{"id":"0fae3289-9fa5-4c29-89e7-232b472de31b","arxiv_id":"2602.16872","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Block-based discrete diffusion can transcribe documents in parallel, roughly matching autoregressive OCR accuracy while cutting inference time by up to about 3x in a lower-accuracy fast variant.","lead":"DODO is a document-OCR model that generates text in parallel blocks using discrete diffusion instead of one token at a time. It reports accuracy close to standard transcription models and up to about 3x faster inference in its fastest, lower-accuracy configuration.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim conflates architecture with data: DODO is fine-tuned on olmOCR-mix-1025, but the Qwen2.5-VL AR baselines are off-the-shelf; no same-data AR control is provided.","rationale":"The reader's weakest assumption is exactly the confound I find most load-bearing. The paper's contribution is architectural, but its central comparison is to an untrained backbone. The ablations control for diffusion variant but not for decoding paradigm with matched data. Adding a same-data AR baseline would settle whether the gains come from block diffusion or simply from OCR fine-tuning. The Vanilla Oracle row makes the need more acute: it shows a vanilla MDM with oracle length reaches 0.100 NED, undermining the 'catastrophic' framing; without a same-data AR control, readers cannot tell if block diffusion is the key or if any model fine-tuned on olmOCR-mix-1025 would do as well. I am not recommending rejection: the method is plausible and the ablations are suggestive. Conditional remains the right verdict until the matched AR control is supplied.","tokens_in":34488,"tokens_out":6994,"duration_ms":65344,"concrete_test":"Fine-tune Qwen2.5-VL-3B-Instruct on olmOCR-mix-1025 with standard autoregressive cross-entropy for the same 200k steps, batch size, and optimizer as DODO; evaluate on OmniDocBench and Fox-Page-EN using the same NED and tokens/sec protocol. If the fine-tuned AR model achieves NED <= 0.066 on OmniDocBench, or if its TPS advantage over DODO-fast disappears under matched decoding settings, then the headline accuracy/throughput claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1 compares DODO, trained for 200k steps on olmOCR-mix-1025 (Section A), against off-the-shelf Qwen2.5-VL models. The ablations in Table 2 compare DODO to a vanilla MDM trained on the same data, but no autoregressive model is fine-tuned on the same corpus. Therefore the headline claim—that DODO outperforms its Qwen2.5-VL autoregressive backbone—cannot separate the block-diffusion architecture from the effect of OCR fine-tuning. This is load-bearing because the central claim is that discrete diffusion is a viable high-performance alternative to autoregressive decoding (Section 5.3); that requires an AR model matched in data and training budget. A secondary internal inconsistency strengthens the need for this control: Table 2 lists 'Vanilla Oracle' at 0.100 NED, while the text says vanilla MDM fails catastrophically and claims a ~10x reduction; measured against 0.100, DODO's 0.066 is only a ~1.5x improvement. The structural-advantage story is thus less clean than presented, and the missing matched AR baseline becomes the decisive gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DODO, a 3B VLM that applies block discrete diffusion to OCR. DODO trains on olmOCR-mix-1025 and is evaluated on OmniDocBench and Fox-Page-EN. It comes in two variants: DODO (full bidirectional attention) and DODO-fast (block-causal attention with exact KV-caching). The paper claims that DODO achieves near-SOTA accuracy while enabling up to 3x faster inference than an autoregressive Qwen2.5-VL backbone, and argues that block-wise generation is necessary to avoid the alignment failures of global masked diffusion on the rigid, deterministic OCR task.","tokens_in":34765,"tokens_out":3059,"duration_ms":27356,"significance":"If the central claims hold, the contribution is valuable: it is the first demonstration that block discrete diffusion can be made to work in a multimodal OCR setting, and the ablation showing vanilla masked diffusion collapses while block training succeeds provides a concrete, falsifiable structural story. The DODO-fast variant also shows a practical speed/accuracy trade-off via exact KV-caching. However, the comparisons against autoregressive baselines are confounded by training data, and the throughput/accuracy framing needs sharper reporting. The paper's methodological core is sound enough to warrant revision rather than rejection.","major_comments":[{"comment":"The headline comparison against Qwen2.5-VL is confounded: DODO is fine-tuned for 200k steps on olmOCR-mix-1025 (Section A), while the Qwen2.5-VL baselines are off-the-shelf. The claim that DODO 'outperforms its autoregressive backbone' cannot separate architecture from data. Table 2 only ablates MDM variants; it does not include an AR model fine-tuned on the same data and training budget. A matched AR baseline is load-bearing for the §5.3 conclusion that discrete diffusion is a viable high-performance alternative to AR decoding.","section":"§5.2, Table 1"},{"comment":"The speedup claim is internally inconsistent. Figure 5 reports DODO at 22.90 TPS vs Qwen2.5-VL-3B at 21.00 TPS (≈1.09x), while DODO-fast reaches 63.18 TPS (≈3.0x). The abstract in the front matter states 'up to 5x faster inference', while the body abstract states '3×'; no configuration supports 5x. Moreover, the 3x speedup belongs to DODO-fast, whose NED is 0.159 vs 0.066 for DODO — a 2.4x error increase. The paper must clearly separate 'near-SOTA