{"id":"6df107dd-16c9-46f6-8649-fe3a488b41ea","arxiv_id":"2504.14666","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Diffusion timestep tokens give large language models a recursive visual language that improves unified multimodal comprehension and generation relative to spatial patch tokens.","lead":"This paper proposes representing images as ordered sequences of discrete visual tokens that gradually fill in missing details, allowing a single large language model to both understand and generate images. The authors report that this approach outperforms other unified multimodal models on several image generation, editing, and visual question answering benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The tokenizer objective in Eq. (1) only requires each prefix to reconstruct x0; it does not force the recursive, semantically ordered token hierarchy that the central claim and the editing gains depend on. The evidence offered is qualitative and needs a decisive ordering ablation.","rationale":"The reader's weakest assumption identifies the same load-bearing issue: nothing in the encoder or quantizer forces a consistent, semantically meaningful hierarchy along diffusion timesteps. My reading sharpens this by pointing to the exact mechanism: Eq. (1) is a nested-conditional reconstruction loss, not a recursive-generation loss, so the 'compensation' property is only weakly encouraged. This is not an internal inconsistency; the method can work even if the narrative is overclaimed, and the empirical results, if reproducible, would still support a conditional accept. The paper's own Appendix G acknowledges limited reconstruction quality on open-domain images, which tempers the SOTA claim but does not change the verdict. Given that the paper already lacks released code, weights, and data, and the reader's CONDITIONAL verdict accounts for that, my concern reinforces the condition rather than moving the verdict. The recommended verdict therefore remains unchanged.","tokens_in":25842,"tokens_out":7311,"duration_ms":82063,"concrete_test":"Retrain the DDT tokenizer with the identical architecture, hyperparameters, and Eq. (1), but reverse the conditioning order: at timestep t, feed the decoder (V_T,...,V_{T-t+1}) instead of (V_1,...,V_t), masking the remaining tokens as before. Then compare ImageNet reconstruction PSNR/FID and the Section 5.4.2 counterfactual-interpolation and progressive-decoding behaviors against the original tokenizer. If the reversed-order tokenizer matches the original on these checks, the diffusion-timestep ordering is not load-bearing and the recursion claim needs revision; if it clearly degrades, the ordering is doing real work and the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DDT tokens are recursive: ft+1(x0) = (ft(x0), Vt+1), with each newly appended token compensating for the attribute loss between timesteps. But Eq. (1) trains the tokenizer only with an expanding-prefix reconstruction loss: at timestep t the decoder sees (V1,...,Vt) and the noised image xt, and must reconstruct x0. This objective requires each prefix to be sufficient; it does not require that Vt+1 encode the incremental attribute lost between t and t+1, nor that the ordering be semantically consistent across images. The encoder sees the clean image x0 and can distribute information across positions in many ways; the only pressure toward incremental encoding is the codebook bottleneck and the nested training schedule, neither of which enforces a linguistic or attribute hierarchy. The qualitative evidence in Sections 5.4.2 (counterfactual interpolation, progressive decoding with subsets) is consistent with the recursion story, but it is also consistent with the decoder exploiting the known noise level t and a few coarse tokens to fill in details, without the token sequence itself carrying a recursive visual language. Figure 1 tests whether an LLM is order-sensitive to the frozen token sequences, not whether the tokenizer actually learned the claimed hierarchy. Since the editing advantages and the 'visual language' narrative are attributed to recursivity, this unforced ordering assumption is the most load-bearing step in the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Discrete Diffusion Timestep (DDT) tokens: an image tokenizer that encodes an image into a sequence of discrete codes, one per diffusion timestep, trained so that the first t tokens together with the noised image x_t reconstruct x_0 (Eq. 1). The authors argue that this yields a recursive visual language whose tokens progressively compensate for attribute loss as the timestep increases, in contrast to spatial tokens that form an 'impossible language' for LLMs. They integrate these tokens into an 8B LLaMA-3-based MLLM trained on 200M image-text pairs and report improved results over other MLLMs on text-to-image generation (GenEval, T2I-CompBench, DrawBench), instruction-based image editing (EVR, MA5K, MagicBrush), and vision-language comprehension benchmarks, despite a tokenizer trained only on ImageNet at 256x256. Additional analyses cover class-conditional ImageNet generation, counterfactual interpolation, progressive decoding with token subsets, a perplexity comparison, an A/B editing test with a 2B Gemma backbone, and preliminary scaling observations.","tokens_in":26131,"tokens_out":4164,"duration_ms":40108,"significance":"If the central claim holds, the paper makes a useful conceptual contribution: replacing spatial token order with a diffusion-timestep order that is claimed to be recursive and semantically structured, while keeping a single next-token-prediction objective for both text and image. The breadth of benchmarks is a strength, and the reported results are generally strong for an MLLM generalist. The paper also presents a controlled A/B comparison (DDT-Gemma vs. MoVQ-Gemma) and an ImageNet class-conditional FID comparison, which are helpful evidence beyond the headline tables. However, the load-bearing property—that the learned token order is genuinely recursive and semantically consistent across images—is currently supported mainly by qualitative demonstrations and by an order-perturbation experiment on the LLM, not by a direct ablation of the tokenizer's ordering. The single-run point estimates in Tables 1-3 and the use of previously published baseline numbers also weaken the confidence in the claimed superiority. The idea is timely and potentially impactful, but the evidence as presented is not yet decisive.","major_comments":[{"comment":"The recursion claim f_{t+1}(x0) = (f_t(x0), V_{t+1}) is not entailed by the training objective in Eq. (1). The loss only requires that the decoder map (x_t, t, V_1, ..., V_t) to x0; since x_t itself preserves most information for small t, the objective can be satisfied by late tokens that are redundant or that encode image-specific statistics rather than the incremental attribute loss between t and t+1. The counterfactual interpolation and progressive-decoding figures are consistent with the claimed hierarchy, but they are also consistent with the decoder exploiting the known noise level t and a few coarse tokens. To make the central claim load-bearing, the authors should provide a decisive ordering ablation: train the tokenizer with (i) the proposed order, (ii) the reversed order, and (iii) a fixed random permutation of timestep indices, holding architecture and data fixed, and then compare reconstruction PSNR and the downstream editing A/B test. If ordering has little effect, the 'recursive visual language' interpretation is a training artifact; if it has a large effect, the size and direction of that effect should be quantified.","section":"Section 3, Eq. (1); Section 5.4.2"},{"comment":"The headline comparisons are based on single-run point estimates without error bars, confidence intervals, or significance tests, and many baseline numbers are taken from published papers rather than rerun in an identical setup. This matters because several headline margins are small, for example GenEval Overall 0.66 vs. Transfusion 0.63 and DrawBench Clip-T 0.267 tied with LaVIT and Lumina-mGPT. The authors should either provide multiple runs (at least for a smaller controlled setting) or state clearly that the differences are not statistically assessed; without this, the claim of systematic superiority over other MLLMs is not fully supported.","section":"Tables 1-3 in Sections 5.1-5.3"},{"comment":"The 'impossible language' argument and the recursion interpretation rely on Figure 1, which shows LLM training loss under different perturbation types for DDT tokens and spatial tokens. The figure lacks experimental detail such as the number of seeds, architecture, training length, and the magnitude of the perturbation; more importantly, order sensitivity of the LLM does not establish that the token order corresponds to a semantically meaningful attribute hierarchy. An LLM can be sensitive to arbitrary positional statistics without the tokens encoding recursive, disentangled attributes. The authors should report quantitative order-sensitivity measures (e.g., final loss gaps averaged over seeds) and connect them to the downstream editing and counterfactual results.","section":"Section 5.4.2 and Figure 1"},{"comment":"The A/B test on image editing is potentially the strongest evidence for the claimed advantage, but the description is incomplete. The paper does not report the number of human raters, the exact instructions, whether ties were allowed, how the test cases were selected, or the total number of cases. The text states that DDT-Gemma outperforms MoVQ-Gemma in 65 editing cases and that MoVQ-Gemma surpasses DDT-Gemma in only 10, but without a total and a protocol the reader cannot judge significance or bias. This