{"id":"f2d66fe3-1fcf-40cc-81f6-c6a33a066772","arxiv_id":"2505.13573","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FreeMesh shows that rearranging coordinates into same-axis groups before byte-pair encoding reduces per-token entropy and sequence length, improving point-cloud conditioned mesh generation quality.","lead":"The paper introduces a training-free metric (PTME) for comparing mesh tokenizers, and a coordinate-merging trick (RMC) that compresses mesh sequences by rearranging coordinates and applying byte-pair encoding. A generalist might read it because it is a simple, plug-in efficiency improvement for auto-regressive 3D mesh generation, cutting sequence length to about 21 percent of the raw size on one tokenizer.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PTME's predictive validity is unsupported: the only controlled evidence is a vocab-size sweep that changes CR and model capacity alongside PTME, and the appendix text is internally contradictory.","rationale":"The reader's verdict is conditional, and my read supports that. The empirical core of the paper—RMC compresses EdgeRunner tokens to 21.2% and improves CD from 0.198 to 0.123 in the same point-cloud-conditioned pipeline—is directly supported by Table 1 and is not the target of my objection. The MC baseline with the same 8192 vocabulary also controls for the obvious vocabulary-size confound. What is not established is the paper's stronger claim that PTME is a valid training-free metric for tokenizer learnability. The derivation in Section 3.2 defines PTME from empirical token frequencies but does not connect it to autoregressive conditional loss; the Appendix C proof applies to a constant-CR setting that the actual RMC changes, and the only empirical validation, Appendix D, is a single univariate correlation over a sweep in which CR and model capacity move with PTME. The manuscript's own text in Appendix D refers to molecular generation rather than mesh generation and misstates the sign of the reported correlation, so the analysis cannot be taken at face value. This is exactly the reader's weakest assumption. I would not reject the paper: the compression and generation improvements are plausible and reproduce internally. But the PTME contribution needs either a corrected, reproducible validation or a clear downgrade from 'theory-driven metric' to 'heuristic that tracks CD in the tested configurations.' Hence the conditional verdict stands.","tokens_in":13728,"tokens_out":16850,"duration_ms":183897,"concrete_test":"Request the per-vocabulary-size table for EDR+RMC (vocabularies 256, 512, 1024, 2048, 4096, 8192) listing PTME, CR, usable-mesh count, and CD, and recompute the partial correlation of PTME with CD after controlling for CR (and vocabulary size). If the partial correlation is near zero or changes sign, PTME has no predictive power beyond compression and the central premise fails; if it remains strong and negative, the paper should restate the Appendix D conclusion as 'lower PTME predicts lower CD' and supply the underlying data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central premise is that a lower PTME means the serialized sequence is easier to learn (Section 3.2), and this premise is what justifies coordinate merging. The only direct, controlled evidence for it is the EDR+RMC vocabulary-size correlation in Appendix D, but that sweep changes vocabulary size, compression ratio, and model embedding capacity simultaneously. The reported univariate r=0.965 between PTME and CD therefore does not show that PTME, rather than CR or model capacity, predicts generation quality. The appendix also states the direction incorrectly: it reports a positive PTME-CD correlation yet concludes that 'a higher PTME score strongly indicates better downstream generation performance measured by lower Chamfer Distance'; a positive correlation says higher PTME goes with higher (worse) CD. The same appendix is written about 'molecular generation' and 'generated molecules,' so the analysis as printed cannot be verified as a mesh-generation experiment. If PTME does not predict CD once CR and data/context capacity are controlled, the theoretical motivation for RMC is unsupported, even though the compression numbers and CD gains in Table 1 remain direct empirical results.