PaperlayerView exact revision on arXiv

Source-linked AI summary

Recursive Code Evolution for Task-Oriented Semantic Communications

Yan Gong, Zheng Chu, Wei Koong Chai, Hari Mohan Pandey, Jingrui Hou, Miao Zhang, Yuanjian Li, Pei Xiao

arXiv:2610.29641v1cs.ITcs.LGeess.SPmath.IT

TL;DR

Task-oriented semantic communication must balance inference accuracy, channel use, and computation, yet its communication program is typically specified before training. ComRSI recursively evolves executable programs under protected evaluation, and the evolved designs improve communication performance in the reported settings. Across ten searches, ComRSI achieved the highest mean performance, though its advantage over the history-equipped agent was statistically inconclusive.

  • Problem

    Task-oriented communication design must choose representations and processing locations that jointly determine inference accuracy, channel use, and computation, while candidates must respect physical-channel and resource constraints.

    Source: §I · pp. 1–2

  • Method

    ComRSI uses a fixed LLM agent to propose recursive edits to executable communication programs, screens them through protected training and evaluation, and retrains candidate weights from scratch.

    Source: §I · p. 2

  • Results

    ComRSI achieved the highest mean PublicScore and hidden-test accuracy across ten searches, significantly outperforming NR-0 but not significantly NR-H.

    Source: A. Discussion · p. 11

  • Takeaways & Limitations

    The reported experiments support recursive program optimization for task-oriented communication within the evaluated settings, including fading-channel performance and transfer to Tiny-ImageNet.

    Source: §VII · p. 12

    Structured reasoning memory and joint evolution of the communication program and proposal agent remain future directions.

    Source: B. Limitations and Future Directions · p. 12

Abstract

from arXiv · show

Semantic communication jointly optimizes inference accuracy, channel use, and computation, yet the underlying communication program is typically specified before training. This paper presents ComRSI (Communications-Oriented Recursive Self-Improvement), a framework for recursively evolving that program through large-language-model-assisted code edits. A fixed agent modifies the encoder, channel adapters, decoder, loss, and training policy; a protected evaluator screens candidates under prescribed budgets and resource constraints. Accepted code and textual feedback inform subsequent proposals, while candidate weights are trained from scratch. In the original ten-generation CIFAR-100 lineage, eight accepted revisions produce an attention-pooled code achieving 74.3% test accuracy over an Additive White Gaussian Noise (AWGN) channel at 32 complex symbols per image, compared with 67.2% at 256 symbols for the initial design. The program also improves fading-channel accuracy and transfers to Tiny-ImageNet after retraining. A repeat study with ten searches per method compares ComRSI against non-recursive agents with and without textual history. Mean final search scores balancing accuracy, transmission rate, and computational cost are 0.7587, 0.7376, and 0.7089, respectively; the mean numbers of accepted revisions per search are 6.0, 5.0, and 2.1. The advantage over the memoryless agent is supported by a two-sided rank test, whereas the difference from the history-equipped agent remains statistically inconclusive at this sample size. Archived proposals show that pooling can yield a large improvement, while retained history supports subsequent refinement. These results support recursive program optimization within the evaluated setting and expose its remaining rate-robustness and compute-placement trade-offs.

I. INTRODUCTION

Task-oriented semantic communication must balance inference accuracy, channel use, and computation through choices about representations and processing. ComRSI recursively evolves executable communication programs under protected evaluation, producing a 74.3% AWGN result at 32 complex symbols per image.

I. INTRODUCTION

  • Task-oriented image communication must choose where semantic processing occurs and which representation crosses the channel, because these choices jointly shape accuracy, rate, and compute.The design space includes encoder depth, transmitter–receiver split, representation dimension, channel adaptation, and training objective.Source: §I · p. 1

A. Learned and Task-Oriented Communication

  • ComRSI studies program evolution as complementary to changing operating variables or learned weights, while prior work motivates changing communication implementations rather than only configurations.ComRSI’s contribution is a procedure for modifying communication implementations, not a claim that pooled representations or channel-aware receivers are new principles.Source: §I · p. 2, A. Learned and Task-Oriented… · p. 2