accuracy' from '3x speedup' and correct the abstract.","section":"§5.4, Figure 5; Abstract"},{"comment":"The text claims DODO reduces error rates by 'nearly 10× compared to the best Vanilla configuration'. The best Vanilla row is 'Vanilla Oracle' at 0.100 NED; DODO at 0.066 is only a ~1.5x improvement. The ~10x figure only holds against the non-oracle Vanilla 8192 row (0.834). This misstatement weakens the structural-advantage narrative and must be corrected with the appropriate baseline clearly specified.","section":"§6.1, Table 2"},{"comment":"Throughput numbers are reported as single point values without measurement conditions (batch size, input length, hardware, number of runs) or error bars. Given that the central claim is an efficiency advantage and the differences between configurations are sometimes small (e.g., 22.9 vs 22.0 TPS in Table 3), the authors should specify the protocol and, where feasible, report variance across runs.","section":"§5.4, §6.2"}],"minor_comments":[{"comment":"The two abstract versions disagree on the speedup factor ('5x' in the header, '3×' in the body). Please unify after correcting the actual supported number.","section":"Abstract"},{"comment":"The 'Vanilla Oracle' row is under-specified: the table lists Max Length and Inference Block Size as '-', and the caption does not explain how the oracle length is provided. Add a footnote.","section":"§5.1, Table 2"},{"comment":"Variant names are inconsistently typeset ('DODOfast', 'DODO fast', 'DODOfast'). Use a single notation. Minor typos such as 'suffers' in §5.4 and 'suffers a sharp accuracy collapse' in §6.2 can be cleaned up.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The reader's take is broadly aligned with my assessment. The missing matched AR baseline is the decisive gap; the abstract inconsistency and the Table 2 '10x' misstatement are concrete, fixable issues. I recommend major revision rather than rejection, since the core block-diffusion ablation is valuable and the central architecture is defensible once the comparisons are properly controlled."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: this is a genuinely new application — block discrete diffusion inside a VLM for OCR — and the core empirical story holds up on the evidence they provide. Global masked diffusion collapses on dense transcription; block-wise training fixes it. That is a useful result, and the ablations support it: vanilla MDM fails even with oracle length, and inference-time blocking alone doesn't rescue it. Scaling the block to 256 and getting competitive OCR accuracy at ~3x throughput is a solid engineering contribution. The limitations section is honest about the accuracy/speed tradeoff, which earns credit.\n\nThe soft spots are real but addressable. The biggest one is the missing matched autoregressive baseline. DODO is fine-tuned on olmOCR-mix-1025 while the Qwen2.5-VL AR numbers are off-the-shelf. The accuracy gap (0.066 vs 0.184) could be mostly fine-tuning, and the paper's rebuttal only shows that vanilla diffusion fails on the same data — it doesn't show that a same-data AR model would do worse. That undercuts the headline claim that diffusion is a high-performance alternative to AR decoding. This is the one experiment an editor should insist on.\n\nThe numbers are also sloppy. The abstract says \"up to 5x\" while the body and Figure 5 say 3x. The text in Section 6.1 claims \"nearly 10x\" error reduction compared to the best vanilla configuration, but the best vanilla (Oracle) is 0.100 NED vs DODO's 0.066 — that's 1.5x. They're presumably comparing against full vanilla 0.834. Both need correction. No error bars or code/weights either, which matters for such heavy experiments.\n\nThat said, the circularity burden is low: they aren't fitting parameters to force the result, and the comparison between vanilla and block training is a clean internal ablation. The method itself is sound and well described. I disagree with any reading that treats this as a refutation — the core result is credible and the limitations are openly stated.\n\nVerdict: send to reviewers. It deserves serious referee time, but the referee should ask for a same-data AR fine-tuned baseline, correction of the internal numbers, and ideally code or weights before acceptance.","headline":"Block diffusion for OCR works in practice and the failure of global masked diffusion on dense text is a real finding, but the paper overclaims the comparison to autoregressive models and has a few sloppy numbers that need fixing.","tokens_in":35286,"tokens_out":2861,"would_cite":true,"duration_ms":28928,"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":"A vision-language model using block discrete diffusion matches autoregressive OCR accuracy while decoding in parallel, achieving roughly three times the throughput.","keywords":["OCR","discrete diffusion","masked diffusion","block diffusion","vision-language model","parallel decoding","document transcription","KV-caching"],"falsifier":"Fine-tune the same autoregressive backbone on the same OCR training corpus and measure normalized edit distance and tokens per second on the same benchmarks; if the autoregressive model matches DODO's accuracy, the paper's central accuracy claim collapses. A second check: run DODO on an OCR set with genuinely ambiguous transcriptions (e.g., handwriting); if errors rise sharply where conditional independence fails, the mechanism is confirmed.","tokens_in":34371,"feed_emoji":"📄","tokens_out":3620,"duration_ms":35240,"temperature":0.7,"pith_summary":"The paper argues that OCR's determinism is the key to parallel decoding: because an image dictates a