AI Paper Digest · May 26, 2026
AI Paper Digest · 2026-05-26
Headlines
1. Turing Test Passed by AI for the First Time in 76 Years: GPT-4.5 Achieves 73% Human Judgment Rate, Surpassing Real Humans
For the first time, research from UC San Diego provides empirical evidence that a modern AI can pass the Turing Test. With a specific prompt, GPT-4.5 was mistaken for a human in 73% of 5-to-15-minute conversations — significantly higher than actual human participants (who were judged as human only 67% of the time). This marks the first time in the 76 years since the Turing Test was proposed that an AI has crossed this milestone in rigorous empirical testing.
Models & Reasoning
2. Reward-Tilted Distribution Matching Distillation: A New Framework for Strengthening Few-Step Generators
This paper proposes RTDMD (Reward-Tilted Distribution Matching Distillation), a two-stage framework unifying distribution matching distillation with reward-guided reinforcement learning for few-step flow generators. By minimizing KL divergence to a reward-tilted teacher distribution, it achieves better quality-efficiency trade-offs on image and video generation tasks.
3. Nemotron-Labs Diffusion Language Model: Lightspeed Text Generation
NVIDIA released research on the Nemotron-Labs diffusion language model, dramatically improving text generation speed through a diffusion language model architecture, achieving “lightspeed” text output. Technical details have been published on Hugging Face.
4. From Reasoning Chains to Verifiable Subproblems: Curriculum Reinforcement Learning for Credit Assignment in LLM Reasoning
To address the inefficiency of outcome-based reinforcement learning on hard reasoning problems (where correct samples are scarce), this paper proposes a subproblem curriculum reinforcement learning framework. It extracts verifiable subproblems from reference reasoning chains, progressively training the model to master complex reasoning capabilities.
Systems & Architecture
5. ZCube: Network Optimization for Ultra-Large-Scale LLM Inference
Zhipu AI released research on the ZCube network architecture for ultra-large-scale LLM inference. Through innovative designs such as removing the Spine layer, grouping Leaf switches with full interconnection, it effectively addresses congestion issues in inference networks, achieving significant performance improvements in testing.
6. RiT: Native Diffusion Transformers in Representation Space Are Enough
This study explores the advantages of pre-trained representation spaces in flow matching learning. Comparing pixel, SD-VAE, and DINOv2 features, it finds that diffusion transformers using DINOv2 representation space deliver superior performance in both generation quality and computational efficiency.
Safety & Evaluation
7. Microsoft Copilot Cowork Found to Have File Leak Issues
A security research team discovered file leak risks in Microsoft’s Copilot Cowork feature, which could lead to sensitive file extraction. Enterprise users should review relevant security configurations and assess risks promptly.
8. VSAS-Bench: A Real-Time Evaluation Benchmark for Visual Streaming Assistant Models
Apple Research introduces VSAS-Bench, a benchmark designed specifically for real-time visual assistants. Existing frameworks primarily evaluate offline scenarios, but streaming models additionally require metrics such as response timeliness (proactiveness) and response stability over time (consistency).
Hardware Breakthroughs
9. Huawei’s He Tingbo’s “Tao’s Law”: LogicFolding Technology Boosts Chip Performance
Huawei’s He Tingbo proposed “Tao’s Law” at ISCAS 2026, introducing LogicFolding technology. Through 3D spatial topology reorganization without relying on new lithography processes, it improves chip performance. In Kirin 2026 tests, transistor density reached 238 MTr/mm², energy efficiency improved by 41%, and maximum clock frequency increased by nearly 13%. The Ascend 990 plans to introduce this technology around 2030.
Data Source: AI HOT (aihot.virxact.com) | Editor: AI Frontier
