acceptodds
Under review as a conference paper at ICLR 2027

Beyond Character Indices: Explicit Geometry for Proportional Text Art

Abstract

A spatial quantity can be fully determined by a model's inputs yet remain useful to supply explicitly. We test this distinction in proportional-font text art, where horizontal position is the cumulative sum of known glyph widths. Pixel-preserving whitespace rewrites change character indices while holding the rendered artwork fixed. Paired Transformers share characters, true widths, target inputs and complete source access, differing only in an extra source scalar: matched index or cumulative position. Canonicalizing every source run in the specified whitespace family during training and evaluation makes each model exactly invariant to those rewrites. Cumulative positions still improve conditional nonspace prediction by 0.0365 bits/target. Three additional paired initializations at a fixed endpoint retain a 0.0410-bit gain. Width-only, calibrated row-scale and pseudo cumulative controls also favor cumulative positions. Frozen models retain a gain on 19 newly acquired, screened event artworks, and larger predictors retain gains with both normalized and raw/augmented inputs. In this controlled natural setting, explicit cumulative position remains useful after the specified encoding invariance is enforced.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
representation learning, positional representations, rendered geometry, canonicalization, conditional character prediction, proportional-font text art

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_tT2kOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether an explicitly supplied cumulative spatial coordinate remains useful when that coordinate is already computable from the model’s inputs. The setting is conditional nonspace-character prediction in proportional-font Shift_JIS text art. Paired Transformers receive the complete previous row, true local glyph advances, target history, target query, and unrestricted query access to source positions; they differ only in whether an additional scalar is a matched character index or the true cumulative rendered position. A source-side canonicalizer maps equivalent U+0020/U+3000 whitespace runs to a unique string, making both models exactly invariant to this specified pixel-preserving rewrite family. Under this controlled setup, cumulative position improves NLL by 0.0365 bits/target in the original normalized stage and by 0.0410 bits/target in three additional fixed-endpoint paired initializations. Additional controls compare against width-only, calibrated row-scale, pseudo-cumulative, dropout, larger normalized, and raw/augmented causal variants, with a separate frozen evaluation on 19 screened event artworks.

Critical strengths
  1. The central comparison is well controlled. W-index and W-scan share the architecture, parameter count, true local widths, target inputs, full source access, initialization, and data order within each seed pair. This makes the scalar-feature intervention cleaner than a standard feature-addition comparison.
  2. The whitespace canonicalization provides a precise intervention. The source-side U+0020/U+3000 canonicalization removes a well-defined encoding redundancy while preserving rendered pixels for the specified rewrite family under the fixed MS PGothic 16px metric.
  3. The authors report several useful controls. Width-only, Linear, Pseudo, dropout, larger-model, and frozen-event experiments help rule out some simple explanations and show that the effect is not confined to a single training configuration.
  4. The paper is transparent about limitations and reproducibility. The appendices document protocol history, reused data, endpoint selection, per-seed results, intervals that cross zero, and detailed replay information.
Critical weaknesses
  1. The significance and generality of the problem are limited. The experiments concern teacher-forced next-character prediction in Shift_JIS text art under one fixed MS PGothic rendering rule, using relatively small custom Transformers. The paper does not show relevance to modern language models, layout-aware systems, other fonts, free generation, or downstream applications. As a result, the work feels more like a carefully controlled niche case study than a broadly significant contribution.
  2. The comparison does not isolate the mechanism behind the gain. The Scan feature provides coordinates in the same metric as the target query, so the improvement may reflect easier source–target coordinate alignment rather than the value of prefix accumulation itself. The paper lacks learned cumulative-position, CoPE-like, or query-relative coordinate baselines that would distinguish these explanations.
  3. The empirical evidence remains statistically and materially narrow. Most experiments reuse the same ordinary additions and blocks, with only three paired fits per stage. Several descriptive fit intervals cross zero. The event evaluation adds useful evidence, but contains only 19 works from one event and does not establish independent-author generalization.
  4. The raw/augmented larger-model comparison is not fully matched. For the causal d128 setting, Index and Scan select different optimization recipes and endpoints. The resulting gap may therefore mix the representation effect with tuning differences.
  5. The practical effect is small. Across the ordinary-material prediction comparisons, the reported gains are approximately 0.0207–0.0410 bits/target, with accuracy gains of 0.200–0.412 percentage points. The frozen event evaluation shows a larger descriptive accuracy gain of 0.886 percentage points, but it uses a separate 19-work population. Without evidence for improved free generation, layout quality, or another downstream metric, the practical importance of the effect remains unclear.
Questions for authors
  1. How can the authors distinguish the benefit of explicit prefix accumulation from the benefit of providing source coordinates in the same metric as the target query?
  2. Why is there no learned cumulative-position, CoPE-like, or query-relative coordinate baseline? Is the claimed invariance strictly source-side? What happens when equivalent whitespace rewrites occur in the target history or current-row prefix?
  3. How should reviewers interpret the positive material bootstrap intervals together with several descriptive three-fit intervals that cross zero? Do the NLL improvements translate into better free generation, layout preservation, or another practically meaningful outcome?
Reviewer_7zGAOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies the predictive value of explicit cumulative coordinates in proportional-font text art. By matching character inputs, true glyph widths, model capacity, and contextual access, while canonicalizing certain equivalent whitespace encodings, the paper find that cumulative coordinates still improve character prediction. The experiments are carefully controlled, however the additional insights and broader impact remain limited.

