PaperlayerView exact revision on arXiv

Source-linked AI summary

Baseline-Relative Counterfactual Refinement for Bit-Aware Visual Token Communication

Jia Guo, Xiaohan Zhao, Changwang Liu, Shuqing He, Chenyang Zhang, Bingchuan Zhao, Jinqi Zhu

arXiv:2608.16192v1cs.AI

TL;DR

Existing visual-token selection does not directly test whether an alternative improves completed reconstruction under the same packet budget. GCR-C addresses this with gated counterfactual evaluation, improving reconstruction quality at active low and medium rates without increasing realized packet rate, at added encoder computation.

  • Problem

    Existing token-selection criteria do not directly determine whether changing the selected tokens improves completed reconstruction under the same packet budget.

  • Method

    GCR-C evaluates diversified feasible alternatives using matched full-budget Local-MDL continuation and accepts a replacement only for a positive baseline-relative gain.

  • Results

    GCR-C improves CIFAR-10 PSNR by 1.503 dB at 0.20 bpp and 0.698 dB at 0.32 bpp without increasing realized packet rate, with positive gains across 12 coded-link conditions.

  • Takeaways & Limitations

    GCR-C provides protocol-scoped evidence for improving reconstruction under fixed budgets across datasets, rates, channel conditions, resolutions, token grids, and vocabularies.

  • Takeaways & Limitations

    The unbatched full-budget evaluator is expensive, limiting the method’s suitability to delay-tolerant systems with sufficient encoder computation.

Abstract

from arXiv · show

Generative visual-token communication reduces transmission load by sending only selected discrete tokens and reconstructing missing content at the receiver. However, existing token-selection criteria based on local uncertainty, importance, or diversity do not directly determine whether changing the current selection improves the final reconstruction under the same packet budget. To address this problem, we propose Gated Counterfactual Refinement for Communication (GCR-C), a rollout-style correction layer over Local-MDL. GCR-C constructs a compact diversified candidate set, evaluates each candidate through matched full-budget Local-MDL continuation, and replaces the baseline action only when a positive baseline-relative reconstruction gain is obtained. Experiments on CIFAR-10, STL-10, a coded 5G-LDPC link, and a limited high-resolution Kodak transfer show that GCR-C consistently improves reconstruction quality at active low- and medium-rate operating points without increasing the realized packet rate, while remaining effective across changes in dataset, channel condition, resolution, token grid, and tokenizer. The results also reveal a clear quality--computation tradeoff due to the additional encoder-side counterfactual evaluation.

I. INTRODUCTION

The introduction frames visual-token communication as a scarce-resource problem in which selecting tokens requires accounting for bit-accurate packet cost and terminal reconstruction value. It proposes GCR-C, a baseline-relative rollout correction over Local-MDL, and reports quality gains without increasing realized packet rate.

  • Motivation: Visual-token communication sends selected discrete evidence and reconstructs missing tokens at the receiver, making it suitable for bandwidth- and power-constrained links.The setting is especially relevant when transmitter-side computation is comparatively available, including satellite and deep-space imaging.
  • Problem formulation: Token selection must account for each token’s contextual effect on completion and the bit cost of serialized positions, not merely token count.A token’s value depends on the protected set and subsequent completion process, while packet cost depends on the selected set.
  • Problem formulation: Existing uncertainty-, importance-, diversity-, and redundancy-based criteria do not directly evaluate baseline-relative terminal reconstruction under the same feasible packet budget.Locally promising actions can be redundant or become inferior after the remaining budget is completed.
  • Method: GCR-C preserves Local-MDL as the default, evaluates a compact diversified candidate set through matched full-budget Local-MDL continuations, and replaces the baseline only when refinement is beneficial.This specializes rollout-style policy improvement to bit-accurate generative visual-token communication, where feasibility follows packet syntax and counterfactual evaluation adds encoder cost.
  • Results: 1.503 dB and 0.698 dB are GCR-C’s PSNR improvements over Local-MDL at 0.20 bpp and 0.32 bpp, respectively, without increasing realized packet rate.These results come from a held-out 500-image CIFAR-10 test split; frozen selections retained positive paired gains across all 12 tested 5G-LDPC/QPSK AWGN configurations.

II. RELATED WORK … A. Sequential baseline-correction problem

The paper situates GCR-C among discrete visual-token generation, token-reduction, semantic-communication, and rollout-policy methods. It formulates sequential baseline correction as selecting feasible alternatives that improve completed receiver reconstruction under the same transmission budget.

