Source-linked AI summary
SMART: Self-Aware Agent for Tool Overuse Mitigation
Cheng Qian, Emre Can Acikgoz, Hongru Wang, Xiusi Chen, Avirup Sil, Dilek Hakkani-Tür, Gokhan Tur, Heng Ji
TL;DR
LLM agents often fail to recognize when parametric knowledge is sufficient, causing unnecessary tool use. The paper introduces SMART, SMART-ER, and supervised-trained SMARTAgent to calibrate tool decisions, reporting lower tool use with improved performance and resource-efficient generalization.
Problem
LLM agents lack reliable awareness of their knowledge boundaries, making it difficult to decide when external tools are necessary.
Method
SMART uses metacognitive decision-making, SMART-ER’s three-domain reasoning chains, and explicit rationales to train agents to balance parametric knowledge with tools.
Results
SMARTAgent reduces tool usage time by 24% on average and improves performance by over 37% across models compared with the best baseline.
Takeaways & Limitations
Strategic tool use can support efficient, low-resource agents whose capabilities reflect both task performance and metacognitive reasoning.
Takeaways & Limitations
The study covers Math, Intention, and Time, while broader architectures and domains remain for future validation.
Abstract
from arXiv · showhide
Current Large Language Model (LLM) agents demonstrate strong reasoning and tool use capabilities, but often lack self-awareness, failing to balance these approaches effectively. This imbalance leads to Tool Overuse, where models unnecessarily rely on external tools for tasks solvable with parametric knowledge, increasing computational overhead. Inspired by human metacognition, we introduce SMART (Strategic Model-Aware Reasoning with Tools), a paradigm that enhances an agent's self-awareness to optimize task handling and reduce tool overuse. To support this paradigm, we introduce SMART-ER, a dataset spanning three domains, where reasoning alternates between parametric knowledge and tool-dependent steps, with each step enriched by rationales explaining when tools are necessary. Through supervised training, we develop SMARTAgent, a family of models that dynamically balance parametric knowledge and tool use. Evaluations show that SMARTAgent reduces tool use by 24% while improving performance by over 37%, enabling 7B-scale models to match its 70B counterpart and GPT-4o. Additionally, SMARTAgent generalizes to out-of-distribution test data like GSM8K and MINTQA, maintaining accuracy with just one-fifth the tool calls. These highlight the potential of strategic tool use to enhance reasoning, mitigate overuse, and bridge the gap between model size and performance, advancing intelligent and resource-efficient agent designs.
1 Introduction
LLM agents combine parametric reasoning with external tools, but often lack awareness of when each is appropriate. SMART, supported by SMART-ER and SMARTAgent, trains agents to balance these modes and reduce unnecessary tool use.
- Motivation: LLM agents alternate between internal reasoning and external tools, raising the question of when each should be used.External tools can address limitations while parametric knowledge advances task solving.
- Tool Overuse: Over 30% of tool invocations are unnecessary because models fail to recognize when their internal knowledge suffices.This behavior increases resource consumption and can degrade performance.
- SMART: SMART applies metacognitive principles to align an agent’s subjective perception with its knowledge boundary when choosing tools or internal reasoning.The approach draws on human awareness of thought processes and problem-solving strategies.
- SMART-ER: SMART-ER spans Math, Time, and Intention, combining easy parametric steps with difficult tool-dependent steps and explicit decision rationales.The dataset targets computational accuracy, outdated knowledge, and user-preference awareness.
- Results: 24% less tool use and over 37% higher overall performance are reported for SMARTAgent across models.SMARTAgent also enables 7B-scale models to match GPT-4 and 70B models and uses one-fifth as many tool calls on out-of-distribution tasks while preserving accuracy.
2 Related Work
Related work studies LLM knowledge boundaries and tool use as complementary capabilities. Prior methods assess whether models know when questions are answerable and use external resources for current or specialized information.
- LM Knowledge Boundary: Knowledge-boundary research examines whether language models recognize the limits of their knowledge.Methods include verbal probing, fine-grained benchmarks, and data augmentation to address hallucinations from unawareness of those limits.
- LM Tool Use: Tool-use research integrates external resources to complement parametric knowledge and overcome limitations such as outdated information or domain-specific expertise.Other work explores tool creation and external module integration.
3 Preliminaries
The preliminaries define Tool Overuse as unnecessary tool reliance and measure it using model competence and tool-invocation behavior. Experiments show this tendency in both standalone LLMs and tool-equipped agents.
- Definition: Tool Overuse is excessive external-tool reliance when an agent could complete a task using parametric knowledge alone.The definition focuses on unnecessary invocation rather than tool use that addresses a knowledge limitation.