I. INTRODUCTION

  • Communication-code search must preserve physical-channel and resource constraints, because increasing public evaluation scores alone cannot establish an improved communication system.The study compares recursive and non-recursive proposals, random search and mutation, and Bayesian optimization under matched nominal proposal and training allocation for the main CIFAR-100 comparisons.Source: §I · p. 2
  • ComRSI evolves the communication program through code edits while a fixed agent and protected evaluation pipeline screen candidates; candidate weights are trained from scratch.Editable components include the encoder, channel adapters, decoder, loss, and training-SNR policy.Source: §I · p. 2
  • 74.3% hidden-test AWGN accuracy at 32 complex symbols per image is the final design’s reported result, alongside improved fading-channel accuracy and transfer to Tiny-ImageNet.The identified structure combines a deeper transmitter, pooled semantic vector, and small SNR-aware receiver.Source: §I · p. 2

B. Agentic and Self-Evolving Wireless Systems

The related work spans agentic wireless design and recursive program search, while ComRSI evolves the communication program itself. Its framework separates editable designs from protected channel and evaluation conditions, then tests selected programs on withheld conditions.

B. Agentic and Self-Evolving Wireless Systems

  • ComRSI differs from agentic wireless workflow design by evolving the communication program, including its transmitter–receiver representation boundary and training objective, while keeping the proposal agent fixed.Evaluation focuses on physical-channel accounting and selected-program behavior under withheld channel conditions.Source: B. Agentic and Self-Evolving Wireless… · p. 2

C. Automated Design and Recursive Program Search

  • Neural architecture search and hyperparameter optimization depend on supplied search spaces, whereas code-based evolution offers a flexible candidate representation; ComRSI evaluates inherited edits rather than improving its fixed optimizer.Its empirical question concerns useful communication-program designs under channel and resource constraints.Source: C. Automated Design and Recursive… · p. 3

A. Framework of ComRSI

  • ComRSI’s editable program and recursive process let an LLM propose communication-design changes, while protected channel models, power normalization, metrics, and resource constraints preserve evaluation conditions.Four candidates are screened, the top two valid candidates receive full-budget training with three seeds, and acceptance requires a score margin plus a paired-bootstrap criterion; accepted code becomes the next parent.Source: A. Framework of ComRSI · p. 3

B. Communication Program and Physical Channel

C. Public Objective and Feasible Programs

PublicScore evaluates communication designs by balancing classification accuracy against transmission rate and computational cost, then searches within parameter and compute limits. The search is heuristic and does not guarantee a global optimum.

C. Public Objective and Feasible Programs

  • PublicScore combines validation accuracy over the public AWGN SNR grid with penalties for communication rate and computational cost.The rate penalty weight is twice the forward-computation penalty weight; the score is a scalar utility, not an accuracy measure.Source: C. Public Objective and Feasible… · p. 4
  • Search maximizes seed-averaged PublicScore subject to parameter-count and FLOP caps.The constraints are Npar(S) ≤ 1.5Npar,0 and F(S) ≤ 2F0.Source: C. Public Objective and Feasible… · p. 4
  • The feasible set requires valid interfaces, code scope, numerical behavior, and integrity, but the heuristic search gives no global-optimality guarantee.Equal training-step allocations do not imply equal training FLOPs or wall-clock time across different architectures.Source: C. Public Objective and Feasible… · p. 4

D. Proposal, Screening, and Inheritance

ComRSI uses public diagnostics and archived history to generate and screen program revisions, then carries accepted code forward while keeping rejected generations from changing the incumbent.

D. Proposal, Screening, and Inheritance

  • The agent uses public diagnostics—including per-SNR accuracy, rate, FLOPs, class errors, confidence, and loss summaries—to form four modification hypotheses and candidate revisions.Each proposal includes a rationale connecting its intended change to observed performance.Source: D. Proposal, Screening, and Inheritance · p. 4
  • The two highest-scoring valid candidates receive full-budget training from scratch with seeds 0, 1, and 2 after screening and proxy training.Candidates are compared against the same incumbent, with cached evaluations used where available.Source: D. Proposal, Screening, and Inheritance · p. 4, Figure 2 · p. 5
  • Only the selected program becomes the next generation’s parent; rejected generations update history but leave the inherited program unchanged.Candidate branches, diffs, hypotheses, and accepted ancestry are archived.Source: D. Proposal, Screening, and Inheritance · p. 5, Figure 2 · p. 5

E. Paired Acceptance Gate and Its Interpretation

ComRSI gates acceptance using paired validation comparisons and additional evaluation checks. Protected evaluation boundaries keep hidden results out of the proposal and selection loop.

