{"id":"b7199eef-11d4-4b17-8c7f-a67d0cd87fb2","arxiv_id":"2607.25527","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A compact unified model that reuses a frozen VLM encoder and hybrid continuous/discrete tokens reaches competitive image understanding and generation with 15.6M training images and about $2,000 in compute.","lead":"Argus-Unified shows that a vision-language model can be turned into a model that both understands and generates images by adding a small image-codebook 'tail' and fine-tuning with about 15 million images at a training cost of around $2,000. The importance is that unified image-text models, normally very expensive to train, may be within reach of small research groups.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing base-VLM baseline: understanding SOTA may be inherited from InternVL3-2B, not achieved by unified training","rationale":"The reader's verdict is CONDITIONAL, and this concern reinforces that condition rather than moving it. The reader's weakest_assumption focused on the frozen-encoder information ceiling in Stage 1; that is a plausible concern but the paper provides substantial counter-evidence (rFID 1.60 on ImageNet, 0.63 with SigLIP2, and Table 11 showing reconstruction quality is not the main driver of downstream performance). The more load-bearing weakness is the missing base-VLM baseline, which the reader mentions in the rationale but does not treat as the central assumption. Without reporting InternVL3-2B's GQA/POPE/VQAv2 scores under identical evaluation, the paper cannot support the unqualified claim of state-of-the-art understanding, nor can it quantify how much understanding is retained when generation capability is added. The proposed test is a single inference run and would settle the question directly. I therefore recommend keeping the CONDITIONAL verdict, with the added condition that the authors report base-VLM baselines and, if necessary, soften the 'state-of-the-art' wording.","tokens_in":23490,"tokens_out":10184,"duration_ms":103621,"concrete_test":"Run the official InternVL3-2B (and InternVL3-1B) checkpoint through the exact same evaluation pipeline used for Argus-Unified on GQA, POPE, and VQAv2, with identical prompts, templates, and preprocessing. Compare those numbers to Table 1 and Table 8. If the base VLM scores are at or above Argus-Unified's scores, the 'state-of-the-art understanding' claim is inherited rather than achieved, and the paper must soften its headline and explicitly report the understanding cost of adding generation; if the base VLM scores are lower, the concern is resolved and the unified training demonstrably improves understanding.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim—state-of-the-art multimodal understanding on GQA, POPE, and VQAv2 with only 15.6M data and ~$2,000—is not anchored to the pretrained VLM used for initialization. Argus-Unified-1.5B is initialized from InternVL3-2B, but InternVL3-2B is never evaluated on these benchmarks. Since Stage 2 (Sec. 3.2) includes understanding SFT on ShareGPT4V, WiT, VFLAN, and MGM-Instruct, any observed understanding score could be (a) inherited from the base VLM, (b) improved by the added SFT, or (c) degraded by joint generation training. The ablation in Table 5 only compares VLM initialization versus separate ViT+LLM initialization; it does not report the untouched InternVL3-2B score, so it cannot answer whether the unified model preserves or improves the pretrained understanding ability. The reader's tokenizer-ceiling concern is real but less decisive: Table 4 shows ImageNet rFID of 1.60 for the frozen default encoder and 0.63 with SigLIP2, and the authors explicitly show better reconstruction does not translate into better unified performance (Table 11). The missing base-model baseline directly undermines the quantitative 'state-of-the-art understanding' claim, which is load-bearing for the central argument that a unified model can be trained economically without sacrificing understanding.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Argus-Unified proposes a two-stage recipe for turning a pretrained VLM into a compact unified image-understanding and image-generation model. A frozen vision encoder is retained from a pretrained VLM; Stage 1 trains a quantizer and an image decoder on top of this frozen encoder, and Stage 2 trains the LLM (initialized from the VLM) jointly on understanding and generation data with a hybrid token design: continuous tokens for understanding and discrete tokens for generation. The paper reports state-of-the-art understanding results on GQA, POPE, and VQAv2 among the compared unified multimodal models, competitive generation results, and a total training cost of about $2,000 using 15.6M public data