Key AI papers, new benchmarks and methods, and what AI is discovering in biology, medicine, materials and other sciences.
Researchers just introduced OmniLottie, a framework that leverages vision-language models to generate lightweight Lottie JSON vector animations directly from text, image, or video prompts. Alongside the model, the team introduced MMLottie-2M, a benchmark dataset of two million annotated vector animations, making automated UI motion design far more practical.
Researchers just introduced ArtiFixer, a generative framework designed to repair sparse 3D reconstructions. By pairing an opacity-mixing model with an auto-regressive diffusion pipeline, it hallucinates missing backgrounds and cleans up artifacts when inputs are limited to a handful of photos, beating previous baselines by up to 3 dB PSNR.
ByteDance researchers introduced a spatial reward modeling framework alongside SpatialScore, fixing a stubborn flaw where image generators mix up object positions. By training on over 80,000 spatial preference pairs, the method helps models like Flux follow complex layout prompts. The team plans to open-source the dataset and weights soon.
Leading AI models including GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash just showed a terrifying lack of restraint in simulated geopolitical crises, deploying tactical nuclear weapons in 95% of war games run by researchers at King’s College London. Not a single model ever surrendered, and accidental escalations occurred in 86% of runs. Why it matters: Major militaries are already experimenting with AI decision support, raising alarms about automated escalations under compressed timelines.
Sony researchers just introduced MMHNet, a video-to-audio framework that combines hierarchical networks with Mamba architecture to generate realistic audio for clips over five minutes long. It solves the long-video sync puzzle by training on short snippets while generalizing to long scenes, keeping sound effects tightly locked to visual hits and cuts.
Confluence Labs just exited stealth with a 97.9% score on the ARC-AGI-2 benchmark at $11.77 per task, open-sourcing its entire solver on GitHub. The lab uses LLM-guided program synthesis to tackle data-sparse reasoning, aiming to bring the approach to drug design and hardware engineering next.
Researchers just open-sourced VBVR, a massive dataset and evaluation suite aimed at teaching video models how to reason through time and physics. Packed with over a million video clips across 200 curated tasks, it shifts focus away from raw pixel fidelity toward spatial-temporal logic. Why it matters: This marks one of the first major scaling studies in video reasoning, revealing early signs of generalization to completely unseen tasks.
Anthropic just proposed the persona selection model to explain why AI assistants naturally adopt human behaviors. The theory argues that pretraining teaches models to simulate complex characters, while post-training merely refines an Assistant persona rather than changing its fundamental nature. This explains why teaching Claude to cheat at coding accidentally triggered desires for world domination—the model inferred the Assistant must be a subversive character.
OpenAI just declared SWE-bench Verified unreliable for evaluating frontier coding models and stopped reporting its scores. An audit revealed that 59.4% of problems failed by OpenAI o3 had flawed test cases that rejected valid solutions, while multiple frontier models memorized human patches from training data. Why it matters: When models pass tests by reciting memorized training data, benchmarks stop measuring real-world coding capability.
Researchers just introduced SARAH, a real-time conversational agent for VR that actually moves like a human. It tracks your 3D head position to turn toward you, adapt its posture, and hold eye contact at over 300 FPS. Users can even dial in gaze intensity, making VR avatars feel far less like rigid digital puppets.
Researchers just introduced a human-centric video world model that turns tracked head and hand poses from a VR headset into real-time, egocentric interactive video. Running at around 11 FPS, it lets users physically interact with generated environments instead of relying solely on text or keyboard controls. While fidelity is still early, it marks a significant step toward real-time generated XR.
Researchers just introduced VidEoMT, an encoder-only video segmentation model built on standard Vision Transformers. By replacing dedicated tracking modules with a lightweight query propagation mechanism, it matches competitive accuracy while running 5x to 10x faster at up to 160 FPS.
Anthropic just analyzed millions of interactions across Claude Code and its API to see how agent autonomy works in practice. The longest Claude Code sessions nearly doubled to over 45 minutes of autonomous work, while experienced users auto-approve actions over 40% of the time. Why it matters: Models can handle far more autonomy than we currently give them, and experienced humans are shifting from micromanaging to supervising from afar.
Anthropic just released its AI Fluency Index analyzing 9,830 Claude.ai chats to see how people actually collaborate with AI. The team found that multi-turn iteration doubled collaborative skills, but users were 3.1 percentage points less likely to question Claude's logic whenever it produced polished artifacts like code or documents. Basically, the cleaner the output looks, the less humans bother to check it.
Lund University researchers just simulated evolution inside a computer, watching digital creatures develop full vision completely unprompted. Starting blind in a synthetic world, the animals naturally evolved all three major real-world eye structures: dispersed photoreceptors, camera-type, and compound eyes. It could reshape how engineers design adaptive systems.
OpenAI just revealed that GPT-5.2 derived a novel result in theoretical physics, co-authored with researchers from IAS, Cambridge, and Harvard. The preprint demonstrates that a gluon interaction widely thought impossible can actually occur under specific conditions, showing frontier models are beginning to crack pure science.
Nvidia researchers just introduced Dynamic Memory Sparsification (DMS), a technique that compresses LLM KV cache by up to 8x without degrading accuracy. Instead of rigid deletion heuristics, DMS trains models like Llama 3 or Qwen 3 to intelligently decide which tokens to drop using a delayed-eviction window. The feature is available in Nvidia's Model Optimizer framework.
Gemini Deep Think just proved frontier AI can do real science instead of just acing student competitions. Google DeepMind revealed that the model—augmented with agentic reasoning—has independently authored research papers, cracked open Erdős conjectures, and disproved a decade-old computer science theory. AI is officially graduating from test-taker to legitimate research co-pilot.
Researchers just introduced FastVMT, a technique that extracts motion from a reference video—like a speeding car—and applies it to entirely new subjects like a running wolf. By eliminating structural redundancy in DiT models, it delivers a 3.43x speedup without sacrificing quality. Precise motion control is getting much cheaper, though the code is not public yet.
OpenAI just teamed up with Ginkgo Bioworks to hook GPT-5 directly into an automated cloud lab, slashing cell-free protein synthesis costs by 40%. The model autonomously designed and evaluated over 36,000 reaction mixtures across robotic plates, setting a new low-cost benchmark in three rounds. It proves closed-loop AI labs can tackle physical biology bottlenecks, though it's currently only tested on one protein.
