Christian Wolf

New papers on recurrent transformers: arxiv.org/pdf/2606.21562 (ours): train full Tr, distill -> rec Tr arxiv.org/pdf/2608.02870 train full and rec jointly, min. coherence arxiv.org/pdf/2606.06479 train rec Tr with predictive obj. arxiv.org/pdf/2608.08888 train full Tr, add re…

arxiv.org

New papers on recurrent transformers: arxiv.org/pdf/2606.21562 (ours): train full Tr, distill -> rec Tr arxiv.org/pdf/2608.02870 train full and rec jointly, min. coherence arxiv.org/pdf/2606.06479 train rec Tr with predictive obj. arxiv.org/pdf/2608.08888 train full Tr, add re…

arxiv.org

New papers on recurrent transformers: arxiv.org/pdf/2606.21562 (ours): train full Tr, distill -> rec Tr arxiv.org/pdf/2608.02870 train full and rec jointly, min. coherence arxiv.org/pdf/2606.06479 train rec Tr with predictive obj. arxiv.org/pdf/2608.08888 train full Tr, add re…

arxiv.org

New papers on recurrent transformers: arxiv.org/pdf/2606.21562 (ours): train full Tr, distill -> rec Tr arxiv.org/pdf/2608.02870 train full and rec jointly, min. coherence arxiv.org/pdf/2606.06479 train rec Tr with predictive obj. arxiv.org/pdf/2608.08888 train full Tr, add re…

arxiv.org

Another relevant paper @phillipisola.bsky.social shared at a different place: arxiv.org/abs/2511.05963

Next-Latent Prediction Transformers Learn Compact World Models arxiv.org

Memory Caching: RNNs with growing memory Behrouz et al. arxiv.org/abs/2602.24281 Take different RNNs and, and at time t don't read out only M_t, but cached previous hidden memories. Neat and effective. Tested on different modern RNN variants.

Memory Caching: RNNs with Growing Memory arxiv.org

IMHO there is no clear difference btw (1) hallucinations (2) Inferring from unobservable information using common sense reasoning. It's the same thing, and (2) is desired, (1) is not. Inferring that there is floor behind the couch on the left is a "desired" hallucination. arxi…

Zero-BEV: Zero-shot Projection of Any First-Person Modality to BEV Maps arxiv.org

AI bubble vs. FIFA world cup vs. Summer Heat wave

For your Embodied AI task you want a recurrent model with constant complexity per step, but you don't want to lose the power of transformers (which store the full obs history and attend to it)? Do not despair, we have your back. We distill transformers into recurrent transformers 1/8

My Dad made fun of me for my use of "please" when talking to Claude. His point: the AI does not care. My point: this is not about the AI but about me.

I declined ACing for ICLR 2027 (but will review!), I am not comfortable with the load of LLM generated text to deal with: over-length reviews, over-length rebuttals. Humans digest machine output, it does not make sense! 😬 ICLR should take a step back, use 1-page pdf rebuttals as CVPR always did.

Excellent Keynote by @akorba.bsky.social at Cap-Rfiap in Montpellier, France, on machine learning and distances between distributions.

Goodhart's law states that "When a measure becomes a target, it ceases to be a good measure". Conference peer-review is a measure, arguably now mostly implemented through LLMs, and will become a target. Ergo, future academic research will target what Claude finds interesting and acceptable.

Finished my AC duties for NeurIPS: absolutely __NO__ LLM used, not even for proof-reading or grammatical checks.

It is 2030 and we ran out of English phrases which we can still use because of they are not frequently employed by LLMs.

ICRA panel on the impact of AI and the paper avalanche on robotics conferences and journals. The same problems arrise in ML and in CV of course.

In 2026 the accepted papers are the ones which were reviewed by the same LLM which the authors had used to check the paper before submission.