{"id":"051b1c44-5e2c-435c-b498-47a760d28e65","arxiv_id":"2504.21356","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Prefilled autoregression in a shared embedding space lets Nexus-Gen unify image understanding, generation, and editing, achieving competitive benchmark scores with a 7B model.","lead":"Nexus-Gen is a unified AI model that handles image understanding, generation, and editing in one shared embedding space, combining a language model with a diffusion image decoder. It introduces a prefilled autoregression trick to reduce errors when predicting image tokens, and reports top scores on several benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Prefilled autoregression requires conditional independence of image tokens, not just the order-invariance cited from Li et al. 2024; the paper never tests final-image quality against naive autoregression.","rationale":"This is the most load-bearing concern because prefilled autoregression is the paper's central novelty; if the conditional-independence assumption fails, the mechanism is not training-inference alignment but a lossy parallel approximation. The concern is not that Li et al. is wrong, but that the cited result is being used for a stronger claim than it establishes. The paper has real strengths: detailed training recipe, multi-stage design, released code/models, and an honest Figure 6 that directly measures the error-accumulation phenomenon. The reader's weakest_assumption points to the same area; I sharpen it by distinguishing order-invariance from conditional independence and add a concrete end-to-end test. I retain the CONDITIONAL verdict because the issue is empirically addressable and the paper otherwise provides substantial evidence. If the required ablation is added and prefilled autoregression wins on end-to-end metrics, the concern is resolved; if not, the paper's central mechanism claim would need to be reinterpreted.","tokens_in":15899,"tokens_out":9654,"duration_ms":108325,"concrete_test":"Run the GenEval benchmark (and a fixed-prompt FID comparison if feasible) using the naive-autoregression checkpoint already trained for Figure 6, keeping the generation decoder, data, and compute identical to Nexus-Gen. Report final image-generation metrics for prefilled vs. naive autoregression. If prefilled does not clearly beat naive on these end-to-end metrics, the paper's central claim that prefilled autoregression eliminates the error-accumulation bottleneck is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The prefilled autoregression strategy (Approach, 'Prefilled Autoregression', Fig. 2) replaces previously generated image tokens with NE fixed learnable embeddings at both training and inference. Its justification is one sentence: 'Prior research (Li et al. 2024) demonstrates that image token prediction is permutation-invariant.' But the property demonstrated in Li et al. 2024 is about invariance to the order in which image tokens are predicted; it does not imply that each token is conditionally independent of all other image tokens given the text and its position. Prefilled autoregression implements the much stronger assumption p(e_i | text, position_i) = p(e_i | text, e_1..e_{i-1}) by discarding all real image tokens from the input. Qwen2.5-VL embeddings are patch-level features with strong spatial and semantic dependencies; if those dependencies matter, prefilling discards information the autoregressive model could exploit. The paper's only direct evidence, Figure 6, reports per-token MSE: prefilled train/test 0.4300 vs. naive inference 0.5543, but naive teacher-forced training is 0.2574. That gap shows the model can predict embeddings much more accurately when it conditions on other image tokens. Prefilled autoregression avoids error recycling by removing the conditioning information, and the net effect on generated image quality is never measured: no GenEval or FID comparison between prefilled and naive autoregression is reported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Nexus-Gen, a unified multimodal model that combines an autoregressive LLM (Qwen2.5-VL-7B) with a diffusion decoder (FLUX.1-Dev) through a shared continuous image embedding space. The core technical contribution is 'prefilled autoregression': during both training and inference, the sequence positions for image tokens are filled with NE learnable embeddings instead of previously predicted image embeddings, which the authors argue eliminates autoregressive error accumulation. The model is trained in three stages on a curated 26.3M-sample dataset spanning understanding, generation, and editing, and is evaluated on standard understanding benchmarks (MME, SEED-Bench, MMMU, TextVQA, VQAv2, RealWorldQA), on GenEval for generation, and on a self-built ImagePulse benchmark for editing. The paper reports strong results, including GenEval 0.81 for the instruction-tuned variant, and releases models, data, and code.","tokens_in":16200,"tokens_out":3183,"duration_ms":34421,"significance":"If validated, the prefilled autoregression idea is a simple and potentially useful alternative to standard teacher-forced autoregressive training for continuous image embeddings, and the unified embedding space connecting an LLM to a diffusion decoder is a plausible direction for joint understanding-generation-editing models. The paper contributes a large-scale training dataset and releases all artifacts, which are valuable for reproducibility and follow-up work. However, the central prefilling claim currently rests on an unverified strong independence assumption and is supported only by a single per-token-MSE plot, while the editing benchmark is constructed by the authors themselves. The claimed state-of-the-art status also needs qualification because the base model outperforms the proposed model on understanding benchmarks, and because the reported GenEval state-of-the-art requires