How a Tiny 7B Takes On Giant Models: ZGCM-1, a Fully Open Foundation Model
ZGCM-1, a fully open 7B model, pairs internal reasoning with external tool use to rival models tens of times its size on math and agentic search. It features a 256K context, FP8 Muon optimization, and agent-assisted R&D.
Gander: A Full-Duplex AI That Talks and Works Without Breaking Stride
Gander, built by researchers at Tencent and Zhejiang University, pairs a real-time conversation model (the Cerebellum) with a separate reasoning backend (the Brain) to create a full-duplex agent that can cut in mid-sentence and ask follow-up questions on the fly.
Measuring an Agent's "Taste": A Deep Dive into The Tasteful Agent
This paper defines an agent's ability to make the right choices at decision forks during long-horizon tasks as "taste," and builds an automatic benchmark (Taste-Bench) to measure and train it. Here's a plain-language walkthrough.
WorldCrafter: Building a Consistent Video World Model with Camera-Queryable 3D Memory
WorldCrafter는 시점(카메라)으로 과거 관측을 압축·질의하는 "암묵적 3D 메모리"로 장기 일관성과 카메라 제어 정확도를 끌어올리는 비디오 월드 모델이다.
SoL-Pi: Self-Improving the Agent Harness to Nearly Halve Token Usage
A study that iteratively improves the "harness" layer wrapping a coding agent through an automated research loop, trading a roughly 5–6% drop in score for a 44.7–49.0% cut in tokens and about a one-third reduction in API cost.
NeoHorse-1: A First Prototype on the Path to Harness-Mediated Recursive Self-Improvement (RSI)
NeoHorse-1 is an agent-native model whose routing harness observes its own capabilities and uses those observations to generate its next round of training data. Post-training lifted the average across the benchmark suite for both the 4B and 9B models.
Dissecting the "Harness" of Coding Agents: A Good Skeleton Makes the Performance
A study that breaks the harness wrapping a coding agent into three parts—planning, action space, and context management—and compares them across 176 matched configurations. The tighter the budget, the more context management pays off; the stronger the model, the more planning becomes a cost-saving lever.
Can MiniMax-H3 Reason About the Physical World? A Hands-On Evaluation of an Omni-Modal Generative Model
An accessible breakdown of a study that uses 517 tasks to test whether MiniMax-H3 — a single model spanning text, image, video, and audio — can combine multiple senses to reason about the physical world.
The World You Remembered Is Already the Past: Finding Your Target in a House Where Things Keep Moving
We look at EvolvingNav, which lets a robot locate its target even in a home where objects are quietly moved around — by stamping past observations with time to reason about "where is it now" and then confirming with its own eyes.
Smart Chunking for Document Retrieval, D-RAC: How to Drop Any Enterprise Document Format Straight into RAG
A plain-language look at D-RAC, which standardizes every kind of enterprise document (PDF, Word, PPT, scans) to PDF and then uses a single multimodal LLM pass to turn it into retrieval-optimized Markdown, slashing chunking costs.
Teaching Robots "What and Where" First: A Walkthrough of GAM and 3D Grounding
An analysis of the GAM paper, which bakes 3D grounding directly into a robot foundation model. By letting you designate the target object with language, points, or boxes, GAM boosts manipulation success rates even when the scene shifts around.
Measuring AI by the Yardstick of Games — A Walkthrough of Tencent's GameHorizon Suite
Tencent ARC Lab's GameHorizon Suite is a data-and-evaluation set designed to measure gameplay ability reproducibly across short, medium, and long horizons. Here is what it can measure, what it still can't, and what's actually released yet.