The InclusionMed Collaboration just released an open benchmark initiative asking which AI models actually serve healthcare needs worldwide! 🩺
"InclusionMed: Benchmarking AI for Inclusive Healthcare"
This paper maps medical AI benchmarks across three healthcare roles and three AI capabilities, then rolls frontier LLM results into a single InclusionMed Index.
It covers four public benchmarks picked from 388 candidates, a private diagnostic reasoning set, Index scores from 17.5 to 49.1 out of 100, rankings that shift by role and capability, top models stronger at patient communication than clinical decisions, and much more!
Read this excellent paper by The InclusionMed Collaboration: paperswithcode.co/paper/2611…
1
16
Cordelia Schmid and Peter Kulits (Inria, Max Planck) just released BrickFlow: give it a pile of LEGO parts and a photo, and it works out how to build the model! 🧱
"BrickFlow: Connectivity-Guided Brick Reconstruction"
This NeurIPS 2026 paper shows how to infer the pose of every part when physical constraints make the problem combinatorial and sensitive to tiny pose errors.
It covers a set-conditioned SE(3) flow model trained on procedurally generated assemblies, image conditioning added with 1% as much data, analytic connectivity-field guidance at test time, build quality that improves with model size, pretraining and guidance, and much more!
Read this excellent paper by Peter Kulits and Cordelia Schmid: paperswithcode.co/paper/2611…
1
28
Models and data are available for research use.
Code: github.com/kulits/BrickFlow
Project page: kulits.github.io/BrickFlow/
8
ITMO University researchers just released a neat finding for anyone fine-tuning small models: the output format you pick decides how often they fail silently! 🍕
"Parsed, Plausible, Wrong: Output Format Controls Silent Errors in Small-Model Semantic Parsing"
This paper shows that writing the target tree in ordinary words, not native s-expressions, makes sub-billion parsers both more accurate and less likely to emit well-formed but wrong output.
It covers six lossless encodings on the PIZZA ordering benchmark, 300 LoRA runs on Qwen3.5-0.8B and LFM2.5-350M, 4.1x fewer tokens, silent errors cut by 2.3x on Qwen and 1.8x on LFM2, why shortening alone never helps, and much more!
Read this excellent paper by Umar Ahmed, Maxim Dunaev and collaborators: paperswithcode.co/paper/2611…
1
37
MIT, Stanford and NTU researchers just released AllocFlow, a generative model that turns raw market data straight into portfolio trades! 📊
"AllocFlow: End-to-End Portfolio Trading Action Generation from Raw Market Data"
This paper shows how flow matching can replace the usual forecast-then-allocate pipeline by generating signed position weights across all assets jointly.
It covers a cross-asset flow Transformer, strategy imitation followed by group-relative RL, annualized net return on hourly crypto perpetual futures rising from 27.79% to 51.38%, a Calmar ratio of 4.90 on six further months of unseen 2026 data, and much more!
Read this excellent paper by Annan Wang, Yuchen Xie, Weisi Lin and collaborators: paperswithcode.co/paper/2611…
1
26
Code: github.com/allocflow/AllocFl…
Project page: allocflow.github.io/
5
Shanghai AI Laboratory and Zhejiang University researchers just released JVT, a strikingly simple recipe for generating video directly in pixel space! 🎥
"JVT: Just Video Transformer makes pixel-space video diffusion simple"
This paper shows that a plain large-patch DiT-style Transformer can do pixel-space video diffusion with no pretrained autoencoder, no cascaded stages and no U-Net-style skip connections.
It covers the design choices that matter (video patch embeddings, register tokens, adaptive layer norm), quality competitive with latent diffusion baselines, joint RGB-disparity generation that improves camera following and yields scene geometry for free, and much more!
Read this excellent paper by Jianjun Zhou, Hua Geng and Haoyu Guo: paperswithcode.co/paper/2611…
1
35
Code: github.com/ZhouTimeMachine/J…
Project page: zhoutimemachine.github.io/JV…
8
Congrats @FluidVoiceApp: Fluid Speech Pro is SOTA on VoxPopuli! 🏆
The 0.6B open-weights model hits 1.7 WER, ahead of NVIDIA's 30B Nemotron 3.5 Transcribe (1.84).
It's built on NVIDIA's Parakeet TDT 0.6B v3, covers 25 European languages and ships under CC BY 4.0.
The best closed model on the board, StepAudio 2.5 ASR, sits at 2.76.
Weights are on Hugging Face: FluidVoice-Labs/fluid-speech-pro
Full VoxPopuli leaderboard: paperswithcode.co/benchmark/…
61
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…
1
41
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
12
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…
Code: github.com/georgecazenavette…
Project page: drifting-distillation.github…
25
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.
1
28
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…
6
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/fscdc/MMFlash
Project page: fscdc.github.io/MMFlash/
1
19