additional instruction tuning. These issues are fixable with additional experiments and more careful framing, so the work is a plausible candidate for publication after major revision.","major_comments":[{"comment":"The justification for prefilled autoregression is that 'Prior research (Li et al. 2024) demonstrates that image token prediction is permutation-invariant,' but permutation invariance is not equivalent to the conditional independence p(e_i | text, position_i) = p(e_i | text, e_1..e_{i-1}) that prefilling implements by discarding all real image tokens from the input. Figure 6 actually undercuts the argument: teacher-forced training with real previous tokens achieves per-token MSE 0.2574, while prefilled training achieves 0.4300, showing the model can predict embeddings much more accurately when it conditions on other image tokens. The paper never measures final image quality (e.g., GenEval or FID) for naive versus prefilled autoregression, so the claimed benefit of 'eliminating error accumulation' is not distinguished from the effect of removing conditioning information. Please add an ablation that compares final generation quality under both training-inference schemes, and either verify the conditional independence assumption on Qwen2.5-VL embeddings or revise the theoretical claim.","section":"Approach, 'Prefilled Autoregression' and Fig. 6"},{"comment":"The editing evaluation is performed on a 1,000-sample test set randomly sampled from the same ImagePulse dataset used to train the editing decoder, making the editing claim circular: the decoder is tuned and evaluated on the same constructed distribution. The paper should also report results on external editing benchmarks such as MagicBrush, HQ-Edit, or Emu-Edit (or at least on independently collected test data), and should compare against unified models such as Emu3, Janus-Pro, or SEED-X on editing if such comparisons are feasible. Without this, the 'state-of-the-art' editing claim in Table 3 is not established.","section":"Experiments, 'Image Editing' and Table 3"},{"comment":"The abstract and conclusion state that Nexus-Gen 'achieves state-of-the-art performance' without qualification, but Table 2 shows that the jointly trained Nexus-Gen (0.77) is below Janus-Pro 7B (0.80) on GenEval; only the instruction-tuned Nexus-Gen* reaches 0.81. Similarly, Table 1 shows that the underlying base model Qwen2.5-VL-Instruct 7B outperforms Nexus-Gen on every understanding benchmark (e.g., MMMU 50.6 vs. 45.7; VQAv2 82.3 vs. 79.3), so the understanding 'state-of-the-art' claim holds only among unified models and at the cost of nontrivial degradation relative to the base. Please qualify the abstract and conclusion to state which comparisons support the SOTA claim and to acknowledge the base-model degradation.","section":"Abstract and 'Main Results', Table 2"},{"comment":"Figure 6 reports a single training/inference loss trajectory with no error bars, no multiple seeds, and no downstream image-quality metric, so the claimed advantage of prefilled autoregression over naive autoregression is not statistically supported. The ablation also omits a comparison of generation and editing results using the naive autoregressive model at inference time, which is necessary to determine whether the reduced per-token MSE translates into better final images. Please provide repeated-run statistics and end-to-end metrics, or explicitly state the limitations of the current evidence.","section":"Experiments, 'The Impact of Prefilled Autoregression', Fig. 6"}],"minor_comments":[{"comment":"The contributions list contains a tense inconsistency: 'We proposed a prefilling strategy' should read 'We propose a prefilling strategy' to match the other bullets.","section":"Introduction, contributions list"},{"comment":"The paragraph contains a duplicated sentence: 'However, these datasets exhibit notable limitations' appears twice in consecutive sentences; please remove the duplicate.","section":"Appendix, 'Dataset Construction Details', Image Editing paragraph"},{"comment":"The notation 'BOI' and 'EOI' in Figure 2 is not defined in the text; please define these tokens in the caption or in Section 3.2.","section":"Figure 2 and Equation (1)"},{"comment":"The table uses abbreviations MME-P, MME-C, RWQA without expansion; please define all metric acronyms either in the caption or in the text before the table.","section":"Table 1"},{"comment":"The description of Nexus-Gen* says it is fine-tuned on Blip3o-60k, but the appendix reports this dataset contains only 0.06M samples; please clarify whether the fine-tuning data includes other sources or whether the 60k size is correct.","section":"Experiments, 'Image Generation' and Table 2"},{"comment":"The formula NE = (H/P) * (W/P) assumes a simple patch grid, but Qwen2.5-VL uses dynamic resolution with additional merging/subsampling; please specify whether NE is exactly this product or an approximation for the actual vision encoder.","section":"Approach, 'Unified Image Embedding Space' and Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing worth knowing about this paper is the prefilled autoregression strategy. Instead of feeding predicted image embeddings back into the autoregressive model, it replaces them with learnable embeddings at both training and inference. That is a simple, elegant way to sidestep error accumulation, and I have not seen it in the prior work they cite. The rest of the architecture—Qwen2.5-VL encoder plus FLUX decoder in a shared embedding space—is well engineered, and the 26.3M-sample dataset and multi-stage training are sensible. The understanding results are respectable: they land at 45.7 MMMU, close to the base Qwen2.5-VL, which is non-trivial when you add a generation head. The GenEval 0.81 after instruction tuning is a strong number. The soft spots are real but not fatal. The stress-test note is right: the justification via permutation