Source-linked AI summary
Beyond Static Tools: Test-Time Tool Evolution for Scientific Reasoning
Jiaxuan Lu, Ziyu Kong, Yemin Wang, Rong Fu, Haiyuan Wan, Cheng Yang, Wenjie Lou, Haoran Sun, Lilong Wang, Yankai Jiang, Xiaosong Wang, Xiao Sun, Dongzhan Zhou
TL;DR
Scientific AI needs computational methods that extend beyond static libraries, which are sparse, heterogeneous, and incomplete for open-ended inquiry. TTE generates, verifies, and evolves executable tools during inference, and achieves state-of-the-art scientific reasoning while supporting cross-domain adaptation. The SciEvo benchmark evaluates this paradigm across 1,590 tasks with 925 evolved tools.
Problem
Static tool libraries are sparse, heterogeneous, and unable to anticipate the bespoke computational primitives required for open-ended scientific reasoning.
Method
TTE synthesizes, verifies, and evolves tools during inference, using TTE-Zero for ab-initio synthesis and TTE-Adapt for cross-domain adaptation.
Results
TTE establishes state-of-the-art scientific reasoning, with TTE-Zero outperforming existing baselines in accuracy and tool-utilization efficiency while TTE-Adapt enables cross-domain adaptation.
Takeaways & Limitations
TTE shifts scientific agents from passive tool selection toward active creation of computational primitives aligned with evolving problem spaces.
Takeaways & Limitations
Generating and verifying tools during inference increases computational overhead and latency, while executing arbitrary generated code creates safety risks requiring stronger semantic verification for real-world scaling.
Abstract
from arXiv · showhide
The central challenge of AI for Science is not reasoning alone, but the ability to create computational methods in an open-ended scientific world. Existing LLM-based agents rely on static, pre-defined tool libraries, a paradigm that fundamentally fails in scientific domains where tools are sparse, heterogeneous, and intrinsically incomplete. In this paper, we propose Test-Time Tool Evolution (TTE), a new paradigm that enables agents to synthesize, verify, and evolve executable tools during inference. By transforming tools from fixed resources into problem-driven artifacts, TTE overcomes the rigidity and long-tail limitations of static tool libraries. To facilitate rigorous evaluation, we introduce SciEvo, a benchmark comprising 1,590 scientific reasoning tasks supported by 925 automatically evolved tools. Extensive experiments show that TTE achieves state-of-the-art performance in both accuracy and tool efficiency, while enabling effective cross-domain adaptation of computational tools. The code and benchmark have been released at https://github.com/lujiaxuan0520/Test-Time-Tool-Evol.
1 Introduction
Scientific reasoning requires executable computational rigor, but static tool libraries are sparse, heterogeneous, and unable to anticipate bespoke primitives for novel inquiries. TTE addresses this limitation by generating, verifying, and evolving tools during inference, supported by the SciEvo benchmark and reported state-of-the-art results.
- Scientific LLM agents need executable computational tools because probabilistic reasoning alone performs poorly on tasks requiring rigorous fidelity.
- Static libraries are difficult to curate comprehensively and cannot anticipate bespoke computational primitives for unseen scientific problems.
- TTE synthesizes, verifies, and evolves executable tools during inference instead of relying on a fixed pre-collected library.
- SciEvo comprises 1,590 scientific evaluation instances supported by 925 evolved tools.
- TTE establishes state-of-the-art scientific reasoning performance, with TTE-Zero improving both accuracy and tool-utilization efficiency and TTE-Adapt enabling cross-domain adaptation.
2 Related Work
Prior tool-augmented agents largely select tools from static libraries, while newer systems generate or accumulate tools dynamically. However, existing dynamic and evolutionary approaches remain separated from inference or focused on non-scientific settings, leaving scientific rigor and domain-specific logic insufficiently addressed.
- General-domain systems such as Gorilla and ToolLLM retrieve tools from massive, pre-defined API libraries.
