We will define every symbol in Eqs. (1)-(2), move procedural phrases such as JSON/CSV provenance and generation-cap audits out of the conceptual narrative, clarify that "cap hit" means reaching the generation-token limit, and replace confusing introductory references with a clear computational schematic.

W2.1-W2.2: Why these updates and the mixer? The problem step evaluates the gradient of an explicit objective at the current state; normalizing its magnitude reduces sensitivity to arbitrary positive objective scaling. sin(gamma) and cos(beta), sin(beta) provide bounded scalar control coefficients; tanh bounds the recurrent state. We used a fixed Hadamard-derived orthogonal mixer as a structured, parameter-free cross-coordinate transform, not because it is uniquely optimal.

Our new paired cross-depth ablation uses 12 independent instances per family/size and five optimizer seeds. Controls are trained at one stage and reused at eight stages (full RAOA: 2 to 16 modeled passes; no-mixer: 1 to 8 problem passes):

12-variable problemFull RAOANo mixer (beta=0)Initial field frozen
Max-Cut0.1750.2520.010
3-SAT0.9800.9740.247
Hypergraph Max-Cut0.9840.9940.143
Partition0.0590.0630.005

The dominant consistent mechanism is state-dependent field recomputation, not the particular Hadamard mixer. No-mixer recurrence can be better and requires fewer potential radio passes; these are equal-stage comparisons, not matched physical-resource budgets. We will explicitly describe mixing as an optional, objective-dependent operation and avoid claiming that alternating mixing is always necessary or superior.

Q1: Could the generic recurrent baseline be used on radio? Potentially yes; our results do not rule that out. The parameter-matched Qwen depth-2 digital recurrent bottleneck obtains WikiText NLL 2.6727 versus RAOA 2.6851. Thus the language-model study establishes that a structured energy-derived recurrent adapter can transfer to pretrained Transformers, not that RAOA uniquely outperforms generic recurrence. A generic recurrent update would require its own operator-to-propagation compilation and physical-cost analysis. We will clarify the distinction and LoRA's separate weight-merging advantage.

Reviewer_yJ4TOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed RAOA, i.e., Radio Alternating Operator Ansatz, to use programmable radio propagation for computation. The RF realization evidence is simulation-based rather than a hardware demonstration.

Critical strengths

The claims are disciplined, and the paper reports negative results openly

Critical weaknesses

The practicality is a major concern. The “Radio” in the title promises much more than the paper delivers. All radio parts are simulated, but there is no prototyping.

There’s also no evidence of efficiency, which is the whole point of computing with radio.

Questions for authors

What is the key benefit of RAOA? Higher accuracy or better energy efficiency?

1 comment

Thank you for highlighting the distinction between a computational proposal and a demonstrated RF system. We agree that our title and framing need to make this distinction much clearer: we do not present a hardware prototype, measured energy efficiency, or a throughput/latency advantage over digital computing. Our central demonstrated benefit is recurrent computation with a fixed, very small learned-control budget, together with an initial constrained propagation-realization study, rather than superior hardware efficiency or accuracy.

In the discrete study, one problem-dependent scalar gamma and one mixer scalar beta are tied across recurrent stages. On 12-variable instances, evaluating unchanged shallow-trained controls at 16 instead of two modeled passes raises mean exact-optimum hit probability from 0.367 to 0.980 on 3-SAT and from 0.0158 to 0.175 on Max-Cut. Additional mechanism ablations show that recomputing the evolving local field is the robust source of these gains; the specific Hadamard mixer is not uniformly beneficial, and we will temper the corresponding claims.

To strengthen the physical evidence, we extended phase-only compilation from nine to 13 augmented transmitter/receiver modes (12-variable Max-Cut), with 128 phase elements per panel, three objectives and three geometries. The new results are:

Passive panelsMean real problem-operator errorFull RAOA exact-optimum success at eight nominal RF passes
10.7640.0178
20.6340.0557
40.4950.0878

These errors remain substantial; the result is partial operator feasibility, not a complete RF implementation. In additional simulations with four panels, success is 0.0878 without receiver noise and 0.0856 at an assumed effective 20-dB readout SNR. Four-bit digital feedback gives 0.0856 at zero drift, versus 0.0878 with unquantized feedback. Both the effective SNR and calibration models are assumptions, not hardware measurements.

Why radio? A programmable propagation path can realize a reconfigurable linear field/mixing operation; receiver feedback and reprogramming make it a candidate substrate for recurrent computation. The nonlinear normalization, saturation, readout, and feedback are still performed digitally in these experiments. We will make that partition explicit and revise the introduction/title language so the paper does not imply a measured RF system or energy savings. We view this as a computational architecture and a radio-feasibility study, not a claim of radio supremacy.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.