invariance from Li et al. is a misreading. Order invariance does not imply that image tokens are conditionally independent given the text, and the paper never measures final image quality for prefilled versus naive autoregression. Figure 6 only reports per-token MSE, and the fact that teacher-forced training gets 0.2574 while prefilled gets 0.43 suggests the model loses some conditioning information. That is a legitimate concern, and the paper should address it with a direct comparison on GenEval or FID. The editing evaluation is weaker: it uses a 1,000-case test set from the same in-house ImagePulse distribution the decoder was trained on. That is a mild circularity, and an external editing benchmark would strengthen the claim. The abstract also overstates the generation result: the base model gets 0.77, below Janus-Pro-7B's 0.80; only the instruction-tuned variant reaches 0.81. And there is a contradiction—the abstract says all datasets are released, but the dataset section says release will happen after security checks. None of this invalidates the core idea. The prefilled autoregression is worth testing further, and the paper's empirical MSE improvement is a hint that it helps in practice. A serious referee should ask for a final-image quality comparison between prefilled and naive autoregression, an external editing benchmark, and a corrected abstract. With those changes, this could be a well-cited paper.","headline":"The prefilled autoregression trick is a genuinely novel idea with promising numbers, but the paper's evidence for it is thinner than it looks and the abstract oversells the generation result; worth a serious referee but needs revision.","tokens_in":718,"tokens_out":938,"would_cite":true,"duration_ms":30350,"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":"One image space unifies understanding, generation, and editing by replacing token-by-token feedback with prefilled placeholders, and the paper reports state-of-the-art scores on all three tasks.","keywords":["unified multimodal model","prefilled autoregression","shared image embedding space","image understanding","image generation","image editing","continuous embeddings","flow matching"],"falsifier":"Train a small autoregressive model on the same continuous embedding space with the prefilled strategy and with naive teacher-forcing, then evaluate generation quality and per-token MSE at inference on a fixed prompt set. If the prefilled variant does not reduce inference loss compared to naive autoregression, or if shuffling the order of ground-truth image embeddings during training changes the predicted embeddings, the permutation-invariance assumption is falsified. A direct test: encode the same image with its tokens in the natural order and in a shuffled order, then check whether the autoregressive model's predictions for subsequent tokens differ measurably; if they do, the embedding space is order-dependent and the prefilling logic collapses.","tokens_in":15716,"feed_emoji":"🖼️","tokens_out":4767,"duration_ms":48324,"temperature":0.7,"pith_summary":"Nexus-Gen is a unified model that claims one architecture can understand, generate, and edit images by routing all visual information through a single continuous embedding space. The paper's central assertion is that an autoregressive language model, a diffusion decoder, and a vision encoder can share this space without forcing the language model to predict discrete tokens, and that a prefilled autoregression strategy prevents the cascade of prediction errors that normally plagues continuous embedding generation. If true, this offers a practical route to a single system that both reasons about images and produces or modifies them, trained jointly on interleaved text-image data. The paper reports state-of-the-art numbers across understanding, generation, and editing benchmarks, including an overall 0.81 on GenEval for the instruction-tuned variant.","feed_headline":"One image space unifies understanding, generation, and editing","feed_subtitle":"Prefilled autoregression avoids the error cascade, hitting state-of-the-art marks on all three tasks.","key_machinery":"The prefilled autoregression mechanism, together with the shared embedding space. During training, input sequences are prefilled with $N_E$ learnable embeddings plus positional encodings at the image-token positions; during inference, the same learned embeddings are inserted as soon as the beginning-of-image token is predicted. This makes training and inference compute over the same input distribution, so errors from one predicted token never get recycled into the next prediction. The composite image loss is mean squared error plus cosine similarity against ground-truth embeddings, and the vision decoder is conditioned on these predicted embeddings through a projector. A second conditioning stream for editing uses 324-token embeddings of the input image to preserve details.","core_discovery":"The core claim is that prefilled autoregression in a shared continuous image embedding space removes the error-accumulation bottleneck that limits previous autoregressive image generators. Instead of predicting image embeddings token by token and feeding those predictions back as input, the model fills the image-token positions with learned placeholder embeddings during both training and inference. This relies on the invariance of image token prediction to ordering: each embedding can be predicted from the text and its positional encoding alone, so the prefilled placeholders lose no information while making train and test behavior identical. Nexus-Gen ties this to a unified embedding space where image understanding, generation, and editing all operate: understanding encodes input images into the space, generation predicts semantic embedding tokens that a diffusion decoder turns into images, and editing feeds both semantic and fine-grained embeddings