- Scientific systems including ChemCrow, CheMatAgent, and ChemMAS rely on fixed expert-curated tools for domain-specific tasks.
- Static tool libraries fail to exhaustively cover open-ended task spaces.
- CREATOR and CRAFT synthesize custom tools, but LATM and ToolMaker separate tool-making from inference, hindering real-time adaptation.
- Voyager, SEAgent, and ToolACE-DEV explore evolving executable skills for embodied, operating-system, or general computer tasks rather than rigorous scientific reasoning.
3 Test-Time Tool Evolution
TTE treats tool creation as an online evolutionary process: it decomposes scientific problems, retrieves or synthesizes tools, verifies and refines them, then executes the resulting sequence. Its greedy updates maintain a library that balances problem-solving utility with expansion costs.
- 3.1 Problem Definition: TTE generates and evolves executable tools during problem-solving rather than relying on tools prepared offline.The framework is formulated as a dynamic alternative to static tool ecosystems.
- 3.1 Problem Definition: The system maintains an evolving library by maximizing cumulative problem-resolution utility while penalizing library expansion.TTE uses a greedy evolution strategy because globally optimizing tool composition is computationally intractable.
- 3.2 Framework Overview: The closed-loop workflow decomposes queries into executable sub-goals, retrieves matching tools, synthesizes tools when retrieval fails, refines validated tools, and executes the sequence.These operations are implemented through five integrated modules.
- 3.5 Generative Tool Synthesis: Generated tools undergo syntax, execution, and domain validation, and only tools passing all checks proceed to refinement.This verification gate precedes library registration.
- 3.6 Atomic Tool Refinement: Atomic refinement decomposes valid tools into reusable units, removes redundancy, and prunes low-usage entries when library capacity is exceeded.The design targets partial reuse and a compact, relevant registry.
- 3.7 Runtime Execution: The executor integrates retrieved or generated tools to synthesize the final answer for the original scientific query.The solution is formalized as S = Solve(P, Lt).
4 The SciEvo Benchmark
SciEvo is built by bootstrapping tools from authentic computational science problems rather than curating a static codebase. The benchmark contains 1,590 evaluation instances and 925 evolved tools spanning 25 sub-disciplines across four scientific fields.
- 4.1 Benchmark Construction: SciEvo tools are generated from scratch with TTE to address authentic scientific reasoning needs.This evolutionary construction distinguishes the benchmark from libraries curated from static codebases.
- 4.1 Benchmark Construction: The seed questions combine SciEval, SciBench, and a specialized materials-science dataset, emphasizing multi-step reasoning and precise numerical solutions.Semantic clustering and stratified sampling are used to broaden problem diversity.
- 4.2 Tool Library Synthesis: The bootstrap process starts with an empty library and permanently inducts atomic Python functions only when they help derive correct ground-truth answers.This process yields a verified library aligned with the benchmark problem space.
- 4.3 Taxonomy: SciEvo organizes its tools into 25 sub-disciplines across Physics, Chemistry, Materials Science, and Mathematics using embedding-based clustering refined by domain experts.The taxonomy combines computational semantics with canonical scientific distinctions.
- 4.4 Data Distribution: SciEvo contains 1,590 evaluation instances supported by 925 evolved tools, distributed as Physics 499, Chemistry 192, Mathematics 171, and Materials 63.Physics contributes the largest tool subset.
- 4.5 Evaluation Metrics: Evaluation measures Accuracy and Tool Reuse Rate under a maximum library capacity of C = 500.TRR@1, TRR@5, and TRR@10 capture increasingly stringent levels of tool reuse.
5 Experiments
Experiments evaluate TTE across SciBench, SciEval, and SciEvo using multiple baselines, models, and tool-reuse metrics. TTE-Zero is reported to outperform baselines in accuracy, while its reuse analysis shows near-perfect utilization on SciEvo.
