AI

Nvidia’s erroneous paper accepted as ICML’s spotlight [D]

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 12:47:26 UTC.

Briefings

A user reports that Nvidia's 'dreamDojo' paper was accepted for ICML despite significant code bugs. The paper claims 44k hours of human data pre-training on a 256 H100 GPU yielded a marginal 0.5 dB PSNR improvement over their prior 'Cosmos 2.5' model. However, the user discovered critical bugs in both pre-training and post-training code, suggesting the results are unreliable. The paper was published on 2026-10-08, and the user notes the authors did not receive feedback on the submission.

Key points

• The paper 'dreamDojo' was accepted for ICML despite the user discovering critical bugs in its pre-training and post-training code.
• The paper claims 44k hours of human data pre-training resulted in a marginal 0.5 dB PSNR improvement over 'Cosmos 2.5'.

Context

Nvidia's 'dreamDojo' is a robotics world model based on 'Cosmos 2.5', cited ~100 times. The user claims the paper was accepted without review due to code bugs, questioning the authors' response to the submission.

Limitations

User claims the paper was accepted without review; specific dates for bug discovery are not provided. Code bugs are described as 'very bad written' and affect pre-training, post-training, and evaluation, but no specific bug details or dates are given.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 08/11/2026, 05:18:25 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeigehmlkug6eulpusjovyljhkyn4mkg4tjxmb3yczdbhetfa2udpli