If a video generation model is allowed to use a non-trivial amount of inference-time compute, how much can it improve generation quality for challenging text prompts?
This is exactly what a new paper from Tsinghua University explores.
RESEARCH
-

Video Generation Models Leverage Inference Compute for Quality
By
–
-

ARC Prize Foundation Launches AGI Competition with $700k Prize
By
–
ARC Prize Foundation launched ARC-AGI-2 – a set of tasks humans find easy but AI struggles with!
— Rowan Cheung (@rowancheung) 25 mars 2025
The foundation's 2025 competition launches soon with a $700k prize for the first to achieve an 85% score within cost limits.
The current best is o3-low's 4%pic.twitter.com/k1Zw4SFkbXARC Prize Foundation launched ARC-AGI-2 – a set of tasks humans find easy but AI struggles with! The foundation's 2025 competition launches soon with a $700k prize for the first to achieve an 85% score within cost limits. The current best is o3-low's 4%
-

Alibaba Qwen2.5-VL-32B-Instruct: Advanced Vision-Language Model
By
–
The Qwen team of Alibaba open-sourced Qwen2.5-VL-32B-Instruct, a new vision-language model —Features enhanced mathematical reasoning and visual capabilities
—Beats Qwen2-VL-72B, Mistral Small 3.1-24B, and GPT-4o-0513
—Apache 2.0 licensed -
DeepSeek V3-0324: Open-Source AI Model with Enhanced Math Capabilities
By
–
DeepSeek released V3-0324, an updated version of its V3 non-reasoning AI
— Rowan Cheung (@rowancheung) 25 mars 2025
—A 641GB model but capable of running on high-end PCs
—Uses MoE, activating only 37B params/token and reducing compute
—Upgraded Math and Coding capabilities
—Open-source under MITpic.twitter.com/sskVI1tIFKDeepSeek released V3-0324, an updated version of its V3 non-reasoning AI —A 641GB model but capable of running on high-end PCs
—Uses MoE, activating only 37B params/token and reducing compute
—Upgraded Math and Coding capabilities
—Open-source under MIT -
AI Startup Reve Dethrones Image Generation Giants
By
–
TODAY'S AI NEWS: A new AI startup, Reve, just dethroned leading image generation giants! Plus, more news from DeepSeek, Qwen, Google, Alibaba, and more. Here's everything you need to know:
-

ARC-AGI-2 Benchmark Shows AI Reasoning at 4% Accuracy
By
–
On the newly launched ARC-AGI-2 benchmark, the best reasoning AI today can only achieve 4% accuracy. https://
arcprize.org/leaderboard -
Medium-Sized Questions in Neuroscience and AI
By
–
On medium-sized questions in neuroscience https://
buff.ly/ezlHepe
#AI #MachineLearning #DeepLearning #LLMs #DataScience -

Qwen2.5-VL-32B-Instruct Released with RL Optimizations
By
–
Qwen2.5-VL-32B-Instruct is here! Optimized with RL, it shows significant improvements in human preference and mathematical reasoning.
-

DeepSeek V3 Latest Version Improves Math Coding Reasoning
By
–
The latest version of DeepSeek V3 is now available! Today’s update improves its math, coding, and general reasoning performance, making it one of the most powerful open-source LLMs available. (1/2)
-

Six Elements of Agent Engineering: Intent, Memory, Planning
By
–
The Six Elements of Agent Engineering compiling answers from @simonw
's definitions thread (thanks @lilianweng @ruthefordml @hwchase17 @DanielMiessler @Dan_Jeffries1 @romechenko
) IMPACT: – Intent (Instructions + Evals)
– Memory (
@gneubig reusable workflows)
– Planning (Editable