{"id":"57ad3e09-557b-48b6-b413-7d84f33e9f4d","arxiv_id":"2601.20430","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A 2.5B vision-language model parses documents into text, formulas, tables, charts, seals, and hierarchy, reporting 5-11x speedups via token-parallel decoding and SOTA OmniDocBench scores.","lead":"Youtu-Parsing is a 2.5-billion-parameter document parsing model that claims top scores on two public benchmarks while using a parallel decoding scheme to cut inference time. The report introduces a masked-token parallel decoding strategy and a region-batching trick, with code, weights, and a demo supposedly released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 10 contradicts the lossless-equivalence claim: FlashAttention Overall score changes with parallelism degree n (92.93→93.30), which cannot happen if outputs are claimed bit-level identical to autoregressive decoding.","rationale":"The reader correctly identified the unstated deterministic/greedy decoding assumption as a risk, but the stronger problem is that the paper's own Table 10 contradicts the losslessness claim even under the most charitable reading. If verification is exact and output is bit-level equivalent to autoregressive decoding, then changing n cannot change the benchmark score; the FlashAttention column in Table 10 shows exactly such a change. This is an internal inconsistency, not merely an unspecified implementation detail. It directly undermines the central contribution, which is the claim that a 5–11× speedup is obtained without any change to the generated output. The speedup figures in the abstract, body, and conclusion are also mutually inconsistent, reinforcing that the efficiency claims have not been reported with sufficient care. Because the core losslessness claim is contradicted by the paper's own experimental evidence, the appropriate verdict moves from CONDITIONAL to REJECT unless the authors can demonstrate that the Table 10 numbers are erroneous or that a deterministic evaluation recovers identical outputs across n. The SOTA accuracy results on public benchmarks are valuable and reproducible in principle, but they do not salvage a central contribution that is internally contradicted.","tokens_in":21618,"tokens_out":6480,"duration_ms":63287,"concrete_test":"Run the released checkpoint on the same OmniDocBench v1.5 pages under greedy/argmax decoding with the same FlashAttention backend, for n=16, n=64, and standard autoregressive decoding (n=1). Compare the output sequences token-by-token and recompute the Overall score. If the n=16 and n=64 outputs differ at any position, or if their Overall scores reproduce 92.93 vs 93.30, the lossless equivalence claim is false. Independently, rerun Table 10 with a fixed seed and identical evaluation code to determine whether the three FlashAttention scores are reproducible or are run-to-run noise.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's headline contribution is the lossless token-parallel decoding scheme, asserted in Sections 2.3.1 and 5 and formalized by the exact-match verification in Eq. (2). If the scheme truly produced output identical to standard autoregressive decoding, the final token sequence—and therefore any deterministic benchmark score—would be invariant to the parallelism degree n and to the attention backend. Table 10 directly contradicts this: under Eager Attention the Overall score is 92.90 for n=16, 32, and 64, while under Flash Attention it rises from 92.93 (n=16) to 93.17 (n=32) to 93.30 (n=64). A verification mechanism that accepts exactly the longest prefix matching the autoregressive model cannot produce different final outputs for different n; a larger n only changes how many tokens are verified per step, not which tokens are ultimately accepted. The observed variation implies either that the two forward passes are not exact-match equivalent (for example, sampling rather than greedy argmax is used, or FlashAttention nondeterminism prevents bit-identical verification), or that the scores come from inconsistent evaluation conditions. Either way, the claimed bit-level equivalence is not supported by the paper's own experiments. The speedup claims are also internally inconsistent: the abstract and Section 2.3.1 state 5–11×, Table 10 reports a maximum 11.13×, Table 11 reports up to 26.82×, and the Conclusion states 10–20×. The SOTA accuracy results may remain valid, but the central 'lossless' contribution is not substantiated as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces