A computational study reports that compact AI models can turn unfamiliar structured inputs into text with fewer judged errors than models of the same size, even when target-domain training text and test references are unavailable. Its DDKD approach trains a smaller model on text generated by a teacher. Across five target domains, the approach beat same-size zero-shot models - models used without target-domain training - and models fine-tuned on WebNLG, a separate collection of graph-text pairs, in Gemma3-1B and Qwen3-1.7B comparisons. The result is most relevant to the narrow problem the researchers set out to test: cross-domain generation when local examples are missing.
The study asks whether text generation can transfer across different data structures and domains without either training text or test reference texts for the target domain. The authors' hypothesis was that DDKD would work best when the teacher had already been fine-tuned for data-to-text generation. That recipe combines broad WebNLG supervision with synthetic examples built around the structure of each target domain.
A test without local reference text
To test the question, QUINTD-1 supplied 100 development inputs and 100 test inputs per domain, with no in-domain reference texts in either set. The source-supervised comparator used LoRA on 40K WebNLG graph-text pairs. DDKD took a different route: it trained a smaller student on teacher-generated target text paired with the structured inputs.
Teaching smaller models from generated examples
The researchers also compared structure-preserving augmentation with simply adding more real target-domain inputs. In the QUINTD-5 control, the real input supply rose from 100 to 500 per domain, a fivefold increase. The synthetic variants were designed to preserve the structure of the inputs while creating additional training examples.
Where the result is strongest
The study's main automatic score was based on the average number of errors per output, so lower values were preferred. On that measure, DDKD had lower error rates than same-size zero-shot and WebNLG-finetuned baselines for both Gemma3-1B and Qwen3-1.7B across all five domains. The reported result is benchmark- and judge-dependent, and the analysis does not provide confidence intervals.
Comparisons with much larger systems are less tidy. The Results section reports that a compact distilled model outperformed Qwen3-32B and GPT-4.1 on four of the five domains. The abstract gives a different summary for a much larger fine-tuned comparison: the small models did better on two domains and performed comparably on the remaining three. Because those two summaries differ, the paper does not offer one clean headline number for its large-model result.
Lower error counts can sometimes reflect sparse output, so the study added a coverage check using GPT-5.1. This language-model measure asked how much of the structured input appeared in the generated text. DDKD-Best recorded a normalized average of 0.8150, above the 1.7B zero-shot value of 0.5795 and the 1.7B SFT value of 0.0000, but below the larger teacher's 0.8908.
A separate human evaluation pointed in the same direction among compact systems: DDKD had the lowest error counts among the 1.7B models and the lowest normalized average overall. That result provides a useful check on the automated scoring, but it does not supply direct reference-based ground truth for the target domains.
The evaluator matters
The evaluator remains a central source of uncertainty. Because target-domain references were unavailable, the study used an LLM-as-a-judge protocol that compared generated text directly with the structured input and classified four error types. The two LLM judges had medium agreement on individual examples, with Pearson correlation r=0.666, but very high agreement on system rankings, at r=0.955 across all domains. That suggests the broad ordering of systems was more stable than the verdict on any one output.
The paper also compares different training recipes. In an ablation of source datasets, WebNLG was the strongest source-domain initialization, with a normalized error score of 0.954 and a normalized coverage score of 1.000. In the real-data comparison, the final normalized error averages were 0.42 for 500-real DDKD versus 0.27 for DDKD-Pert under a zero-shot teacher; under a WebNLG-SFT teacher, the corresponding values were 0.20 for 500-real DDKD and 0.10 for DDKD-Mixed. The cited comparisons favored the augmented variants, but they were nonrandomized.
What the result does not establish
Taken together, the findings do not establish that synthetic data generally beats collecting more real target-domain examples. The evidence is limited to reference-free evaluation on five QUINTD target domains, mainly compact models and two LLM judges. The augmentation depends on heuristic choices about atomic units and exchangeability, while agreement is weaker for subjective misleading errors. The study therefore supports a benchmark-specific conclusion: DDKD can be tested as a way to investigate cross-domain generation when references are missing, but broader validation across model families, domains and human or reference-based evaluations remains open.
Paper data and sources
Original title: Cross-Domain, Multi-Task Data-to-Text Generation without In-Domain Training Data
Authors: Yifei Song, Kun Efimov-Zhang, Claire Gardent
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: Not available
Original paper · Full text