{"id":"7c143bcb-3140-40f0-b83b-ccae59cbdfa2","arxiv_id":"2608.07279","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A token communication framework for MLLMs sends compressed image latents and reconstructs them as both an image prior and adapter-injected visual tokens, improving task performance at low bitrates.","lead":"This paper designs a way to compress and transmit images to a multimodal AI model so the model understands them using far fewer bytes. A small adapter at the receiver turns the compressed data into visual tokens the model can use directly, and tests show it often beats sending a reconstructed image.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Bridge baseline is modified with the proposed reconstruction prior and pooling head, so the claimed superiority over existing MLLM-oriented coding schemes is not established.","rationale":"The reader's weakest_assumption concerns whether adapter token injection via addition preserves the frozen tokenizer's feature manifold and whether ImageNet SigLIP alignment transfers to diverse benchmarks. That is a plausible and important assumption, but the empirical evaluations already include diverse tasks (MME, POPE, SeedBench, COCO Caption) and show consistent gains, so the transfer concern is partially addressed by the reported experiments. The more directly load-bearing threat to the stated central claim is the Bridge baseline modification: the comparison against 'an existing MLLM-oriented coding scheme' is not a comparison against the published Bridge, but against Bridge's loss combined with A-ImageTC's architecture. If the original Bridge cannot converge in this setting, then the paper's abstract overstates the evidence. The conditional verdict is still appropriate because this is fixable by running the unmodified baseline or by tempering the claim, and because the reconstruction-based comparison remains favorable to A-ImageTC. I do not see a reason to move beyond conditional acceptance; the concern is about an unmet verification condition, not an internal inconsistency in the proposed method.","tokens_in":19783,"tokens_out":6303,"duration_ms":64058,"concrete_test":"Obtain or retrain the unmodified Bridge from [32] (adapter branch only, CE loss, no reconstruction prior, no attention pooling head) on Qwen3-VL-8B with the same DCVC-RT codec and ImageNet data. If it converges, plot its MME/POPE/SeedBench/COCO Caption scores against A-ImageTC at the same bpp points as Fig. 5a. If the original code or checkpoints are unavailable, explicitly mark the Bridge comparison as unverified and remove the claim of outperforming existing MLLM-oriented schemes, or release the modified Bridge code so the community can judge the delta.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract; Section V-B, Fig. 5) is that A-ImageTC outperforms both reconstruction and existing MLLM-oriented coding schemes at equal bpp. For the Qwen3-VL experiments, however, the Bridge baseline is not the published Bridge. Section V-A states: 'our reproduction shows that its original training method cannot converge. Therefore, in the experiments related to Qwen3-VL, we keep its loss function, but adopt the proposed architecture with a reconstruction prior and use the attention pooling head to extract visual features.' That is, the comparison Bridge is Bridge's loss fitted into A-ImageTC's receiver architecture, including the very reconstruction prior and attention pooling that the paper argues are beneficial. The reported margin over Bridge may therefore reflect the added components, not the proposed adapter-injection framework. Because the abstract and conclusion claim superiority over 'other image processing schemes,' this baseline modification is load-bearing. The paper also does not release code, checkpoints, or error bars, which compounds the difficulty of verifying the comparison. If the original Bridge cannot be trained at all on Qwen3-VL, the honest claim is that no converging Bridge baseline exists for this setup, not that A-ImageTC beats it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes A-ImageTC, a token communication framework for multimodal large language models. A neural codec is integrated into the vision tokenizer so that only compressed latents are transmitted. At the receiver, a decoder reconstructs a pixel-domain image as a reconstruction prior, while a lightweight adapter converts the same latents into visual tokens that are injected into an intermediate Transformer block of the frozen