samples.","tokens_in":23846,"tokens_out":4356,"duration_ms":46517,"significance":"The main contribution is the empirical claim that a unified multimodal model can be obtained economically by adapting an existing pretrained VLM rather than training components from scratch. If the claim survives scrutiny, the hybrid-token design and the two-stage recipe are practically valuable, and the paper's emphasis on public data and low compute is a useful counterpoint to the large-scale trend in this area. The paper includes multiple ablations (initialization, token types, data volume, CFG, tokenizer data, resolution), which strengthen the evidence for the proposed pipeline. The absence of the base-VLM baseline, however, makes the headline 'state-of-the-art understanding' claim insufficiently supported.","major_comments":[{"comment":"The headline understanding claim is not anchored to the pretrained VLM used for initialization. Argus-Unified-1.5B is initialized from InternVL3-2B, and Stage 2 SFT includes standard understanding data (ShareGPT4V, WiT, VFLAN, MGM-Instruct). The paper never reports the performance of untouched InternVL3-2B on GQA, POPE, VQAv2, or MME-P under the same evaluation protocol. Therefore the observed numbers could reflect the base VLM's inherited ability, the added SFT data, or degradation from joint generation training. Table 5 only contrasts initialization from a pretrained VLM against initialization from separate pretrained ViT and LLM; it does not isolate the effect of unified training relative to the base model. This is load-bearing for the abstract's 'state-of-the-art multimodal understanding' claim and for the paper's central economic argument. I request adding the untouched InternVL3-2B","section":"§4.2, Tab. 1, Tab. 8; §3.2"},{"comment":"The comparison set for 'state-of-the-art multimodal understanding' omits current strong understanding-only VLMs beyond the base model. Table 8 includes MobileVLM, LLaVA, InstructBLIP, Qwen-VL-Chat, and IDEFICS, but not more recent and stronger compact VLMs such as InternVL3-2B itself, Qwen2.5-VL, or SmolVLM. The paper's SOTA language is broad ('state-of-the-art multimodal understanding on GQA, POPE, and VQAv2'), but the evidence only supports 'best among the UMMs tabulated and among the older understanding-only models tabulated.' Adding the base VLM and at least one or two contemporary compact VLMs would let the reader assess whether the unified training preserves the pretrained understanding ability, which is the central claim.","section":"§4.2, Tab. 1 and Claims in Abstract"}],"minor_comments":[{"comment":"The table heading says 'training with (w/) and without CFG', but classifier-free guidance is typically an inference-time mechanism. Please clarify whether CFG dropout is used during training, and specify the guidance scale used for the main results in Table 1, since reproduction of generation scores depends on this.","section":"§4.3, Tab. 3"},{"comment":"The cost arithmetic should be stated unambiguously. Table 2 says '800×H100' under 'GPU Hours', while the footnote quotes an 8×H100 node rate of $20.3/hour. 800 GPU-hours at the per-GPU rate would be about $16,000; 100 node-hours at $20.3/hour gives about $2,032. If the latter is the intended calculation, say so explicitly and label the table column 'GPU-hours' or 'node-hours' consistently.","section":"§4.1, Tab. 2 and footnote 1"},{"comment":"Janus-Pro has no MJHQ-30K or VQAv2 entry, so the statement that generation quality is 'competitive compared to ... Janus-Pro' rests on GenEval alone. Please state this limitation in the comparison or fill in the missing values if available.","section":"Tab. 1"},{"comment":"The inverted axis for MJHQ-30K is explained but the caption could be clearer that only the axis is reversed and that lower original FID is better. Consider labeling the axis with the original FID scale directly instead of an inverted scale.","section":"Fig. 1 caption"},{"comment":"The paper does not mention whether code, model checkpoints, or the exact training/evaluation scripts will be released. Given the paper's reproducibility emphasis and public data usage, stating the release plan would strengthen the contribution.