should be documented, together with inter-annotator agreement if multiple raters were used.","section":"Section 5.4.4 and Figure 9(b)"}],"minor_comments":[{"comment":"There are typos: 'two weaks' should read 'two weeks' (appears twice), and the reference 'Coyo [ ? ]' in Appendix B.1 is incomplete.","section":"Appendix A.1 and B.1"},{"comment":"Main-text figure references are inconsistent with the numbering: Section 5.1 refers to 'Figure 10' for qualitative T2I examples, and Section 5.4.2 refers to 'Figure 16' and 'Figure 17' for counterfactual interpolation and subset decoding, while the corresponding main-paper figures are numbered 3, 6, and 8.","section":"Section 5.1 and 5.4.2"},{"comment":"In the GenEval subtask list, 'single object (TwoObj)' should be 'single object (SingObj) and two objects (TwoObj)'.","section":"Appendix B.2"},{"comment":"The captions contain article errors: 'an blue cell phone' and 'an purple suitcase' should be 'a blue cell phone' and 'a purple suitcase'.","section":"Figure 12 captions"}],"recommendation":"major_revision","confidential_remarks":"The paper's central contribution is the claim that the timestep ordering of the discrete tokens induces a recursive, semantically consistent visual language. That claim is currently supported by qualitative evidence and by an LLM order-sensitivity experiment, but not by a direct tokenizer ordering ablation. I would ask the authors to add such an ablation before acceptance; if the ordering is shown to be unimportant, the paper would need reframing as an empirical study of a particular tokenization scheme rather than a demonstration of recursion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know about 2504.14666. First, the central idea—ordering discrete visual tokens by diffusion timestep instead of spatial position—is genuinely new and deserves attention. Second, the headline claim that these tokens form a 'recursive visual language' is not yet demonstrated; the objective doesn't force it, and the evidence offered is mostly qualitative.\n\nWhat the paper does well: the DDT tokenizer is a clean design. An encoder maps a clean image to T=480 query vectors; each is quantized to a 65K codebook; a rectified-flow decoder reconstructs x0 from the noisy image xt and the first t tokens. Training with expanding prefixes is a sensible way to encourage coarse-to-fine information. The order-perturbation figure is a nice observation: autoregressive loss on DDT tokens is more order-sensitive than on spatial tokens. The empirical scope is large—T2I, editing, comprehension, plus a controlled Gemma-2B A/B editing test—and the ImageNet-only tokenizer transferring to open-domain tasks is notable. The appendix limitation is honest about reconstruction lagging MOVQ.\n\nSoft spots, in order of seriousness. First, no code, weights, or data; every benchmark table is single-run point estimates, and Table 4 leans on published FID values rather than re-run baselines. Second, the recursivity claim: Eq. (1) always gives the decoder (V1..Vt) and xt, so each prefix must be sufficient; nothing forces Vt+1 to encode the incremental attribute loss. The counterfactual and progressive-decoding figures are consistent with the story but also consistent with the decoder exploiting the known t and a few coarse tokens. What is missing is an ordering ablation—permute the timestep-to-token assignment during tokenizer training, or measure how much each appended token changes the output. Without that, 'recursive' is an interpretation, not a property. Third, the 'impossible language' conclusion rests on one loss-curve figure. Fourth, the contribution list says SOTA; on GenEval they beat other MLLMs but trail SD3, so the claim is selective. Minor: the same-group MorphTokens reference is not differentiated.\n\nI largely agree with the reader's conditional verdict. The paper deserves a serious referee—the idea is novel, the system is substantial, and a reviewer can push for the ordering ablation and release. I would not accept as-is, but desk rejection would be wrong. It's a paper for anyone working on visual tokenization or unified MLLMs; bring it to reading group and argue about whether the recursion is learned or imposed.","headline":"Novel diffusion-timestep tokenization with strong but under-verified results; the recursivity claim needs a decisive ordering ablation.","tokens_in":26699,"tokens_out":4534,"would_cite":true,"duration_ms":40837,"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":"The paper claims that ordering visual tokens by diffusion timesteps instead of spatial position gives multimodal LLMs a recursive visual language that makes unified comprehension and generation work.","keywords":["discrete diffusion timestep tokens","recursive visual tokens","multimodal large language models","unified comprehension and generation","vector quantization","image editing","text-to-image