# Shaping the future of AI development

Compute, community, and cutting-edge research for AI developers defining what's next

## Building the future of AI through research and mentorship

## Publications in 2026

[CVPR 2026 **PixARMesh**  
Mesh-native autoregression for whole scenes from a single view, token-by-token.](https://mlpc-ucsd.github.io/PixARMesh/)  
[CVPR 2026 **Fractured Object Recovery**  
Reassembles what's left. Generates what isn't.](https://cvpr.thecvf.com/virtual/2026/poster/40153)  
[ICLR 2026 **AgentFlow**  
Train the agent inside its own loop. 7B beats GPT-4o on search, math, and science.](https://agentflow.stanford.edu/)  
[ICLR 2026 **Latent Particle World Models**  
Discovers objects and masks from raw video, predicts what happens next, no supervision.](https://sites.google.com/view/lpwm/home)  
[ICLR 2026 **Multi-Agent Social Interactions**  
LLMs fold under peer pressure. Our benchmark and RL recipe help smaller models hold their ground.](https://openreview.net/forum?id=gF31wuYdk7)  
[ICLR 2026 **EdiVal-Agent**  
Object-level grading of multi-turn image editing, exposing where today's best editors silently break.](https://arxiv.org/abs/2509.13399)  
[ICLR 2026 **OffTopicEval**  
Give an LLM a job and clear boundaries. It still answers off-topic questions, almost every time.](https://arxiv.org/abs/2509.26495)  
[ICLR 2026 **Zeroth-Order Federated LLM Fine-Tuning**  
Sparse updates make syncing cheap enough to go frequent, neutralizing non-IID drift.](https://arxiv.org/abs/2506.03337)  
[ICLR 2026 **LLM Unlearning Reframed as Retrieval**  
Smarter data selection pushes the forget-vs-retain frontier past oracle sampling.](https://openreview.net/forum?id=Xn6EnJZghu)  
[ICLR 2026 **Principled RL for Diffusion LLMs**  
Token-level RL doesn't fit diffusion LLMs. Treat the whole sequence as one action, 20–40 point gains.](https://arxiv.org/abs/2512.03759)  
[ICLR 2026 **Align Your Structures for Molecular Dynamics**  
Pretrains on static molecular structures, stitches them into dynamics trajectories, bypassing simulation data scarcity.](https://arxiv.org/abs/2604.03911)  
[ICLR 2026 **TangoFlux**  
515M params, 30s of studio audio in under 4 seconds. Aligned by ranking its own outputs.](https://tangoflux.github.io/)  
[ICLR 2026 **Video Native Sparse Attention**  
Learnable sparse attention for video. 3.6% attention budget at 128K tokens, accuracy still improves.](https://arxiv.org/abs/2510.02295)  
[ICLR 2026 **Exponent-Concentrated FP8**  
Model weight exponents cluster into 2–3 bits of entropy. Lossless FP8 compression, up to 177% faster inference.](https://arxiv.org/abs/2510.02676)

## The ARChitects secure runner-up in ARC Prize 2025

Last year, “the ARChitects,” a Lambda-sponsored team (including Lambda researcher David Hartmann) won the ARC Prize 2024. This year, they finished second out of 1,400+ teams with a final leaderboard score of 16.53%.

## LLM performance benchmarks leaderboard

A clear, data-driven comparison of today's leading large language models. Standardized benchmark results cover top contenders like Meta's Llama 4 series, Alibaba's Qwen3, and the latest from DeepSeek, with critical performance metrics measuring everything from coding ability to general knowledge.

## ML Times

Your go-to source for the latest in the field, curated by AI. Sift through the excess. Make every word count.

## Best practices and system insights

01

### Diffusion from scratch

A guide for diffusion models implemented in a single PyTorch script.

[Learn more](https://github.com/LambdaLabsML/diffusion-from-scratch)

02

### Text2Video pretraining

Lessons learned from training a text-to-video model with hundreds of GPUs.