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The core recipe is promising and the paper is transparent about many implementation details. The missing base-VLM baseline is the key technical issue: it directly affects the central 'state-of-the-art understanding at low cost' claim, but it is easily addressable by evaluating InternVL3-2B and InternVL3-1B under the same protocol and adding them to the comparison tables. I therefore recommend major revision rather than rejection. I would also encourage the authors to moderate the SOTA wording unless the augmented comparison set supports it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Main thing you should know: the paper's core result is plausible and useful—you can turn a pretrained VLM into a unified model with hybrid continuous/discrete tokens, a frozen encoder, and roughly $2k in compute—but the headline \"state-of-the-art understanding\" is not currently supported because the base VLM (InternVL3-2B) is never evaluated on the same benchmarks. That missing baseline matters more than the tokenizer-ceiling concern the reader flagged: Table 11 actually shows better reconstruction does not translate into better unified performance, so the frozen-feature ceiling is not the load-bearing weakness.\n\nWhat is genuinely new: the hybrid-token design from a single frozen encoder, with a quantizer/decoder trained for reconstruction and the LLM warm-started from a VLM. That is a legitimate variation on Janus/VILA-U/UniTok, and the paper gives credit to those. The ablations are unusually informative: VLM init beats separate ViT+LLM init (Table 5), hybrid tokens beat pure discrete (Table 6), and more SFT data helps while more PT data hurts (Table 10). The dataset is public and the cost estimate is transparent. That is a solid engineering contribution.\n\nThe soft spots are real but mostly fixable. The missing base-VLM baseline is the big one. If InternVL3-2B already scores around the same numbers on GQA/POPE/VQAv2, then the \"SOTA\" is inherited, not earned by unified training. The paper needs to report the untouched base model under the same evaluation protocol. Second, the comparison set in Table 1 is wide but omits the base VLM and several stronger VLMs; \"state-of-the-art\" is too strong a phrase. Third, the numbers are inconsistent across tables—GQA for the 1.5B model appears as 62.8, 63.0, and 62.95 in different places. That looks like sloppy table-building, but in a paper whose claim rests on those digits it needs to be cleaned up. Fourth, no code or full hyperparameters are released; a \"low-cost recipe\" paper should ship those. The tokenizer-ceiling concern is minor in light of Table 11: the authors explicitly show that higher reconstruction fidelity does not produce better generation or understanding.\n\nWho is this for: people building small unified models on a limited budget; it is a strong baseline and a useful data/training recipe. It deserves a serious referee, but the referee should insist on the base-model baseline and number alignment before acceptance.","headline":"Argus-Unified is a genuinely cheap recipe for unified multimodal modeling, but its SOTA understanding claim is unanchored without the base VLM baseline.","tokens_in":24353,"tokens_out":2067,"would_cite":true,"duration_ms":19694,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Argus-Unified shows that a unified image understanding-and-generation model can be trained for about $2,000 on 15.6 million public examples by bootstrapping a pretrained vision-language model with hybrid continuous/discrete visual tokens, r","keywords":["unified multimodal models","image understanding","image generation","hybrid visual tokens","pretrained vision-language models","autoregressive generation","vector quantization","efficient training"],"falsifier":"Swap the frozen vision encoder for one whose features are global-pooled (no spatial detail) and retrain only the Stage-1 quantizer and decoder; if the ImageNet reconstruction FID degrades sharply and Stage-2 generation benchmarks (GenEval, MJHQ-30K) collapse while understanding stays high, the frozen-encoder information ceiling—not the LLM—is what bounds generation.","tokens_in":23383,"feed_emoji":"🖼️","tokens_out":4965,"duration_ms":49833,"temperature":0.7,"pith_summary":"Argus-Unified claims that a unified image understanding and generation model can be trained for roughly $2,000 on 15.6 million public examples—about 10 times less compute and 5 times less data than comparable unified models—by starting from a pretrained vision-language model instead of aligning modalities from scratch. Its central idea is a hybrid token design: the same frozen vision encoder feeds continuous tokens to the language model for understanding and discrete tokens to a learned quantizer/decoder for generation. On GQA, POPE, and VQAv2 the 1.5B model beats unified models with dedicated generation encoders and larger backbones, while matching their generation quality on GenEval and MJHQ-30K. If correct, this shows the main cost driver for unified models is re-learning vision-language alignment, which can be inherited from existing VLMs.","feed_headline":"Unified image model trained for $2,000 tops three benchmarks","feed_subtitle":"Starting from a pretrained vision-language model, 15.6M data and hybrid tokens match far larger