Source-linked AI summary
S1-NexusAgent: a Self-Evolving Agent Framework for Multidisciplinary Scientific Research
NexusAgent Team
TL;DR
Scientific agents face challenges in long-horizon planning, robust goal maintenance, and learning from execution across complex, tool-intensive research workflows. S1-NexusAgent combines hierarchical Plan-and-CodeAct execution with dynamic tools, sparse context management, and reusable skills. It achieves state-of-the-art performance across three scientific benchmarks.
Problem
Multidisciplinary scientific workflows require long-horizon planning, specialized tools, and management of large data and context, areas where existing agents have limitations.
Source: §1 · p. 1, §7.1 · p. 18Method
S1-NexusAgent couples global planning and subtask-level CodeAct execution with MCP-compatible tool retrieval, sparse context management, and trajectory-based skill distillation.
Source: §1 · p. 2Results
S1-NexusAgent achieves state-of-the-art performance across Biomni-Eval1, ChemBench, and MatSciBench in life sciences, chemistry, and materials science.
Source: §8 · p. 20Takeaways & Limitations
The framework provides a foundation for research on agentic learning, cross-domain scientific skill transfer, and AI systems resembling iterative scientific reasoning.
Source: §8 · p. 20
Abstract
from arXiv · showhide
Modern scientific research relies on large-scale data, complex workflows, and specialized tools, which existing LLMs and tool-based agents struggle to handle due to limitations in long-horizon planning, robust goal maintenance, and continual learning from execution. To address these issues, in this work, we propose S1-NexusAgent, a self-evolving agent framework designed for multidisciplinary scientific research. S1-NexusAgent adopts a hierarchical Plan-and-CodeAct execution paradigm, decoupling global scientific planning from subtask-level tool execution through a dual-loop architecture, thereby enabling stable modeling of complex research workflows. The system natively supports the Model Context Protocol (MCP), integrates up to thousands of cross-disciplinary scientific tools, and achieves efficient orchestration of heterogeneous research tools via intention-aware dynamic tool retrieval and hot-plug mechanisms. To address long-context and large-scale data challenges in scientific settings, S1-NexusAgent introduces object-reference-based sparse context management, which enables sub-task context isolation and intermediate result compression. Building on this, a Critic Agent automatically evaluates complete execution trajectories and distills high-quality research paths into reusable Scientific Skills, forming a closed loop for continuous self-evolution, which is valuable for sustainable and long-horizon scientific research. Experiments on authoritative scientific benchmarks involving long-horizon planning and complex specialized tool orchestration, including biomini-eval (biology), ChemBench (chemistry), and MatSciBench (material science), demonstrate that S1-NexusAgent achieves state-of-the-art performance, validating its effectiveness and generalization capability in complex scientific tasks.
1. Introduction
Scientific research increasingly involves multidisciplinary, data-intensive workflows requiring specialized tools, while existing agents face limitations in long-horizon planning, flexible tool use, context management, and continual evolution. S1-NexusAgent combines hierarchical planning, executable tool use, sparse context management, and trajectory-based skill reuse; experiments report state-of-the-art performance across three scientific benchmarks.
1. Introduction
- The framework addresses challenges arising from increasingly complex research workflows spanning disciplines, intermediate feedback, specialized tools, and heterogeneous data sources.The introduction identifies gaps in long-horizon planning, flexible tool operation, context management, and continual evolution.Source: §1 · pp. 1–2
- The framework uses hierarchical Plan-and-CodeAct execution and an inner–outer dual loop to connect high-level research objectives with executable subtasks.CodeAct couples reasoning with programmatic interaction with scientific tools.Source: §1 · p. 2
- S1-NexusAgent supports up to thousands of cross-disciplinary tools through MCP and intent-aware retrieval that loads relevant tools on demand.This targets the reasoning interference associated with large tool collections.Source: §1 · p. 2
- Its sparse context management combines object references, subtask isolation, trajectory compression, and planning-aware augmentation for long, data-intensive workflows.The framework aims to preserve decision-relevant evidence while suppressing context expansion and reasoning noise.Source: §1 · p. 2
- A Critic Agent evaluates execution paths and distills high-quality ones into reusable Scientific Skills for continued experience accumulation across tasks.The authors state that skill acquisition and reuse progressively improve decision-making efficiency over time.Source: §1 · p. 2
- S1-NexusAgent achieves state-of-the-art performance across biology, chemistry, and materials-science benchmarks.The evaluation includes biomini-eval, ChemBench, and MatSciBench.Source: §1 · p. 3
2. Related Work
S1-NexusAgent is positioned as a cross-domain scientific agent whose explicit dual-loop control, flexible tool orchestration, and reusable skills distinguish it from domain-specific and code-enhanced systems.
