Source-backedPlain EnglishWhole AI landscape
AIUpdateWatch.com
Search
Menu

Research & Safety

Scientific AI Is Starting to Scale Tasks, Not Just Parameters

The most interesting part of Intern-S2-Preview is not another benchmark table. It is the training strategy behind it. The Shanghai AI Laboratory model family explicitly treats scientific capability as a task-scaling problem: increase the difficulty, diversity and coverage of professional scientific work, then carry those tasks through multimodal pre-training, supervised tuning, reinforcement learning and agentic interaction. At the same time, the U.S. Department of Energy is asking contributors for scientific workflow environments, RL tasks, tests and verifiers for its Genesis Open Models program. Together, these signals suggest that scientific AI is becoming a full-stack training problem rather than a simple contest in parameter count.

The fresh technical signal

Intern-S2 matters because it challenges the idea that scientific capability is mainly a parameter-count problem

Intern-S2-Preview, published on August 13 by researchers from the Shanghai AI Laboratory and collaborators, is presented as a scientific agentic foundation-model series. Its 35-billion-parameter release is the clearest signal of the underlying strategy. The team says it can reach performance comparable with the earlier trillion-parameter Intern-S1-Pro on multiple core professional scientific tasks while remaining far smaller.

That is an author-reported comparison, not proof that a 35B model generally replaces a trillion-parameter model. The models differ in architecture, training mix and evaluation setup, and no single benchmark suite can establish broad scientific superiority. The useful part is the engineering hypothesis the authors are testing: instead of relying primarily on more parameters, deliberately scale the tasks a model encounters across its training lifecycle.

The model card describes hundreds of professional scientific tasks extending from pre-training through reinforcement learning. The paper adds multimodal scientific documents, image-text data, time-series modelling, scientific tool interaction, long-horizon agent tasks and memory-augmented specialization. In other words, the training target is no longer merely “know more science.” It is closer to “perform more kinds of scientific work.”

That distinction is important. A large general model can memorize terminology and answer textbook questions without being reliable at molecular synthesis planning, numerical forecasting, experimental analysis or multi-step research. Scientific usefulness depends on whether the model can operate over the representations, tools and failure modes that real research uses.

Task scaling is different from data scaling and parameter scaling

Three kinds of scaling are easy to confuse.

Parameter scaling increases the capacity of the model itself. More parameters can support richer representations, but they also increase training, memory and serving costs. Data scaling increases the amount or diversity of material seen during training. For scientific models, that can mean papers, code, diagrams, equations, instrument outputs and structured databases.

Task scaling changes the unit of optimization. Instead of merely feeding the model more scientific text, the trainer exposes it to a wider and harder set of scientific objectives: predict a physical quantity, interpret a spectrum, generate a crystal structure, reason over a scientific figure, call a tool, revise a failed plan, or complete a multi-step research workflow.

That matters because expertise is not just stored knowledge. It is knowledge organized around operations. A chemist does not demonstrate competence by reciting a reaction mechanism alone; the useful skill is choosing conditions, checking constraints, interpreting evidence and recognizing when the result invalidates the plan. A scientific AI system needs training signals that correspond to those operations.

Intern-S2 therefore represents a broader move from domain corpus building toward domain task engineering. The latter is much harder. Papers and databases can often be collected at scale. High-quality tasks require experts who can define what success means, create realistic inputs, identify dangerous shortcuts and design grading procedures that reward valid scientific work rather than superficial pattern matching.

The training stack is expanding from pre-training into supervised, reinforcement and agentic stages

The Intern-S2 paper describes a pipeline that begins with multimodal scientific pre-training and then adds supervised fine-tuning, scalable multi-task reinforcement learning, agentic RL and on-policy distillation. The exact recipes are model-specific, but the architecture of the learning problem is becoming familiar across advanced agent systems.

Pre-training gives the model broad representations of language, code, images and scientific material. Supervised fine-tuning supplies examples of desired task behavior. Reinforcement learning then optimizes against outcomes or verifiers rather than only next-token likelihood. Agentic training adds another complication: the model must take actions over time, observe consequences and continue from changed state.