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two contributions for autoregressive mesh generation: a training-free metric, Per-Token-Mesh-Entropy (PTME), intended to measure tokenizer quality, and a plug-and-play coordinate merging technique, called Rearrange & Merge Coordinates (RMC), that rearranges serialized coordinate sequences and then applies BPE-style merging to compress them. The method is evaluated by plugging RAW, AMT, and EDR tokenizers into a unified point-cloud-conditioned autoregressive pipeline at 7-bit quantization. The main empirical result is that EDR+RMC achieves a compression ratio of 0.212 and improves Chamfer distance from 0.198 to 0.123 and Hausdorff distance from 0.408 to 0.280 compared to EDR alone (Table 1). The authors also report that RMC increases the number of meshes fitting within a fixed 9,000-token context window (Figure 6). The paper claims PTME is validated by a strong correlation with downstream Chamfer distance, but that validation is internally contradictory and appears to reuse text from a different domain.","tokens_in":13956,"tokens_out":4932,"duration_ms":51742,"significance":"If the empirical gains are reproducible, RMC is a simple, effective, and practical compression technique for coordinate-level mesh tokenizers: it is training-free on the target model, integrates into existing tokenizers, and consistently improves generation quality across three tokenization schemes in a unified pipeline. The paper also introduces the reasonable idea of using information-theoretic quantities to compare mesh serializations without training. However, the central scientific claim that PTME is a valid predictor of learnability is not established: the only direct validation (Appendix D) contains a sign error, appears to be copied from a molecular-generation context, and is confounded by simultaneous changes in compression ratio and model capacity. The theoretical derivation in Section 3.2 and Appendix C also conflates PCME and PTME. As a result, the motivation for RMC rests on an unvalidated premise, even though the compression and generation results stand as direct empirical observations. The paper would be considerably stronger with corrected validation or a reframed contribution that explicitly treats RMC as an empirically motivated heuristic.","major_comments":[{"comment":"The correlation analysis in Appendix D is internally contradictory and cannot serve as evidence for PTME. The text reports 'a strong positive linear correlation: r=0.965' between PTME and CD, then concludes that 'a higher PTME score strongly indicates better downstream generation performance measured by lower Chamfer Distance.' A positive correlation means higher PTME is associated with higher (worse) CD, not lower; the stated conclusion is the opposite of what r=0.965 implies. In addition, the paragraph and Table 2 repeatedly refer to 'molecular generation' and 'generated molecules,' indicating that the analysis as printed was not performed on the mesh-generation experiments in this paper. As the only direct empirical validation of the learnability assumption in Section 3.2, this analysis must be corrected and re-run on mesh data before the PTME claim can be evaluated.","section":"Appendix D"},{"comment":"The reported PTME-CD correlation is confounded. The vocabulary-size sweep (256 to 8192) simultaneously changes the compression ratio, the sequence length, the number of tokens consumed per mesh, and the embedding capacity of the model. Therefore, r=0.965 does not identify PTME as the causal predictor of generation quality; the same correlation could be driven by CR or by model capacity alone. The paper claims in Section 4.3 that 'PTME shows a stronger correlation with generation quality than the compression ratio,' but no controlled comparison is provided. A proper test would hold vocabulary size fixed and vary only the rearrangement or merging scheme, or compute partial correlations controlling for CR. Without such a test, the metric's predictive validity is unsupported.","section":"Section 4.3 and Appendix D"},{"comment":"There is a mismatch between the metric used in the experiments and the quantity analyzed in the proof. The main text defines PTME in Eq. (8) as PTME = (Hs/l) * CR, but Appendix C derives a condition for PCME (Eq. 10) to decrease after merging, assuming CR is constant. Table 1 shows that the baseline MC reduces CR yet increases PTME (e.g., RAW PTME goes from 6.742 to 6.943), which directly contradicts the intuition expressed in Eq. (5) that 'Imerged < Itotal.' The proof in Appendix C only shows a condition for PCME reduction under a simplified probability model, and it does not establish the corresponding statement for PTME. The paper should either provide a direct theoretical guarantee for PTME under RMC or explicitly state that PTME reduction is an empirical property of the proposed rearrangement, not a proven consequence.","section":"Sections 3.2 and Appendix C"},{"comment":"Table 1 reports single values with no error bars, standard deviations, or significance tests for any of the generation metrics. Several comparisons are visually small (e.g., EDR+MC CD 0.204 vs. EDR CD 0.198; AMT+RAC CD 0.226 vs. AMT CD 0.219), so without measures of variance it is not possible to assess whether the observed improvements are meaningful. The paper's central empirical claim that RMC improves generation quality would be substantially strengthened by reporting means and standard deviations over multiple training runs or at least over multiple evaluation samples.","section":"Table 1"}],"minor_comments":[{"comment":"There is a typo in the first sentence of the RAW representation paragraph: 'In theRA Wrepresentation' should be 'In the RAW representation.'","section":"Section 3.1"},{"comment":"The caption begins with a lower-case 'w' after a period: 'while the baseline Merge Coordinates (MC) method...' should be 'While the baseline Merge Coordinates (MC) method...'