2.1. Domain-Specific Scientific Agents
- Domain-specific scientific agents can perform strongly within fixed scopes, while S1-NexusAgent targets cross-domain tasks with unified system-level mechanisms.The contrast concerns long-horizon stability, heterogeneous tool orchestration, context structuring, and self-evolution.Source: §2.1 · p. 3
2.2. Complex General Scientific Agent Systems
- Against monolithic or loosely structured multidisciplinary systems, S1-NexusAgent separates task-level planning from subtask-level executable reasoning.The authors describe this structure as supporting goal stability, localized trial-and-error, replanning, and trajectory-level evaluation.Source: §2.2 · pp. 3–4
2.3. Code-enhanced and programmatic scientific agents
- Unlike code-enhanced agents that use code as a program or pipeline, S1-NexusAgent embeds iterative CodeAct execution within explicit global planning.It also evaluates trajectories and distills Scientific Skills so successful executions can be accumulated and reused.Source: §2.3 · p. 4
3. Architecture
S1-NexusAgent turns high-level scientific requests into coordinated workflows through tool retrieval, sandboxed CodeAct execution, and planner-led feedback. Its architecture combines dual-loop control, sparse context management, and trajectory-based skill reuse.
3.1. Overview
- The overview transforms high-level scientific intents into executable workflows through planning, execution, reflection, and evolution.Intent recognition structures objectives, pre-planning outlines stages, and the Planner dispatches subtasks to CodeAct before updating global state from outcomes.Source: §3.1 · pp. 4–5
3.2. An Extensible Scientific Tool Ecosystem
- The tool ecosystem combines an intelligent retrieval engine with a standardized, modular interface for navigating thousands of specialized instruments.The tools include domain-specific scientific tools, literature retrieval, and general utilities, accessed through CodeAct in a secure sandbox.Source: §3.2 · p. 6, §3.2.2 · p. 7
3.2.1. Intent-Aware Dynamic Hot-Plugging (DHP)
- DHP parses structured intent, filters irrelevant tool domains, then semantically matches and hot-plugs relevant tools into the active context.The on-demand retrieval design aims to reduce context explosion and reasoning noise.Source: Figure 2 · p. 6, §3.2.1 · p. 6
3.2.2. The Scientific Tool Ecosystem and Execution Interface
- The execution interface uses CodeAct to couple reasoning with programmatic tool interaction in a secure sandbox.MCP compatibility and the three tool classes support integration of diverse scientific resources.Source: Figure 3 · p. 7, §3.2.2 · p. 7
3.3. Unifying Global Planning and Local Exploration via Inner–Outer Dual Loops
- The outer Planner maintains research objectives and adjusts decomposition, while the inner CodeAct loop validates hypotheses through tool use, correction, and tuning.Structured subtask results return to the Planner, which chooses whether to expand exploration, backtrack, or converge; execution depth and planning granularity scale with task complexity.Source: §3.3 · pp. 8–9
3.4. Context Management for Scientific Tasks
- Sparse Context Management controls context growth through object references, subtask isolation, trace compression, and planner-aware augmentation.Large artifacts are lazily loaded, subtask contexts remain independent, and compressed outcomes inform subsequent planning.Source: §3.4 · pp. 9–10
3.5. Self-Evolution via Trajectory Evaluation and Scientific Skills
- TE-SE evaluates complete trajectories and distills high-quality paths into reusable Scientific Skills stored as structured experience triples.The accumulated trajectories are described as a foundation for later Planner policy optimization and cross-domain skill transfer.Source: §3.5 · pp. 10–11
4. Agentic Training
S1-NexusAgent trains its inner-loop CodeAct executor while leaving the outer-loop Planner untrained, using supervised fine-tuning followed by reinforcement learning. The stages target structured execution behavior and environment interaction.
4. Agentic Training
- Agentic training targets the inner-loop CodeAct module, leaving joint inner–outer-loop learning for future work because execution behaviors are structured and recurrent.The outer-loop Planner handles task-level strategy and is already supported by strong foundation models.Source: §4 · p. 11
4.1. SFT
- SFT first maximizes token-level likelihood of expert execution trajectories to establish executable reasoning and structured action patterns before reinforcement learning.Trajectories come from rejection sampling or curated instructions.Source: §4.1 · p. 12
- SFT data combine 500 teacher-selected scientific trajectories with 1,500 code-interaction samples and 3,000 general-dialogue samples.For each scientific task, eight DeepSeek-V3.2 rollouts are generated and the highest-reward trajectory is selected.Source: §4.1 · p. 12
4.2. Agentic RL
- RL starts from GRPO to improve on-policy exploration and interaction in real execution environments, with asymmetric clipping intended to stabilize policy updates.The clipping ranges are 0.2 for the lower bound and 0.28 for the upper bound; the authors say this prevents premature policy collapse while encouraging novel execution paths.Source: §4.2 · pp. 12–13
- RL rewards correct final results with 0.9 and protocol-compliant code and solution formatting with 0.1.The formatting reward requires sandbox actions in <code>...</code> and final answers in <solution>...</solution>.Source: §4.2 · p. 13
5. Quantitative evaluation
Across biology, chemistry, and materials science benchmarks, S1-NexusAgent reports state-of-the-art performance. The evaluation also examines executor scale and agentic training, with gains reported from training smaller executors.
