# alphaXiv - shadow discussion board for arxiv

Source: https://www.alphaxiv.org
Kiwi News: https://news.kiwistand.com/stories/alphaXiv-shadow-discussion-board-for-arxiv?index=0x66ddb302e7fea7fc1a853ef14bca3538e15a5209252dca019949c3cf635ce1bbd7a71b31
Submitted by m-j-r.eth on 2024-09-08 UTC, 7 upvotes, 1 comment

## Summary

This page lists several recent arXiv papers surfaced through alphaXiv, a discussion board for arXiv preprints, though the text itself does not state how the papers were selected. Three are described: a semantically steerable image tokenization method that can prioritize semantic fidelity alongside pixel accuracy when compressing images, a large flow language model that drafts several tokens in parallel and verifies them autoregressively to generate text faster without sacrificing comparable quality, and an evolving agent interfaces approach in which computer-use agents improve their own tools from unlabeled task experience. The listing notes that no more papers match the current filters.

## Article

A Fix for Softmax and Cross-Entropy

More accurate softmax and cross-entropy gradients prevent training instabilities at high confidence and stop distillation students from drifting when they match their teachers.

Kimi K3.1 Pro ranks fourth on Artificial Analysis’s Intelligence Index and leads Terminal-Bench 4.0, with open weights planned for release.

Agentic World Action Models

Combining reusable code-based skill plans with visual foresight helps robots perform complex manipulation across parallel-gripper and dexterous-hand tasks.

There are no more papers matching your filters at the moment.

## Comments

**macbudkowski.eth**: Wow. It's fascinating that they saw this huge library of AI content (arxiv), and decided to permissionlessly build a community around it (alphaxiv). The concept is kind of similar to Kiwi, but we have to bootstrap the crypto content library ourselves :)