  • A. Discrete visual representations and generative recovery: Discrete visual representations and masked generation enable sequence-model access to images and receiver-side recovery of untransmitted tokens.Learned compression and probabilistic priors connect rate budgets with descriptions of missing content.
  • B. Visual-token reduction, contextual selection, and semantic communication: Visual-token reduction research includes progressive compression, diversity-aware selection, clustering, representation-shift analysis, redundancy modeling, skipping, pruning, and adaptive allocation.Contextual, language-guided, training-free, and inference-optimal selectors broaden the signals available for token selection.
  • B. Visual-token reduction, contextual selection, and semantic communication: Communication-oriented work studies text-guided and adaptive semantic tokens alongside matchable subspaces, self-compression, reading-twice pruning, energy-aware pruning, robustness, and wireless constraints.These directions expose coupled robustness, throughput, and latency considerations for compressed visual-token transmission.
  • B. Visual-token reduction, contextual selection, and semantic communication: GCR-C evaluates a compact proposal through full-budget Local-MDL completion and accepts an alternative only when its baseline-relative advantage is positive.Selected tokens are packetized and transmitted through the coded link, while receiver reconstruction is reserved for offline evaluation.
  • C. Rollout-style policy improvement and self-positioning: GCR-C specializes rollout-based policy improvement to bit-accurate generative visual-token communication with packet-syntax and position-dependent feasibility.The paper does not claim the rollout principle itself as new because candidates are taken once and completed with a known base policy.
  • C. Rollout-style policy improvement and self-positioning: Unlike signals optimized for pruning efficiency, inference, or adaptive transmission, GCR-C tests whether feasible replacements improve receiver reconstruction after matched-budget completion.The method adapts the published DivPrune max–min diversity principle as a matched external selector under its packet interface.
  • A. Sequential baseline-correction problem: The communication objective is to construct a feasible transmitted set maximizing completed reconstruction quality subject to B(S) ≤ Bmax.Because direct global subset search is combinatorial, the paper instead sequentially compares a default Local-MDL action with feasible same-budget alternatives.

B. Tokenized image and packet budget

Images are patch-tokenized, and packets encode selected token values together with their positions using a budget-aware bitmap-or-gap representation. Packet accounting separates core transmission, forward-error protection, realized packet rate, and, for coded links, physical-layer rate.

  • Tokenization and selection: Images are divided into N non-overlapping patches and quantized into tokens from a vocabulary of size V; selected tokens are transmitted with their zero-based positions.The packet begins with a mode bit followed by either an N-bit bitmap or a gap list.
  • Core packet format: The mode bit selects the shorter bitmap or gap-list representation, while the core packet contains a 32-bit header, selected-token codewords, and a 16-bit CRC.For CIFAR-10/STL-10 with V = 32, each selected token contributes 5 bits.
  • Serialization: The CIFAR-10 encoder sorts positions before coding, and empty, singleton, and full selections have uniquely decodable syntax.The coded-link packet uses Btx = ⌈1.25Bcore⌉ as its packet-accounting proxy.
  • Rate accounting: The coded-link system distinguishes selection rate Rpacket = Btx/(HW) from physical-layer rate RPHY = n/(HW), with realized code rate Rc = k/n = 0.8.At 0.20 bpp, formal packets have Bcore = 159–163 and Btx = 199–204; at 0.32 bpp, Bcore = 258–261 and Btx = 323–327.
  • Budget constraint: Selector comparisons enforce packet-budget feasibility using Eqs. (1)–(4) and report the realized packet rate Rpacket.The coded-link experiment additionally reports the realized physical-layer rate RPHY.

C. Channel abstraction and cognitive transmitter state

The section abstracts channel effects through the receiver-available token set and defines a cognitive transmitter state for deterministic, budget-aware selection. In coded-link tests, CRC success preserves the selected set while failure removes it, whereas clean experiments use no channel feedback.

  • Channel abstraction: The channel abstraction maps each transmitted selection S to a receiver-available set Cω(S), supporting erasures and whole-frame decoding failures.The receiver reconstructs from the available set using the same masked prior.
  • Coded channel: In the coded-link experiment, CRC success yields Cω(S) = S, while CRC failure yields Cω(S) = ∅.The packet uses QPSK, approximately 0.8-rate 5G-NR LDPC coding, 20 belief-propagation iterations, AWGN, and block-Rayleigh fading with perfect receiver CSI.
  • Cognitive transmitter state: The transmitter state tracks the protected token set, remaining packet budget, local recoverability or receiver-state features, and optional channel context.The state is written using St, Brem,t, ut, and ct, respectively.
  • Decision loop: The decision loop proceeds through perception, proposal construction and QB completion, baseline-relative gating, and adaptation of the set, budget, and eligibility state.Clean experiments use constant null channel context without feedback, and the loop is deterministic rather than reinforcement learning.

