Christian Wolf

NeurIPS 2026 accepted 102 workshops from 454 valid submissions, spread across Sydney, Paris, and Atlanta venues.
Acceptance rates were 21.5% in Sydney, 25.4% in Paris, and 23.6% in Atlanta, with each proposal getting at least two reviews.
New safeguards cap workshops at 8 organizers and bar any individual from appearing on more than 2 proposals.

View on Bluesky →

↻ Christian Wolf reposted

Phillip Isola @phillipisola.bsky.social

Students sometimes ask me if it still makes sense, in this accelerating age, to pursue a PhD in AI. Perhaps counterintuitively, I think it's a great time to do so. I wrote up some thoughts on this here: web.mit.edu/phillipi/www...

On the Value of Doing a PhD in the Age of AI web.mit.edu View on Bluesky →

↻ Christian Wolf reposted

@pdejorge.bsky.social

1/6 Excited to share that our paper on model merging was accepted at ECCV 2026! 🎉 We introduce an efficient, decoder-free proxy that makes model selection faster, simpler and practical across vision tasks. 📄 arxiv.org/abs/2604.12935 🌐 europe.naverlabs.com/task-alignment 🧵👇

Task Alignment: A Simple Proxy for Practical Model Merging Across Diverse Vision Tasks arxiv.org

A NAVER Labs Europe paper accepted at ECCV 2026 introduces a 'task alignment proxy' for selecting model-merging hyperparameters without training the decoder for each candidate.
The method targets heterogeneous vision tasks with trainable decoders, moving beyond the CLIP image classification setup that dominated prior model merging work.
The authors claim the proxy speeds up hyperparameter selection by orders of magnitude while retaining downstream performance.

View on Bluesky →

↻ Christian Wolf reposted

NeurIPS Conference @neuripsconf.bsky.social

The NeurIPS 2026 August Newsletter is now available on our blog: blog.neurips.cc/2026/09/05/n... If you would like to receive these newsletters directly by email, please subscribe to the newsletter mailing list under your profile: neurips.cc/Profile/subs...

NeurIPS Newsletter – August 2026 – NeurIPS Blog blog.neurips.cc

NeurIPS 2026 runs across three cities in December: Sydney (Dec 6-12), Atlanta (Dec 8-13) and Paris (Dec 9-13).
Organizers accepted 102 workshops out of 454 valid submissions: 48 in Sydney, 28 in Paris, 26 in Atlanta.
Registration opened September 3 for Sydney and Atlanta; Paris will open after separate system testing.

View on Bluesky →

Having fun with Blender ... Distilling context transformers into recurrent transformers: arxiv.org/abs/2606.21562 @weinzaepfelp.bsky.social @mbsariyildiz.bsky.social

Compressing Observation History into Agent Memory: Distilling Transformers into Recurrent Transformers arxiv.org

Naver Labs Europe researchers propose distilling full-history transformers into recurrent variants for long-horizon streaming vision and robotics.
A teacher model compresses its observation history into a fixed-size bottleneck representation that directly supervises the student's recurrent memory.
The approach achieves linear-time complexity on the Mem-RPE task and is also validated on streaming visual question answering.

View on Bluesky · ♥ 22 ↻ 5 ↩ 1 · 2 from the directory shared this · 12d ago

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…

View on Bluesky · ♥ 43 ↻ 4 ↩ 2 · 2 from the directory shared this · 54d ago

View on Bluesky · ♥ 43 ↻ 4 ↩ 2 · 2 from the directory shared this · 54d ago

View on Bluesky · ♥ 43 ↻ 4 ↩ 2 · 2 from the directory shared this · 54d ago

View on Bluesky · ♥ 43 ↻ 4 ↩ 2 · 2 from the directory shared this · 54d ago

FEEL (Force-Enhanced Egocentric Learning): A Dataset for Physical Action Understanding Dessalene et al., UMD #ECCV2026 oral arxiv.org/abs/2603.15847 Very nice!

FEEL (Force-Enhanced Egocentric Learning): A Dataset for Physical Action Understanding arxiv.org

FEEL contains approximately 3 million force-synchronized frames of unscripted kitchen manipulation captured from custom piezoresistive gloves and egocentric video.
45% of frames involve hand-object contact; the paper reports state-of-the-art temporal contact segmentation without manual contacted-object annotations.
Using force prediction as a self-supervised pretraining objective improved transfer to EPIC-Kitchens, SomethingSomething-V2, EgoExo4D, and Meccano.

View on Bluesky · ♥ 17 ↻ 2 ↩ 1 · 2 from the directory shared this · 38d ago

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

NextLat adds a self-supervised auxiliary loss that trains a transformer to predict its own next latent state given the next token.
The authors argue these latents provably converge toward belief states, a compact summary of history needed to predict the future.
Reported gains span world modeling, reasoning, planning and language modeling, with inference sped up by up to 3.3x via self-speculative decoding.

View on Bluesky · ♥ 5 ↻ 0 ↩ 1 · 2 from the directory shared this · 54d ago

Recurrent Looped Transformer Zhang et al. Combine context/history transformer with recurrence: have a causal encoder with a recurrent decoder. arxiv.org/abs/2610.07591

Recurrent Looped Transformer arxiv.org

Two RLT splits generalize parity from 40 training bits to 256 bits at 100% accuracy in every seed; an eight-layer Transformer stays at chance.
On swap-based S_5 permutation tracking at eight times training length, RLT reaches 97% final-state accuracy versus under 1% for the Transformer.
Removing per-token feedback drops parity and S_5 to chance; chunking feedback every four tokens cuts length-64 swap S_5 from 100% to 20%.