Youtu-Parsing, a 2.5B-parameter document parsing vision-language model built on a three-stage decoupled architecture: a NaViT-based shared visual feature extractor, a layout analysis stage, and region-prompted decoding with Youtu-LLM-2B. The central technical contribution is a dual-track parallel decoding scheme. Token parallelism appends up to n=64 mask tokens per step, generates candidate tokens in one forward pass, and verifies them with a second pass using the exact-match rule of Eq. (2); the authors claim this is lossless and mathematically equivalent to autoregressive decoding, yielding 5--11x speedups. Query parallelism batches up to five bounding-box queries in one forward pass, claiming an additional ~2x speedup with equivalent output quality. The paper reports state-of-the-art accuracy on OmniDocBench v1.5 (overall 93.22) and olmOCR-bench (80.5 +/- 0.9), plus fine-grained results on text, table, formula, chart, and seal recognition. It also describes hybrid masked training (80/20), a three-stage training recipe with GRPO, an iterative data refinement pipeline, and open releases of code, model weights, and a demo.","tokens_in":21837,"tokens_out":4962,"duration_ms":41669,"significance":"If the central claims held, this would be a practically significant result: a single 2.5B model that outperforms dedicated pipelines and much larger VLMs on two public document-parsing benchmarks, while also delivering a theoretically lossless and empirically fast decoding scheme, is genuinely useful for high-throughput document intelligence. The open release of code, model weights, and a demo is a clear strength, and the use of external benchmarks (OmniDocBench, olmOCR-bench) provides independent grounding for the accuracy claims. However, the lossless-equivalence assertion is not supported by the paper's own ablations: reported scores vary with parallelism degree n and with the attention backend, which is impossible if outputs are bit-identical to autoregressive decoding. Because losslessness is the stated basis for claiming zero accuracy degradation and mathematical equivalence, this inconsistency is load-bearing. The accuracy results on public benchmarks may well survive independently, but the efficiency and equivalence claims need substantial correction or re-evidence.","major_comments":[{"comment":"The claimed lossless equivalence is contradicted by the reported OmniDocBench v1.5 scores. Under the exact-match acceptance rule of Eq. (2) with greedy deterministic decoding, the accepted token sequence—and therefore any deterministic benchmark score—cannot depend on the number n of mask tokens or the attention backend. Table 10 shows Flash Attention Overall rising from 92.93 (n=16) to 93.17 (n=32) to 93.30 (n=64), while Eager Attention stays at 92.90 and differs from Flash at every n. Please explain this variation (e.g., sampling, nondeterministic attention kernels, or inconsistent evaluation conditions) and either provide bit-identical outputs across all configurations or revise the 'mathematical equivalence' and 'bit-level equivalence' claims in Sections 2.3.1 and 5.","section":"§4.3.1, Table 10"},{"comment":"The losslessness guarantee is only valid under a deterministic decoding rule. The verification accepts a candidate when it exactly equals the token from the second forward pass; this guarantees identity with standard autoregressive decoding only if tokens are selected by argmax and if the two forward passes are bitwise deterministic. The paper never states the sampling temperature, whether greedy decoding was used in all experiments, or how nondeterminism from FlashAttention is handled. If sampling is used for output diversity, exact-match verification does not preserve the target distribution, and the claimed equivalence in Sections 2.3.1 and 5 fails. Please specify the exact decoding configuration and demonstrate determinism, or narrow the claim.","section":"§2.3.1, Eq. (2)"},{"comment":"The speedup numbers are internally inconsistent. The abstract and Section 2.3.1 state 5--11x, Table 10 reports a maximum of 11.13x, Table 11 reports up to 26.82x, and the Conclusion states 10--20x. Moreover, Table 11's caption defines the baseline as n=2, but the table includes n=1 rows with 1.00x speedup. Please report one consistent measurement protocol, specify the baseline, and explain how the per-scenario speedup of 26.82x relates to the claimed 5--11x aggregate.","section":"§4.3, Tables 10-11, Conclusion"},{"comment":"Query Parallelism