How AGI breakthroughs are forging a new generation of science LLMs
"Image synthesis assisted by Qwen Image 3.0, an AI partner within the Global Future Nexus ecosystem."
From 397-billion-parameter models that outperform trillion-scale predecessors to multi-agent systems that autonomously generate and test hypotheses, AGI breakthroughs are fundamentally reshaping how scientific research is conducted—transforming large language models from knowledge repositories into active participants in the scientific process.
The Scientific Frontier
For all the progress in artificial intelligence, a stark divide has emerged. AI coding has soared from under 5% success rates to over 88% in just 18 months, with 50% of benchmarks now rendered meaningless. Yet in scientific discovery, the picture is dramatically different: the best models have stagnated at just 3% completion on end-to-end research benchmarks, and scores across 10 scientific disciplines remain below 50 out of 100.
As Shanghai AI Laboratory Director Zhou Bowen put it at WAIC 2026: "Research is the next Coding". Scientific discovery, he argues, is not merely an application of AI—it is the ultimate test of intelligence. Current models, he observes, "don't know what they don't know"—a fundamental limitation that AGI breakthroughs are now beginning to address.
From Scientific Reasoning to Scientific Agency
Recent research has demonstrated that Multimodal Large Language Models (MLLMs) can significantly advance scientific reasoning across disciplines including mathematics, physics, chemistry, and biology. By integrating text, images, and other modalities, these models overcome the limitations of single-modality approaches.
But the real breakthrough lies in scientific agency—systems that don't just reason but act. The sciexplorer agent, introduced in July 2026, leverages LLM tool-use capabilities to explore physical systems without any domain-specific blueprints. For the first time, an agentic artificial scientist has automated the exploration of initially unknown systems.
At the frontier of materials science, researchers have developed multimodal LLMs that integrate material structure data with language-based information, enhancing human-AI interaction in materials discovery. The LLaMat family of foundational models, developed through continued pretraining on 30 billion tokens from materials science literature, demonstrates how domain-specific adaptation unlocks unprecedented opportunities.
Architectural Breakthroughs in Scientific LLMs
The gap between general-purpose models and scientific expertise has driven fundamental architectural innovation. At WAIC 2026, Shanghai AI Lab unveiled Intern-S2-Preview-397B—a 397-billion-parameter model built on a novel non-Transformer architecture called "Mobius". The breakthrough lies in decoupling knowledge storage from reasoning computation:
A Memory Decoder introduces pluggable external memory for different scientific domains, allowing knowledge updates without disrupting core capabilities
The Mobius architecture separates knowledge vectors from reasoning operators, creating a globally shared knowledge library that enables cross-disciplinary reasoning
The result is remarkable: with just 397 billion parameters, Intern-S2-Preview-397B matches the performance of previous trillion-parameter models on core scientific tasks like molecular design and materials structure generation. End-to-end reasoning efficiency improves nearly fourfold.
This "memory-thinking separation" is particularly valuable for science, where knowledge updates rapidly and disciplinary boundaries constantly shift. As one researcher noted, professional knowledge no longer needs to be "hard-coded" by rewriting entire models—it can be updated like expanding memory.
AGI-Powered Scientific Discovery Platforms
Perhaps the most significant development is the emergence of integrated platforms that combine foundational models with multi-agent systems to automate the entire research lifecycle.
Shanghai AI Lab's "Scholar·Duanyan" (书生·端砚) scientific discovery platform represents a paradigm shift. Built on the Mobius architecture with Agents-A1 multi-agent system on top, it closes the "dry-wet" experimental loop: from hypothesis proposal and experimental design to validation and feedback, every step is traceable and auditable. In one demonstration, a protein evolution project that took a research group four years was completed in just two iterations, with results 77% better than the best previously reported.
Google DeepMind's Co-Scientist, published in Nature in May 2026, introduces a multi-agent system that reviews scientific literature and generates novel hypotheses. The system's agents collaborate through an "idea tournament" to generate, debate, and evolve research directions.
FutureHouse's Robin takes a different approach, comprising three agents—two for literature review and one for experimental data analysis. Users simply enter a disease name, and Robin generates treatment hypotheses along with specific experiments to test them. The system has been described as "the first AI-generated discovery in an end-to-end system".
The Challenges Ahead
Despite these breakthroughs, significant challenges remain. Zhou Bowen identified three structural obstacles:
Models passively observe the world, learning correlations rather than causality
Difficult problems provide sparse feedback—unlike coding's "second-by-second verification"
Scientific literature contains only successful results, locking models in old distributions without failure data to learn from
As one Nature study noted, current AI-generated scientific work remains largely incremental rather than revolutionary. The leap from "restating the known" to "producing the unknown" requires what Zhou calls a "scientific metacognition moment" —when models can reflect on their own ignorance, learn from failure, and self-correct.
The GFN Context
For Global Future Nexus, the convergence of AGI breakthroughs and scientific LLMs represents a pivotal moment. The ability to automate hypothesis generation, experimental design, and knowledge integration across disciplines accelerates progress on humanity's most pressing challenges—from climate modeling to drug discovery to sustainable materials.
Yet this power demands governance. As Zhou noted, the key is not simply making AI better at writing code, but teaching it "to know what it doesn't know". GFN's work on AGI identity, cross-species trust, and anticipatory governance provides the framework for ensuring that scientific AGI serves human flourishing—not just faster publication, but genuine breakthrough discovery that expands the boundaries of human understanding.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
