Next-Concept Modeling (NCM) is a text-generation framework designed to combine the speed of parallel generation with the quality of sequential, autoregressive models. It addresses limitations of diffusion language models: quality falls sharply when sampling steps are reduced, each step processes the full sequence, and strong performance can require many training passes over data. NCM first uses a multiscale residual vector-quantized autoencoder to turn text into a hierarchy of discrete codes. A transformer then generates the hierarchy from coarse to fine, predicting every code at a given level in parallel. This structure aims to reduce the amount of sequential computation while letting broad concepts guide the details produced later.
In matched-parameter comparisons, NCM has better generative perplexity than the strongest diffusion baseline on LM1B and runs 21 times faster. On OpenWebText it is competitive with that baseline at 10 times the speed, while on XSum it matches the baseline’s ROUGE-L score and runs 31 times faster. The authors also test Sudoku as a reasoning task: NCM outperforms autoregressive and diffusion methods across all difficulty levels while using fewer function evaluations. Taken together, the results support coarse-to-fine, level-by-level generation as a way to achieve substantial inference speedups without sacrificing benchmark quality, and potentially to improve performance on structured reasoning problems.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Researchers tested 16 LLMs on all 55 TIS-100 puzzles, giving each one attempt and time to refine its code. The models found 133 new performance tradeoffs across 18 puzzles and set three outright cycle records, though human players improved on most solutions within days. (tis-100.wincon.dev)
Australia is considering a new criminal offence for AI companies that fail to take reasonable steps to prevent harm caused by their models. The government is also weighing an AI safety watchdog with power to impose significant fines, following a hack of a Medicare statistics portal involving OpenAI. (abc.net.au)
AI is increasingly becoming a gateway through which people find information. Audiences still resist having AI produce their news. (niemanlab.org)
Nvidia holds a central position in the AI industry, making disruption to the company potentially catastrophic for the wider ecosystem. The race for advanced AI is also intensifying economic and geopolitical conflict, while concentrated corporate power and infrastructure face growing political resistance. (france24.com)
LLM inference can require reading about 26 GB of model weights to generate each token from a 13-billion-parameter model. The explanation shows how model size drives memory bandwidth demands during text generation. (youtube.com)
Deletion Tuning matches full fine-tuning across three LLM post-training methods by zeroing fewer than 0.05% of model weights. Its compact masks also preserve prior-task accuracy better than full fine-tuning, LoRA, and random sparse tuning. (arxiv.org)
Today's best Hacker News stories, summarized and screenshotted, one email a day.