vision tokenizer by addition. The adapter is trained in two stages: first with a distillation loss that matches the intermediate features of the uncompressed image branch, then with a SigLIP-based semantic alignment loss, with a FiLM-based QP-adaptive variant so that one adapter supports multiple codec rates. Experiments compare the method against reconstruction-based coding and the Bridge baseline on MME, POPE, SeedBench, and COCO Caption, including a Rayleigh fading wireless setting, and ablations study the roles of the adapter, the reconstruction prior, and the SigLIP loss.","tokens_in":20025,"tokens_out":5387,"duration_ms":52931,"significance":"If the reported results are reproducible, the contribution is practically relevant: the method avoids end-to-end LLM training, adds only about two million adapter parameters, works across two neural codecs and two MLLM scales, and shows consistent gains over reconstruction at low bit rates on several external benchmarks. The use of external benchmarks that are not part of the training objective is a real strength, as is the QP-adaptive design that avoids training one adapter per rate point. However, the main empirical claim is currently weakened by the modified Bridge baseline for Qwen3-VL, the absence of error bars, and the lack of released code or checkpoints, so the significance can only be assessed after these points are addressed.","major_comments":[{"comment":"The Qwen3-VL Bridge baseline is not the published Bridge: Section V-A states that the original training method cannot converge and that the authors therefore keep Bridge's loss function but adopt the proposed architecture with a reconstruction prior and attention pooling head. As a result, Figs. 5 and 10 compare A-ImageTC against a hybrid that already contains components the paper argues are beneficial, so the reported margin over Bridge may be due to those components rather than to adapter-based token injection. The abstract and conclusion claim superiority over existing MLLM-oriented coding schemes, but this is not established for the Qwen3-VL experiments. Please either run the original Bridge with documented convergence fixes, or relabel the baseline (e.g., 'Bridge loss plus proposed receiver') and restrict the claims accordingly.","section":"V-A and V-B"},{"comment":"The ablation evidence is incomplete in two ways. First, Table I reports no values for A-ImageTC (w/o SigLIP) and Section V-C states only that the model cannot converge properly; this does not quantify the role of the SigLIP loss and leaves open whether any semantic-alignment loss with a different schedule would work. Second, none of the performance curves in Figs. 5, 7, 8, 9, and 10 carry error bars or multiple-seed results, and some reported gaps are small relative to typical benchmark variance, for example POPE in Fig. 5a at several rate points. Please provide a converged no-SigLIP variant or state the convergence difficulty as a limitation, and add error bars or seed-level results at least for the main comparisons.","section":"V-C and Figs. 5-10"},{"comment":"The semantic alignment in Eq. (44) is trained with ImageNet class labels as positives, yet the evaluation benchmarks (MME, POPE, SeedBench, COCO Caption) are not class-label tasks. The paper should justify empirically that class-level SigLIP alignment transfers to general multimodal understanding and captioning, for example by showing the effect of the text embedding table source or by comparing with an alignment objective built on captions. This is not a circularity issue because the evaluation benchmarks are external, but it is a load-bearing assumption about transfer that currently has no direct evidence.","section":"IV-C, Eq. (44), V-B"},{"comment":"The manuscript does not release code, checkpoints, or evaluation scripts, and the modified Bridge baseline is not fully specified: training hyperparameters, convergence criterion, and which components are frozen are only described textually. Given that the headline comparison relies on a nonstandard baseline modification, this makes the central empirical claim difficult to verify. Please release the artifacts or provide a complete recipe for reproducing every baseline.","section":"V-A"}],"minor_comments":[{"comment":"In Table II, 'evalutation' is a typo for 'evaluation'.","section":"Table II"},{"comment":"The Bridge variants d1, d2, and d3 in Fig. 9 are not defined in the text or caption.","section":"Fig. 9"},{"comment":"The notation section introduces