{"id":"0ebff384-ad72-462c-9c25-84390a251d06","arxiv_id":"2508.05628","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":2,"one_line_summary":"H-Net++ uses hierarchical dynamic chunking to remove the tokenizer from byte-level language modelling, reporting gains over BPE-based GPT-2-fa on Persian.","lead":"This paper introduces H-Net++, a language model that learns its own word chunks from raw bytes instead of using a fixed tokenizer. On Persian, it reports better compression, better benchmark scores, and stronger resistance to a common orthographic error than a BPE-based comparison model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains may be due to extra modules and unmatched baselines, not to learned dynamic chunking; ablations are missing.","rationale":"The reader identified the weakest assumption as the capacity of a byte-level dynamic-chunked model to represent Persian morphology as well as BPE. My concern is adjacent but more specific: even if the model has sufficient capacity, the reported empirical gains are not causally tied to the chunking component because no ablations or matched baselines are described. The F1 73.8% provides some evidence that learned chunks are morphology-aligned, but it lacks a comparison point (e.g., random chunks) and does not demonstrate that this alignment drives the compression or ParsGLUE improvements. Since the paper is abstract-only, I cannot verify whether such ablations exist in the full text; therefore the reader's UNVERDICTED verdict remains appropriate. My concern does not change the verdict, but it identifies a concrete evidence gap that the full text should fill.","tokens_in":699,"tokens_out":2951,"duration_ms":27355,"concrete_test":"Run a controlled comparison on the same Persian corpus: (a) H-NET++ as described; (b) the same architecture but with random or fixed-size byte chunks (e.g., every 4 bytes) and no learned segmentation; (c) the same architecture with gold morphological boundaries as chunks. If (b) achieves comparable BPB and ParsGLUE to (a), or if (c) does not improve over (a), the claimed advantage of learned dynamic chunking is not supported. Also report total trainable parameters and training FLOPs for H-NET++ and the GPT-2-fa baseline to verify the comparison is matched.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of H-NET++ is that hierarchical dynamic chunking of raw bytes yields state-of-the-art compression and ParsGLUE performance while learning morphology-aligned chunks. The strongest empirical support is a 0.159 BPB reduction versus BPE-based GPT-2-fa, a 5.4pp ParsGLUE gain, and 73.8% F1 on gold morphology. The load-bearing assumption is that these gains are causally attributable to the dynamic chunking mechanism. The abstract does not provide any ablation isolating the chunking component. The architecture also includes a 1.9M-parameter context mixer, a two-level hyper-prior, and curriculum training; these could account for part or all of the improvements. The F1 73.8% is reported without baseline comparison; because Persian orthography uses explicit spaces and ZWNJ, fixed or random chunking might already align with morphology much better than chance. Without an ablation replacing learned chunks with fixed/random/gold chunks, and without matched model size and compute for the GPT-2-fa baseline, the central claim is not well-supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript (abstract only) introduces H-NET++, a byte-level, tokenizer-free language model for Persian that uses hierarchical dynamic chunking to learn linguistically-informed segments. Key components are a 1.9M-parameter Transformer context-mixer for cross-chunk attention, a two-level latent hyper-prior for document-level consistency, specialized ZWNJ handling, and curriculum training with staged sequence lengths. On a 1.4B-token Persian corpus, the authors report a 0.159 BPB reduction over BPE-based GPT-2-fa (12% better compression), a 5.4 percentage-point gain on ParsGLUE, 53% improved robustness to ZWNJ corruption, and 73.8% F1 on gold morphological boundaries, claiming state-of-the-art tokenizer-free performance for morphologically-rich languages.","tokens_in":984,"tokens_out":1977,"duration_ms":19983,"significance":"If the reported gains are real and causally attributable to dynamic chunking, the work would be a meaningful step toward tokenizer-free LMs for MRLs, addressing a known computational bottleneck. The use of an external benchmark (ParsGLUE) and a concrete BPE baseline is a strength, as is the explicit morphological-boundary F1 metric, which makes the segmentation claim falsifiable in principle. The reported numbers are substantial, but the abstract alone provides no derivation, no protocol details, no error bars, and no ablations, so the significance cannot be assessed beyond the claims themselves.","major_comments":[{"comment":"Central causal claim is unsupported: the reported gains (0.159 BPB, +5.4pp ParsGLUE) are attributed to hierarchical dynamic chunking, but the architecture also contains a 1.9M-parameter context mixer, a two-level latent hyper-prior, and curriculum training. Without ablations that replace the learned chunker with fixed, random, or gold chunks while holding other components constant, any of these additions—or their interaction—could account for the improvements. The abstract provides no such ablation.","section":"Abstract"},{"comment":"The 73.8% F1 on gold morphological boundaries is reported without a baseline. Persian orthography already uses spaces and ZWNJ to mark many morpheme boundaries, so fixed or random byte chunking may achieve non-trivial alignment. A