An arXiv preprint dated 20 August 2026 reports that a system called UniLang had the highest displayed score in every reported recommendation metric on MovieLens-20M and in every reported legal-precedent metric across three LePaRD subsets. UniLang extends Llama-3.2-1B-Instruct with machine-native tokens, but the study did not examine whether that symbolic fine-tuning affects natural-language generation quality, fluency or coherence.
A second vocabulary for structured output
The idea is to make structured prediction part of language generation itself. UniLang treats natural-language tokens and machine-native symbols as first-class generative units inside the same pretrained large language model, asking whether the model can generate symbols directly instead of verbalizing them or abandoning the pretrained model.
To build the symbol set, the implementation uses an RQ-VAE discretization pipeline with three quantization levels and a 256-entry codebook at each level, producing a machine-native vocabulary of 1,024 tokens. It aligns machine-token sequences with embeddings of corresponding text descriptions using an InfoNCE-style contrastive loss that keeps matching representations close and separates them from embeddings of other items.
For downstream tasks, Llama-3.2-1B-Instruct is expanded with the 1,024 machine tokens. The original and grounded embeddings are frozen, while LoRA adapters are trained with supervised fine-tuning.
The reported scores
Evaluation covered six benchmark configurations: Amazon Beauty, MovieLens 1M, MovieLens 20M and three LePaRD splits, spanning sequential recommendation and legal precedent prediction. Recommendation used Recall and NDCG at k=5 and k=10 with leave-one-out evaluation and full ranking; legal prediction used Recall@1, Recall@10 and NDCG@10 with a 90%/5%/5% train, validation and test split.
MovieLens-20M produced UniLang scores of 0.1911 for Recall@5, 0.1382 for NDCG@5, 0.2597 for Recall@10 and 0.1603 for NDCG@10—the highest displayed model score in each row. The table lists improvement entries of 114.96%, 151.73%, 73.95% and 99.88%, respectively, but does not define what those percentages are measured against.
A separate summary of three independent MovieLens-20M runs reported means ± standard errors of 0.1908 ± 0.00016 for Recall@5, 0.1378 ± 0.00017 for NDCG@5, 0.2596 ± 0.00014 for Recall@10 and 0.1600 ± 0.00014 for NDCG@10. Those figures describe run-to-run variability for MovieLens-20M only.
The same pattern appeared in legal precedent prediction. UniLang had the highest displayed value for all three reported metrics in each LePaRD subset—10k, 20k and 50k. Recall@1 was 0.2938, 0.2486 and 0.1676, respectively, with table-listed improvement entries of 49.36%, 48.51% and 36.15%.
What the tests leave unanswered
A component-removal check on a fixed validation subset offered a more mixed view of the design: NLRemoved improved at first but degraded after 40,000 steps; NoTypeDelim remained stable but below the full model; and NoWarmup stayed at zero NDCG and Recall.
The paper reports no confidence intervals or significance tests for the MovieLens table or LePaRD comparisons, and only MovieLens-20M received independent-seed run-to-run estimates. Because the evidence covers six benchmark configurations across two task types, the findings are best read as results for those reported benchmark settings.
The study also did not evaluate whether symbolic fine-tuning affects natural-language generation quality, fluency or coherence. That leaves a central question about the approach unanswered even as its benchmark rankings were strong.
Paper data and sources
Original title: When Machines Speak: A Unified Generative Framework for Integrating Machine-Native Symbols into Pretrained Large Language Models
Authors: Su Yan, Rakesh Iyer
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text