D. Receiver-side completion … B. Baseline-relative fixed-budget value

The receiver uses a fixed one-pass masked-prior completion so comparisons isolate token-selection policy. GCR-C corrects Local-MDL through baseline-relative, fixed-budget counterfactual evaluation, accepting a proposal only when it improves the matched continuation.

  • D. Receiver-side completion: The masked prior receives known tokens and a mask, then deterministically completes untransmitted positions in one pass for image reconstruction.This completion rule is shared by Local-MDL, GCR-C, and Random comparisons.
  • D. Receiver-side completion: The fixed completion rule keeps experiments focused on selection policy rather than changes in decoder scheduling.
  • IV. GATED COUNTERFACTUAL REFINEMENT: GCR-C operates as a correction layer: Local-MDL supplies the default action, while a compact proposal adds spatially complementary alternatives.Each proposal is tested as the next transmission, with the remaining budget completed by Local-MDL.
  • IV. GATED COUNTERFACTUAL REFINEMENT: GCR-C replaces the default only when the best candidate has positive full-budget advantage, then returns all later actions to Local-MDL after one accepted intervention.
  • A. Local-MDL backbone: Local-MDL scores local recoverability with masked negative log-likelihood and repeatedly selects tokens until the packet budget is full.It also serves as GCR-C’s fallback when the correction gate is off.
  • B. Baseline-relative fixed-budget value: A candidate’s horizon-h value applies it once and uses Local-MDL for h −1 further actions until the fixed budget is completed.The baseline-relative advantage compares this matched continuation with the action the system would otherwise take.
  • B. Baseline-relative fixed-budget value: Subtracting the baseline removes image- and rate-dependent variation in absolute PSNR and makes usefulness operational: candidates must beat the existing action.The evaluator measures matched completion under the declared receiver policy, not a globally optimal set.
  • B. Baseline-relative fixed-budget value: Individual candidate advantages may be negative, whereas exhaustive first-action headroom HB(S) is nonnegative because the Local-MDL action has AB(aL | S) = 0.HB(S) enumerates feasible first actions with fixed Local-MDL continuation; HB(S) − eHB(S) defines proposal regret, not global subset optimality.

C. Multi-source proposal and gate

GCR-C reduces exhaustive action evaluation by using a diversified, fallback-preserving candidate proposal and a strict baseline-relative gate based on matched Local-MDL continuation. The gate is restricted to selected-count states with r(S_t) ≤ 0.30, permits at most one intervention per image, and remains a non-learned correction policy.

  • Proposal construction: The candidate proposal removes duplicates, includes the Local-MDL action a_L(S) explicitly, and uses complementary-coverage candidates, with the P4 union evaluated as a source-ablation reference.This ensures the fallback remains available while comparing alternative proposal sources.
  • Eligibility: Counterfactual gating is eligible only when r(S_t) ≤ 0.30, corresponding to selected-count states 0–19 on CIFAR-10 and 0–43 on STL-10.This eligibility restriction is distinct from the one-intervention-per-image limit.
  • Gating rule: GCR-C selects arg max over proposed actions only when max_a A_B(a | S) > δ; otherwise, it retains a_L(S), with δ = 0 and at most one intervention per image.The strict gate is baseline-preserving under the same evaluator because a_L(S) belongs to the proposal set.
  • Baseline preservation: Under exact matched completion, GCR-C either reproduces the baseline action and continuation or selects a candidate with strictly larger matched-completion value.Both candidate and baseline values use the same Local-MDL continuation policy from their respective post-action states.
  • Limitation: GCR-C is a non-learned correction policy, while learning a lower-cost approximation of its proposal and evaluator pipeline is left for future work.The paper identifies the proposal/evaluator pipeline as the target for future cost reduction.

V. EXPERIMENTAL SETUP AND EVALUATION PROTOCOL · A. Data, model, and held-out evaluation

The evaluation uses discrete visual-token systems with held-out development and validation splits, and tests a fixed GCR-C configuration at selected operating rates under state-based eligibility constraints.