5.1. Setups
- The benchmarks cover multi-step reasoning, structured tool invocation, and iterative feedback across life sciences, chemistry, and materials science.Experiments use a unified MCP-based interface and sandbox; stochastic results are averaged over multiple runs.Source: §5.1 · p. 13
5.2. Experiments on Biomni-Eval1
- 76.07 average accuracy on Biomni-Eval1 with Claude-4.5-Sonnet exceeds Biomni-R0-32B’s 66.9, with improvements across nearly all task categories.The reported comparison uses the strongest previously reported agent as the baseline.Source: §5.2 · p. 14, Table 1 · p. 14
- With the same Qwen3-8B backbone on Biomni-Eval1, S1-NexusAgent scores 33.63 versus 31.8 for the non-agent baseline; Qwen3-8B-Sci reaches 42.42.The authors attribute the pattern to complementary domain-aware execution training and agentic control, rather than backbone capacity alone.Source: §5.2 · p. 14
5.2. Experiments on Biomni-Eval1; 5.3. Experiments on ChemBench; 5.4. Experiments on MatSciBench
- S1-NexusAgent reports state-of-the-art performance on Biomni-Eval1, ChemBench, and MatSciBench.The reported comparisons cover diverse biomedical tasks, chemistry subfields, and text-only materials-science categories.Source: Table 1 · p. 14, §5.2 · p. 14, Table 2 · p. 15
5.3. Experiments on ChemBench
- On ChemBench, reported gains are associated with multi-step planning, retrieval, and tool use across chemistry tasks; agentic training improves Qwen3-8B performance by 6.5 points.Reducing the executor from DeepSeek-V3.2 to Qwen3-8B drops performance by 10.6 points, especially on in-depth coding and retrieval tasks.Source: §5.3 · p. 15
5.4. Experiments on MatSciBench
- On MatSciBench, the authors report state-of-the-art results and improvements in formula calculation and failure-mechanism analysis with tool use.Reducing the executor from DeepSeek-V3.2 to Qwen3-8B drops performance by 6.6 points; agentic training improves it by 3.0 points.Source: Table 3 · p. 16, §5.4 · p. 16
6. Case studies of S1-NexusAgent
A GWAS variant-prioritization case study follows the agent from intent recognition and hierarchical planning through iterative execution and verification. It concludes that rs748196 is the most promising candidate and provides a biological explanation tied to PEMT.
Case study
- The case asks the agent to prioritize one of 11 SNPs associated with 2-Arachidonylglycerophosphocholine, requiring cross-database evidence synthesis and iterative refinement.The task is presented as a test of complex, long-horizon, tool-intensive scientific work.Source: §6 · p. 16
Intent Recognition and Hierarchical Planning
- Intent recognition loads GWAS and dbSNP tools, and the Planner divides the goal into data retrieval, correlation analysis, and functional annotation.The case is identified as Figure 7.Source: §6 · pp. 16, 18, Figure 7 · p. 17
Code-Driven Iteration and Strategy Adaptation; Cross-Stage Verification and Strategic Refinement; Synthesis and Conclusion
- After exact-name retrieval fails, the Code Agent switches to synonym expansion, then identifies rs748196 in the FADS1/2/3 cluster through subsequent cleaning and correlation-ranking iterations.A second planning loop invokes dbSNP for deeper characterization; the report concludes rs748196 is most promising and links its biological impact to the PEMT pathway.Source: §6 · p. 18
7. Discussion
The discussion evaluates the dual-loop architecture and staged SFT-plus-RL training on Biomni-Eval1. Results favor the integrated loops and RL-based training, while motivating further study of joint execution- and strategy-layer optimization.
