Pterodactyl reports a 1020 Group WMT26 system for segment-level quality prediction and error-free segment detection, built as a low-cost LLM-as-a-judge setup tuned with prompt engineering and ROC analysis. Alongside that automated evaluation work, there is a paper on word-level alignment between OCR output and ground truth under zoning and reading-order errors, and another asking whether a trained 27B hybrid-attention model can be converted to pure byte-level tokenization while retaining most of its ability.

Pterodactyl: Pushing the Limits of Efficient Automated Translation Evaluation

Use a prompt-based LLM judge only for segment-level triage and error-free detection, because this submission is tuned for ultra-low-cost binary decisions rather than full quality assessment.

In this paper, we present 1020 Group's submission to the WMT26 shared task on automated quality evaluation. Our system, Pterodactyl, was engineered specifically for subtasks 2 and 3, that is, segment-level quality score prediction and the detection of error-free segments. Pterodactyl is a highly customizable, ultra-low cost system optimized for binary classification in translation triage and data curation. It is based on an LLM-as-a-judge approach, with prompt engineering and model selection guided by ROC analysis. We also developed several variants of Pterodactyl that were specifically adapted for subtasks 3.

Word-Level Alignment of OCR and Ground Truth under Zoning and Reading-Order Errors

For OCR alignment, anchor methods need explicit recovery from reading-order permutations, and unique-anchor approaches fail when repeated content leaves too few anchors.

OCR post-processing needs OCR text aligned word by word with its ground truth. Data sets rarely provide that mapping, and a monotone aligner can lose its place when zoning errors permute the reading order. We revisit this problem, the Zoning Problem of our 2019 paper, with a protocol that tests correctness under such permutations. We apply OCR-side reading-order permutations with construction-time word-level gold and score correct-match F1, cascade length and recovery. We compare the 2019 location-aware aligner, a re-anchoring extension of it, a RETAS-style baseline (RETAS-lite), a multi-region aligner that combines unique-word and unique n-gram anchors with chunk-to-window assignment, genalog's RETAS and the original Java RETAS. Experiments use synthetic character noise with block moves, paragraph swaps and column interleaving, historical newspaper reading-order permutations, geometry-derived gold for Tesseract OCR, and a small word-box set. Synthetic tests show where local tracking fails after large displacements and where unique-anchor methods collapse when unique anchors are scarce. On our synthetic boilerplate page the 2019 method had the highest F1 among those tested. Historical newspaper and geometry-derived evaluations carry the limitations of approximate gold and the article-order convention. The first multi-region version has a quadratic scaling cost. Local tracking remains fragile under large reading-order changes; the tested anchor-based methods handle many of those cases but fail when unique anchors are scarce. Approximate gold, the article-order convention, missing residual annotation, and the lack of modern multi-column word-level ground truth remain limitations.

One Byte, One Token: Byteifying a 27B Hybrid-Attention LLM with Self-Generated Data

Byteifying a large Chinese LLM can preserve most capability with only 384 vocabulary entries, but it still needs 4.77 GB of supervised bytes and loses on some reasoning and instruction tasks.

Byte-level BPE fragments Chinese text (for Qwen3.8 on Chinese Wikipedia, 3.84 bytes per token on average; 64% of tokens are single characters) and hides spelling from the model. We ask whether an already-trained 27B subword model can be converted directly into a pure byte-level model (one byte, one token), without a patching structure or a local encoder, while keeping most of its ability. Starting from Qwen3.8-27B (48 Gated DeltaNet layers + 16 full-attention layers), we shrink the vocabulary from 248,320 to 384 entries (the 256 byte symbols already in the BPE vocabulary plus special ids), leave every other weight untouched, and continue training in a two-stage curriculum (encoding first, then reading) on 4.77 GB of supervised bytes: all dialogue/behavioral supervision is generated on-policy by the original model itself (character dictionary, word dictionary, general dialogue, coding sessions), and knowledge reading uses encyclopedia text (54% of supervised bytes). Training takes 178.6 h on 8xH200. On the same GPU and serving stack, the byte model's gap to the original falls within the paired bootstrap interval on MATH-500, HumanEval+, MMLU-Pro and non-thinking GSM8K, and it loses 4-6 points on C-Eval, thinking-mode GSM8K and non-thinking IFEval. Without thinking it leads on every character-level task (CUTE 75.2 -> 84.0, EXECUTE-zho 81.6 -> 93.9, our Chinese character suite 54.4 -> 64.0), but lookup knowledge about Chinese characters such as pronunciation and stroke count drops, and thinking does not recover it. The cost is throughput: per-step speed is identical, but each step yields 3-5x less output; Chinese characters per second is about 1/4 of the original, and prefill at 64K bytes is 6.5x slower. We also document four byte-model failure modes (greedy runaway, sampling breaking UTF-8, ineffective penalty terms on a small vocabulary, repetition) and their fixes.

Papers announced by arXiv on Tuesday, 10 November 2026, covering submissions from 9 November to 10 November 2026.