complex-valued quantities and Hermitian/conjugate operators that do not appear in the main derivations; consider trimming them to avoid confusion.","section":"Notations"},{"comment":"The 'Uncompressed' reference is shown without markers or confidence bounds in several figures; please state the exact score values or add them to the captions.","section":"Figs. 5-10"},{"comment":"The BD-rate values in Table I are computed relative to Recon (ELIC), while several earlier figures use DCVC-RT as the default codec; please clarify this choice in the text.","section":"V-C"},{"comment":"The initialization QP for the QP-adaptive adapter is described only as a 'middle-rate point'; please give the numerical QP index used.","section":"V-A"}],"recommendation":"major_revision","confidential_remarks":"The main issue for the editor is the Bridge baseline modification: the authors explicitly adopt their own reconstruction prior and pooling head inside the Bridge baseline for Qwen3-VL, so the quantitative claim of superiority over existing MLLM-oriented coding schemes is not yet supported. If the original Bridge cannot be trained on Qwen3-VL, the honest fix is to reframe the claim and present the comparison as 'Bridge loss plus proposed receiver components.' The paper otherwise fits the journal's scope, and the method is interesting, but the reproducibility package and error-bar reporting need improvement before the empirical claims can be fully trusted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the token communication paper from Ying et al. The core idea is neat: instead of sending pixels or task-specific coded representations, you train the transmitter's vision tokenizer to output compression latents, then at the server you reconstruct both a pixel image and adapter-generated visual tokens that are added into an intermediate transformer block of the frozen MLLM. That is a genuinely different integration of neural codec and MLLM, and the QP-adaptive FiLM module plus the two-stage distillation/SigLIP training is a sensible way to make it work. The experiments show consistent gains over plain reconstruction across MME, POPE, SeedBench, and COCO Caption with two Qwen3-VL sizes and LLaMA-Adapter, which is a decent chunk of evidence. The ablation table, apart from the missing w/o SigLIP entry, supports the roles of the reconstruction prior, the adapter, and the SigLIP loss. The wireless simulation is a plus.\n\nThe stress-test note about Bridge is accurate. Section V-A tells you that the Bridge baseline for Qwen3-VL was not the published Bridge. The authors say the original Bridge training did not converge, so they kept its loss but adopted the proposed architecture with the reconstruction prior and attention pooling head. That means the Bridge curves in Figs. 5 and 10 compare against a hybrid that contains the very components the paper argues are beneficial. The margin over \"Bridge\" could be from the reconstruction prior and pooling head, not from the adapter-injection idea. On LLaMA-Adapter, where Bridge appears original, A-ImageTC is roughly comparable to Bridge (d3), not clearly better. So the claim of superiority over MLLM-oriented coding is inflated. This is a load-bearing caveat, not a minor footnote.\n\nAlso, no error bars, no code or checkpoints. For rate-distortion comparisons where curves often cross within a few points, single runs are hard to judge. The authors are honest about the non-convergence, so I don't think it's a deliberate distortion, but it does mean the evidence cannot support the abstract's sweeping sentence.\n\nMy take: the framework is worth a serious referee and likely worth publishing after revision, with the baseline clearly labeled or replaced, error bars added, and artifacts released. The reader for this is someone working on source-channel coding for MLLM serving, or edge-cloud inference. I wouldn't cite it in its current form, but I'd want to see a revised version.","headline":"The receiver-adapter idea is promising and the ablation story is mostly clean, but the claimed edge over existing MLLM-oriented coding is not proven because the Bridge baseline was rebuilt with the proposed architecture.","tokens_in":20553,"tokens_out":2930,"would_cite":false,"duration_ms":27715,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural codec inside the vision tokenizer plus a receiver-side adapter delivers MLLM-grade visual tokens at far fewer transmitted bits, beating reconstruction-based coding at the same bitrate.","keywords":["token