  • A. Data, model, and held-out evaluation: CIFAR-10 uses a 32-entry vector-quantized codebook over non-overlapping 4 × 4 RGB patches.The masked prior has 64 token positions.
  • A. Data, model, and held-out evaluation: 64 token positions and a 32-token codebook imply 5 payload bits per transmitted token.The passage specifies N = 64 and V = 32.
  • A. Data, model, and held-out evaluation: 2,000 images train the codebook and prior, 500 images form the selector-development pool, and 100 images are reserved for validation.The configuration is fixed after proposal, evaluator, gate, and operating-rate schedule development.
  • A. Data, model, and held-out evaluation: The selected configuration combines PC (Top-3 Local plus Top-5 coverage), QB, δ = 0, and ρmax = 0.30.It permits at most one intervention.
  • A. Data, model, and held-out evaluation: Gate evaluation is eligible at 0.20 and 0.32 bpp, but not at 0.44 bpp.Eligibility is evaluated from r(St) = |St|/N before the current outer action.
  • A. Data, model, and held-out evaluation: Eligible states are 0–19 for CIFAR-10 and 0–43 for STL-10.At each eligible state, the gate evaluates the deduplicated proposal and stops after the first accepted intervention or budget termination.

B. Validation studies and baselines · C. Metrics and computation accounting · D. Link-level and high-resolution transfer protocols

The paper validates GCR-C with matched token-selection baselines, explicit reconstruction and computation metrics, and reproducible accounting. It further evaluates frozen selections over coded links and checks transfer at higher resolution with LlamaGen VQ-16.

  • B. Validation studies and baselines: The evaluation audits packet encode–decode consistency, candidate alignment, bit feasibility, deterministic Local-MDL actions, mask consistency, and explicit Local inclusion.Headroom uses 50 images per rate, while validation studies preserve packet accounting, continuation, decoder, and initial-state settings.
  • B. Validation studies and baselines: DivPrune is adapted as a matched external baseline using cosine distances between raw 4 × 4 RGB patches and packet-feasible max–min selection.Each step is restricted to the exact feasible set F(S, Bmax).
  • C. Metrics and computation accounting: Reported metrics include PSNR, SSIM, realized packet bpp, selected-token count, interventions, and counters for proposals, counterfactuals, rollouts, priors, decoders, and accepted interventions.Nprop, Ncf, Nroll, Nprior, Ndec, and Nint respectively track these computation stages or outcomes.
  • C. Metrics and computation accounting: 160 evaluations are the CIFAR-10 worst-case bound, versus 352 for STL-10, with realized counts reduced by early termination, deduplication, or accepted interventions.Statistical comparisons use image-level bootstrap resampling and percentile 95% CIs.
  • D. Link-level and high-resolution transfer protocols: The link experiment reuses frozen clean selections from 500 CIFAR-10 evaluation images without re-optimizing QB decisions for SNR.Packets use CRC-protected 5G-NR LDPC coding near rate 0.8, QPSK, and 20 belief-propagation decoding iterations.
  • D. Link-level and high-resolution transfer protocols: Complex AWGN and single-tap block-Rayleigh fading are tested using frozen validation-pilot operating points selected by nearest FER targets and evaluated on 500 held-out images.The design evaluates a frozen selection over a coded PHY link rather than a channel-aware selector or over-the-air system.
  • D. Link-level and high-resolution transfer protocols: The high-resolution check uses LlamaGen VQ-16 with 16-pixel downsampling, V = 16,384, embedding dimension 8, and a 24×24 grid for 384×384 Kodak-24 crops.A six-layer, six-head masked prior is trained on 3,200 DIV2K crop sequences, with 20 separate DIV2K images for validation.

VI. RESULTS … D. Proposal-size quality–compute frontier

GCR-C has substantial counterfactual headroom at low rates, but its quality gains depend on full-budget evaluation, diversified proposals, and a compact proposal budget. The results position GCR-C as a gated correction layer with visible encoder-side computation and K = 8 near the quality–compute knee.