E. Paired Acceptance Gate and Its Interpretation

  • Acceptance uses paired-bootstrap validation score differences, with 2,000 resamples preserving candidate-incumbent image pairings and the fifth percentile as a one-sided lower bound.All three seeds must also have valid full-budget evaluations.Source: E. Paired Acceptance Gate and Its… · p. 5
  • The gate reduces acceptance of small, image-sampling-sensitive changes; hidden evaluation and fresh seeds further check selected programs, while ten searches per agent variant measure search-level variability.Source: E. Paired Acceptance Gate and Its… · p. 5

F. Protected Evaluation and Information Flow

  • The data, channel, trainer, evaluator, selector, and budget constants are integrity-checked, while agent access is restricted and code checks prohibit specified imports, I/O, and RNG reseeding.Source: F. Protected Evaluation and… · p. 5
  • Integrity tests replace received code with noise, check permutation equivariance, and detect a fixture that stores encoder features for later receiver access.The stated thresholds are max(5/C, 1/C + 0.05) for noise-only accuracy and below 10^-2 for normalized transmitter and receiver permutation errors.Source: F. Protected Evaluation and… · p. 5
  • Programs are registered before hidden evaluation, and hidden scores are never returned to proposal agents.The hidden evaluation targets include accepted ComRSI generations and comparator incumbents.Source: F. Protected Evaluation and… · pp. 5–6
  • The recursive procedure generates four candidates, screens and proxy-trains them, full-trains up to two, applies the paired-bootstrap gate, and updates the program, diagnostics, and history.If no candidate is accepted, the current program remains in place.Source: F. Protected Evaluation and… · p. 6

IV. EXPERIMENTAL PROTOCOL

The experimental protocol uses fixed benchmark splits and a shared initial communication system, then compares recursive code inheritance with non-recursive agents and configuration-search alternatives. The comparison methods differ in their use of history, inherited code, and search capability.

A. Data and Initial Communication System

  • CIFAR-100 is the primary benchmark, with 45,000 training, 5,000 public-validation, and 10,000 hidden-test images; Tiny-ImageNet uses 90,000 training, 10,000 public-validation, and 10,000 hidden images.CIFAR-10 is used only for development and excluded from reported comparisons.Source: A. Data and Initial Communication System · p. 6
  • The common starting system is a ResNet-style transmitter with a 256-complex-symbol latent and a receiver that globally pools and classifies.On CIFAR-100, it has 3.63 million parameters and 0.694 GFLOPs per image, using cross-entropy and uniform training SNR from −3 to 9 dB.Source: A. Data and Initial Communication System · p. 6

B. Compared Methods

  • ComRSI modifies the current accepted program using source code, proposal history, and public feedback; accepted edits are inherited, while model weights are retrained from scratch.Edits may change network structures, semantic representations, channel adapters, training objectives, or configuration values.Source: B. Compared Methods · p. 6
  • NR-H starts each generation from the original program while retaining textual descriptions of prior proposals and outcomes, isolating direct code inheritance with shared interfaces and nominal budgets.The external selector still retains the best accepted program for evaluation and comparison.Source: B. Compared Methods · p. 6
  • NR-0 starts every generation from the original program without proposal history or access to earlier candidate workspaces, while the selector tracks the best accepted incumbent.Source: B. Compared Methods · p. 6
  • Random mutation recursively inherits accepted configuration changes, whereas random search samples independently and TPE uses prior scores, all within predefined configuration spaces.The configuration methods cannot express every program change available to unrestricted code editing, so comparisons assess procedures with different search capabilities rather than proposer quality at equal expressiveness.Source: B. Compared Methods · pp. 6–7

C. Training and Search Budgets

ComRSI’s search uses matched training budgets and fresh-seed, withheld evaluation to assess accepted program changes. The resulting lineage moves from a spatial representation to a 32-symbol pooled code with an SNR-aware receiver.

C. Training and Search Budgets

  • The protected trainer uses SGD with fixed core settings, while learning rate and weight decay remain editable within protected ranges.CIFAR-100 proxy and full training use 2,500 and 10,000 optimizer steps; both budgets are doubled for Tiny-ImageNet.Source: C. Training and Search Budgets · p. 7
  • Each CIFAR-100 search uses 700,000 nominal candidate-training steps, with four proposals per generation and three full-budget seeds for each of two promoted candidates.The total excludes initial-system evaluation, checker runs, and post-search retraining.Source: C. Training and Search Budgets · p. 7

D. Withheld Evaluation Conditions and Retraining

  • Withheld evaluation retrains selected programs from scratch on three fresh seeds and tests them across AWGN, Rayleigh, and Rician channels; Tiny-ImageNet transfer also retrains from scratch.Public and hidden channel random seeds are disjoint, and hidden-test accuracy is reported separately from PublicScore.Source: D. Withheld Evaluation Conditions and… · p. 7