Critical strengths

They carefully control factors that could affect the comparison, including access to local widths, parameter counts, the scale of positional scalars, and access to source-row positions. Multiple experiments support the reported gains. And they also transparently disclose evaluation-data reuse, different sources of uncertainty, and unfavorable results, and clearly delimit the scope of their conclusions.

Critical weaknesses

My main concern is the significance of the paper's contribution. Several prior studies have already shown that explicitly providing task-relevant spatial representations can facilitate learning. It remains unclear what important open question this study resolves or how it advances the field’s existing understanding. The experiments and conclusions also do not develop the finding into a more explanatory account or establish its broader applicability.

Questions for authors

Please see the above Weaknesses.

Reviewer_ksn8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper investigates whether providing a model with an explicit, pre-calculated spatial quantity enhances predictive performance, even when the raw inputs already inherently embed this information. Using proportional-font Japanese text art as a controlled setting, the experiments validate that explicit spatial metrics provide an essential inductive bias that mitigates the optimization hurdles of implicit arithmetic. To isolate this, the authors compare paired Transformers—one utilizing ordinal character indices (W-index) and the other pre-calculated cumulative coordinates (W-scan). By canonicalizing whitespace to remove encoding redundancies, the results demonstrate that the explicit coordinate model consistently achieves superior predictive likelihood across diverse evaluations, confirming that supplying explicit spatial quantities yields a persistent advantage even when information is fully derivable.

Critical strengths
  • The paper utilizes character art as an suitable experimental setting to isolate spatial geometry from complex textual noise, providing a clean and intuitive proof for the theory that computational feasibility does not equal utilization efficiency.
  • The study conducts solid comparisons across multiple dimensions, featuring not only the normalized tests for the main models (Table 1) but also a series of ablation experiments (Width-only, Linear, Pseudo, etc.) that rule out alternative explanations like having an arbitrary extra input channel or relying on coarse linear scaling.
  • The authors transparently disclose all overfitting phenomena, intervals crossing zero, and negative results regarding large language model transfer attempts.
  • The "Design Map" in the supplemental systematically organizes all material assignments, easy for readers to locate.
Critical weaknesses
  • Limited Generalization: The positive claims are entirely restricted to single-character prediction on a single font; as the authors candidly admit in Appendix L.4, attempts to transfer this framework to other fonts, NetHack, or pretrained LLMs (Qwen and Gemma) all failed, showing that it is not yet a universal representation learning principle.
  • Entangled Underlying Mechanisms: Because the target query qq already uses true coordinates, the advantage of W-scan may stem from "source–target metric consistency" rather than purely sparing the model from computing prefix sums.
  • Scaling Degradation: Performance actually drops when scaling model capacity from d=128d=128 to d=256d=256 (Table 5), and the causal scaling experiments use asymmetric training recipes and update steps between arms.
  • Lack of Modern Baselines: The study lacks comparisons against modern learned cumulative position encodings such as CoPE, and evaluations on end-to-end multi-line generation quality.

CoPE: A Lightweight Complex Positional Encoding https://arxiv.org/abs/2508.18308

Questions for authors
  • Since both models are given the pre-computed target coordinate qq, W-scan benefits from direct source–target metric consistency. What happens if the target-side coordinate is removed, forcing both models into asymmetric metric spaces where target queries use only the ordinal column cc? Does the performance advantage of W-scan over W-index still persists?
  • Given the severe late-epoch overfitting and asymmetric training configurations in the causal d=128d=128 experiment (where W-Index and W-Scan used different update budgets and recipes), could the authors report test contrasts under strictly matched training recipes, update counts, and with residual dropout applied?
  • The appendix notes that attempts to generalize this framework to a second font (Osaka), NetHack, or pretrained LLMs (Qwen and Gemma) yielded negative results. Could the authors elaborate on why these transfer attempts failed, and what this implies for the practical scope of explicit spatial features?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.