is claimed to maintain output quality equivalent to standard decoding with deterministic sequence decomposition and mapping back to layout elements. Yet Table 12 reports the Overall score changing from 88.77 (m=1) to 90.12 (m=5). If the model's token-level outputs were identical across m, a deterministic scorer would give the same Overall score. The variation implies either different evaluation workloads, nondeterministic decoding, or that query batching changes the outputs; in any case the 'quality equivalent' claim needs qualification or correction.","section":"§4.3.2, Table 12"}],"minor_comments":[{"comment":"In the bullet list of contributions, 'åToken Parallelism' appears to contain a stray character; please correct the typo.","section":"§1, Contributions"},{"comment":"Table 5 reports 'Edit Distance Performance' with values around 98-99 for handwritten text, yet the surrounding text calls these 'recognition accuracies.' Please state explicitly whether lower or higher is better and whether the values are 1 - normalized edit distance.","section":"Table 5 and §4.2.1"},{"comment":"In the definition of CSS, the symbol d_edit is described only as character-level edit distance; please specify the normalization convention and the handling of empty cells consistently.","section":"§4.2.4, Eq. (5)"},{"comment":"The fine-grained evaluation relies on in-house datasets that are not released; please state whether these datasets will be made available to enable independent verification of the tables.","section":"§4.2.1-§4.2.5"},{"comment":"The resolution notation '12288×32×32' in Table 2 is not explained; please clarify the meaning of the three factors.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The contradiction between the claimed lossless equivalence and the n-dependent scores in Table 10 is the key issue; I would ask for a reproducibility script or raw evaluation logs before accepting. The SOTA accuracy claims on public benchmarks may be independent of the decoding claims and could survive a revised paper that either proves determinism or weakens the equivalence statement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version first. Youtu-Parsing is a 2.5B document parser that covers text, formulas, tables, charts, seals, and hierarchy in one model, and it hits SOTA on OmniDocBench v1.5 (93.22) and olmOCR-bench (80.5±0.9), with code and weights linked. That alone makes it a useful resource. The decoding scheme—mask tokens drafted in parallel and then verified by a second forward pass—is a legitimate variant of speculative decoding, and the query-parallel batching of up to five regions is a pragmatic addition. If the external numbers hold up on reproduction, this is a meaningful engineering contribution.\n\nThe soft spot is the 'lossless' claim, and it is not minor. Table 10 is supposed to demonstrate that output fidelity is preserved as the parallelism degree n changes. Under Eager attention the Overall score is 92.90 for n=16, 32, and 64—exactly what you would expect if the output is bit-identical. Under Flash attention the score moves from 92.93 to 93.17 to 93.30 as n grows. If the verification step really accepts only the longest prefix that matches the autoregressive model, the final sequence cannot depend on n. So either the decoding uses sampling rather than greedy argmax, or the attention implementation introduces nondeterminism that breaks exact-match verification, or the evaluation was run under inconsistent conditions. The paper never states the sampling temperature or addresses FlashAttention nondeterminism, so the bit-level equivalence asserted in Sections 2.3.1 and 5 is not supported by the evidence in the same document. That is a load-bearing problem for the central contribution.\n\nThere are other, smaller inconsistencies. The abstract and Section 2.3.1 say 5–11×, the conclusion says 10–20×, and Table 11 reports up to 26.82× for table scenarios. These can probably be reconciled if you state what is being measured (token throughput vs end-to-end latency, with and without query parallelism), but right now the numbers just don't agree. The fine-grained evaluation also leans on in-house datasets and author-proposed metrics (CSS, Edge-F1) for several subclaims, which is fine for a technical report but gives those subclaims a circularity burden; the public-benchmark SOTA claims are on firmer ground. The olmOCR-bench lead over PaddleOCR-VL is 80.5 vs 80.0, within the reported error bars, so that particular SOTA claim should be phrased carefully.