- Definition: The Tool Overuse Rate quantifies unnecessary tool use as the product of intrinsic reasoning capability α and invocation fraction β.α measures the fraction of questions answerable without tools, while β measures tool invocation on that subset.
- LLM Experiments: Llama-3.1-8B and Mistral-7B are evaluated on GSM8K with and without tools to compare autonomous tool decisions against tool-free reasoning.Figure 2 reports significant tool overuse, including a rate above 50% for Llama.
- LLM Experiments: Tool use can introduce unnecessary complexity and degrade performance in some cases.This finding accompanies the observed excessive reliance on tools.
- Agent Experiments: XAgent and AgentGPT invoke tools unnecessarily on 50 GSM8K queries that can be answered without tools, significantly slowing problem solving.Both systems are instructed to use tools only when necessary.
4 Method
SMART-ER constructs compositional reasoning chains that alternate between parametric knowledge and external tools, with explicit justifications for each decision. SMARTAgent is trained on these chains to balance tool reliance and internal reasoning interactively.
- SMART-ER Dataset: SMART-ER spans Math, Time, and Intention, combining subproblems models handle well with subproblems requiring external tools.Its queries mix simple arithmetic with complex math, slow- and fast-changing facts, and model-resolvable content with user-specific intentions.
- Reasoning Chain Construction: Each query is decomposed into n subgoals, and every subgoal is annotated as tool-reliant or solvable with parametric knowledge.Queries containing only one type of subgoal are discarded to preserve the intended contrast.
- Reasoning Chain Construction: Tool-reliant subgoals are mapped to Code, Search, or AskUser, while other subgoals produce reasoning steps from parametric knowledge.The process iteratively incorporates prior reasoning steps and tool outputs into subsequent subgoals.
- Reasoning Chain Construction: SMART-ER enriches each reasoning step with a justification explaining whether internal knowledge suffices or tool use is necessary.These explicit natural-language rationales are intended to calibrate strategic switching between internal reasoning and external tools.
- Quality Control and Training: SMART-ER reasoning chains pair parametric steps with tool invocations and outputs, with human supervision applied to 5% of data and a pass rate above 95%.The resulting instruction pairs support iterative inference up to the next tool invocation or final solution.
- Quality Control and Training: SMARTAgent models are instruction-tuned from Llama, Mistral, and Nemo model families to interactively balance tools with parametric knowledge.The trained family includes models ranging from 7B to 70B-scale configurations.
5 Experiment
Experiments evaluate SMARTAgent across in-domain and out-of-distribution tasks, measuring accuracy alongside tool use and decision confidence. Results indicate that SMARTAgent improves performance while reducing unnecessary or excessive tool calls, including for smaller models and OOD settings.
- Main Results: SMARTAgent reduces tool usage time per query by 24% on average while improving performance by over 37% across models versus the best baseline.
- Main Results: 7B- and 8B-scale SMARTAgent models outperform GPT-4o and its 70B counterpart in Time and Intention while using fewer tool calls.In Math, SMARTAgent lags behind larger models but remains competitive with same-architecture baselines.
- OOD Results: SMARTAgent generalizes to OOD benchmarks, achieving better overall performance while making only one-fifth as many tool calls as the base model in MINTQA.
- Main Results: SMARTAgent maintains near-optimal tool use on Time and Intention tasks, averaging close to one call for Time and three calls for Intention queries.Time queries ideally require at least one tool call, while Intention queries contain two to four missing details.
- Tool Overuse: SMARTAgent reduces unnecessary tool calls by up to 50% on GSM8K and Math test data compared with prompting the base model with tool access.Tool overuse nevertheless persists and is examined through error analysis.
- Confidence Validation: Confidence experiments show higher confidence on correct internal-reasoning or tool-invocation decisions than on incorrect decisions.The evaluation sampled 50 decision steps from Time and Intention test splits using token logits.
6 Discussions
The discussion examines improper tool use, its mechanisms, and remaining limitations. SMARTAgent improves strategic and interpretable tool selection, but feedback neglect, excessive caution, and broader resource-allocation trade-offs remain relevant concerns.
- Improper Tool Usage: Tool overuse may arise from uncertainty about model capabilities and the perceived ease of external lookups, while tool underuse may reflect misjudged task complexity.
- Case Study: SMARTAgent’s case study shows logical planning, context corroboration, and explicit awareness of knowledge boundaries alongside reduced tool-use overhead.
- Future Directions: Confidence probing, self-checking modules, reinforcement learning from feedback, and iterative or in-context learning are proposed as possible ways to refine tool-use thresholds.
7 Conclusion
The paper concludes that metacognitive calibration can help agents recognize knowledge boundaries and balance parametric reasoning with tool use. SMARTAgent’s results support strategic tool use as a route toward efficient, low-resource agent development.