generation","next-token prediction"],"falsifier":"Take a complete DDT token sequence and decode it after shuffling the token order, then compare reconstruction quality with the ordered case: if the recursion claim holds, shuffled tokens should reconstruct the image substantially worse, whereas if the tokens just encode the image redundantly, a shuffled but complete set would still reconstruct well. A second check is to train the tokenizer with a reversed timestep order under the same loss and see whether the editing and generation advantages disappear.","tokens_in":25630,"feed_emoji":"🖼️","tokens_out":9550,"duration_ms":78236,"temperature":0.7,"pith_summary":"The paper argues that spatial visual tokens—image patches read in spatial order—lack the recursive structure of language, so an LLM trained on them is memorizing an impossible language rather than learning to see. It replaces that ordering with one inherited from the diffusion process: as noise is added to an image in increasing timesteps, an encoder emits one new discrete token per timestep that compensates for the attributes lost so far. These Discrete Diffusion Timestep (DDT) tokens form an expanding prefix, so the diffusion decoder can reconstruct the original image from the noised image plus the first $t$ tokens. The authors build an 8-billion-parameter multimodal LLM around this tokenizer, train it with next-token prediction on 200M image-text pairs, and report that it surpasses other generalist MLLMs on text-to-image generation, zero-shot image editing, and several visual comprehension benchmarks even though the tokenizer was trained only on ImageNet.","feed_headline":"Images become recursive tokens when ordered by diffusion timesteps","feed_subtitle":"A single ImageNet-trained tokenizer lets an 8B model beat larger MLLMs at generation and editing.","key_machinery":"The load-bearing object is the DDT tokenizer: a transformer encoder with $T=480$ learnable query tokens, a vector-quantization codebook of 65,536 entries, and an MMDiT diffusion decoder. During training, the noised image is sampled by Rectified Flow interpolation, $x_t = t\\epsilon + (1-t)x_0$; the decoder receives $x_t$, the timestep $t$, and only the first $t$ quantized tokens, and is trained to return $x_0$. This expanding-prefix reconstruction loss is what forces each new token to carry the information removed by the extra noise. The paper also uses an order-perturbation experiment, borrowed from the impossible-language test, to show that DDT sequences lose predictive performance when shuffled, while spatial token sequences are nearly order-insensitive.","core_discovery":"The central claim is that images can be tokenized into a discrete, recursive visual language by aligning token order with diffusion timesteps. For a clean image $x_0$, the encoder $f$ produces a sequence $(V_1,\\ldots,V_T)$; at timestep $t$, the first $t$ tokens together with the noised image $x_t = t\\epsilon + (1-t)x_0$ are enough for the diffusion decoder to reconstruct $x_0$. Because each appended token accounts for the attributes destroyed by the additional noise, the sequence satisfies $f_t(x_0) = (f_{t-1}(x_0), V_t)$, which gives the token stream a language-like recursive structure. The paper supports this with three kinds of evidence: autoregressive training curves that degrade when DDT token order is perturbed, as human language does; counterfactual interpolation in which swapping a subset of tokens changes only the corresponding attributes instead of cutting and mixing image regions; and progressive decoding in which the earliest tokens fix fine details while later tokens build contours and color. With this tokenizer, the MLLM learns to translate between text and DDT tokens under a single next-token prediction objective, and the same diffusion decoder serves both generation and comprehension.","pith_inferences":["Any signal with a natural degradation process—video, audio, or 3D scenes under corruption—could be tokenized along its own timestep ordering; the design principle of recovering lost attributes with expanding discrete prefixes is not image-specific.","The learned attribute hierarchy is encouraged by the reconstruction loss rather than enforced by a structural constraint, so a direct test of the visual-language claim would measure how well the ordering transfers to out-of-domain images, where the paper reports strong results but does not quantify attribute disentanglement.","A testable extension is to train the tokenizer with reversed or shuffled timestep order under the same loss: if the advantages in editing and generation persist, recursion is not the active ingredient; if they vanish, the diffusion-timestep ordering is doing the work.","Because editing quality improves when the backbone grows from 2B to 8B, the DDT formulation may offer a direct path to attribute-level controllable