communication","multimodal large language model","neural image compression","vision-language semantic alignment","adapter-based coding","vision tokenizer","rate-adaptive neural codec","feature-wise linear modulation"],"falsifier":"Run the same evaluation on text-dense images such as documents or OCR scenes, where the paper argues the reconstruction prior matters most; if A-ImageTC falls below the reconstruction baseline there, the claimed advantage of adapter token injection does not generalize.","tokens_in":19580,"feed_emoji":"📡","tokens_out":6491,"duration_ms":53336,"temperature":0.7,"pith_summary":"Multimodal large language models (MLLMs) consume enormous numbers of visual tokens, and transmitting images to them is expensive. This paper claims that the right place to compress is inside the vision tokenizer: encode the image into a compact latent, transmit only that bitstream, and at the receiver rebuild both a pixel image and a set of visual tokens that are injected into an intermediate layer of the frozen tokenizer. Under the same number of transmitted bits, this adapter-based token communication scheme outperforms reconstruction-based coding and an existing MLLM-oriented baseline on MME, POPE, SeedBench, and COCO Caption, with the largest gains at low bitrates. If the claim holds, a small, trainable adapter can cut transmitted data substantially without retraining the billion-parameter MLLM.","feed_headline":"Small adapter turns compressed images into MLLM-ready tokens","feed_subtitle":"Adapter-built visual tokens beat reconstruction coding at equal bitrate on MLLM benchmarks.","key_machinery":"The load-bearing mechanism is the injection operation $T_{\\mathrm{vis}} = V^{n:N_V}\\left(T_{\\mathrm{vis}}^{\\mathrm{rec},n-1} + \\alpha T_{\\mathrm{vis}}^{\\mathrm{apt}}\\right)$, where adapter-produced visual tokens are added into the intermediate features of the frozen vision tokenizer after $n-1$ blocks. The adapter is a one-block Vision Transformer with patchification of the decoded latent, optionally modulated by FiLM with a QP embedding; training combines a distillation objective that matches fused tokens to original-image features and a SigLIP (sigmoid language-image pre-training) pairwise alignment loss between attention-pooled visual semantics and a precomputed text embedding table. This two-stage alignment is what makes compression-oriented latents usable as native MLLM tokens.","core_discovery":"The central discovery is that a neural codec and a lightweight Transformer-block adapter can together be viewed as an enlarged vision tokenizer split across a channel. The sender outputs compression-oriented latents, and the receiver uses a decoder to produce a reconstruction prior plus an adapter to produce visual tokens, which are added into the tokenizer's intermediate features at the third block with weight $\\alpha = 0.1$. A two-stage training scheme first distills the injected tokens toward the original-image intermediate features, then aligns the final visual semantics with text semantics through a SigLIP loss, so the injected tokens remain informative to the frozen MLLM. The result is that for equal bits per pixel, A-ImageTC improves task performance over reconstruction and the Bridge baseline, especially in the low-rate regime, and a QP-adaptive FiLM adapter matches per-QP fixed adapters with one set of weights.","pith_inferences":["If the alignment-transfer assumption holds beyond ImageNet, the same adapter recipe could be applied to other frozen visual backbones, turning any codec into an MLLM-aware compressor without joint LLM training.","The paper's separated digital transmission pipeline still shows a cliff effect at low SNR; a natural extension is to make the adapter robust to channel-induced bit errors, or to replace the separated pipeline with joint source-channel coding that feeds soft latents into the adapter.","The FiLM conditioning could be extended to condition on task metadata or prompt embeddings, not just the quantization parameter, potentially steering the injected tokens toward task-relevant semantics.","Since the paper observes the counter-intuitive CIDEr result where compression beats uncompressed input on COCO Caption for the 32B model, a testable implication is that adapter-injected tokens act as a regularization that suppresses over-detailed captioning; measuring factual consistency rather than n-gram overlap would clarify whether this is a real gain."],"forward_implications":["Under equal bits per