
When data, not compute, is the limit, the authors find latent feedback is the better use of a FLOP. Code: github.com/ifm-ai/LoRM

Trending AI research papers with code, datasets and leaderboards.

When data, not compute, is the limit, the authors find latent feedback is the better use of a FLOP. Code: github.com/ifm-ai/LoRM

CMU, USC and MBZUAI's Institute of Foundation Models just released a careful answer to a big question: what should you do with extra compute once high-quality data runs out? 📈 "Latent Feedback Is a Compute and Data Amplifier" This paper tests latent feedback, which feeds each token's top-layer state into the next token's input, against a vanilla Transformer that spends the same compute on more epochs. It covers a ladder of model sizes, unique data and epochs, savings of up to 48% of the compute and 47% of the unique data, one scaling law for both models that predicts a held-out 778M model's loss to within 0.3%, and much more! Read this excellent paper by @huskydogewoof and collaborators: paperswithcode.co/paper/2611…


Antonio Torralba, Vincent Sitzmann and collaborators at MIT just released a fresh take on dataset distillation, a NeurIPS 2026 paper! 🧪 "Dataset Distillation via Drifting" This paper shows how to compress a large dataset into a small set of synthetic samples by matching each class's data distribution directly, instead of optimizing for a classifier. It covers "drifting" samples in a feature extractor's latent space toward the real data and away from each other, far better coverage of the real distribution, synthetic sets that still train strong classifiers, a visualization tool for interpreting models, extensions beyond images, and much more! Read this excellent paper by @GCazenavette and collaborators: paperswithcode.co/paper/2611…

Introduced in 2025 in "Diffusion Transformers with Representation Autoencoders" (1.13 FID on ImageNet 256×256), and since extended to text-to-image, video latents and world models. Full explainer: paperswithcode.co/methods/re…

What is a Representation Autoencoder (RAE)? 🖼️ Diffusion transformers usually work in a VAE's latent space, learned only for reconstruction. RAE swaps the VAE for a frozen encoder like DINOv2 plus a trained ViT decoder: semantically rich latents, still high-quality reconstructions. Now on Papers with Code: 101 papers use it.


National University of Singapore researchers, with Xinchao Wang, just released MMFlash, a faster way to run speculative decoding on vision-language models! ⚡ "MMFlash: Faster Speculative Decoding for Vision-Language Models" This paper shows why drafters struggle on image-heavy outputs: token acceptance falls from 46% to 22% as a token's dependence on the visual input rises. It covers a KL-based metric of each token's visual dependence, upweighting the visually dependent tokens a drafter predicts poorly, sparse attention for long videos, lossless speedups of up to 3.58x, gains of up to 195.4% over EAGLE-3 and 17.1% over DFlash, and much more! Read this excellent paper by Sicheng Feng, Zigeng Chen and collaborators: paperswithcode.co/paper/2611…

Code: github.com/H-furkan/redivers… Project page: rediverse-distillation.githu…

UIUC and Johns Hopkins researchers, with James Rehg, just released a sharp fix for one of the most annoying problems in fast image and video generators: diversity collapse! 🎨 "Restoring Diversity in Distilled Image and Video Generators by Probing Activations" This paper shows why few-step distilled models produce near-identical samples from different seeds: they keep the teacher's total activation variance but squeeze it into far fewer directions. It covers where the collapse happens (one intermediate block at the first denoising step), ReDiverse-FT, a lightweight adapter that regresses student activations onto the teacher's, ReDiverse-TF, a training-free fix that steers along the teacher's principal directions, and much more! Read this excellent paper by Furkan Horoz, Ozgur Kara and collaborators: paperswithcode.co/paper/2611…


Fudan, SJTU and VoxDance Lab researchers just released VoxMel, a text-to-speech model that generates plain mel spectrograms, with no learned audio tokenizer! 🔊 "VoxMel TTS Technical Report" This paper shows how a fixed, exactly invertible Haar transform makes autoregressive mel modeling stable. It covers a language-model backbone with a flow-matching head that generates coarse mel frames, non-autoregressive tiers that predict the residuals, a 0.6B model trained on 100K hours of Emilia, 1.15% WER and 0.753 speaker similarity on test-en after GRPO, 2- and 4-step distilled students, and much more! Read this excellent paper by Wenhao Guan, Jiaxin Ye and collaborators: paperswithcode.co/paper/2611…

Their takeaway: dedicated vision capacity, trained jointly with the backbone and placed early within it. Project page: unbundling.github.io/

Meta FAIR and NYU, with Saining Xie, Lucas Beyer and Luke Zettlemoyer, just released a sharp controlled study on how multimodal models should see images! "Unbundling Multimodal Vision Encoding" Do you need a pretrained vision encoder? They pretrain every model from scratch on the same data at matched compute to find out. It covers why the visual prior isn't necessary, why encoder-free models still trail, how vision modules inside early backbone layers can lower time to first token, and much more! Read this excellent paper by @_ellisbrown and collaborators at @AIatMeta: paperswithcode.co/paper/2611…

Introduced in 2025 in "Mean Flows for One-step Generative Modeling", now a common baseline for one-step image, audio and policy generation. Full explainer: paperswithcode.co/methods/me…

What is MeanFlow? ⚡ Flow matching learns the instantaneous velocity, so sampling takes many steps from noise to data. MeanFlow learns the average velocity over a time interval instead, so a single network evaluation jumps straight from noise to data. Trained from scratch, no distillation: 3.43 FID on ImageNet 256×256 in one step. Now on Papers with Code: 140 papers use it.