.","section":"Figure 2 caption"},{"comment":"In Eq. (2), the symbol c is used both for the coordinate sequence (in p(ci | c1:i−1, c; θ)) and for the point-cloud condition c. This overloading is confusing and should be resolved by using a different symbol for one of the two quantities.","section":"Equation (2)"},{"comment":"The caption of Table 1 states that metrics are computed using '10K sampled points per mesh,' while Section 4.1 says the point cloud encoder processes 8,192 sampled points. These numbers are inconsistent and should be reconciled.","section":"Table 1 caption vs. Section 4.1"},{"comment":"Appendix D's PTME vs PPL discussion refers to 'molecular generation' and 'generated molecules' rather than meshes. This appears to be a copy-paste error from a different project and should be rewritten for the mesh-generation context, along with the corresponding tables.","section":"Appendix D"},{"comment":"The point cloud encoder is attributed to 'Michelangeo' (typo for 'Michelangelo') and the reference is listed as Zhao et al., 2024b. Please correct the spelling.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The empirical core of the paper — the RMC compression and the consistent CD/HD improvements in Table 1 — appears sound as a direct engineering contribution, and the paper is well within the scope of the venue. My main concern is that the paper's headline contribution, the PTME metric, is validated by an analysis that is both internally contradictory (Appendix D sign error) and confounded (vocabulary-size sweep changes multiple variables at once). These issues are fixable in a revision, but they are load-bearing for the paper's central claim. I would be willing to look at a revised version that either provides a clean, controlled validation of PTME or honestly reframes the contribution as an empirical compression method without the unverified theoretical motivation. Also note the odd domain mismatch in Appendix D, which should be corrected regardless."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The part worth taking seriously is RMC, the rearrange-and-merge coordinate scheme. Applying BPE after grouping coordinates by axis is simple, non-obvious, and in your Table 1 it consistently improves all three tokenizers you tried: compression ratio drops to 0.212 for EDR+RMC, and Chamfer distance improves from 0.198 to 0.123. Those are direct empirical results from a controlled within-pipeline comparison, and they are the paper's real contribution. The derivation of PTME itself is straightforward and the training-free intent is reasonable, but the metric is oversold as a predictive tool.\n\nThe soft spots are real and need to be fixed. The central premise—lower PTME means easier to learn—is only validated by the Appendix D correlation, and that correlation is confounded: you sweep vocabulary size and compression ratio together with PTME, so r=0.965 does not isolate PTME as the causal factor. The appendix also states the direction incorrectly: a positive PTME-CD correlation says higher PTME goes with higher (worse) CD, yet the text claims the opposite. And the paragraph is written about 'molecular generation' and 'generated molecules,' which is clearly leftover from another draft; as printed, that analysis cannot be verified as a mesh experiment. Those are not minor typos. They undermine the only evidence for the metric's validity.\n\nThat said, the main method does not collapse. Even if PTME is removed entirely, RMC still stands as a practical compression technique with measurable generation gains. The paper would be more honest if it dropped the strong claim that PTME predicts generation quality and instead presented RMC as an effective compression trick, with PTME as a descriptive statistic. You should also add error bars or multiple seeds to Table 1—the gains are consistent but nothing rules out variance of the same magnitude—and address the limitation you already acknowledge: the method is only tested at 7-bit quantization and may not survive at 1024.