Source-linked AI summary
SAIL: Scientific Agentic Intelligence via a Science-Aware Loop
SAIL Model Team, Boyuan Sun, Bryan Dai, Che Liu, Chi Liu, Derek Li, Hongming Piao, Mengzhuo Chen, Xidong Wang, Yan Shu, Yinda Chen, Ziyang Zeng
TL;DR
Scientific agents need to make judgments throughout literature research, coding, and multi-step workflows, but training resources must be organized around the capabilities a model lacks. SAIL uses frontier-model agents to diagnose failures and construct targeted tasks, then trains a 35B-total, 3B-active-parameter model that achieves the highest SciCode and ArxivDIGESTables scores in the comparison and competes with substantially larger open-weight models across scientific workflows.
Problem
Training an open model for scientific work requires organizing scientific papers and code repositories around the capabilities it still lacks so it can make judgments throughout execution.
Source: §1 · p. 3Method
Frontier-model agents diagnose SAIL’s failures and construct targeted problems, demonstrations, executable tasks, tools, and environments, which are used in a loop of training and revised priorities.
Source: §1 · p. 3Results
With 35B total and 3B active parameters, SAIL achieves the highest SciCode and ArxivDIGESTables scores in the comparison and competes with substantially larger open-weight models across scientific workflows.
Source: §6.3 · p. 14Takeaways & Limitations
The authors release SAIL and most of its training data to support research on scientific agents and development of more capable AI scientists.
Source: §1 · p. 3SAIL ranks behind stronger larger models on repository execution and long research workflows, while full-text question answering has a weaker relative ranking despite a score of 91.67 and a 3.57-point gap to the leader.
Source: §6.2 · p. 14
Abstract
from arXiv · showhide
We introduce SAIL, an open model with 35B total and 3B active parameters for literature research, scientific coding, and multi-step research workflows. SAIL is developed through a science-aware improvement loop: agents built on frontier AI models analyze its task failures and construct training tasks that address the underlying capability gaps. The diagnosis examines search and evidence selection in literature tasks, scientific assumptions and reasoning in coding, and planning and revision in longer investigations. The agents draw on paper collections and scientific code repositories to build problems, interaction trajectories, and executable tasks with the required environments and tools. We repeat this loop over multiple development cycles and train SAIL through supervised fine-tuning, specialist training, multi-teacher on-policy distillation, and agentic reinforcement learning. SAIL achieves competitive performance across scientific research tasks with substantially fewer parameters than leading open-weight models.
1 Introduction
SAIL addresses the challenge of training an open model to make scientific judgments throughout literature research, coding, and tool-using workflows. It uses frontier-model agents to diagnose failures and construct targeted training tasks, then trains a 35B-parameter model that leads the comparison on SciCode and ArxivDIGESTables.
1 Introduction
- Scientific training tasks should target the capability a failure reveals, such as evidence selection, scientific assumptions in code, or revising a plan after intermediate findings.A program can run without errors yet rest on an invalid assumption, while a multi-step investigation can fail to use a finding to change course.Source: §1 · p. 3
- Frontier-model agents diagnose SAIL’s responses and trajectories, select scientific resources, and construct problems, demonstrations, executable tasks, and the tools and environments to practice targeted skills.Training priorities are revised across development cycles using the updated model’s behavior.Source: §1 · p. 3
- SAIL is an open model for literature retrieval and synthesis, scientific coding, and research workflows involving tool use.The paper also presents specialist training, MOPD, and agentic reinforcement learning as a recipe for consolidating scientific capabilities in one model.Source: §1 · p. 3
- With 35B total and 3B active parameters, SAIL achieves the highest SciCode and ArxivDIGESTables scores in the comparison and competes with substantially larger open-weight models across scientific workflows.Source: §1 · p. 3
2 Science-Aware Improvement Loop
The science-aware improvement loop uses observed capability gaps to guide task construction, training, and re-evaluation. It tailors scientific resources, supervision, and interactive environments to the skill being trained, then uses updated model behavior to set the next priorities.