  • A. Rate-dependent counterfactual headroom: 2.45–2.58 dB mean exhaustive first-action headroom occurs at 0.20 bpp, where 84–86% of states exceed 0.10 dB.Mean headroom falls to 0.73–0.81 dB at 0.32 bpp and 0.27 dB at 0.44 bpp.
  • A. Rate-dependent counterfactual headroom: P4 regret is 0.95 dB at 0.20 bpp, 0.45–0.57 dB at 0.32 bpp, and 0.18–0.22 dB at 0.44 bpp, limiting compact-candidate optimality.Coverage candidates recover gaps left by Local and set-gain sources, so GCR-C is interpreted as a gated correction layer rather than an optimal subset selector.
  • B. Full-budget value versus short-horizon approximations: Q1, Q2, and Q4 recover at most 14.1% of the QB gain at 0.32 bpp and can underperform Local after final-budget continuation.Their confidence intervals include zero at the active rates.
  • B. Full-budget value versus short-horizon approximations: +0.091 dB is achieved by QB at 0.44 bpp, with 95% CI [0.023, 0.183], while costing about 2.61 s/image.QB is the only horizon with a positive interval at both active rates and is used as the reference-quality evaluator.
  • B. Full-budget value versus short-horizon approximations: Nprop = Ncf = 7.6 on average, while full-budget evaluation adds 93.1, 221.3, and 372.4 counterfactual Local rollout actions across the three rates.Each formed proposal candidate is evaluated exactly once in the initial-state diagnostic.
  • C. Proposal-source ablation: Local+Coverage improves over the previous P4 union by 0.250 dB at 0.20 bpp and 0.206 dB at 0.32 bpp.The direct paired Local+Coverage versus Random-control difference is +0.000 dB at 0.20 bpp and +0.200 dB at 0.32 bpp, with intervals reported in the passage.
  • D. Proposal-size quality–compute frontier: K = 8 lies near the main knee of the proposal-size quality–compute frontier.K = 12 improves over K = 8 by 0.202 dB at 0.20 bpp, 0.108 dB at 0.32 bpp, and 0.049 dB at 0.44 bpp while increasing synchronized latency by approximately 27–33%.
  • D. Proposal-size quality–compute frontier: K = 16 adds at most 0.044, 0.009, and 0.012 dB beyond K = 12, supporting the compact K = 8 point for the final proposal.Within that budget, the source study selects a coverage-aware diversification allocation, yielding PC rather than the previous P4.

E. Held-out CIFAR-10 performance · F. Frozen selections over a coded wireless link · G. Realized rate and encoder-side cost

Across held-out CIFAR-10 and coded-link tests, GCR-C improves reconstruction at active rates without increasing realized packet payload. The gains remain positive across tested channels, while requiring substantial encoder-side counterfactual computation.

  • E. Held-out CIFAR-10 performance: E. Held-out CIFAR-10 performance: GCR-C improves over Local-MDL by 1.503 dB at 0.20 bpp and 0.698 dB at 0.32 bpp.Paired win rates are 97.4% and 90.4%, respectively, with strictly positive paired intervals.
  • E. Held-out CIFAR-10 performance: E. Held-out CIFAR-10 performance: At 0.44 bpp, GCR-C follows Local-MDL because the selected operating rule does not evaluate the gate.The two methods are identical at this rate.
  • E. Held-out CIFAR-10 performance: E. Held-out CIFAR-10 performance: Entropy-greedy and Coverage-greedy do not reproduce GCR-C’s active-rate pattern.Entropy-greedy gives +0.902 dB at 0.20 bpp but +0.024 dB at 0.32 bpp with a CI crossing zero; Coverage-greedy gives +0.472 and −0.748 dB.
  • E. Held-out CIFAR-10 performance: E. Held-out CIFAR-10 performance: GCR-C exceeds adapted max–min diversity by +2.017 dB at 0.20 bpp and +1.684 dB at 0.32 bpp.The corresponding 95% CIs are [+1.777, +2.258] and [+1.459, +1.904].
  • F. Frozen selections over a coded wireless link: F. Frozen selections over a coded wireless link: GCR-C has positive paired PSNR gains at all 12 tested rate-channel conditions.Gains range from +0.378 to +1.416 dB, and every image-level 95% bootstrap interval remains above zero.
  • F. Frozen selections over a coded wireless link: F. Frozen selections over a coded wireless link: At 0.20 bpp, gains range from +0.946–+1.416 dB on AWGN and +0.697–+1.304 dB on block-Rayleigh.At 0.32 bpp, the corresponding ranges are +0.378–+0.647 and +0.394–+0.630 dB.
  • G. Realized rate and encoder-side cost: G. Realized rate and encoder-side cost: GCR-C selects nearly the same average token counts as Local-MDL while adding counterfactual evaluations.GCR-C selects 12.01 versus 12.02 tokens at 0.20 bpp and 29.11 versus 29.11 at 0.32 bpp; it executes 13.198 and 26.876 counterfactual evaluations per image.