generators.","key_machinery":"Hybrid visual tokens from a single frozen vision encoder. Continuous tokens (raw encoder features) are projected into the LLM for understanding; the same features are mapped through a multi-codebook vector quantizer into discrete codes for generation, and an image decoder reconstructs pixels from those codes. A two-stage training pipeline first fits the quantizer/decoder with reconstruction, perceptual, adversarial, and vector-quantization losses, then fine-tunes the VLM-initialized LLM with next-token prediction on mixed data. The load-bearing mechanism is that freezing the encoder and inheriting VLM weights removes the need for large-scale vision-language alignment.","core_discovery":"The paper claims that pretrained vision-language models can be turned into unified multimodal models with a two-stage recipe. Stage 1 trains a vector-quantized tokenizer (quantizer plus image decoder) on top of a frozen vision encoder using only reconstruction losses; Stage 2 initializes the LLM from a pretrained VLM and fine-tunes it with a next-token objective on mixed understanding and generation data. The hybrid token scheme keeps continuous features for understanding and quantized features for generation, avoiding the semantic loss of discrete-only models and the parameter overhead of dual-encoder designs. The authors report state-of-the-art understanding on GQA, POPE, and VQAv2 and gen","pith_inferences":["If this recipe holds, any strong open vision-language model could be converted into a unified image model at modest cost, making unified image models accessible to small labs and startups.","The frozen-encoder assumption implies a ceiling on generation: the paper's own ablation with a different encoder and codebook suggests that encoder choice matters, and future work may close the gap to dedicated tokenizers.","The finding that more pretraining data can degrade performance when starting from a pretrained VLM may generalize to other VLM-initiated fine-tuning, inverting the usual from-scratch data scaling recipe.","The method's scope as stated is compact models; extending to image editing or video generation would stress the frozen-encoder information ceiling and could reveal where the hybrid token design breaks down."],"forward_implications":["Unified multimodal training can be done at micro-budget; the cost and data barrier drops by an order of magnitude relative to current unified models.","A VLM's continuous visual features are rich enough to also drive generation through a learned quantizer, so dedicated generation encoders are not necessary.","Hybrid tokens substantially improve understanding over discrete-only tokens while preserving generation quality, as shown by the ablation.","A 0.5B-LLM variant outperforms several 7B+ unified models on understanding benchmarks, suggesting LLM scale is not the main driver in this regime.","Data scaling behaves differently when starting from a pretrained VLM: more supervised fine-tuning data helps, but more pretraining data can hurt."],"fun_headline_variants":["$2k and 15.6M data train a unified vision model that tops 3 benchmarks","Hybrid tokens enable $2k unified vision model to top 3 benchmarks","Pretrained VLM + hybrid tokens: $2k unified model beats larger ones","Two-stage recipe turns pretrained VLM into unified image model for $2k"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that a frozen semantic vision encoder retains enough fine-grained visual detail for a learned quantizer and decoder to reconstruct high-quality images, and the paper only claims this in the compact model regime.","fun_headline_variants_meta":{"raw":{"variants":["$2k and 15.6M data train a unified vision model that tops 3 benchmarks","Hybrid tokens enable $2k unified vision model to top 3 benchmarks","Pretrained VLM + hybrid tokens: $2k unified model beats larger ones","Two-stage recipe turns pretrained VLM into unified image model for $2k"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001471,"raw_usage":{"total_tokens":5779,"prompt_tokens":797,"completion_tokens":4982,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":4901}},"tokens_in":541,"tokens_out":4982,"duration_ms":34988,"temperature":1.0,"reasoning_tokens":4901,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T02:09:06.088204+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Swap the frozen vision encoder for one whose features are global-pooled (no spatial detail) and retrain only the Stage-1 quantizer and decoder; if the ImageNet reconstruction FID degrades sharply and Stage-2 generation benchmarks (GenEval, MJHQ-30K) collapse while understanding stays high, the frozen-encoder information ceiling—not the LLM—is what bounds generation.","supporting_citations":[],"review_version":1}