2 Science-Aware Improvement Loop
- The loop repeats evaluation, diagnosis, task construction, and training, with frontier-model agents guiding diagnosis and construction and the updated model’s behavior setting later priorities.The energy-drift example illustrates how a diagnosed scientific reasoning failure can motivate training on relevant conservation principles.Source: §2 · p. 3, Figure 4 · p. 6
2.1 Diagnosing Capability Gaps
- Diagnosis examines searches, scientific reasoning and computational outputs, and multi-stage investigations to identify the specific skill a later task should train.The diagnosis set combines manually collected and generated tasks separate from benchmark evaluation sets, and agents inspect reasoning, tool calls, and observations leading to failure.Source: §2.1 · p. 4
2.2 Constructing Tasks from Scientific Resources
- Tasks pair a diagnosed gap with a prioritized topic, drawing on papers, scientific repositories, or both, and use focused problems, demonstrations, or executable interaction to train the objective.Longer investigations link actions so later decisions depend on earlier findings.Source: §2.2 · p. 5
2.3 Environments, Tools, and Feedback
- Interactive environments expose the observations needed for practice: retrieval results for search, execution tools and scientific libraries for coding, and persistent workspaces for longer investigations.Generated material is checked against sources and task-specific criteria, including tool availability and supervision or feedback.Source: §2.3 · p. 5
2.4 Training and Re-Evaluation
- After updates, targeted and broader development-set evaluation assesses improvement, transfer, and retention, and the model’s new responses and trajectories guide the next diagnosis and task construction.Training stages depend on the capability being addressed.Source: §2.4 · p. 5
3 Training Recipe
SAIL’s training recipe starts from a shared supervised fine-tuning checkpoint, develops capability-focused specialists, transfers their supervision to the student through MOPD, and then uses task feedback for agentic reinforcement learning.
3 Training Recipe
- The recipe combines supervised fine-tuning, specialist training, multi-teacher on-policy distillation, and agentic reinforcement learning to train SAIL on tasks from the improvement loop.The SFT checkpoint initializes both specialists and the student, and the final stage trains multi-step execution using scientific-environment feedback.Source: §3 · p. 7
3.1 Supervised Fine-Tuning
- The base model is fine-tuned on general and scientific data covering reasoning, literature tasks, coding, and tool interaction, with demonstrations that include tool calls and subsequent actions.The resulting checkpoint initializes specialist training and the student used in MOPD.Source: §3.1 · p. 7
3.2 Specialist Training
- Specialists start from the shared SFT checkpoint and receive capability-focused SFT, reinforcement learning, or both, with their training adapted to the available data and feedback.They are trained independently and later teach the common student during MOPD.Source: §3.2 · p. 7
3.3 Multi-Teacher On-Policy Distillation
- MOPD has the student generate responses or interactive trajectories, then uses the relevant specialist to score student-generated tokens under the contexts encountered, including tool observations.The objective masks task inputs and environment observations so only model-generated tokens contribute to the divergence.Source: §3.3 · p. 7
3.4 Agentic Reinforcement Learning
- After MOPD, agentic reinforcement learning trains the student on executable task families using task-specific verifiers, completion criteria, or other feedback signals.Training exposes the model to states resulting from its own actions, including tool failures and unsuccessful intermediate attempts.Source: §3.4 · p. 7
Specialist training
Figure 5 depicts the training progression from a shared SFT checkpoint through specialist training and MOPD to the final SAIL model, followed by agentic reinforcement learning.
Specialist training
- Figure 5 distinguishes model initialization or training progression with solid arrows from teacher supervision with dashed arrows.Specialists provide MOPD supervision before the student undergoes agentic reinforcement learning to produce the final SAIL model.Source: Figure 5 · p. 8
- The student’s agentic reinforcement learning uses task outcomes to supervise decisions across episodes, with rewards from verifiers, completion criteria, or other feedback signals.The model encounters states from its own actions, including tool failures and unsuccessful intermediate attempts, to improve multi-step task completion.Source: Specialist training · p. 8
- The final SAIL checkpoint is obtained after agentic reinforcement learning.This stage follows the student’s interaction with executable task environments.Source: Specialist training · p. 8
4 Infrastructure for Scalable On-Policy Training
The infrastructure separates rollout, scientific task execution, and optimization while sharing one student-trajectory collection pipeline across agentic RL and MOPD. It preserves episode and context information through collection so the two training methods can apply their respective supervision.