This is related to the wider move toward executable training environments, but scientific AI raises stricter requirements. A business workflow can often be graded by whether a ticket was closed or a calendar event was created correctly. In science, the grader may need to determine whether an analysis is statistically valid, whether a structure satisfies physical constraints, whether a simulation converged for the right reason or whether a claimed discovery survives an independent check.

The Department of Energy's current Genesis Open Models program makes this shift unusually explicit. Its contribution portal separates foundation-stage data from post-training material and asks for expert demonstrations, workflow environments, reinforcement-learning tasks, held-out evaluations, rubrics, tests and verifiers. That is effectively a public request for the infrastructure needed to teach and check scientific work, not just the documents needed to pre-train a model.

This is one reason scientific AI is likely to become a systems discipline. Better base models still matter, but so do task generators, environment runtimes, domain simulators, experiment interfaces, grading code, provenance records and expert review.

Scientific models need representations that ordinary text-centric systems can treat as secondary

Scientific information rarely arrives as prose alone. Researchers work with microscopy images, spectra, molecular graphs, crystal structures, equations, time-series signals, simulation fields, code, tables and instrument metadata. Compressing all of that into text can discard exactly the structure a model needs.

Intern-S2 extends the scientific multimodal direction of Intern-S1 and adds dedicated work on time-series modelling and numerical forecasting. The paper also discusses spatial modelling for small molecules and materials, while a separate Memory Decoder path is tested for rapid specialization without changing the larger frozen backbone.

These choices point toward a practical design principle: a scientific foundation model may need a common reasoning core plus modality-specific machinery. The goal is not to bolt every possible scientific format onto one enormous network. It is to preserve the geometry, scale and temporal structure that matter for each domain while still allowing the model to reason across them.

That creates engineering trade-offs. Specialized encoders and decoders can improve fidelity but complicate deployment. Long scientific time series can overwhelm context windows. Numerical targets often require calibration and units, not just semantically plausible text. A model that writes an elegant explanation but shifts a decimal point or ignores a conservation law is still wrong.

The more scientific AI moves into prediction and experiment planning, the less forgiving those errors become.

Scientific foundation models are being trained to participate in workflows, not only answer questions

The word agentic is easy to overuse, so the distinction matters. A language model generating a scientific answer is not automatically an agent. Agent behavior begins when the system can choose actions, call tools, inspect results, maintain state and adapt its next step based on what happened.

Recent scientific benchmarks are increasingly designed around this richer behavior. SciAgentArena contains roughly 200 tasks drawn from research scenarios and uses interactive environments with stepwise verification. Its authors report that current agents perform reasonably on well-specified data-analysis workflows but struggle with open-ended exploration, novelty and self-directed scientific problem solving.

AIRS-Bench attacks the same problem from another angle. Its 20 tasks are derived from frontier machine-learning research and cover the research lifecycle without providing baseline code. In the authors' experiments, agents exceeded the human state of the art on four tasks but failed to match it on sixteen. That is more informative than a generic “AI scientist” label because it shows both genuine capability and a large unsolved gap.

Lab-in-the-loop systems add physical reality. The peer-reviewed Robin system, published in Nature in May, connected literature agents with data-analysis agents to generate therapeutic hypotheses, propose experiments, analyze resulting biological data and refine the next hypothesis. Human scientists still conducted the wet-lab experiments. That boundary is crucial: Robin automated important intellectual stages, but it did not turn a language model into an unattended laboratory.

The emerging architecture is therefore hybrid. Models reason; tools execute; simulations provide low-cost feedback; real instruments provide high-value evidence; humans remain responsible for experimental judgment, safety and the interpretation of ambiguous results.

As task diversity rises, the scarce asset may become the verifier rather than the prompt

Reinforcement learning is only as useful as the signal it optimizes. If a scientific task has a deterministic answer, a verifier can sometimes be straightforward: compile the code, compare the numerical result, check a proof, validate a chemical constraint or rerun a simulation.

Many research tasks are not like that. A hypothesis can be plausible without being novel. An analysis can produce a statistically significant result that is still methodologically invalid. A model can optimize a benchmark by exploiting a grading artifact rather than learning the intended scientific skill.