7.1. Ablation on the Dual-Loop Architecture
- The dual-loop design separates task-level planning from execution-level exploration and error recovery, comparing the full system with single-loop variants.Only Outer-loop uses a single-pass CodeAct call; Only Inner-loop disables global decomposition and replanning.Source: §7 · p. 18, §7.1 · p. 19
- 42.42% average is the full Dual-loop system’s best overall performance, versus 38.99% for Only Inner-loop and 37.23% for Only Outer-loop.The comparison evaluates isolated loop configurations against the integrated framework on Biomni-Eval1.Source: §7.1 · p. 19, Table 4 · p. 19
7.2. Ablation on RL-based Staged Training
- Staged training initializes CodeAct with SFT for protocol adherence, then uses RL policy optimization to refine decisions under uncertainty.The training-strategy comparison fixes the base Qwen3-8B backbone and compares Only SFT with SFT + RL.Source: §7 · p. 18, §7.2 · p. 19
- 42.42% overall success rate with SFT + RL exceeds 38.40% with SFT alone, including 50.0% versus 24.0% on GWAS Catalog.SFT + RL is the full Dual-loop NexusSciAgent configuration.Source: §7.2 · p. 19, Table 5 · p. 20
- The discussion proposes exploring joint optimization of execution and strategic layers to support cross-level cooperative learning and long-term decision consistency.This is presented as a future research direction for long-horizon scientific tasks.Source: §7.2 · p. 20
8. Conclusion
S1-NexusAgent combines hierarchical planning and CodeAct execution with an extensible tool ecosystem, sparse context management, and trajectory-based self-evolution. Experiments on three scientific benchmarks report state-of-the-art performance across life sciences, chemistry, and materials science.
8. Conclusion
- S1-NexusAgent decouples global scientific planning from subtask execution through an inner–outer dual-loop architecture, while adding MCP-enabled tool retrieval and sparse context management.The framework targets long-horizon, tool-intensive, multidisciplinary research workflows.Source: §8 · p. 20
- A Critic Agent evaluates complete execution trajectories and distills high-quality scientific behaviors into reusable Scientific Skills.The design forms a loop of execution, evaluation, and capability growth across tasks and domains.Source: §8 · p. 20
- Experiments on Biomni-Eval1, ChemBench, and MatSciBench report state-of-the-art performance across life sciences, chemistry, and materials science.The authors present these results as evidence of effectiveness and generalization in complex scientific tasks.Source: §8 · p. 20
- The authors describe the framework as a step toward scalable, adaptive scientific agents capable of long-term reasoning and learning.They identify agentic learning, cross-domain skill transfer, and systems resembling iterative scientific practice as future research directions.Source: §8 · p. 20
9. Appendix
The appendix illustrates tasks and workflows spanning biology, genetics, and chemistry. Its chemistry example compares two candidate compounds using computed properties and oral-drug-likeness criteria.
9.1.1. Sample Task
- The biology sample asks for a rare-disease diagnosis from patient phenotypes and a candidate-gene list, with disease name and OMIM ID as the requested output.The appendix also gives a GWAS task to prioritize the most promising variant for a phenotype.Source: §9.1.1 · p. 26
9.2.1. Sample Task
- The chemistry sample frames selection between two compounds for further drug development in an early virtual-screening setting considering oral availability and molecular profile.The workflow uses cheminformatics tools, including RDKit, to calculate molecular properties.Source: §9.2.1 · p. 26, §9.2.2 · p. 26
9.2.2. Sample Workflow(summarized)
- Compound B’s TPSA is 76.30 Å2 versus 39.72 Å2 for Compound A; both meet Lipinski and Veber rules, while A has two chiral centers and B has none.The passage describes B’s polarity as within the stated 40–90 Å2 oral-absorption range and notes that A’s chiral centers add synthesis complexity.Source: §9.2.2 · p. 26, Table 6 · p. 27
Phase 3: Reasoning and Decision
The sample reasoning favors Compound B based on polarity and synthesis considerations. A separate electrochemical task asks for the temperature corresponding to a specified Zn/Pb cell potential, and its workflow invokes the Nernst equation.
Phase 3: Reasoning and Decision
- Compound B is favored for its TPSA in the stated ideal range and lack of stereocenters, while Compound A’s lower TPSA may limit solubility and absorption.The passage also describes the absence of stereocenters as simplifying synthesis.Source: Phase 3: Reasoning and Decision · p. 27
9.3.1. Sample Task
- The electrochemical sample asks what temperature gives a Zn/Pb cell potential of +0.568 V at specified Zn2+ and Pb2+ concentrations.The workflow identifies the Nernst equation and lists E, E°, R, n, F, and Q=[Zn2+] as given quantities.Source: §9.3.1 · p. 27, §9.3.2 · p. 28
9.3.2. Sample Workflow
- The workflow says the requested temperature is expected in Kelvin, but the supplied passage does not state its numerical value.The expectation of Kelvin units is attributed to common calculation conventions in the passage.Source: §9.3.2 · p. 29