\n\nWho is this for? Anyone building autoregressive mesh generators who wants a cheap way to fit more faces in a context window. The RMC result deserves a serious referee even though the metric section needs heavy revision. I would engage with it, cite the compression result if I worked in the area, and send it to review with a clear request to fix the appendix and rescope the claims.","headline":"The coordinate-merging trick is a genuinely useful plug-in for mesh tokenizers, but the PTME metric is oversold and its validation in Appendix D is currently broken.","tokens_in":14484,"tokens_out":1667,"would_cite":true,"duration_ms":20364,"reading_group":"maybe","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 proposes Per-Token-Mesh-Entropy as a training-free measure of mesh-tokenizer quality, then uses it to justify a rearrange-and-merge coordinate compression that cuts sequence length by 58% and lowers Chamfer distance from 0.198…","keywords":["mesh generation","autoregressive transformer","tokenization","entropy metric","byte-pair encoding","coordinate compression","point cloud conditioning","3D shape generation"],"falsifier":"Hold the model, data, and context window fixed while comparing two tokenizations with equal compression ratio but different PTME; if the lower-PTME tokenization does not produce lower Chamfer distance in a controlled run, the PTME-learnability claim is false.","tokens_in":13553,"feed_emoji":"🧊","tokens_out":9449,"duration_ms":88116,"temperature":0.7,"pith_summary":"The paper tries to establish that mesh tokenizers—the front ends that turn a 3D mesh into a sequence of coordinate tokens for autoregressive generation—can be scored without training, and that this score can be lowered on purpose. It proposes Per-Token-Mesh-Entropy (PTME), which multiplies the entropy per merged coordinate substring by the tokenizer's compression ratio; lower PTME is claimed to mean an easier-to-learn sequence. Building on PTME, it introduces Rearrange & Merge Coordinates (RMC): group each face's coordinates axis by axis (all x's, then all y's, then all z's) before byte-pair merging, so the repeated patterns the merge finds actually co-occur. Applied to the best-performing baseline tokenizer, RMC cuts the compression ratio from 0.505 to 0.212 and improves point-cloud conditioned generation, with Chamfer distance falling from 0.198 to 0.123. The payoff would be that any autoregressive mesh pipeline can be strengthened by a tokenizer-agnostic post-processing step rather than a new architecture or retraining from scratch.","feed_headline":"Coordinate merging shrinks mesh token streams by 58 percent","feed_subtitle":"A training-free entropy metric makes it work, cutting generation error from 0.198 to 0.123.","key_machinery":"The central object is Per-Token-Mesh-Entropy, $PTME = (H_s / l) \\cdot CR$, with $H_s$ the Shannon entropy of the merged coordinate substrings, $l$ their average length measured in coordinates, and $CR$ the compression ratio relative to the raw coordinate stream. It converts \"how learnable is this sequence\" into one number computable from token frequencies alone. The companion mechanism is RMC (Rearrange & Merge Coordinates), a rule that reserializes every face's coordinates from the interleaved order $x_1,y_1,z_1,x_2,y_2,z_2,x_3,y_3,z_3$ into $x_1,x_2,x_3,y_1,y_2,y_3,z_1,z_2,z_3$, with special handling for subsequences shorter than nine tokens and for direction words, and only then applies byte-pair merging. The rearrangement does not change the information content of the sequence—the paper shows PTME is essentially unchanged after rearrangement alone—but it makes frequent multi-coordinate substrings discoverable, which is what allows merging to reduce PTME.","core_discovery":"On the paper's terms, the discovery is that the quality of a mesh tokenizer is measurable, before training, by $PTME = (H_s / l) \\times CR$, and that this quantity can be minimized by matching the merge operation to the geometry. A naive merge of adjacent coordinates increases PTME because byte-pair merging pairs up numbers that do not repeat when read in the interleaved x-y-z order of the raw serialization. RMC rearranges each group of nine coordinates into contiguous x-runs, y-runs, and z-runs (and shifts direction tokens ahead of coordinates) before merging, so the most frequent substrings are genuinely repeated. The paper reports that with vocabulary size 8192, the rearrangement plus merge lowers PTME from 3.139 to 2.231, lowers the compression ratio from 0.505 to 0.212, and improves Chamfer distance from 0.198 to 0.123 in a point-cloud conditioned mesh generation pipeline, while also admitting more training meshes into the same 9,000-token context window.","pith_inferences":["A natural extension is to compute PTME for other autoregressive 3D representations, such as point clouds or surface patches, by defining the coordinate unit and substring scheme; the paper's derivation does not depend on triangles specifically.","Appendix