single correct transcription, masked diffusion's assumption that tokens can be predicted independently actually holds, unlike in captioning. It then shows why vanilla masked diffusion still fails at OCR: the rigid, exact-match output cannot tolerate length mismatches or positional drift that flexible text generation absorbs. To fix this, DODO decomposes generation into causally anchored blocks, so each block denoises in parallel against a committed prefix. The result is near-autoregressive accuracy on document benchmarks, with a block-causal variant using KV-caching achieving about 3x the throughput of the autoregressive backbone.","feed_headline":"Block diffusion reads documents 3x faster than autoregressive OCR","feed_subtitle":"A 3-billion-parameter diffusion VLM decodes whole text blocks in parallel while matching autoregressive accuracy on dense document benchmark","key_machinery":"Block discrete diffusion: the sequence is split into contiguous blocks of up to 256 tokens; each block is decoded by masked diffusion in parallel, conditioned on the already-decoded prefix blocks, replicating autoregressive causality at block granularity. The block-causal attention variant freezes prefix representations and enables exact KV-caching; the bidirectional variant recomputes representations for maximal information flow. Ablations show that this structure, not the OCR data alone, is responsible for the gains — full-sequence masked diffusion fails even with oracle length guidance.","core_discovery":"DODO is the first vision-language model to apply block discrete diffusion to OCR. The central discovery is that block-level autoregressive anchoring resolves the structural instability of global masked diffusion: length estimation errors and absolute positional misalignment, which are catastrophic for exact transcription, are bounded within each block, while tokens inside a block are still decoded in parallel. Using 256-token blocks and training with block-causal masking, DODO achieves normalized edit distances of 0.066 on OmniDocBench and 0.041 on Fox-Page-EN, surpassing its autoregressive backbone, and DODO-fast reaches about 63 tokens per second, roughly 3x the backbone's throughput.","pith_inferences":["If the result generalizes, any deterministic serialization task with exact-match targets (structured data extraction, code from screenshots, formula transcription) is a candidate for block diffusion with similar speedups.","The fairness caveat: the reported accuracy superiority over the autoregressive backbone may partly reflect training data rather than architecture; an autoregressive model fine-tuned identically is the natural check.","The accuracy/throughput tradeoff with block size suggests that adaptive block sizing or recurrent refinement of low-confidence tokens could push both metrics; the paper leaves this implicit.","DODO-fast's inverse block-size trend (smaller blocks better with frozen context) hints at a testable hypothesis about representation drift that could guide future cache-friendly architectures."],"forward_implications":["OCR no longer requires one forward pass per token; latency decouples from sequence length.","Block-causal training makes exact KV-caching compatible with diffusion, unlocking roughly 3x throughput.","Diffusion VLMs can match autoregressive accuracy on dense document transcription, closing a gap that prior diffusion VLMs left open.","Larger block sizes than used in text-only block diffusion are viable because OCR's low-entropy posterior supports conditional independence.","The structural benefit comes from training: applying block decoding at inference time to a vanilla masked diffusion model does not recover performance."],"fun_headline_variants":["Diffusion-based OCR decodes in blocks, runs 3x faster","First block-diffusion OCR matches accuracy at 3x speed","Parallel diffusion OCR: 3x faster, same exact match","DODO: discrete diffusion OCR with parallel block decoding","OCR via block diffusion: 3x throughput, accuracy maintained"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The empirical comparison assumes that the off-the-shelf autoregressive baselines are fair stand-ins for DODO, even though they were not fine-tuned on the same OCR corpus; without that assumption, the accuracy and throughput claims may partly reflect data rather than the block-diffusion design.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion-based OCR decodes in blocks, runs 3x faster","First block-diffusion OCR matches accuracy at 3x speed","Parallel diffusion OCR: 3x faster, same exact match","DODO: discrete diffusion OCR with parallel block decoding","OCR via block diffusion: 3x throughput, accuracy maintained"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000272,"raw_usage":{"total_tokens":1460,"prompt_tokens":728,"completion_tokens":732,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":472,"completion_tokens_details":{"reasoning_tokens":645}},"tokens_in":472,"tokens_out":732,"duration_ms":6500,"temperature":1.0,"reasoning_tokens":645,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T22:23:59.911881+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fine-tune the same autoregressive backbone on the same OCR training corpus and measure normalized edit distance and tokens per second on the same benchmarks; if the autoregressive model matches DODO's accuracy, the paper's central accuracy claim collapses. A second check: run DODO on an OCR set with genuinely ambiguous transcriptions (e.g., handwriting); if errors rise sharply where conditional independence fails, the mechanism is confirmed.","supporting_citations":[],"review_version":1}