"A Fix for Softmax and Cross-Entropy"

Turns out two of the most fundamental functions in deep learning have numerical bugs that can actually destabilize training.

This paper proposes simple fixes to softmax and cross-entropy that prevent tiny gradients from incorrectly rounding onzero when they should vanish, as these errors can cause sudden loss spikes and interfere with knowledge distillation.

On small-scale grokking benchmarks, the fixes eliminate Slingshot loss spikes and improve test accuracy from 94.3% to 100% in one setting. They also prevent student models from drifting away from their teachers during distillation, without requiring higher numerical precision.

alphaxiv.org/abs/2610.31642

The good news: you can mid-train yourself to the frontier!

At a fraction of the cost! Without distillation!

What we did: We took Qwen3.5 14B and mid- and post-trained it to beat every other open model and perform as well as the very best frontier models.

This is no simple ng, we are training the model to understand and reason over DNA sequences. Bio is the next frontier for AI and universal access will be crucial to make progress.

The fun part: because we train the model on both natural language and DNA sequences you can actually 'chat to your DNA': Ask it questions, ponder the impact of mutations, or ask it to generate a small sequence.

We'll soon see the frontier labs move away from math and cyber towards bio. It already started, and we believe researchers should have access to capable models.

We are releasing the full recipe, so you can do it, too!

Model: huggingface.co/HuggingFaceBio…

Blog: huggingface.co/spaces/Hugging…

Looking for a fun weekend read?

Introducing the Illustrated NeurIPS 🌎

We indexed all 9000+ NeurIPS papers and built a visual way to explore the whole landscape

Search your favorite topics, open a cluster, and dive right in

Check it out here! alphaxiv.org/conferences/ne…

how much can models remember?

there’s a neat perspective on neural networks that deserves to be more widely known

here’s how you could have invented every layer of the transformer from memory

(1/n) from my new blog post!

🏆 The #3DV2027 Citations Grand Prix continues! P3, with 2793 citations. "Deeper Depth Prediction with Fully Convolutional Residual Networks".
arxiv.org/abs/1606.00373

📢 Looking for #PhD? Ψάχνεις για #διδακτορικό? My team #3DI has 3 open positions in our #ERC project on perceiving 3D human-object interaction; a step towards world models for avatars/physical-AI!
👉Apply: dtzionas.com/apply/ →Early 🐦 eats 🪱
📍@Auth_University z help spread the word

Interested in world models?

Join us for the last keynote talk at the MELEX #ECCV2026 Workshop today by Prof. Jean Ponce on "Self-supervised learning for world modeling" in Malmömassan C1.

@eccvconf @bazile @lanalpa

Today at 9/9 at #ECCV26:

9:50-10:20: I’ll give a keynote at the MELEX Workshop (More EXploration, Less EXploitation, sites.google.com/view/melex-202… , located in Malmömässan C1) on “Targeted Scientific Discovery in Ecology”

(more below)

Our second talk at MELEX #ECCV2026 Workshop by Prof. Sara Beery on "Targeted Scientific Discovery in Ecology"

Join us at Malmömassan C1 for this amazing talk!

@eccvconf @sarameghanbeery

First talk by Prof. Josef Kittler on "Intriguing Properties of Foundation Model Representations"

#ECCV2026 @eccvconf

#ECCV MELEX Workshop happening right now at MalmöMässan C1!

See you there!

@eccvconf @bazile @lanalpa @ghcarneiro

The 3rd MELEX workshop (More Exploration, Less Exploitation) will be held at @eccvconf. Thrilled to have Josef Kittler, @orussakovsky, and Jean Ponce as invited speakers.

9 September 2026, 09:00–12:30, Malmömässan, Malmö #ECCV2026

Organizing with: @bazile @ghcarneiro @lanalpa

Model-agnostic Retrieval-Augmented Extended Forecasting for time series

Juan Pablo Villa Serna, Rohan Asthana, Vasileios Belagiannis
arxiv.org/abs/2608.14054 [𝚌𝚜.𝙻𝙶]

CfP: Join us for the 3rd Workshop on More Exploration, Less Exploitation (MELEX 2026) at #ECCV2026 (@eccvconf ) in Malmö! 🇸🇪
Tired of minor SOTA-chasing tweaks? We are looking for bold, highly innovative and exploratory computer vision ideas! (1/3)

Submission Deadline: June 22, 2026 ⏰
Accepted papers will be published in the ECCV proceedings.

Submit your manuscript + 1-page review response letter here: 🔗 sites.google.com/view/melex-202… (2/3)

#ComputerVision #MachineLearning #ECCV #DeepLearning

🔥 Invited Speakers: David Forsyth, Danica Kragic, Josef Kittler, Olga Russakovsky, & Jean Ponce!

Turns out you can edit images with a VQ‑VAE autoregressive model faster than diffusion – they “nudge” the logits using a mask from the source token map. Same quality, ~10× speedup. Who knew a little logit push could beat diffusion? 🚀🤯

arxiv.org/abs/2604.14591

Just as we say goodbye to Vancouver (YVR)🇨🇦🥹, it’s time to reveal our next destination! ✈️

From the Pacific coast to the Mediterranean sea...
We are thrilled to announce that #3DV2027 is heading to Thessaloniki, Greece (SKG)! 🇬🇷🥳

See you there in Mid-April 2027! 🌍

[new CVPR'26 paper]
🔄 SSL works great when you have tons of data.
But in 3D… we don’t.
High-quality 3D scans are expensive, slow, and hard to scale. So what if we could pretrain 3D models without any real 3D scans? 1/

The @IEEE Transactions on Robot Learning (T-RL) will launch on March 30! Co-EiCs: Todd Murphey and Vincent Vanhoucke
ieee-ras.org/publications/t…