into a dual-conditioned decoder. Empirical sections report state-of-the-art results on MMMU (45.7), GenEval (0.81 after instruction tuning), and editing metrics on the ImagePulse test set.","pith_inferences":["If the permutation-invariance assumption holds in this continuous embedding space, the same prefilling trick could be applied to other continuous token domains such as audio or video, where autoregressive feedback drift is even more severe; this is a testable extension the paper does not explore.","The observed trade-off between 81 tokens and 324 tokens suggests a capacity limit in autoregressive prediction of dense embeddings; one could adaptively choose token budgets based on image complexity or saliency rather than a fixed count.","The paper's own limitations section admits sensitivity to prompt variations and unexplored visual reasoning; a natural next step is to evaluate whether the unified embedding space serves as a scratchpad for step-by-step vision-language reasoning, where prefilled placeholder tokens could act as intermediate reasoning states.","The editing decoder's joint attention layer for fusing semantic and detail streams is a reusable architectural pattern for other multi-condition diffusion tasks, such as guided inpainting or subject-driven generation."],"forward_implications":["A single 7B-parameter model can match or surpass specialized single-task models on image understanding, generation, and editing without task-specific heads, because all tasks share one embedding interface.","Interleaved text-image corpora become directly usable for joint training, since visual inputs and outputs are expressed in the same continuous space as the language model's own representations.","The prefilled autoregression strategy promises to remove the training-inference gap for any continuous-valued autoregressive prediction, not just images, since it eliminates the feedback of biased predictions.","For editing, dual conditioning that combines coarse semantic tokens with fine-grained detail tokens preserves unedited regions better than a single-stream generation decoder, pointing to a reusable two-stream design.","Bilingual generation and editing emerge naturally from re-annotating a portion of the training data, indicating that the unified space adapts to language diversity without architectural changes."],"supporting_citations":[{"why":"Supplies the permutation-invariance claim that prefilled autoregression relies on, allowing learned placeholders to replace sequential image tokens.","marker":"Li et al. 2024"},{"why":"Provides the pretrained vision encoder and autoregressive model whose embedding space and weights initialize Nexus-Gen.","marker":"Bai et al. 2025"},{"why":"Supplies the diffusion transformer (FLUX.1-Dev) used as the vision decoder for generation and editing.","marker":"Labs 2024"},{"why":"Defines the autoregressive-with-diffusion paradigm that Nexus-Gen extends, and serves as a baseline for generation comparisons.","marker":"Ge et al. 2024"},{"why":"Provides Janus-Pro, a strong unified-model baseline for understanding and generation, and a comparison point for the shared-space design.","marker":"Chen et al. 2025b"},{"why":"Supplies the GenEval benchmark used to measure text-to-image alignment in the generation experiments.","marker":"Ghosh, Hajishirzi, and Schmidt 2023"},{"why":"Provides the cosine similarity loss component for image embeddings and the CLIP metrics used in editing evaluation.","marker":"Radford et al. 2021"},{"why":"Presents MetaMorph, a representative autoregressive-with-diffusion unified model that serves as a baseline and motivation for the prefilling fix.","marker":"Tong et al. 2024b"}],"fun_headline_variants":["Prefilled autoregression unifies image tasks in one space","Shared embedding space ends error accumulation in image AI","Nexus-Gen: one space for understanding, generation, editing","Image model uses prefilled tokens to stop error cascade"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole prefilling design depends on the claim that image-token prediction is permutation-invariant, meaning each image embedding can be predicted from the caption and its position alone; if that invariance does not hold in the continuous embedding space, prefilled placeholders would discard order information the model needs.","fun_headline_variants_meta":{"raw":{"variants":["Prefilled autoregression unifies image tasks in one space","Shared embedding space ends error accumulation in image AI","Nexus-Gen: one space for understanding, generation, editing","Image model uses prefilled tokens to stop error cascade"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1344,"prompt_tokens":946,"completion_tokens":398,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":331}},"tokens_in":562,"tokens_out":398,"duration_ms":4038,"temperature":1.0,"reasoning_tokens":331,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:04:49.696554+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a small autoregressive model on the same continuous embedding space with the prefilled strategy and with naive teacher-forcing, then evaluate generation quality and per-token MSE at inference on a fixed prompt set. If the prefilled variant does not reduce inference loss compared to naive autoregression, or if shuffling the order of ground-truth image embeddings during training changes the predicted embeddings, the permutation-invariance assumption is falsified. A direct test: encode the same image with its tokens in the natural order and in a shuffled order, then check whether the autoregressive model's predictions for subsequent tokens differ measurably; if they do, the embedding space is order-dependent and the prefilling logic collapses.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the diffusion transformer (FLUX.1-Dev) used as the vision decoder for generation and editing."}],"review_version":1}