- 5.1 Experimental Setup: TTE is evaluated on SciBench, SciEval, and SciEvo to assess solution accuracy and tool-evolution efficiency.The evaluation includes both problem-solving and library-quality objectives.
- 5.1 Experimental Setup: TTE-Zero is compared with chain-of-thought, program-of-thought, and agentic tool-use baselines, while TTE-Adapt includes No Tool and Source Only comparisons.The baselines span reasoning-only and tool-enabled paradigms.
- 5.2 Implementation Details: Experiments use GPT-4o, Qwen2.5-7B-Instruct, and GPT-3.5-turbo, with main results reported using GPT-3.5-turbo at temperature 0.3.Retrieval uses bge-m3 embeddings and bge-reranker-v2-m3 reranking with top-k = 3.
- 5.3 Accuracy: TTE-Zero consistently outperforms all baselines in accuracy across the evaluated benchmarks.Table 2 reports the cross-benchmark accuracy comparison without providing the row-level values here.
- 5.2 Implementation Details: Final answers are judged with relative numerical tolerance 10^-5 and exact canonical matching for symbolic expressions, alongside Accuracy and TRR metrics.Semantic deduplication rejects new tools when cosine similarity exceeds τ = 0.8.
6 Results and Analysis
TTE-Zero improves scientific reasoning accuracy and tool reuse, while sub-goal decomposition strengthens tool-augmented performance and TTE-Adapt transfers tools across domains.
- Comparative Analysis on Scientific Benchmarks: 0.62 accuracy on SciEvo exceeds CheMatAgent (0.56) and KTCE (0.55), while TTE-Zero reaches 0.45 on SciBench versus KTCE (0.37).TTE-Zero also improves over Basic-COT by +0.29 on SciEvo.
- Analysis of Tool Evolution Quality: TRR@1 reaches 0.99 for TTE-Zero on SciEvo, compared with Creator (0.17) and KTCE (0.31), indicating highly effective tool reuse.At k = 10, TTE-Zero retains reuse rates of 0.41 on SciEvo and 0.21 on SciBench, versus Creator’s 0.02 and 0.01.
- Ablation Study: Both “Q+Tools” and “S+Tools” outperform “No Tool call” across evaluated models, with “S+Tools” consistently achieving the highest accuracy.With a library size of 100 on Qwen2.5-7B, “S+Tools” scores 0.364 versus 0.313 for “Q+Tools”.
- Cross-Domain Adaptation: TTE-Adapt outperforms “No Tool” and “Source Only” baselines in cross-domain settings by pruning irrelevant source tools and consolidating target-domain primitives.TRRtrans@1 decreases from 0.26 to 0.23 in Chemistry, while TRRevol@1 reaches 0.24 in Chemistry and 0.32 in Physics.
7 Conclusion
The paper presents Test-Time Tool Evolution as a way to address static-tool limitations by allowing agents to create computational primitives during inference. Its evaluations report improved reasoning accuracy and cross-domain tool adaptation.
- Conclusion: TTE shifts scientific agents from passive tool selection to synthesizing, verifying, and evolving computational primitives during inference.The framework is intended to keep the tool space aligned with the evolving scientific problem space.
- Conclusion: The evaluations report new state-of-the-art reasoning accuracy and robust tool adaptation across diverse scientific domains.The conclusion frames autonomous tool evolution as important for general-purpose scientific AI.
8 Limitations
The current TTE framework has higher inference cost, depends on the coding ability of its backbone LLM, and introduces safety challenges when executing generated code.
- Inference Latency and Computational Cost: TTE incurs higher computational overhead and latency than static retrieval because tools are synthesized and verified during inference.The authors suggest lightweight meta-models to skip evolution for trivial queries.
- Dependency on Base LLM Coding Capability: TTE performance degrades with smaller open-source models below 7B parameters that struggle to generate syntactically correct Python primitives.The reported strongest results use high-capacity backbone models.