- SMART uses SMART-ER reasoning chains with curated strengths and weaknesses to train SMARTAgent to balance parametric knowledge and tool use.
- The paper reports that this data-driven approach calibrates model awareness and supports efficient, low-resource agent development.
Limitations
The study evaluates SMARTAgent primarily in Math, Intention, and Time, while acknowledging broader domain and architecture coverage as open areas for validation.
- The study focuses on three domains: Math, Intention, and Time.
- Long-tail knowledge and domain-specific expertise remain outside the study’s evaluated scope.The authors note that these areas may require external resources.
- The findings’ generalizability could be further assessed across architectures including Qwen, DeepSeek, and different model sizes.
A.2 Model Experiment Details
The preliminary model experiments compare normal reasoning with tool-enabled reasoning, using prompts that ask models to reason independently and invoke tools only when necessary.
- Llama-3.1-8B-Instruct and Mistral-7B-Instruct are each prompted twice for every GSM8K question.One condition uses normal reasoning; the other provides tools and lets the model decide whether to use them.
- The normal condition asks models to solve tasks autonomously using knowledge and reasoning while articulating reasoning steps.
- The tool-enabled condition instructs models to reason independently, use tools only when necessary, and continue step by step after identifying a knowledge gap.
- Code snippets are executed when generated, and the resulting output is returned so the model can continue reasoning.
B.1 Data Selection
SMART-ER is constructed by selecting mixed-difficulty questions across Math, Time, and Intention, assigning relevant tools and explicit rationales, then applying automated adaptation and supervised fine-tuning.
- Data Selection: Math questions combine straightforward and challenging reasoning steps by decomposing difficult MATH problems and inspecting their ground-truth chains.
- Data Selection: Time questions combine fast-changing facts with slow-changing facts through augmented, multi-hop queries.
- Data Selection: Intention questions exclude vague queries and verify that selected questions are generally answerable without tools.
- Tool Assignment: The constructed tool set includes Code, Search, and AskUser, with Code mainly used for Math, Search for Time and Intention, and AskUser mainly for Intention.
- Training Preparation: Each reasoning step receives a parametric or tool-dependent treatment plus a justification explaining why internal knowledge suffices or a tool is required.
- Fine-Tuning: Supervised fine-tuning uses Alpaca-format outputs across the listed base models, with LoRA applied across all model layers.
C.1 Data Setting
The evaluation uses adapted in-domain samples and unmodified or filtered out-of-domain benchmarks, with iterative inference that parses tool calls, executes them, and re-prompts the model.
- Data Setting: In-domain testing samples 400 Math instances, 100 Time instances, and 100 Intention instances from adapted data.The Math sample covers testing categories and five difficulty levels; Time instances combine fast- and slow-changing aspects.
- Data Setting: Out-of-domain testing uses the full GSM8K test set and a sampled MINTQA subset requiring multi-hop reasoning with old and new knowledge.
- Baselines: The Normal Reasoning Trained baseline uses domain-specific reasoning chains from MATH or GPT-4o-generated chains guided by annotations.
- Interactive Inference: Interactive inference identifies final responses or tool calls through rule-based matching, executes detected calls, appends their outputs, and re-prompts the model.
- Additional Results: Llama-3.3-70B-Instruct results are compared with the Llama-3.1-70B-Instruct SMARTAgent, with no significant performance difference found between the two 70B versions.
- Confidence Validation: Decision evaluation tests whether the model chooses a tool or parametric knowledge for the nth step given the preceding n −1 steps within a full m-step chain.
D.1 Evaluation on Advanced Reasoning Dataset
SMARTAgent is evaluated on AMC’23 and on time-domain question answering with distilled models, comparing against always-available tool prompting and examining both overuse and underuse. It outperforms the baseline on AMC and improves distilled-model performance while promoting strategic tool engagement.
- AMC’23 evaluation: SMARTAgent outperforms always-available tool prompting across both tested model backbones on the AMC’23 benchmark.The evaluation uses Llama-3.1-8B-Instruct and Mistral-Nemo-Instruct, with AMC’23 selected for mathematical complexity and nuanced problem solving.
- Distilled-model behavior: The distilled Llama and Qwen variants exhibit tool underuse rather than the tool overuse targeted primarily by SMART.Tool underuse is attributed to possible overfitting to parametric reasoning and is measured on Time and AMC tasks.
- Distilled-model adaptation: SMARTAgent improves both distilled models on the time-domain QA benchmark after fine-tuning with the SMART paradigm.The adaptation tests whether SMART can address behavior beyond overuse in distilled models.
- Broader applicability: The results indicate that SMARTAgent promotes strategic tool engagement and applies across diverse reasoning paradigms and model types.This conclusion extends the reported effectiveness from mitigating overuse to addressing underuse.