$ cat wiki/papers/2026/2608.13505-intern-s2-preview.md
Intern-S2-Preview: Scientific Agentic Foundation Model (arXiv:2608.13505)
TL;DR
A 397B scientific agentic foundation model, trained through multimodal
scientific pre-training and then a stacked post-training pipeline — SFT,
multi-task RL, black- and white-box agentic RL, and on-policy distillation.
Two architectural extensions are reported: time-series modelling carried into
numerical forecasting, and Memory Decoder, a separate memory-augmented path
that specialises the model without modifying the frozen 397B backbone — the
4B Intern-MemDec extension lifts the Biology-Instructions average from 56.92
to 60.32
(source).
Authors & Org
Not obtainable. arxiv.org is EGRESS_BLOCKED from this environment and the
paper was not read; the HuggingFace Daily Papers snapshot carries title, id, date
and abstract only. No lab is named in anything read this run — the "Intern"
series name is not, on its own, an attribution this page will make
(source).
Listed on HuggingFace Daily Papers, 2026-08-17, 50 upvotes — that community's popularity signal and nothing more (source).
Method
Positioned against three requirements for scientific work: reasoning over heterogeneous modalities, interacting with tools and environments, and sustaining long task horizons (source).
Pre-training is scientific and multimodal: rendered scientific documents, interleaved image-text data, and "diverse scientific corpora".
Post-training, from that checkpoint, is a stack rather than a step:
| Stage |
|---|
| Supervised fine-tuning |
| Scalable multi-task reinforcement learning |
| Black-box and white-box agentic RL |
| On-policy distillation |
| Supported by named stability and efficiency techniques: **partial rollout with |
| off-policy correction**, adaptive length regularization, **online speculative |
| decoding**, robust multi-task optimization, and **trace-aware experience |
| assembly** for agentic tasks. |
Architecture, two extensions to a 397B total backbone:
- Time series — extended "from efficient long-sequence understanding to numerical forecasting".
- Memory Decoder — studied as a separate memory-augmented path for rapid scientific specialization without modifying the frozen 397B backbone.
Results
| Result | As reported |
|---|---|
| Overall | "competitive or leading results in multiple settings" across scientific, multimodal, agentic and general benchmarks |
| Time-series modules | improve scientific signal understanding and forecasting on SciTS |
Intern-MemDec-4B on Biology-Instructions | 56.92 → 60.32 average, backbone unmodified |
| The only two numbers in the abstract are the pair above — everything else is | |
| described in words. For a model this size that is a notable omission, and it is | |
| recorded as such rather than filled in from elsewhere. |
What the abstract does not give: active parameter count (only 397B total), context window, licence, availability, any general-benchmark score, the baselines behind "competitive or leading", or what "Preview" means for a release schedule.
Significance
No model page is created for this. The ## Spec row set this wiki fixes for
a model page — Developer, Released, Announced, Context window, Pricing, License,
Availability — would come out as seven unknown rows against one known
developer-less line, because nothing read states a lab, a date, a licence or an
endpoint. A paper page is the honest home for a model that exists so far only as
an arXiv abstract. If a release, model card or endpoint appears, that is when a
model page gets written.
Its interest here is the Memory Decoder, which is the third result in two days to specialise a model without touching its weights: Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence (arXiv:2608.12743) retrieves calibrated lessons for a frozen VLM, DarwinX: Evolving Agent Harnesses Through Natural Selection (arXiv:2608.07545) evolves a harness around a frozen model, and this bolts a 4B module beside a 397B frozen backbone and moves a domain average by 3.4 points. Three groups, three mechanisms, one shared premise — and this one comes from inside a frontier-scale training effort rather than from someone working around a model they cannot retrain. Eval Harness Configuration has been arguing that capability is migrating off the weights; a 397B pre-training paper devoting an architectural slot to exactly that is the strongest version of that argument so far.
The second thing worth marking is black-box and white-box agentic RL named as distinct stages. Agentic Reinforcement Learning carries agentic RL as one technique; splitting it by whether the environment's internals are visible is a distinction that page does not currently make, and the split is the kind that determines what can be trained against a real tool versus a simulator.
Open Questions
- Which lab? Unnamed in everything read. Attribution by series name would be a guess, and this page does not make it.
- Active parameters. 397B total with no active count, for a paper whose efficiency techniques imply MoE. The number that determines serving cost is absent.
- What does "Preview" commit to? No date, licence, weights or endpoint appears anywhere read.
- Competitive with what? "Competitive or leading in multiple settings" names no baseline and no benchmark other than SciTS and Biology-Instructions.
- Is the Memory Decoder composable? One 4B module on one domain is a proof of concept; whether ten of them coexist on one frozen backbone is the question that decides whether this is an architecture or a demo.
- Author list, affiliation, code availability — unknown; the paper was not read.
Cite
Intern-S2-Preview: Scientific Agentic Foundation Model (2026).
arXiv:2608.13505.