generation at scale, since token-level interventions become semantically meaningful."],"forward_implications":["An MLLM can be built without a pretrained CLIP-style vision encoder and still match or beat comprehension-specialized models on captioning, VQA, and MLLM benchmarks.","Zero-shot instruction-based image editing improves because editing becomes a matter of changing the specific tokens that carry the modified attributes; DDT-LLaMA outperforms both generalist and specialist editors on EVR, MA5K, and MagicBrush.","Counterfactual interpolation with DDT tokens changes only the substituted attributes, whereas spatial tokenizers behave like CutMix, producing a patch-wise blend rather than a semantic one.","Scaling the LLM backbone from 2B to 8B parameters and increasing training compute improves visual quality, consistent with scaling laws for the DDT-based formulation.","The paper's stated limitation is that tokenizer reconstruction quality, not token ordering, is the bottleneck: a tokenizer trained only on ImageNet at 256x256 restricts aesthetic quality for open-domain generation, so scaling tokenizer data is the clear next step."],"supporting_citations":[{"why":"Supplies the order-sensitivity test used to argue spatial tokens are an impossible language for LLMs while DDT tokens behave like language.","marker":"[34]"},{"why":"Provides the Rectified Flow noise schedule $x_t = t\\epsilon + (1-t)x_0$ used in the tokenizer's reconstruction loss.","marker":"[54]"},{"why":"Supplies the MMDiT architecture reused as the diffusion decoder and the two-transformer encoder design.","marker":"[19]"},{"why":"Supplies the vector-quantization module that turns encoder features into discrete codebook entries.","marker":"[76]"},{"why":"Provides the empirical basis for the low-dimensional 16-dim projection that improves codebook usage.","marker":"[84]"},{"why":"Serves as the spatial tokenizer baseline for reconstruction and class-conditional generation comparisons.","marker":"[18]"},{"why":"Serves as the MoVQ spatial tokenizer baseline in the A/B editing test and the reconstruction comparison.","marker":"[94]"},{"why":"Provides the main next-token-prediction MLLM baseline for generation and comprehension comparisons.","marker":"[77]"},{"why":"Provides the pretrained 8B LLM backbone whose vocabulary is expanded with the visual codes.","marker":"[17]"},{"why":"Supplies the 200M image-text pairs used for multimodal pretraining.","marker":"[68]"}],"fun_headline_variants":["Diffusion timesteps turn images into a recursive visual language","Visual tokens ordered by diffusion timesteps form a recursive language","Recursive image tokens from diffusion timesteps boost MLLMs","Diffusion timesteps give images a recursive token structure","Language-like visual tokens via diffusion timestep ordering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single fixed ordering of tokens along diffusion timesteps induces a consistent, semantically meaningful attribute hierarchy across all images, so that the first $t$ tokens together with the noised image suffice to reconstruct the original image.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion timesteps turn images into a recursive visual language","Visual tokens ordered by diffusion timesteps form a recursive language","Recursive image tokens from diffusion timesteps boost MLLMs","Diffusion timesteps give images a recursive token structure","Language-like visual tokens via diffusion timestep ordering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000599,"raw_usage":{"total_tokens":2827,"prompt_tokens":996,"completion_tokens":1831,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":1747}},"tokens_in":612,"tokens_out":1831,"duration_ms":11087,"temperature":1.0,"reasoning_tokens":1747,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:43:07.065097+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a complete DDT token sequence and decode it after shuffling the token order, then compare reconstruction quality with the ordered case: if the recursion claim holds, shuffled tokens should reconstruct the image substantially worse, whereas if the tokens just encode the image redundantly, a shuffled but complete set would still reconstruct well. A second check is to train the tokenizer with a reversed timestep order under the same loss and see whether the editing and generation advantages disappear.","supporting_citations":[{"cited_title":"Movq: Modulating quantized vectors for high- fidelity image generation","cited_arxiv_id":null,"evidence_quote":"Serves as the MoVQ spatial tokenizer baseline in the A/B editing test and the reconstruction comparison."},{"cited_title":"Laion-5b: An open large-scale dataset for train- ing next generation image-text models","cited_arxiv_id":null,"evidence_quote":"Supplies the 200M image-text pairs used for multimodal pretraining."}],"review_version":1}