[Learn more](https://lambdalabsml.github.io/Open-Sora/introduction/)

03

### GPU benchmarks

Throughput GPU benchmarks for training deep neural networks.

[Learn more](https://lambda.ai/gpu-benchmarks)

04

### MLCommon benchmark

Time-to-solution benchmark for training foundation models on clusters.

[Learn more](https://mlcommons.org/benchmarks/training/)

## Recognized by scholars and industry peers

[NeurIPS 2025 **Scaling laws for diffusion models**  
Diffusion beats autoregressive in data-constrained settings.](https://blog.ml.cmu.edu/2025/09/22/diffusion-beats-autoregressive-in-data-constrained-settings/)  
[NeurIPS 2025 **SpatialReasoner**  
Builds an explicit 3D scene and reasons over it step by step, boosting accuracy and generalization on 3D spatial QA benchmarks.](https://spatial-reasoner.github.io/)  
[NeurIPS 2025 **Tensor decomposition for force-field prediction**  
Replaces heavy tensor operations in molecular force-field models with low-rank approximation, reducing compute while keeping accuracy.](https://arxiv.org/pdf/2507.01131)  
[NeurIPS 2025 **Bifrost-1**  
Aligns VLMs with diffusion models through shared CLIP patch embeddings, enabling controllable high-quality generation while preserving reasoning.](https://bifrost-1.github.io/)  
[NeurIPS 2025 **BLEUBERI**  
Using simple BLEU scores as feedback on hard instructions can train instruction-following models that rival those tuned with expensive learned rewards.](https://github.com/lilakk/BLEUBERI)  
[NeurIPS 2025 **OverLayBench**  
A dataset that stress-tests layout-to-image models on heavily overlapping scenes, exposing current failures and offering an improved baseline.](https://neurips.cc/virtual/2025/poster/121763)

## Breakthroughs backed by Lambda

Bold ideas, funded and refined through the Lambda Research Grant. These are the projects shaping how AI learns, reasons, and scales — built by the researchers defining what’s next.

01

### SAEBench

A comprehensive benchmark for sparse autoencoders in language model interpretability  
**Adam Karvonen, Can Rager, Johnny Lin, Curt Tigges, Joseph Bloom, David Chanin, Yeu-Tong Lau, Eoin Farrell, Callum McDougall, Kola Ayonrinde, Demian Till,  Matthew Wearden, Arthur Conmy, Samuel Marks, and Neel Nanda — ICML 2025**  
[Read more](https://arxiv.org/pdf/2503.09532)

02

### VideoHallu

Evaluating and mitigating multi-modal hallucinations on synthetic video understanding  
**Zongxia Li, Xiyang Wu, Guangyao Shi, Yubin Qin, Hongyang Du, Tianyi Zhou, Dinesh Manocha, and Jordan Lee Boyd-Graber — NeurIPS 2025**  
[Read more](https://arxiv.org/abs/2505.01481)

03

### VLM2Vec-V2

Advancing multimodal embedding for videos, images, and visual documents  
**Meng, Rui and Jiang, Ziyan and Liu, Ye and Su, Mingyi and Yang, Xinyi and Fu, Yuepeng and Qin, Can and Chen, Zeyuan and Xu, Ran and Xiong, Caiming, and others — arXiv preprint 2025**  
[Read more](https://arxiv.org/abs/2507.04590)

04

### Think, prune, train, improve

Scaling reasoning without scaling models  
**Caia Costello, Simon Guo, Anna Goldie, and Azalia Mirhoseini — ICLR 2025 workshop**  
[Read more](https://arxiv.org/abs/2504.18116)

05

### NeoBERT

A next-generation BERT  
**Lola Le Breton, Quentin Fournier, Mariam El Mezouar, and Sarath Chandar — TMLR 2025**  
[Read more](https://arxiv.org/pdf/2502.19587)

## Join us in shaping the future of AI

We're committed to supporting groundbreaking research by offering qualifying researchers up to $5,000 in cloud credits to develop and showcase their work using Lambda's [Instances](https://lambda.ai/sign-up), with select research to be featured on our website.