H. Cross-dataset replication on STL-10 · I. High-resolution transfer on Kodak-24

GCR-C remains effective on STL-10 after changes to dataset, resolution, token grid, and independently trained prior, improving PSNR at both tested budgets. On Kodak-24, it transfers to a substantially larger discrete representation with negligible packet-rate change and positive PSNR gains, while serving as a transfer diagnostic rather than a codec benchmark.

  • H. Cross-dataset replication on STL-10: STL-10 uses native 96 × 96 RGB images, 8 × 8 patches, N = 144 positions, and a newly trained V = 32 codebook/prior.The setup changes dataset, spatial resolution, token grid, and prior relative to the baseline experiments.
  • H. Cross-dataset replication on STL-10: STL-10 evaluation uses a stratified 500-image held-out test subset with budget caps of 350 and 650 bits.Eligibility remains |St|/N ≤0.30, corresponding to selected-count states 0–43 for N = 144.
  • H. Cross-dataset replication on STL-10: +0.621 dB paired gain occurs at the lower budget, with 95% CI [0.559, 0.687] and 99.6% win rate.The interval comes from 10,000 image-level paired bootstrap resamples.
  • H. Cross-dataset replication on STL-10: +0.204 dB paired gain occurs at the middle budget, with CI [0.169, 0.243] and 94.2% win rate.Held-out images are not used for checkpoint, budget, or policy selection, providing limited external-validity evidence.
  • I. High-resolution transfer on Kodak-24: Kodak-24 enlarges the token grid to 24 × 24 (N = 576) and the vocabulary to V = 16,384 for high-resolution transfer.The experiment is defined as a high-resolution transfer diagnostic rather than a semantic-codec benchmark.
  • I. High-resolution transfer on Kodak-24: 83.3% and 70.8% image-level win rates are reported at HR-Low and HR-Mid, respectively.These results provide protocol-scoped evidence that baseline-relative correction transfers to a substantially larger discrete visual representation.
  • I. High-resolution transfer on Kodak-24: Kodak-24 results do not establish that the lightweight high-resolution receiver is a competitive image codec.Separate DINOv2/LPIPS diagnostics are reported in supplementary material and do not establish a consistent conclusion.

VII. LIMITATIONS AND REPRODUCIBILITY … AI USE DISCLOSURE

The paper identifies scope, channel, metric, and protocol limitations, while documenting artifacts and procedures for reproducibility. It concludes that GCR-C improves held-out reconstruction quality under fixed budgets, with AI tools used only for language editing and polishing.

  • A. Limitations: CIFAR-10/STL-10 use compact tokenizers, while lightweight Kodak results do not establish transfer to full LlamaGen, VQGAN, diffusion, or multimodal tokenizers.The Kodak prior also produces weak low-rate reconstructions.
  • A. Limitations: Kodak evaluation assumes AWGN or single-tap block-Rayleigh channels with perfect CSI and excludes mobility, burst errors, feedback, multiuser effects, and OTA impairments.Matched receiver-side tokenizer, prior, and decoder access, along with model-mismatch effects, were not studied.
  • A. Limitations: QB is PSNR-specific: Kodak DINOv2 and LPIPS do not show consistent advantage, so pixel-domain baseline preservation is not semantic-metric preservation.Task-aware utilities, stronger priors, learned or batched evaluation, OTA validation, and channel-aware Qch B selection remain future work.
  • A. Limitations: The adapted max–min result is a protocol-matched adaptation rather than a reproduction of the original LMM pipeline.This limitation concerns the scope of the reported protocol comparison.
  • B. Reproducibility and Data Availability: The reproducibility package includes checkpoints, split manifests, fixed policies, scripts, result records, calibration files, figures, computation traces, and paired-bootstrap summaries.Policies were frozen before held-out evaluation, coded-link trials used shared seeds, and public sources covered CIFAR-10, STL-10, and Kodak-24.
  • VIII. CONCLUSION: GCR-C improves held-out CIFAR-10 PSNR by 1.503 and 0.698 dB at 0.20/0.32 bpp, while the 0.44-bpp policy follows Local-MDL.Under frozen coded-link selections, paired gains are positive at all 12 AWGN/block-Rayleigh conditions; STL-10 and limited Kodak transfer provide protocol-scoped evidence.
Loading 2608.16192v1…