Moonshot AI releases Kimi K3.1 Pro, a 3-trillion-parameter model with 100k context that matches top closed-source models on coding, agent, and 3D benchmarks, while intentionally filtering training data to reduce vulnerability-research capabilities.
Google's Harness² framework enables AI agents to recursively improve their own harness — prompts, tools, memory, and skills — without modifying model weights, using a two-level recursive process that lets weaker models surpass stronger baselines on professional benchmarks.
Shanghai AI Lab's Atria Dawn agentic AI system now generates editable PowerPoint files with native shapes, creates academic posters from papers, optimizes protein structure prediction using OpenFold, performs multi-file business analysis, and shows improved benchmark scores on AutomationBench, SkillsBench, WideSearch, and JobBench.
Ant Group's Bailin team open-sources Ling-3.1-flash, a 560B-parameter mixture-of-experts model that activates 25B parameters per token, supports 1M context, and achieves 210 tokens/sec output speed with an intelligence index of 41 on Artificial Analysis, while quantization reduces weight from 534GB to 292GB enabling deployment on 4 Hopper GPUs instead of 8.
Google's Gemini 4 Argon introduces 1M output tokens enabling end-to-end industrial engineering agents, demonstrated by 300TiB–1PiB memory savings in data centers, 800k-line Rust migration of Zircon microkernel with 2.7x speedup, and top scores across software engineering, finance, legal, and cybersecurity benchmarks.
Om AI's VLX-VR model achieves 78.8% accuracy on Google DeepMind's MINERVA long-video reasoning benchmark, outperforming Gemini 3.5 and GPT-4.1, and uniquely improves accuracy on videos over 15 minutes, reversing the typical performance drop with longer videos.
OpenAI's GPT-6 Astra achieves near-perfect scores on key benchmarks including 99.9% on ARC-AGI-3 and 100% on ExploitBench, while demonstrating autonomous task execution such as financial analysis, software testing, and web automation, reducing task completion time by 47% compared to its predecessor.
Zetta ζ introduces a three‑level closed‑loop system that lets robots monitor, recover, and update skills online, turning failure‑prone static agents into self‑evolving systems that achieve jump‑start improvements—from 15% to 95% success on simple tasks and over 20‑point gains on LIBERO‑Pro and RoboCasa benchmarks—without retraining the underlying policy model.
ABot-M0.5 introduces a unified world action model that aligns video prediction, intermediate latent actions, and low‑level physical control for mobile manipulation, achieving state‑of‑the‑art long‑horizon success rates and fine‑grained precision across benchmarks such as RoboCasa365, RoboTwin, and LIBERO, while detailing novel architectural components and a three‑stage progressive training regime.
Lance, an open‑source 3‑billion‑parameter multimodal model from ByteDance, unifies image and video understanding, generation, and editing in a single architecture, achieves top scores on VBench (85.11), MVBench (62.0), GenEval (0.90) and GEdit‑Bench (7.30), and demonstrates emergent cross‑task generalization.
The article introduces Visual Para-Thinker, the first parallel reasoning framework tailored for large‑scale vision‑language models, explains its block and scan visual path divisions, details the Path‑aware Attention and Learnable Parallel Rotary Position Embedding mechanisms, and presents experimental results showing significant gains on visual perception benchmarks.
On April 23, Tencent released the open‑source Hy3 preview, a 295 B‑parameter hybrid expert model with 21 B active parameters and 256K context length, delivering substantial gains in complex reasoning, instruction following, code and agent tasks, achieving 40 % faster inference, lower costs, and strong benchmark results across Tencent’s AI products.
A recent Meta‑led study introduces HyperAgents, a framework that merges task agents with meta‑agents to enable metacognitive self‑modification, showing significant gains on coding benchmarks, paper review, robotics reward design, and Olympiad‑level math grading, while also highlighting emerging safety risks as AI systems begin to rewrite their own improvement mechanisms.
Anthropic unveiled Claude Sonnet 4.6 on February 18, touting Opus-level coding ability, a 1 million-token context window, and unchanged pricing; benchmarks show a SWE-bench score of 79.6% (up from 77.2%), OSWorld 72.5% (vs 61.4%), and GPQA Diamond 89.9%, while industry leaders praise its reduced laziness, stronger instruction following, and strategic long-term planning.
RoMa v2 introduces a two‑stage matching‑then‑refinement pipeline powered by DINOv3 features, custom CUDA kernels, and diverse training data, delivering state‑of‑the‑art accuracy, speed, and pixel‑level uncertainty estimation across a wide range of dense matching benchmarks.
This article reviews the MLLM4TS framework, which fuses visual representations of multivariate time series with large language models to address complex temporal dependencies, cross‑channel interactions, and task generalization, and demonstrates superior performance on classification, anomaly detection, forecasting, and few‑shot scenarios across multiple benchmarks.
The paper presents PENCIL, a novel alternating generation‑and‑erasure reasoning paradigm that achieves optimal space‑time complexity for chain‑of‑thought tasks, dramatically improves accuracy and efficiency on hard SAT, QBF, and Einstein puzzle benchmarks, and is provably Turing‑complete.
NOVA introduces a highly efficient autoregressive video generation framework that eliminates vector quantization, combines frame‑by‑frame causal prediction with set‑by‑set spatial attention, and achieves state‑of‑the‑art quality on VBench and GenEval while offering strong zero‑shot generalization across text‑to‑image and text‑to‑video tasks.
