MiniMax-M2: A New Model Series from China

Chinese AI company MiniMax has announced a new model series, MiniMax-M2. As the title "Mini Activations Unleashing Max Real-World Intelligence" suggests, it aims to deliver strong performance on real-world tasks with a small number of activation parameters. Detailed benchmark results await the release of the paper, but the M series appears positioned to further advance MiniMax's existing model lineup.

MiniMax has long offered models with strengths in long-context processing and has attracted attention as one of the notable Chinese LLM providers. We look forward to future detailed reports on what architecture the M2 series adopts.

Soap2Soap: Transforming Videos into Pixar and Anime Styles

Soap2Soap is a technique that uses multi-agent collaboration to transform live-action videos into a variety of styles, including Pixar, Disney, LEGO, anime, and claymation. Rather than simply applying filters, its key feature is maintaining consistent characters, environments, and cinematic composition across shots.

In Video-to-Video generation, maintaining character consistency across long-form footage has been a major challenge. Soap2Soap addresses this by having multiple agents divide the work among themselves, and it has the potential to fundamentally change video production workflows.

RT-Lynx: Leveraging "Activation Sparsity" to Accelerate Diffusion Models (ICML 2026)

RT-Lynx, scheduled to be presented at ICML 2026, proposes an interesting approach to accelerating diffusion models (Diffusion Transformers). Model acceleration has typically relied on "weight sparsity," but RT-Lynx instead focuses on "activation sparsity."

The authors' finding is simple but important: activations in DiTs (Diffusion Transformers) are inherently sparse and far more tolerant of 2:4 semi-structured sparsity than weights are. RT-Lynx exploits this property by combining online activation sparsification, norm-based correction, and a lightweight LoRA branch, achieving roughly 1.2x end-to-end speedup and up to 1.55x on linear layers while preserving quality.

It has been validated on Qwen-Image, FLUX.1-dev, and Z-Image, suggesting practical real-world impact.

AKBE: Cutting Tool Usage in Agentic RL by 18%

AKBE (Agentic Knowledge Boundary Enhancement) is a method that improves tool-use efficiency in LLM agents. In conventional agentic reinforcement learning, agents wasted calls invoking tools even in situations where they weren't needed.

During training, AKBE runs two paths — with and without tools — simultaneously, dynamically probing the model's own knowledge boundary. This enables the model to determine, for each question, whether tools are needed and what the minimum number of tool calls is. In experiments across seven QA benchmarks, it improved accuracy by an average of +1.85 while reducing tool calls by 18%, and tool productivity improved by 25%.

In production agent deployments, reducing API call costs and latency is a major challenge, which gives this approach real practical significance.

OpenAI Takes On an 80-Year-Old Unsolved Math Problem

AI is now seriously tackling difficult problems in pure mathematics. The "unit distance problem," posed by mathematician Nelson in 1946 — how many colors are needed at minimum to color all points on a plane so that no two points at distance 1 share the same color — is a conjecture that went unsolved for 80 years.

OpenAI's work is said to have shed new light on this problem. How far AI can go in the field of mathematical reasoning is one of the most closely watched topics in the research community today. Not as mere acceleration of computation, but as a test of emergent reasoning ability, future developments will be worth watching closely.

Trajel: Auditing Multi-Agent Hallucinations at the Trajectory Level

When running LLMs as autonomous agents, evaluating only the final output means missing hallucinations that occur in intermediate steps. Trajel proposes a framework for trajectory-level hallucination auditing in multi-agent industrial workflows.

It introduces five hallucination categories (factual, referential, logical, procedural, and scope-based) and builds a dataset annotated by experts. A VLM is fine-tuned to perform detection via Chain-of-Thought reasoning. Among the key findings: roughly half of hallucination trajectories contain multiple types simultaneously, and existing benchmarks overlook the most common failure modes.

Trajectory-level auditing is essential for ensuring agent safety, making this an important step toward real-world deployment.

---