comparison with character-level, fixed-window, or BPE segmentation is needed to establish that the learned chunks carry morphological signal beyond trivial orthographic cues.","section":"Abstract"},{"comment":"Baseline fairness is not established. The comparison to 'BPE-based GPT-2-fa' does not state whether model parameter count, training data, compute, and evaluation protocol are matched. The proposed model adds a 1.9M-parameter context mixer and a curriculum schedule; if the baseline is a smaller model or trained with fewer tokens, the reported BPB and ParsGLUE differences may reflect capacity or training budget, not the chunking mechanism.","section":"Abstract"},{"comment":"No statistical reliability is reported. There are no error bars, number of seeds, significance tests, or details on ParsGLUE splits and evaluation conditions. A single run on one dataset is insufficient to support state-of-the-art claims, especially for the small ParsGLUE gain (5.4pp) and the robustness metric (53%), whose variance is unknown.","section":"Abstract"}],"minor_comments":[{"comment":"The term 'state-of-the-art' is used without citing or discussing other tokenizer-free models on the same benchmark(s); the comparison set is limited to one BPE baseline.","section":"Abstract"},{"comment":"The 'two-level latent hyper-prior' is mentioned as a key innovation but is not defined in abstract; a journal reader would need at least a formal description or a reference.","section":"Abstract"},{"comment":"The phrase 'specialized handling of orthographic artifacts' is vague; it should specify whether this is architectural, data-augmentation-based, or a decoding rule.","section":"Abstract"}],"recommendation":"uncertain","confidential_remarks":"This review is based solely on the abstract; no full text was available. The reader's take and stress-test note correctly identify missing ablations and baseline-matching concerns, but these cannot be resolved at the abstract level. If the full manuscript contains the required ablations and protocol details, the paper may be viable; if not, the central claim would need substantial revision. I recommend obtaining the full text before a definitive decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the abstract promises a useful contribution to byte-level language modelling for Persian, but the only evidence we have is the abstract's reported numbers. The causal story is plausible, not established. I'd send this to a knowledgeable referee rather than dismiss it.\n\nWhat's actually new: the combination of hierarchical dynamic chunking, a 1.9M-parameter context-mixer, and a two-level latent hyper-prior is a real architectural bundle, and the evaluation targets are well chosen for the claim. ParsGLUE gives a downstream check, the gold-morphology F1 directly tests whether learned chunks line up with linguistic structure, and the ZWNJ-robustness test addresses an orthographic artifact that matters in Persian. If the numbers hold, this is a solid subfield advance.\n\nSoft spots: the abstract does not include ablations, so we cannot attribute the 0.159 BPB reduction or the 5.4pp ParsGLUE gain to the chunking component. The context-mixer, hyper-prior, and curriculum training could plausibly carry most of the benefit. The GPT-2-fa baseline is not described as matched in size or compute, which matters for a 12% compression claim. The 73.8% F1 on morphology has no baseline; since Persian already uses spaces and ZWNJ as delimiters, a fixed or random chunking baseline could be high. I'd want to see an ablation replacing learned chunks with fixed/random/gold chunks, plus error bars or at least multiple seeds.\n\nThese are standard referee requests, not fatal flaws. If the full paper contains that material, the contribution is credible. The overlap with the earlier H-Net work should come out in the paper, but the '++' alone isn't a problem.\n\nThe stress-test note worries about attribution and unmatched baselines; I think that concern is fair and central. My own verdict is 'unverified' rather than 'wrong' — the abstract is too thin to judge soundness, but the direction and evaluation design show honest engagement with the problem.\n\nWho this is for: researchers working on tokenizer-free LMs, morphological segmentation, and low-resource or morphologically rich languages. It deserves serious peer review. Recommendation: send it out. If the full text backs the abstract's claims, publishable; if not, the referee will catch it. I wouldn't cite it until I see the ablations.","headline":"H-NET++ has a well-motivated architecture and sensible Persian-language evaluations, but the abstract alone doesn't show the gains come from dynamic chunking; worth refereeing, not desk-rejecting.","tokens_in":1385,"tokens_out":1938,"would_cite":false,"duration_ms":21865,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"H-NET++ claims a tokenizer-free, byte-level language model can beat BPE on Persian by learning dynamic chunks that align with morphology.","keywords":["tokenizer-free","byte-level language model","dynamic chunking","Persian","morphologically-rich languages","compression","ZWNJ","language modelling"],"falsifier":"Measure the agreement between H-NET++'s dynamic chunk boundaries and gold morphological boundaries on a held-out Persian test set, and compare it with random chunk boundaries of the same length distribution. If the agreement is not significantly above chance, the paper's claim that chunks align with morphology through end-to-end training is falsified.","tokens_in":652,"feed_emoji":"🧩","tokens_out":4579,"duration_ms":42313,"temperature":0.7,"pith_summary":"This