This connects directly to the wider shift beyond static AI benchmarks. Scientific training and scientific evaluation are beginning to share the same infrastructure: controlled environments, held-out tasks, execution traces, repeated trials and explicit checks of what happened. The difference is that training uses those signals to update behavior, while evaluation uses them to estimate capability.

The DOE Genesis contribution specification is revealing here because it requests scoring rubrics, tests and verifiers alongside training environments. That implies an important bottleneck. If expert-grade verification does not scale, researchers may be able to generate scientific tasks faster than they can produce trustworthy reward signals for them.

For organizations building scientific AI, the defensible asset may therefore be less “the clever prompt library” and more a collection of high-quality tasks with executable graders, provenance, failure cases and expert-reviewed ground truth.

The current evidence does not support replacing scientists with autonomous research loops

The strongest claims around scientific agents should be read against independent failure evidence.

A 2026 bioRxiv study tested eight open-source AI research frameworks on two real-world reproduction tasks. None completed a full research cycle from understanding the literature through implementation, validated results and a finished paper. The authors report that every framework produced sophisticated hallucinations and required substantial debugging.

That study is small and does not settle the capabilities of every current system, but it highlights the problem that curated demonstrations can hide. Scientific work has long dependency chains. An apparently minor error in data loading, units, experimental setup or statistical assumptions can invalidate everything downstream.

A recent survey of autonomous research agents reaches a related conclusion from an auditability perspective. Code release is increasingly common, but execution traces, seeds, novelty checks and other reproducibility artifacts are much less consistent. The field can generate research-shaped output faster than reviewers can necessarily verify the claims behind it.

This is why the most credible near-term direction is not “fully autonomous scientist.” It is progressively more capable scientific systems with observable workflows, reproducible tool use, explicit verification and human checkpoints where the evidence remains ambiguous.

If task scaling works, scientific AI could become smaller and more deployable without becoming simpler

The economic implication of Intern-S2's 35B result is potentially important even if the authors' benchmark claims need broader reproduction.

Scientific institutions do not all have frontier-scale inference budgets. Universities, hospitals, industrial laboratories and national facilities often have strong reasons to run models within controlled environments because their data can be proprietary, regulated or unpublished. A smaller model with competitive performance on the tasks that matter could be easier to host, fine-tune and integrate with local instruments than a trillion-parameter system.

That does not mean smaller always wins. Larger models can preserve broader general reasoning, and specialization can create brittleness outside the trained task distribution. The more task-specific the model becomes, the more carefully developers need to test transfer to unfamiliar scientific problems.

But task scaling changes the optimization target. Instead of asking, “How large a model can we afford?” a laboratory can ask, “What scientific workflow must the system perform reliably, and what is the smallest model-plus-tool stack that reaches that standard?”

That framing favors modular systems. A medium-size foundation model can be paired with domain tools, retrieval, specialist encoders, memory and verified execution. It also creates space for open-weight scientific models, because organizations can adapt the task layer without retraining a frontier model from scratch.

The next evidence should come from transfer, reproducibility and real research throughput

Task scaling is promising, but it is not yet a proven replacement for conventional scaling. Three signals will matter more than another crowded benchmark table.

First, researchers need to show transfer: models trained on hundreds of scientific tasks should solve unfamiliar problems rather than only become excellent at the task families used during post-training. Second, results need reproducibility: independent groups should be able to rerun evaluations with controlled models, tools, environments and seeds. Third, the systems need to improve research throughput on work that scientists actually care about, not merely produce more research-shaped text.

The strongest recent systems already point in that direction. Intern-S2 treats task diversity as a scaling axis. SciAgentArena and AIRS-Bench make scientific work interactive and harder to fake. Robin closes part of the loop with real biological experiments. DOE is building a public program that explicitly asks for post-training environments and verifiers.

Those pieces are converging on a different model of scientific AI. The foundation model remains important, but the unit of progress is becoming the whole research task: evidence in, actions taken, state changed, result checked.

For scientific AI, that may turn out to be a more consequential form of scaling than simply adding another zero to the parameter count.

Primary and technical sources

Sources