More than three thousand years ago, diviners in ancient China carved questions about harvests, warfare, and royal fortunes onto turtle shells and ox bones. Today those inscriptions survive as fragments of cracked, degraded surfaces captured in stone rubbings, and the ability to read them rests in the hands of a small community of specialist paleographers. A new study published in Discover Artificial Intelligence shows that the artificial intelligence systems being built to help decipher these texts are failing in a way their creators may never have noticed: they appear impressive on average while quietly collapsing on the rarest and, arguably, most interesting characters.
The research team, led by Dan Sui and Yuhang Xing of Anyang Normal University alongside Penglan Liu of the University of Jyväskylä and corresponding author Bang Li, took aim at a task they describe as the entrance gate to AI-assisted interpretation: detecting and recognizing the visible characters on oracle bone rubbing images. Oracle bone inscriptions constitute the earliest known mature Chinese writing system, and reading them automatically requires two things at once. A system must first locate each visible character on the rubbing, then assign it an identity from a very large repertoire of glyph categories.
That second requirement is what transforms an apparently ordinary computer vision problem into something far more treacherous. The study builds on the Oracle Bone Inscription Multi-modal Dataset, known as OBIMD, which for the first time provides both bounding boxes and glyph category labels on the same rubbing images. With more than 1,700 character categories spread over roughly 9,100 training images and more than 82,000 annotated visible characters, OBIMD is the only existing testbed for full-image, large-vocabulary oracle character recognition. It is also brutally imbalanced: over 60 percent of its categories fall into a few-shot regime, meaning most glyphs appear only a handful of times across the entire corpus.
The headline finding of the paper is a diagnostic one. A competitive direct multi-class YOLO detector, trained in the standard way, reaches an overall average precision at 50 percent intersection-over-union (AP50) of 0.63 across three random seeds. Yet the same system scores only 0.06 AP50 on the few-shot class group. In other words, the aggregate number that most papers report conceals a near-total failure on rare glyphs, the categories where automated help is arguably needed most because experts see them too infrequently to build easy intuition.
To make this hidden failure measurable, the authors constructed a frequency-aware stratified evaluation protocol, dividing categories into head, medium, long-tail, few-shot, and unseen groups. Sixty-nine of the 1,730 categories appear only in validation or test data, contributing 74 test annotations that a closed-set recognizer can never label correctly. This kind of per-frequency breakdown is standard practice in long-tailed detection research generally, but applying it to oracle bone reading is what exposes the rare-class collapse and, the authors argue, what should become the reporting norm for this task.
Armed with that protocol, the team tested a structural hypothesis. In a conventional closed-set detector, localization and recognition are optimized together in a single dense head. Localization, however, pools supervision from every visible character regardless of category; it is category-independent and, the experiments show, already strong, achieving 0.85 AP50 and 0.80 recall. Recognition, by contrast, depends on class-specific evidence whose frequency varies by several orders of magnitude. Forcing those two objectives into the same head, the authors reason, couples a robust capability to a fragile one.
Their remedy is deliberately minimal. The proposed framework, called FAR-YOLO for Frequency-Aware Recognition YOLO, splits the problem in two. A class-agnostic YOLO11s localizer trained with a single foreground class called oracle character finds character regions. Each detected crop is padded to a square canvas and passed to a separate recognizer, here a ResNet18 classifier, trained with frequency-aware sampling that multiplies the exposure of long-tail categories by three and few-shot categories by eight, combined with morphology-preserving augmentation. That augmentation is carefully constrained: rotation within 14 degrees, translation up to 8 percent, modest scale and shear changes, and mild brightness and contrast jitter, all chosen to avoid altering the stroke topology that distinguishes one ancient glyph from another.
The results validate the design. FAR-YOLO keeps overall AP50 essentially unchanged while improving few-shot AP50 by 10.5 percentage points and class-balanced AP50 by 3.7 points over the direct multi-class detector, with consistent tail gains replicated across seven different recognition backbones including ConvNeXt, EfficientNetV2, and a Vision Transformer. Notably, the obvious alternative fails: naive inverse-frequency reweighting of the detector head actually lowers overall performance without rescuing the tail. Other long-tailed methods show lopsided trade-offs. RIDE achieves the best aggregate scores but trails FAR-YOLO by 0.05 to 0.06 on tail metrics, while PaCo and EQLv2 win on specific frequency strata only by sacrificing substantial aggregate accuracy. FAR-YOLO ranks second on all four primary frequency-stratified metrics, a balanced operating point no other method matches.
Perhaps most revealing is what the framework chooses not to say. Detector confidence and recognizer probability are multiplied into a fused score, and when that score falls below a threshold the system abstains rather than forcing a guess. With a ConvNeXt-Tiny recognizer, the fused confidence distinguishes seen from unseen categories with an AUROC of 0.77, and a threshold of 0.5 rejects over 90 percent of unseen instances. The authors frame this abstention as a route toward candidate ranking, retrieval-assisted recognition, and paleographer-in-the-loop verification rather than closed-set forced classification, positioning the machine as a collaborator that flags uncertainty for expert judgment.
The study is equally candid about what remains unsolved. Qualitative analysis of full rubbings shows three recurring error modes: localization misses under dense character layouts, confusion between visually similar glyphs, and above all large intra-class variation, since a single oracle character category can encompass several structurally distinct allographic forms, each with limited training examples. In one clean example a medium-frequency character is correctly localized by both systems yet misread as a similar glyph, an error rooted in the intrinsic variability of the script rather than in data scarcity. Future work, the team suggests, may pair glyph images with component descriptions, structural decompositions, and expert-authored interpretations through vision-language alignment. For now, the paper delivers a quieter but more consequential message for the field: in a domain shaped by a 3,000-year-old long tail, the average is exactly where the truth is not.
Subject of Research: Long-tailed detection and recognition of oracle bone script characters using frequency-aware deep learning
Article Title: Beyond aggregate metrics: diagnosing the long-tailed failure of oracle bone character detection and recognition
Article References: Beyond aggregate metrics: diagnosing the long-tailed failure of oracle bone character detection and recognition. (n.d.). https://doi.org/10.1007/s44163-026-02210-2
Image Credits: AI Generated
DOI: 10.1007/s44163-026-02210-2
Keywords: oracle bone inscriptions, OBIMD, long-tailed recognition, few-shot learning, object detection, YOLO, character recognition, machine learning, digital humanities, paleography, frequency-aware sampling, FAR-YOLO
News Source: Denise Maddox. (October 12, 2026). AI Reads Ancient Oracle Bones: New Framework Exposes Hidden Failures in Rare Glyph Recognition. Scienmag.