paper tries to establish that byte-level language models do not need a fixed tokenizer in morphologically-rich languages: instead, the model can learn to split the raw byte stream into variable-sized chunks during training, and those chunks turn out to line up with morpheme boundaries. On a 1.4B-token Persian corpus, the proposed model (H-NET++) reports a 0.159 BPB reduction over a BPE-based GPT-2-fa baseline (12% better compression), a 5.4-point gain on ParsGLUE, 53% better robustness to ZWNJ character corruption, and 73.8% F1 on gold morphological boundaries. If these numbers hold, the result matters because it removes the need for language-specific tokenizer engineering, which is especially costly for morphologically-rich languages. The work is a proposal of a new method, with the compression and downstream gains presented as evidence for the method's core claim.","feed_headline":"Tokenizer-free model beats BPE on Persian by 12%","feed_subtitle":"H-NET++ learns word boundaries on its own and gains 5.4 points on ParsGLUE.","key_machinery":"The central mechanism is hierarchical dynamic chunking: the byte stream is split into variable-length chunks, and a lightweight Transformer context-mixer (1.9M parameters) attends across chunk boundaries to exchange information. A two-level latent hyper-prior keeps document-level consistency, and specialised handling of the Persian ZWNJ character (a zero-width non-joiner marking morpheme boundaries) lets the model treat this orthographic artifact as a segmentation signal. Curriculum training with staged sequence lengths helps the model learn short-range patterns first, then longer-range structure. The claimed effect is that these components—not a fixed subword vocabulary—produce the compress","core_discovery":"The central discovery claimed is that hierarchical dynamic chunking lets a byte-level language model learn linguistically-informed segmentation purely from the language-modelling objective, without tokenizer priors or explicit morphological supervision. The paper describes H-NET++, which processes raw bytes in hierarchical, dynamically-sized chunks, applies a lightweight 1.9M-parameter Transformer context-mixer for cross-chunk attention, and uses a two-level latent hyper-prior to maintain document-level consistency. On the reported Persian benchmark, H-NET++ achieves state-of-the-art results on compression, ParsGLUE, robustness to ZWNJ corruption, and morphological boundary identification. T","pith_inferences":["Editorial inference: If the same dynamic-chunking objective transfers, byte-level models could become viable for low-resource languages where training a tokenizer is difficult because of scarce text.","Editorial inference: The ZWNJ-specific component suggests part of the gain may come from exploiting one transparent orthographic boundary marker; on languages without such a marker, the advantage over BPE could shrink.","Editorial inference: A direct test on another morphologically-rich language (e.g., Turkish or Finnish) would separate a general benefit of dynamic chunking from Persian-specific artefacts, and would also reveal how much the context mixer contributes versus simple longer context."],"forward_implications":["The reported compression gain implies byte-level models can be made competitive in morphologically-rich languages without a tokenizer, removing language-specific subword engineering.","The 73.8% F1 on morphological boundaries suggests the learned chunks can serve as an unsupervised morphological segmenter for Persian.","The 53% robustness gain to ZWNJ corruption implies greater tolerance of orthographic noise, which matters for real-world Persian text from social media and user input.","The small 1.9M-parameter context mixer suggests that cross-chunk attention can be kept cheap, so the approach remains computationally practical at scale."],"supporting_citations":[],"fun_headline_variants":["H-NET++ beats BPE on Persian by 12% without a tokenizer","Learns Persian morphology from raw bytes, no tokenizer","H-NET++: +5.4 ParsGLUE, 12% better compression than BPE","Robust to Persian ZWNJ corruption, H-NET++ beats BPE"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that byte-level dynamic chunking plus a 1.9M-parameter context mixer can represent Persian morphology as well as a BPE-based model, and that the reported gains over GPT-2-fa come from the learned chunking rather than from extra parameters, the curriculum schedule, or the ZWNJ-specific preprocessing.","fun_headline_variants_meta":{"raw":{"variants":["H-NET++ beats BPE on Persian by 12% without a tokenizer","Learns Persian morphology from raw bytes, no tokenizer","H-NET++: +5.4 ParsGLUE, 12% better compression than BPE","Robust to Persian ZWNJ corruption, H-NET++ beats BPE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000615,"raw_usage":{"total_tokens":2688,"prompt_tokens":732,"completion_tokens":1956,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":476,"completion_tokens_details":{"reasoning_tokens":1868}},"tokens_in":476,"tokens_out":1956,"duration_ms":15647,"temperature":1.0,"reasoning_tokens":1868,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T23:10:41.599323+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the agreement between H-NET++'s dynamic chunk boundaries and gold morphological boundaries on a held-out Persian test set, and compare it with random chunk boundaries of the same length distribution. If the agreement is not significantly above chance, the paper's claim that chunks align with morphology through end-to-end training is falsified.","supporting_citations":[],"review_version":1}