- Safety and Sandboxing in Open-Ended Evolution: Generating and executing arbitrary code creates risks of excessive resource use or unsafe operations, requiring stronger semantic safety verification beyond sandboxing and timeouts.The experiments use a strictly sandboxed environment with timeout constraints.
9 Ethical Statement
The paper describes safeguards for dynamically generated tools, including sandboxed execution, conservative registration, structured interfaces, and manual review of released artifacts.
- Ethical Statement: The released tool library underwent manual review to exclude tools enabling harmful applications, with responsible-disclosure practices applied.The authors identify dual-use risks in chemistry and materials science, including possible misuse for toxin synthesis.
- Data and Artifacts: SciEvo uses public scientific repositories, screens for personally identifiable information and offensive content, and releases code and data under the MIT License.The authors acknowledge that the system may reflect biases present in the underlying data.
- Implementation Safeguards: The pipeline distinguishes retrieval acceptance from deduplication through separate thresholds, while structured JSON or XML outputs support robust parsing.The same evolution loop covers TTE-Zero and TTE-Adapt with different library initialization states.
- Safety and Robustness: TTE rejects tools that fail syntax, runtime, or domain verification and can fall back to reasoning-only execution or partial tool chains.The implementation primarily registers only verified tools and uses a lightweight fallback after failed generation.
C Subject-wise Results on SciEvo
Subject-wise evaluation shows that tool augmentation improves direct inference, while retrieving tools from decomposed sub-questions is generally more robust than question-level retrieval. The largest decomposition gains occur where queries require multiple domain-specific operations.
- Tool augmentation provides clear benefits over direct inference across all model–subject pairs.
- Sub-question retrieval generally outperforms original-question retrieval, especially in Chemistry and Physics.Tool selection in these subjects is sensitive to units, constants, and domain-specific formulas.
- Chemistry: Chemistry gains most from structured decomposition because queries often combine conversions, gas relations, and stoichiometry.Decomposition provides a sharper retrieval signal and reduces irrelevant tool selection.
- Materials science: Materials science shows smaller incremental gains from decomposition because it often has higher baseline performance and fewer atomic operations per problem.
- The best retrieval configuration depends on both the model and the subject, and larger inventories do not always improve question-level retrieval.
D Analysis of Tool Reusability
Tool-reuse analysis distinguishes TTE from baselines by showing a shift from disposable, query-specific scripts toward reusable computational primitives. TTE redistributes tool usage toward moderate-to-high reuse frequencies, supporting transfer across scientific problems.
- Baseline tools are concentrated in the lowest hit-count bins, typically receiving only 1–2 uses.This pattern is associated with monolithic scripts, redundancy, and poor transferability.
- TTE shifts tool-usage density toward moderate-to-high frequencies, including roughly 10–50+ uses.
- The shift indicates a move from ad hoc solutions toward atomic computational primitives such as canonical formulas and unit conversions.These generalized modules can solve diverse problems through composition.
E.1 Metrics for TTE-Zero
TTE-Zero evaluates tool evolution through reuse, retrieval, and library-growth analyses, alongside scientific case studies and benchmark coverage. The results support reusable decomposition, while also identifying tool overload and maintenance requirements as scaling boundaries.
- Metrics for TTE-Zero: TRR@1 measures whether generated tools are executed at least once, while TRR@2 measures whether they transfer across multiple distinct queries.Higher values indicate less dead code and stronger immediate transferability.
- Atomic decomposition: Decomposition guarantees higher expected reusability when partial subproblems have positive probability and non-zero marginal utility.
- Tool overload: Under the stated overlap assumption, retrieval success decreases strictly as library size increases without improved retrieval mechanisms.
- Scaling boundaries: Library size converges asymptotically to a stable equilibrium, but future systems still require pruning, indexing, verification, and uncertainty-aware generation.