Research Briefing — August 5, 2026

LongCat Sparse Attention is a hardware-algorithm co-designed framework that integrates streaming-aware, cross-layer, and hierarchical indexing to speed up sparse attention for LLMs. It achieves significant speedups for million-token contexts while maintaining model quality.

This technical report details DiffusionGemma, an open-weight text model using discrete block diffusion to generate blocks of 256 tokens in parallel. Fine-tuned from the mixture-of-experts Gemma 4 model using less than 10% of the original training budget, it combines bidirectional denoising SFT with reinforcement learning and sampler distillation. This demonstrates how discrete diffusion can accelerate inference speeds while preserving language model quality.

ETA introduces a planner-centered, closed-loop architecture for embodied AI that integrates large language models as planners and robot capabilities as tools. This approach achieved 90% task success on LIBERO manipulation tasks using a lightweight interface.

Paul Christiano returns to the Alignment Research Center (ARC) as executive director to focus on mechanistic explanations for neural network behavior. The organization plans to scale up research aiming to detect and address AI misalignment directly.

Latent Reward Registers propose a mechanism to estimate preference from intermediate noisy latents in diffusion models. This addresses the temporal credit-assignment problem in alignment by providing dense, differentiable signals throughout the denoising process.

ReflectRL proposes using failed trajectories from expert models, termed 'Golden Negative Trajectories', as a valuable training signal. It argues that reflecting on these flawed paths is often more effective than solving problems directly from scratch for hard cases.

Qwen-3D is a generalist 3D Vision-Language Model that uses geometry-aware decoding to improve spatial reasoning. It addresses the limitations of frame-centric tokenization by fusing observations from multiple views and time steps.

LoopMTP enhances looped transformers by using multi-token prediction to guide intermediate latent states across loops. This structural alignment mitigates latent overthinking and provides forward-looking supervision for parameter-efficient depth scaling.

A new deep learning method published in Nature Machine Intelligence enables molecular dynamics simulations to surpass the femtosecond time scale constraint while retaining physical property accuracy.

This study identifies deletion avoidance in AI coding models, showing that top models on the SWE-bench Verified leaderboard fail to remove obsolete or incorrect code lines in over 48% of cases. Instead, models rely on additive Guard-and-Go workarounds that wrap code in conditional guards, passing existing test suites while polluting codebases with technical debt. The paper exposes benchmark evaluation flaws and outlines strategies to incentivize proper code deletion.

This research tests a fundamental assumption in long-horizon agents: that cached KV entries remain informative after their source observations are deleted. By omitting specific observations, the authors demonstrate that downstream events often rely on these 'semantic materialized' rows, revealing a new memory contract for sparse attention systems.

JoyAI-Video-Edit presents a 16B-parameter autoregressive diffusion framework capable of real-time, open-ended video editing. The system achieves 30 FPS on an Nvidia B200 GPU, outperforming existing streaming editors in both quality benchmarks and deployment efficiency.