A. Search Trajectory and Retraining Stability

  • 0.666 to 0.785: ComRSI’s search PublicScore rises over ten generations, with eight accepted revisions; fresh-seed retraining yields 0.784.Five consecutive improvements occur at G1–G5, followed by further improvements at G8–G10.Source: A. Search Trajectory and Retraining… · p. 7
  • At G8, the spatial latent representation becomes a 64-symbol pooled code; G9 reduces it to 32 symbols, and G10 adds learned attention pooling.The encoder has d = 256 features across 64 spatial locations; three attention heads and a global-average branch form pooled features, paired into 32 complex symbols.Source: A. Search Trajectory and Retraining… · p. 7, B. The Communication Structure… · p. 8

B. The Communication Structure Produced by Search

  • The final receiver concatenates channel-output real and imaginary parts with two SNR features, then uses a 512-dimensional hidden vector and a linear classifier.The resulting design has 5.29 million parameters and 1.14 GFLOPs, within protected caps.Source: B. The Communication Structure… · p. 8

C. Inheritance and Configuration-Space Comparisons

ComRSI’s recursive program search outperforms the listed non-recursive and configuration-space comparators on fresh-seed PublicScore. Hidden-test results also show a rate–robustness trade-off between G8 and G10.

C. Inheritance and Configuration-Space Comparisons

  • 0.784 versus 0.715 for NR-H and 0.722 for NR-0: ComRSI achieves the higher fresh-seed PublicScore.NR-0 retains a pooled 128-symbol program through G10, while NR-H accepts four revisions and ends at 384 symbols.Source: C. Inheritance and Configuration-Space… · p. 8
  • 0.710 for random mutation, 0.722 for random search, and 0.727 for Optuna TPE: these are the configuration-space methods’ fresh-seed PublicScores.ComRSI’s broader program-editing search includes structural changes such as the pooled-code architecture discovered at G8.Source: C. Inheritance and Configuration-Space… · p. 8
  • 0.098 higher fresh-seed PublicScore than the hand-tuned reference accompanies a reduction from 256 to 32 channel uses; versus Optuna TPE, ComRSI scores 0.784 versus 0.727 with 32 versus 384 uses.The earlier-agent variant scores 0.704 and restricting editable source files yields 0.699, compared with 0.784 for the main run.Source: C. Inheritance and Configuration-Space… · p. 8

D. Hidden-Image and Fading-Channel Performance

  • 74.3% AWGN-5 accuracy at 32 complex symbols per image compares with 67.2% at 256 symbols for the initial system; G8 reaches 67.1% under Rayleigh fading versus 64.0% for G10.G10 halves the communication rate relative to G8 while maintaining comparable average AWGN-5 accuracy; G8 performs better at low SNR and under fading.Source: D. Hidden-Image and Fading-Channel… · p. 8

E. Rate, Robustness, and Compute Placement

The evolved lineage trades channel robustness against transmission rate, while shifting more computation to the transmitter. The CIFAR-100 program also transfers to Tiny-ImageNet after retraining and outperforms the shorter direct search there.

E. Rate, Robustness, and Compute Placement

  • G8 introduces a 64-symbol representation, G9 reduces it to 32 symbols, and G10 improves accuracy through learned attention pooling.Fig. 6 plots hidden-test AWGN-5 accuracy against complex symbols per image on a logarithmic rate axis.Source: E. Rate, Robustness, and Compute… · p. 8, Figure 6 · p. 11
  • 0.0139 of PublicScore comes from reducing transmission rate from 64 to 32 symbols; at −6 dB AWGN, G8 scores 71.6% versus 63.9% for G10.G10’s 32-symbol representation maintains comparable average AWGN accuracy while G8 has stronger low-SNR performance.Source: E. Rate, Robustness, and Compute… · p. 8
  • The final G10 design shifts computation toward the transmitter: it uses 5.203 million transmitter parameters and 0.086 million receiver parameters, with 1.145 GFLOPs at the transmitter and 0.17 MFLOPs at the receiver.The source describes this as concentrating semantic processing at the transmitter while producing a compact receiver and 32-symbol code.Source: E. Rate, Robustness, and Compute… · pp. 8–9

F. Program Transfer and Search on Tiny-ImageNet

  • The transferred CIFAR-100 G10 program scores 0.615 PublicScore and 58.1% hidden AWGN-5 accuracy on Tiny-ImageNet, versus 0.550 and 55.7% for the transferred random-search program.It reaches 47.1% under Rayleigh fading at 32 complex symbols per image.Source: F. Program Transfer and Search on… · p. 9
  • Three accepted revisions in the five-generation Tiny-ImageNet search raise fresh-seed PublicScore from 0.511 to 0.539 and hidden AWGN-5 accuracy from 51.2% to 54.5%.The transferred CIFAR-100 G10 program performs better than programs found in this shorter Tiny-ImageNet search.Source: F. Program Transfer and Search on… · p. 9

