Tiny AI Models Prove Hard Theorems, Video Mashups Get Smarter

🧠 AI for Mathematical Reasoning

Tiny AI Tackles Olympiad Math

The Rundown

Researchers have developed QED-Nano, a compact 4-billion parameter model that demonstrates impressive capabilities in generating proofs for Olympiad-level mathematics. This development challenges the notion that only massive proprietary models can handle complex reasoning tasks. QED-Nano's training pipeline involves three key stages: supervised fine-tuning using distilled knowledge from DeepSeek-Math-V2 to imbue effective proof-writing styles, followed by reinforcement learning (RL) with rubric-based rewards. The final stage expands RL by incorporating a reasoning cache, which breaks down long proofs into iterative summarize-and-refine cycles. This approach allows for stronger test-time reasoning. QED-Nano surpasses larger open models like Nomos-1 and GPT-OSS-120B in proof generation performance. It even approaches the capabilities of proprietary models such as Gemini 3 Pro, all while operating at a significantly lower inference cost. The team is releasing the full QED-Nano pipeline, including models and code, to foster further research in open mathematical reasoning.

The details

• QED-Nano, a 4B parameter model, achieves competitive reasoning performance on Olympiad-level math problems.
• Its training pipeline includes supervised fine-tuning, reinforcement learning with rubric-based rewards, and a reasoning cache.
• The model surpasses proof-generation performance of larger open models like Nomos-1 and GPT-OSS-120B.
• QED-Nano approaches the performance of proprietary models like Gemini 3 Pro at a fraction of the inference cost.

Why it matters

This notable advance democratizes advanced AI reasoning capabilities. Startups can now leverage smaller, more accessible models for complex tasks like theorem proving or scientific discovery, reducing reliance on expensive proprietary systems and accelerating R&D cycles.

🎬 Generative Video

DIRECT: Hierarchical Video Mashups

The Rundown

Creating compelling video mashups requires intricate orchestration across semantic, visual, and auditory elements. Existing automated editing frameworks often struggle with this cross-level multimodal coordination, leading to disjointed sequences. To address this, researchers introduced DIRECT, a hierarchical multi-agent framework that treats video mashup creation as a Multimodal Coherency Satisfaction Problem. DIRECT simulates a professional production pipeline, decomposing the task into three cascade levels: the Screenwriter anchors the global structure, the Director instantiates adaptive editing intent, and the Editor performs fine-grained shot sequence editing. This approach ensures professional-grade fluidity with improved visual continuity and auditory alignment. The framework is validated on Mashup-Bench, a new benchmark with tailored metrics. Experiments show DIRECT significantly outperforms current best baselines in both objective metrics and human subjective evaluations, offering a more coherent and engaging approach to automated video recomposition.

The details

• DIRECT uses a hierarchical multi-agent framework for video mashup creation, simulating a professional production pipeline.
• The framework decomposes the task into Screenwriter, Director, and Editor agents for structured editing.
• It addresses the Multimodal Coherency Satisfaction Problem for seamless visual and auditory integration.
• DIRECT significantly outperforms baselines on the new Mashup-Bench benchmark for visual continuity and auditory alignment.

Why it matters

This framework offers a significant upgrade for content creation tools. Startups in media and marketing can leverage DIRECT to automate complex video editing tasks, enabling faster production of high-quality promotional content and social media assets with greater creative control.

🤖 Agents

MemMachine: Ground-Truth Memory for Agents

The Rundown

Large Language Model (LLM) agents need persistent memory for personalization and long-horizon reasoning, but standard context windows and retrieval-augmented generation (RAG) degrade over time. MemMachine introduces an open-source memory system designed to preserve ground truth across multi-session interactions. It integrates short-term, long-term episodic, and profile memory within a unique architecture that stores entire conversational episodes, reducing lossy LLM-based extraction. MemMachine employs contextualized retrieval that expands nucleus matches with surrounding dialogue turns, improving recall accuracy. Benchmarks show strong accuracy-efficiency tradeoffs: on LoCoMo, it achieves 0.9169 accuracy using gpt4.1-mini. On LongMemEvalS, optimizations like retrieval depth tuning and context formatting yield significant gains. Compared to Mem0, MemMachine uses approximately 80 percent fewer input tokens under matched conditions. A companion Retrieval Agent adaptively routes queries, achieving high accuracy on benchmarks like HotpotQA-hard.

The details

• MemMachine is an open-source memory system for LLM agents that preserves ground truth across sessions.
• It integrates short-term, long-term episodic, and profile memory, storing entire conversational episodes.
• Contextualized retrieval improves recall by expanding matches with surrounding dialogue turns.
• On LongMemEvalS, retrieval-stage optimizations improved accuracy by up to 4.2%.
• MemMachine uses roughly 80% fewer input tokens than Mem0 under matched conditions.

Why it matters

Robust, persistent memory is crucial for building truly personalized and reliable AI agents. Startups developing AI assistants or customer service bots can use MemMachine to create more engaging and context-aware experiences, leading to higher user retention and satisfaction.