\n\nWho should read this: anyone building or benchmarking production document-parsing systems, and anyone working on accelerated decoding for structured generation. Despite the flawed lossless framing, the model and the public results are worth serious attention. For peer review, I would send it out, but the authors need to explain Table 10, state the decoding distribution and determinism assumptions, align the speedup numbers, and release the eval subsets. Without that, the core claim is unsubstantiated.","headline":"A genuinely useful document-parsing model with released weights, but the paper's headline lossless-decoding claim is contradicted by its own ablation table and needs a major fix before it is credible.","tokens_in":22538,"tokens_out":3901,"would_cite":true,"duration_ms":32943,"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":"Youtu-Parsing claims that a 2.5B-parameter document parser can decode with a 5–11× speedup via lossless token parallelism while achieving top scores on OmniDocBench v1.5 and olmOCR-bench.","keywords":["document parsing","token parallelism","query parallelism","lossless speculative decoding","NaViT vision encoder","Hybrid Masked Training","OmniDocBench","olmOCR-bench"],"falsifier":"Run the released model on any page of OmniDocBench v1.5 with token-parallel decoding under greedy selection and compare its output to the same model's sequential greedy output; any page where the two differ would falsify the lossless claim. Alternatively, an independent re-run of the benchmark that does not reproduce the reported overall score of 93.22 would also falsify the central claim.","tokens_in":21330,"feed_emoji":"📄","tokens_out":8966,"duration_ms":70395,"temperature":0.7,"pith_summary":"Youtu-Parsing claims that a single, decoupled vision-language model can outperform both modular pipeline tools and far larger generalist VLMs on full-document parsing, while running 5–11× faster than standard autoregressive decoding through a lossless token-parallel scheme. The model splits parsing into shared visual feature extraction, layout analysis, and region-prompted decoding, then accelerates the final decoding with two complementary parallelisms: predicting up to 64 candidate tokens at once and verifying them in a second forward pass, and querying up to five detected regions per pass. If these claims hold, high-throughput document digitization no longer needs to choose between recognition fidelity and inference speed. The paper reports state-of-the-art overall scores of 93.22 on OmniDocBench v1.5 and 80.5±0.9 on olmOCR-bench with a 2.5B-parameter model.","feed_headline":"Token-parallel decoding tops OCR benchmarks at 5–11x speed","feed_subtitle":"Rejects the speed-accuracy tradeoff: a single model outparses big VLMs and runs an order of magnitude faster.","key_machinery":"The load-bearing mechanism is the two-pass token-parallel decoding loop: the model extends the current context with $n=64$ <mask> tokens, produces $n$ candidate tokens in one forward pass, then runs a second forward pass over the candidate sequence and accepts only the tokens up to the first mismatch, guaranteeing bit-level equality with standard autoregressive decoding under greedy selection. The supporting object is the shared NaViT visual feature map, which lets layout analysis and region-prompted decoding reuse the same encoded representation and makes query parallelism natural. A Hybrid Masked Training objective (80% masked samples, 20% unmasked) instills the look-ahead ability while preserving baseline autoregressive integrity.","core_discovery":"The central discovery is that document parsing output is structured and predictable enough that a model can be trained to guess multiple future tokens at once, and a cheap verification pass can reject any wrong guesses, making the parallel decoding provably identical to sequential greedy decoding. The paper demonstrates this through Hybrid Masked Training, where 80% of fine-tuning samples receive random mask tokens so the model learns multi-token look-ahead, and reports an average acceptance of 10–20 tokens per iteration, translating to a 5–11× speedup (up to ~27× on tables), consistent with the theoretical speedup $S\\approx k/2$ where $k$ is the average accepted tokens per iteration. Together with query parallelism, which batches up to five region prompts into one forward pass, the scheme yields end-to-end per-page latency comparable to much smaller specialized models while achieving higher