4 Infrastructure for Scalable On-Policy Training
- The system decouples rollout, task execution, and optimization, while agentic RL and MOPD share student-trajectory collection with policy weights synchronized between rounds.Task harnesses control scientific-environment execution, and the training system records actions and model-call context before attaching task feedback or specialist supervision.Source: §4 · p. 8, Figure 6 · p. 9
4.1 System Overview
- The framework uses separate optimization and rollout engines, with Uni-Agent abstractions and task-specific adapters connecting harnesses to sandboxes, tools, and environments.A model gateway routes harness requests to the rollout engine and records token-level information for training.Source: §4.1 · p. 9
4.2 Scientific Execution and Trajectory Collection
- Scientific tasks run in isolated, episode-specific workspaces, while persistent runtimes retain artifacts and environment state across tool calls and model-context compaction.The environment state remains independent of the model context.Source: §4.2 · p. 9
- The collector preserves generated token IDs, log probabilities, and training masks; compacted histories start new segments that remain tied to the same episode and outcome.Episode identifiers and segment boundaries allow rewards, advantages, and losses to be aggregated across the original trajectory.Source: §4.2 · p. 10
4.3 Synchronous On-Policy Training
- Synchronous rounds synchronize the student policy before collection, hold it fixed during episode rollouts, then optimize and publish updated parameters for the next round.Rollout and optimization can use different parallelism and memory configurations, with sandbox capacity scaled separately from model serving.Source: §4.3 · p. 10
4.4 Shared Training Interface for RL and MOPD
- RL and MOPD collect fresh trajectories from the current student policy, then attach different supervision: task-outcome rewards for RL and specialist-teacher signals for MOPD.Harness routing and teacher routing are independent, allowing execution harnesses and specialist supervision to be reconfigured separately.Source: §4.4 · p. 10
5 Evaluation and Analysis
The evaluation compares SAIL across scientific literature, coding, data analysis, and research workflows against its base and other open-weight models. SAIL leads several benchmark comparisons and improves over its base across all 13 evaluation settings.
5.1 Evaluation Setup
- Evaluation covers 13 settings across literature, coding, data analysis, and research workflows, comparing SAIL with its base and similarly sized and larger models in shared environments.Source: §5.1 · p. 11, §5 · p. 11
5.2 Main Results
- SAIL leads SciCode and ArxivDIGESTables, improves all 13 evaluations over its base, and leads compared 35B models on 12 settings.Source: §5.2 · p. 11, Table 1 · p. 12, Table 2 · p. 13
- 59.76 on the 12-task unweighted aggregate (excluding LitQA2-FullText) ranks SAIL second, behind GLM-5.2 (63.30) and ahead of DeepSeek-V4-Flash-0731 (59.15).Source: §5.2 · p. 11
5.3 Literature Retrieval and Analysis
- LitQA-search rose from 48.00 to 85.33 and PaperFindings from 22.19 to 33.25, ranking SAIL second on both; ArxivDIGESTables led the comparison at 35.24 versus DeepSeek-V4-Flash-0731's 35.13.Source: §5.3 · p. 13
5.4 Scientific Coding and Execution
- SciCode reaches 50.35 (+10.45 over base; +2.78 over next-highest), while DS-1k reaches 74.30 (+17.10); repository tasks rank first among compared 35B models.Source: §5.4 · p. 13
5.5 Data Analysis and Research Workflows
- On E2E-Bench, SAIL scores 89.46 on Basic and 77.27 on Hard, ranking fourth overall while leading compared 35B models on both.It ranks second on DiscoveryBench at 37.48 and fourth on DRB2 at 42.61.Source: §5.5 · p. 13
6 Discussion and Conclusion
The discussion describes how failures guide the use of scientific resources and training, while identifying performance gaps and dependencies for future development. The conclusion presents SAIL and reports its strongest comparative results on SciCode and ArxivDIGESTables.
6.1 Scientific Resources and Adaptive Training
- Agents use SAIL’s failures to choose how scientific resources support training, from broader literature search or selective synthesis to focused coding or multi-experiment investigations.They then construct corresponding tasks and environments.Source: §6.1 · p. 14
- Specialist training focuses on selected capabilities, MOPD transfers that supervision to the student’s own trajectories, and agentic RL trains with task feedback.Shared infrastructure collects trajectories across tools and environments while retaining each model action’s context.Source: §6.1 · p. 14
6.2 Remaining Challenges
- SciCode and ArxivDIGESTables are SAIL’s strongest relative results, while full-text question answering scores 91.67, 3.57 points behind the leader; larger models lead on repository execution and long workflows.DiscoveryBench shows a smaller improvement over the base model.Source: §6.2 · p. 14
- Evidence interpretation, repository execution, and longer investigations are identified as priorities for further development.Source: §6.2 · p. 14
- Future development cycles depend on resource-collection coverage and diagnosis and feedback quality; execution alone may not check scientific assumptions or experimental interpretations.Source: §6.2 · p. 14
6.3 Conclusion
- SAIL combines a science-aware improvement loop with SFT, specialist training, MOPD, and agentic RL to support literature research, scientific coding, and multi-step tool use.The conclusion reports the highest SciCode and ArxivDIGESTables scores in the authors’ comparison and competition with substantially larger open-weight models across scientific workflows.Source: §6.3 · p. 14