Hermann Kumbong and colleagues test a way to generate text in successive levels, from a rough representation to finer detail. Their next-concept model predicts many codes at each level in parallel, then decodes the finished representation into text.
- On the LM1B text dataset, the authors report generation 21 times faster than the DUO diffusion baseline, with a better score on their model-based text-quality metric.
- On the XSum summarization dataset, they report 31 times faster generation than AURORA-LM at comparable ROUGE, a measure of overlap with reference summaries. Both timings use one H100 GPU.
- On hard Sudoku puzzles, the model solves 86.1% completely, versus 81.3% for the strongest reported diffusion baseline. This is a narrow reasoning test.
Caveat: The main experiments use small, separately trained models for each dataset. These are author-reported comparisons with specific baselines, not proof of a speedup over frontier coding or chat models. A separate tokenizer also needs training, and good text-overlap scores do not guarantee factual accuracy.
The 26-comment thread on Hacker News raises deployment and comparison questions:
What the thread adds
- pinyen_cheng — “a brand new tokenizer and training method need a lot of effort to do the retraining.”
- cubefox — “they hardly compare against the conventional next-token-prediction objective,” questioning how it compares with standard language models.
- swiftcoder — “Who’s going to train a nice big coding model based on this?”
Handles are pseudonymous; HN provides no per-comment scores, and its ranking gives only a slice of the discussion, not a consensus.