AIAny
Icon for item

Intern-S2-Preview: Scientific Agentic Foundation Model

Supports multimodal scientific understanding, long-horizon agentic workflows and scientific tool interaction using a unified pipeline of multimodal pretraining, supervised fine-tuning and scalable multi-task reinforcement learning. Distinctive features include time-series modules for signal forecasting and a separate Memory Decoder that enables rapid domain specialization without changing the frozen 397B backbone.

Introduction

Scientific progress increasingly depends on systems that can reason across heterogeneous evidence, interact with domain tools, and sustain multi-step investigations. Intern-S2-Preview focuses on that gap by treating agentic scientific workflows as a first-class objective and by scaling a full-chain pipeline from multimodal pretraining to on-policy agentic RL.

Key Findings
  • Unified full-chain training for scientific agents — combines rendered scientific-document pretraining, interleaved image-text corpora, supervised fine-tuning, scalable multi-task RL, black-/white-box agentic RL and on-policy distillation; this means the model is trained end-to-end for both understanding and multi-step action under scientific constraints.

  • Architectural and task extensions for science signals — adds time-series modules that extend the model from long-sequence understanding to numeric forecasting, improving performance on scientific signal benchmarks; practical implication: better handling of measurement/temporal data in earth sciences and experiments.

  • Memory Decoder for rapid specialization — a separate Memory Decoder path enables domain-specific improvements (e.g., Biology-Instructions gain) without modifying the frozen 397B backbone, so teams can adapt the model for niche scientific tasks with a much smaller specialized module.

  • Practical RL and stability techniques — introduces partial rollout with off-policy correction, adaptive length regularization, online speculative decoding and trace-aware experience assembly to stabilize and scale agentic RL; this reduces rollout inefficiency and improves training robustness for long-horizon tasks.

Who it's for and trade-offs

Great fit if you need a research-grade foundation model tailored to scientific workflows: multimodal document understanding, long-horizon planning, tool-assisted experiments and time-series forecasting. It is useful for labs building autonomous or semi-autonomous scientific agents, or teams wanting a large frozen backbone plus a lightweight specialization path.

Look elsewhere if you require a small, easily deployable model for low-resource production (the paper centers on very large backbones such as 397B and the associated training pipeline), or if you need a turn-key, hosted API — this work emphasizes research architecture, training methods and benchmarks rather than lightweight deployment convenience.

Where it fits

Positions itself between general-purpose multimodal LLMs and domain-specific scientific tools by explicitly optimizing for agentic, long-horizon scientific tasks and by offering a route to specialize via Memory Decoder modules without re-training the whole backbone.

Information

  • Websitearxiv.org
  • AuthorsLei Bai, Jiaqi Cao, Chiyu Chen, Guanzhou Chen, Kai Chen, Guangran Cheng, Erfei Cui, Xuanlang Dai, Shengyuan Ding, Shangheng Du
  • Published date2026/08/13

Categories

More Items

Treats agent self-improvement as natural selection over a population of harnesses (prompts, tools, skills, control flow), evolving a frozen-model agent by selecting harness edits that extend capability without regressing others. Uses a preserve-and-extend contract, lineage archive, and verifier-driven fitness (no gold solutions) to recombine complementary edits and transfer across benchmarks.

Evaluates whether LLM-driven storyteller agents preserve long-horizon logical consistency under adversarial player interventions. Introduces NCP-Bench (100 movie-derived narrative environments) with structured trajectory/commitments and automatic violation checks; finds strong LLMs often contradict themselves across multi-turn interactions.

Uses a stronger 'builder' model at inference time to construct executable harnesses that boost weaker target models without parameter updates, mainly by turning unstable reasoning into deterministic code, routing, and strict answer-format enforcement.