Translated title. Original: 自己回帰生成モデルの熱力学を構築 ―大規模言語モデルに内在する「世界モデル」の定量的理解に向けて―
A thermodynamics framework for autoregressive generative models ties irreversibility to entropy generation, introducing sentence-level entropy, and showing through open-language models that sequential event order influences entropy. The work aims to quantify the model's internal world representations and is published in PRX Intelligence with institutional support from Tokyo-tech and JST.
Future Chunk Prediction improves autoregressive video generation by predicting future chunks, enhancing temporal evolution modeling without extra encoders and with lower training cost, while preserving inference and yielding more structured representations.
The paper proves that, in autoregressive transformers, the prompt's influence on future tokens decays over time under broad conditions, with a quantified rate. It introduces a conditional no-hair theorem and provides an exactly solvable model to illustrate and bound the forgetting dynamic, showing dilution is necessary and sufficient under observed regimes.
MetaExp presents a scalable framework to train a transferable experience-learning capability for language-model agents using verifiable cross-task task families, with two designs (Shared-Tool and Shared-Workflow), large-scale environment generation, trajectory curation, and fine-tuning that yields performance gains across multiple benchmarks.
Translated title. Original: Språkforsker fant «typiske KI-klisjeer» i Horsts siste roman
A language researcher identified recurring AI-style wording and a sharp rise in dashes in Horst’s latest crime novel. The text links these changes to his use of AI.
The article reports that AI data-labeling workers in countries including Venezuela and Pakistan can earn very low pay despite long hours. Workers describe scarce tasks and unpaid time that reduce their effective wages, while Appen says it is working to improve task access.
The post promotes an Orthogonal Model of Emotions (OME) for sentiment analysis, built with Transformers/PyTorch and an MIT-licensed dataset. It claims superior accuracy on diverse text, notes unique dataset coverage, and highlights long development time while inviting others to check it out.
A method to convert pretrained raster-scan autoregressive image models into any-order generators using future-aware positional embeddings and an ordering curriculum, achieving better inpainting/outpainting than MaskGIT with modest extra compute while keeping the AR inference pipeline intact.
Uno adds a diffusion-based adapter path to draft token blocks, while the original autoregressive model verifies them and preserves its output distribution. In reported H200 tests, Uno exceeded the base model and two speculative-decoding methods in aggregate throughput, though results depend on workload and concurrency.
A-RSI separates when to refresh tasks from how to generate them, refreshing the task pool based on reward variation and training accuracy without external data or human labels, while auditing answers and verifying reward/refresh logic; it improves benchmark performance by adding fresh tasks.