pixel, MLLM task performance moves closer to the uncompressed-image upper bound, with the largest gains in the low-bitrate regime.","A single QP-adaptive FiLM adapter can serve multiple codec rates, removing the need to train and store one adapter per rate point.","The framework is codec-agnostic: the same adapter works with DCVC-RT, ELIC, and VQ-based codecs, so future codec improvements can be dropped in without retraining the MLLM.","Receiver-side complexity stays nearly constant: the adapter adds about two million parameters and negligible processing time to the full pipeline.","The semantic alignment loss appears essential: ablations show training fails to converge without the SigLIP term, and replacing it with cross-entropy hurts all benchmarks."],"supporting_citations":[{"why":"Supplies DCVC-RT, the main rate-adaptive neural codec used in the proposed framework.","marker":"[11]"},{"why":"Supplies ELIC, the fixed-rate codec used to show the framework works across codec architectures.","marker":"[12]"},{"why":"Provides the SigLIP loss that forms the textual-semantic alignment objective of the second training stage.","marker":"[17]"},{"why":"Defines the Bridge baseline that the paper reproduces and compares against; the paper also adopts its architecture when evaluating with Qwen3-VL.","marker":"[32]"},{"why":"Provides the scale-hyperprior entropy model used for rate estimation and arithmetic coding of compressed latents.","marker":"[34]"},{"why":"Introduces Qwen3-VL, the main MLLM whose frozen vision tokenizer and language model are used in the experiments.","marker":"[4]"}],"fun_headline_variants":["Neural codec + adapter beats reconstruction for MLLM tokens","Split vision tokenizer: compress, then adapt for MLLM","Adapter-injected visual tokens outdo reconstruction on MLLM","Token communication: adapter beats reconstruction at same bitrate","Compress-then-adapt: better MLLM tokens from fewer bits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The design assumes that simply adding the adapter's visual tokens to the frozen tokenizer's intermediate features keeps the feature manifold intact, and that semantic alignment learned on ImageNet class labels transfers to the diverse tasks in MME, POPE, SeedBench, and COCO Caption.","fun_headline_variants_meta":{"raw":{"variants":["Neural codec + adapter beats reconstruction for MLLM tokens","Split vision tokenizer: compress, then adapt for MLLM","Adapter-injected visual tokens outdo reconstruction on MLLM","Token communication: adapter beats reconstruction at same bitrate","Compress-then-adapt: better MLLM tokens from fewer bits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000166,"raw_usage":{"total_tokens":1263,"prompt_tokens":963,"completion_tokens":300,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":211}},"tokens_in":579,"tokens_out":300,"duration_ms":3303,"temperature":1.0,"reasoning_tokens":211,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T11:09:05.161217+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same evaluation on text-dense images such as documents or OCR scenes, where the paper argues the reconstruction prior matters most; if A-ImageTC falls below the reconstruction baseline there, the claimed advantage of adapter token injection does not generalize.","supporting_citations":[{"cited_title":"Towards practical real-time neural video compression,","cited_arxiv_id":null,"evidence_quote":"Supplies DCVC-RT, the main rate-adaptive neural codec used in the proposed framework."},{"cited_title":"ELIC: Efficient learned image compression with unevenly grouped space- channel contextual adaptive coding,","cited_arxiv_id":null,"evidence_quote":"Supplies ELIC, the fixed-rate codec used to show the framework works across codec architectures."},{"cited_title":"Sigmoid loss for language image pre-training,","cited_arxiv_id":null,"evidence_quote":"Provides the SigLIP loss that forms the textual-semantic alignment objective of the second training stage."},{"cited_title":"Bridging compressed image latents and multimodal large language models,","cited_arxiv_id":null,"evidence_quote":"Defines the Bridge baseline that the paper reproduces and compares against; the paper also adopts its architecture when evaluating with Qwen3-VL."},{"cited_title":"Variational image compression with a Scale Hyperprior,","cited_arxiv_id":null,"evidence_quote":"Provides the scale-hyperprior entropy model used for rate estimation and arithmetic coding of compressed latents."}],"review_version":1}