C's proof implies a prospective rule stronger than frequency-only merging: a merge is entropy-reducing mainly when the pair's co-occurrence probability and pointwise mutual information are both high, which could be used to choose merges rather than diagnosing failures after the fact.","The PTME-to-generation link currently rests on one pipeline, so a cheap test is to apply RMC under a different context length, model capacity, or dataset and check whether Chamfer distance still tracks PTME.","At higher coordinate quantization, repeated coordinate patterns become rarer, so pattern-based merging may need to become adaptive or byte-level; the paper names this as its main limitation, suggesting a concrete next experiment at 1024-level quantization."],"forward_implications":["Any existing coordinate-level mesh tokenizer can be compressed further by the same plug-in step, because RMC is applied after serialization and does not require retraining the original tokenizer.","Within a fixed context window, a lower compression ratio translates directly to more faces per generated mesh, and the paper shows the RMC variants admit tens of thousands more training meshes under a 9,000-token limit.","Compression ratio alone is not a reliable proxy for tokenizer quality; PTME orders two comparable baselines differently from their compression ratios, and that ordering matches the measured generation quality.","Sequence order is not sacred: pure coordinate rearrangement leaves generation nearly unchanged, so future work can optimize serialization order without paying a training penalty.","PTME can be used to choose vocabulary size without training, because the reported gains from merging plateau once the vocabulary reaches roughly 8,192 entries."],"supporting_citations":[{"why":"Supplies the RAW coordinate serialization and the coordinate-level next-coordinate generation paradigm that PTME measures.","marker":"Chen et al., 2024a"},{"why":"Supplies the Adjacent Mesh Tokenization (AMT) with shared-edge vertex reuse, one of the three tokenizers compared.","marker":"Chen et al., 2024c"},{"why":"Supplies the EdgeRunner tokenization with half-edge direction tokens, the baseline that reaches the best RMC results.","marker":"Tang et al., 2024a"},{"why":"Provides the subword tokenizer used to run the byte-pair-encoding merge on coordinate streams.","marker":"Kudo & Richardson, 2018"},{"why":"Introduces byte-pair encoding for subword vocabularies, the greedy merging algorithm that MC and RMC apply to coordinates.","marker":"Sennrich et al., 2016"},{"why":"Supplies the information entropy definition from which PCME and PTME are constructed.","marker":"Shannon, 1948"},{"why":"Provides the minimum-entropy principle that lower sequence entropy means easier sequence learning, the premise PTME relies on.","marker":"Su, 2018"},{"why":"Motivates the dense-context tokenization view behind selecting the 8,192-entry vocabulary.","marker":"Weng et al., 2024b"}],"fun_headline_variants":["FreeMesh: coordinate merging cuts mesh token streams by 58%","Training-free metric PTME guides better mesh tokenization","Merge coordinates to shrink mesh tokens and improve generation","New entropy metric boosts mesh generation without training","Coordinate merging lowers mesh generation error from 0.198 to 0.123"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a lower Per-Token-Mesh-Entropy means the sequence is easier to learn and thus generates better meshes; the paper validates that link only with one correlation measured across vocabulary sizes in a single pipeline.","fun_headline_variants_meta":{"raw":{"variants":["FreeMesh: coordinate merging cuts mesh token streams by 58%","Training-free metric PTME guides better mesh tokenization","Merge coordinates to shrink mesh tokens and improve generation","New entropy metric boosts mesh generation without training","Coordinate merging lowers mesh generation error from 0.198 to 0.123"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000796,"raw_usage":{"total_tokens":3483,"prompt_tokens":906,"completion_tokens":2577,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":2495}},"tokens_in":522,"tokens_out":2577,"duration_ms":19995,"temperature":1.0,"reasoning_tokens":2495,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:18:49.881195+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold the model, data, and context window fixed while comparing two tokenizations with equal compression ratio but different PTME; if the lower-PTME tokenization does not produce lower Chamfer distance in a controlled run, the PTME-learnability claim is false.","supporting_citations":[{"cited_title":"and Richardson, J","cited_arxiv_id":null,"evidence_quote":"Provides the subword tokenizer used to run the byte-pair-encoding merge on coordinate streams."}],"review_version":1}