How can a science-aware loop build an efficient scientific agentic model?

The loop turns failures into training tasks

The paper's central question is not how to collect more scientific text, but how to select and organize scientific resources around the capabilities a model still lacks. SAIL answers this by looking at the model's behavior on training and development-validation tasks, then letting agents built on frontier AI models diagnose the underlying gap and build new training material for it.

The diagnosis is deliberately specific. Two unsuccessful literature searches can call for different interventions: one may need broader exploration, while another needs better selection among papers already retrieved. In scientific coding, a program may run successfully even though the chosen method rests on an invalid assumption. In a longer investigation, the model may execute each step correctly yet fail to use an intermediate finding to revise its plan. Each failure type points to a different kind of task, and the loop repeats across multiple development cycles as the updated model's behavior revises the training priorities.

What the frontier agents actually build

The construction agents draw on paper collections and scientific code repositories to produce three kinds of material. Focused problems exercise individual reasoning skills; demonstrations show how observations should change subsequent actions; executable tasks let the student practice through its own interaction. The same resource collection can therefore support different training objectives as the model's weaknesses change.

For interactive tasks, the agents also assemble the tools and execution state alongside the questions. Search tasks expose intermediate retrieval results so the model can revise its queries and selection. Coding tasks provide execution tools and scientific libraries so the model can inspect computational outputs. Longer investigations use persistent workspaces, allowing later actions to build on earlier artifacts and findings. Generated material is validated against its sources and task-specific checks, including execution where applicable, covering scientific content, tool availability, and the supervision used for training.

Four training stages from SFT to agentic RL

SAIL is trained in four stages. Supervised fine-tuning on a mixture of general capability data and the loop-constructed scientific data establishes the capabilities and interaction formats needed later; for interaction-oriented tasks, demonstrations include tool calls, environment observations, and subsequent model actions. This checkpoint then initializes both the specialist models and the student policy.

Specialists are trained independently on focused capability distributions, using additional supervised fine-tuning, reinforcement learning, or both, so that data mixtures and optimization settings can be adapted to each capability. Multi-Teacher On-Policy Distillation (MOPD) then transfers specialist supervision to the generalist student on the student's own trajectories: the student generates responses and interaction trajectories, and the routed specialist scores every student-generated token under the same context the student observed, including tool observations and compacted history. A mask excludes task inputs and environment observations so that only model-generated tokens contribute.

Finally, agentic reinforcement learning trains multi-step execution using rewards derived from task-specific verifiers, completion criteria, or other available feedback signals. Training spans multiple task families and exposes the model to states resulting from its own actions, including tool failures and unsuccessful intermediate attempts. The final SAIL checkpoint is obtained after this stage.

Why teacher routing is separate from execution

Agentic RL and MOPD share the same rollout and collection pipeline: fresh trajectories are generated by the current student policy through a common rollout interface, and the two methods differ only in the supervision attached to the student-generated actions after collection. For RL, verifiers or reward functions evaluate the episode outcome; for MOPD, task metadata routes each trajectory to a specialist teacher.

Harness routing and teacher routing are deliberately decoupled. The harness determines how a task is executed and how the student interacts with its environment, while the teacher determines which specialist provides supervision for the resulting trajectory. A task can therefore use different execution harnesses under the same supervision rule, and specialist teachers can be reconfigured without modifying the interaction stack. The infrastructure also separates model optimization, rollout serving, and environment execution, so rollout and optimization can use different parallelism and memory configurations and sandbox capacity can scale separately from model-serving capacity.

Where SAIL leads and where it does not

The paper reports SAIL's strongest relative results on SciCode and ArxivDIGESTables. SAIL scores 50.35 on SciCode, above the 47.57 of the 744B-total GLM-5.2 and the 43.10 of the comparable-scale Apodex-1.0-mini. On ArxivDIGESTables it scores 35.24, above DeepSeek-V4-Flash-0731's 35.13 and GLM-5.2's 34.21. These are the two benchmarks the discussion names as SAIL's strongest relative results.

The picture is more mixed elsewhere. On full-text question answering, SAIL scores 91.67 but has a weaker relative ranking, a 3.57-point gap to the leader. On repository execution and long research workflows, the strongest larger models retain a lead, and DiscoveryBench shows a smaller improvement over the base model. The paper's own reading is that evidence interpretation, repository execution, and longer investigations remain priorities for further development.

Within the comparable 35B-total, 3B-active group, SAIL's margins vary by task. It reports 33.25 on LitQA search against 28.36 for BigBang-v1, 85.33 on ScholarQA CS2 against 76.00 for Apodex-1.0-mini, and 86.51 on LitQA2 against 74.32 for Apodex-1.0-mini. On E2E-Bench Hard it reports 89.46 against 83.83 for Nex-N2.5-mini, while on E2E-Bench Basic its 77.27 sits below Nex-N2.5-mini's 70.02 only in the sense of being higher; the benchmark-level ordering differs by task family, which is why the paper does not claim uniform superiority.

What the 3B active-parameter comparison means

SAIL has 35B total parameters and 3B active parameters, the same total/active configuration as the comparable-scale group that includes BigBang-v1, Apodex-1.0-mini, Nex-N2-mini, Nex-N2.5-mini, Agents-A1, and Qwen3.6-35B-A3B. The larger-scale group ranges from 124B to 1600B total parameters. The paper's claim of competitive performance with substantially fewer parameters is therefore a comparison against that larger group, not a claim of matching every benchmark leader.

The distinction matters for readers comparing models. SAIL's SciCode and ArxivDIGESTables scores exceed every listed model in both groups, but its full-text QA score of 91.67 trails the leader by 3.57 points, and larger models retain the lead on repository execution and long research workflows. A reader choosing a model should therefore read the benchmark family, not a single aggregate impression: the paper's evidence supports strength in scientific coding and table-oriented literature synthesis, with remaining gaps in evidence interpretation and extended investigations.

What limits the loop and the claims

The paper states that the next development cycles depend on the coverage of the resource collections and on the quality of diagnosis and feedback. It also notes that scientific assumptions and experimental interpretations remain difficult to check through execution alone, which is a direct constraint on how far the loop's validation can go: executable checks can confirm that code runs, but not that the underlying scientific assumption is valid.

A second boundary is scope. The reported comparisons cover literature research, scientific coding, and multi-step research workflows on the benchmarks listed in the paper; the results do not establish behavior on scientific domains or task types outside those collections. The paper also releases the model and most of its training data, but the diagnosis and construction steps rely on agents built on frontier AI models, so the loop's task quality is tied to those agents rather than to SAIL alone.

The priority the results point to

The paper's own priority ordering is clear: SciCode and ArxivDIGESTables are where SAIL's relative advantage is strongest, while evidence interpretation, repository execution, and longer investigations are where the strongest larger models still lead and where further development is targeted. For a reader assessing whether the science-aware loop is worth adopting, the decision-relevant split is between tasks whose failures can be diagnosed and verified through execution and tasks whose failures hinge on scientific assumptions that execution cannot check. SAIL's results support the first category more strongly than the second, and the paper's stated next steps follow that same boundary.

Sources used in this answer