1) Experimental Setup:

  • The ten-search comparison evaluates ComRSI, NR-H, and NR-0 on CIFAR-100 with matched proposal model, candidate allocation, training budgets, and acceptance criteria.Each final program is retrained using three random seeds and evaluated under the same hidden-test protocol.Source: 1) Experimental Setup · pp. 10–11

2) Overall Performance and Continued Improvement:

Across ten searches, ComRSI attained the highest mean final PublicScore, with a significant advantage over NR-0 but not NR-H. Pooling appeared across all three methods, while subsequent refinement was more consistent when prior discoveries could be retained.

2) Overall Performance and Continued Improvement:

  • 0.0553 mean Grest for ComRSI exceeds 0.0447 for NR-H and 0.0133 for NR-0; ComRSI also accepts 6.0 revisions on average, versus 5.0 and 2.1.The paper interprets these Grest results as evidence of improvement through successive accepted revisions.Source: 2) Overall Performance and Continued… · p. 11

3) Pooling Transitions and Subsequent Refinement:

  • Pooling emerges in 7/10 ComRSI searches, 4/10 NR-H searches, and 5/10 NR-0 searches, with mean first-pooling gains of 0.0443, 0.0415, and 0.0447.The first accepted pooling revision yields substantial gains across all three methods.Source: 3) Pooling Transitions and Subsequent… · p. 11
  • After pooling, all ComRSI and NR-H searches improve further (7/7 and 4/4), versus 1/5 for NR-0; mean additional PublicScore gains are 0.0234, 0.0239, and 0.0040.ComRSI inherits accepted code, NR-H can reuse textual history, and NR-0 lacks access to earlier search outcomes.Source: 3) Pooling Transitions and Subsequent… · p. 11

A. Discussion

  • The discovered 32-symbol design improves classification accuracy and lowers communication overhead relative to the initial system while maintaining strong performance under fading channels.The discussion also reports successful transfer to Tiny-ImageNet and concludes that repeated searches give ComRSI the highest mean PublicScore and hidden-test accuracy among the evaluated methods.Source: A. Discussion · p. 11
  • The 32- and 64-symbol designs illustrate a trade-off between communication efficiency and channel robustness, so different requirements can guide program evolution toward different operating needs.The supplied table caption identifies results over ten independent searches per method, reported as mean ± sample standard deviation across searches, with fresh-seed metrics averaging three retrainings.Source: A. Discussion · pp. 11–12, Table V · p. 12

2) Overall Performance and Continued Improvement:

B. Limitations and Future Directions

The conclusion reports improvements on CIFAR-100 and points to broader evaluations and deployment-aware optimization as future work. The current evaluation remains focused on image classification under simulated wireless channels.

B. Limitations and Future Directions

  • The current evaluation focuses primarily on image classification under simulated wireless channels, leaving broader communication tasks and realistic channel conditions for future assessment.The paper also identifies diverse datasets as a direction for assessing general applicability.Source: B. Limitations and Future Directions · p. 12
  • Deployment-aware optimization remains future work, including latency, energy consumption, and device-specific computational constraints.Future deployment evaluations may also incorporate channel estimation, pilot overhead, digital modulation, finite-blocklength coding, and adaptive rates under latency and energy budgets.Source: B. Limitations and Future Directions · p. 12
  • Future directions include structured reasoning memory for reusing editing experiences and joint evolution of the communication program and its proposal agent.The conclusion also points to multitask communication, adaptive channel-aware optimization, and deployment-oriented systems.Source: B. Limitations and Future Directions · p. 12, §VII · p. 13

VII. CONCLUSION

  • ComRSI is presented as a promising approach to automated semantic-communication design, with accepted code modifications progressively optimizing architectures, training strategies, and transmission configurations under protected evaluation constraints.The conclusion describes these as findings from the evaluated work and identifies broader communication settings as future directions.Source: §VII · pp. 12–13
  • The discovered 32-symbol program outperforms the 256-symbol initial design on CIFAR-100, and repeated searches give ComRSI the highest mean performance among evaluated approaches.The program also shows robustness under fading channels and transferability to Tiny-ImageNet.Source: §VII · p. 12

Related papers

Loading 2610.29641v1…