parsing accuracy.","pith_inferences":["The verification scheme is a general recipe: any task whose output is greedy-decoded and cheaply re-scored could adopt the same candidate-then-verify loop, so the idea likely transfers beyond documents to code generation or structured data serialization.","The claimed bit-level equivalence is contingent on deterministic greedy decoding; if a deployment switches on sampling for diversity, the exact-match verification no longer preserves the target distribution, and the speedup would come with an uncontrolled distribution shift.","Hybrid Masked Training could be tested on larger base models; if acceptance rates grow with model scale, the 5–11× speedup would become a lower bound for future systems."],"forward_implications":["A single 2.5B parameter model reaches 93.22 overall on OmniDocBench v1.5, surpassing both pipeline tools (Marker, MinerU-pipeline) and VLMs up to 241B parameters.","Token-parallel decoding yields 5–11× higher token throughput on typical pages, rising to ~27× on tables, with output identical to sequential greedy decoding.","Query parallelism adds ~2× end-to-end speed by batching up to five region queries per forward pass, with accuracy stable or slightly improved.","The model covers text, formulas, tables, charts, seals, and hierarchical structures, so a single deployment can replace multiple task-specific components."],"supporting_citations":[{"why":"Foundational work on non-autoregressive sequence generation whose error-recovery idea informs the verification step.","marker":"[Ran et al., 2020]"},{"why":"MaskGIT masked token prediction, the training-time pattern behind Hybrid Masked Training.","marker":"[Chang et al., 2022]"},{"why":"Medusa's multiple decoding heads, the speculative decoding approach token parallelism extends.","marker":"[Cai et al., 2024]"},{"why":"NaViT vision encoder that yields the shared feature map used by both layout analysis and region decoding.","marker":"[Tschannen et al., 2025]"},{"why":"Youtu-LLM-2B, the base language model that is fine-tuned into the parser.","marker":"[Lab, 2025]"},{"why":"OmniDocBench v1.5, the primary benchmark where the overall 93.22 score is claimed.","marker":"[Ouyang et al., 2025]"},{"why":"olmOCR-bench, the second benchmark where the 80.5 pass rate is claimed.","marker":"[Poznanski et al., 2025]"},{"why":"PaddleOCR-VL, the prior SOTA baseline on OmniDocBench that the paper must outperform.","marker":"[8]"},{"why":"MinerU2.5, a strong decoupled document parser used as a key comparison model.","marker":"[7]"}],"fun_headline_variants":["Parallel token decoding makes OCR 5–11x faster","Token-guessing parses docs 5–11x faster, same accuracy","High-parallelism decoding accelerates document parsing","Youtu-Parsing: 5–11x speedup via token parallelism","Break speed-accuracy tradeoff: 5–11x faster OCR parsing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The lossless guarantee of token parallelism holds only when decoding is greedy (argmax) and the two forward passes produce numerically identical logits; any sampling temperature or nondeterminism in attention kernels would break the exact equivalence.","fun_headline_variants_meta":{"raw":{"variants":["Parallel token decoding makes OCR 5–11x faster","Token-guessing parses docs 5–11x faster, same accuracy","High-parallelism decoding accelerates document parsing","Youtu-Parsing: 5–11x speedup via token parallelism","Break speed-accuracy tradeoff: 5–11x faster OCR parsing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00033,"raw_usage":{"total_tokens":1870,"prompt_tokens":1009,"completion_tokens":861,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":770}},"tokens_in":625,"tokens_out":861,"duration_ms":7291,"temperature":1.0,"reasoning_tokens":770,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:39:29.953196+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released model on any page of OmniDocBench v1.5 with token-parallel decoding under greedy selection and compare its output to the same model's sequential greedy output; any page where the two differ would falsify the lossless claim. Alternatively, an independent re-run of the benchmark that does not reproduce the reported overall score of 93.22 would also falsify the central claim.","supporting_citations":[{"cited_title":"Maskgit: Masked generative image transformer","cited_arxiv_id":null,"evidence_quote":"MaskGIT masked token prediction, the training-time pattern behind Hybrid Masked Training."}],"review_version":2}