The Tool-Overuse Illusion: Why Does LLM Prefer External Tools over Internal Knowledge?

Yirong Zeng; Shen You; Yufei Liu; Qunyao Du; Xiao Ding; Yutai Hou; Yuxian Wang; Wu Ning

cs.AIMar 3, 2026

The Tool-Overuse Illusion: Why Does LLM Prefer External Tools over Internal Knowledge?

Authors: Yirong Zeng, Shen You, Yufei Liu, Qunyao Du, Xiao Ding, Yutai Hou, Yuxian Wang, Wu Ning,

Organizations: Harbin Institute of Technology, SCIR · Huawei Technologies Co., Ltd · Peking University

Abstract

Equipping LLMs with external tools effectively addresses internal reasoning limitations. However, it introduces a critical yet under-explored phenomenon: tool overuse, the unnecessary tool-use during reasoning. In this paper, we first reveal this phenomenon is pervasive across diverse LLMs. We then experimentally elucidate its underlying mechanisms through two key lenses: (1) First, by analyzing tool-use behavior across different internal knowledge availability regions, we identify a knowledge epistemic illusion: models misjudge internal knowledge boundaries and fail to accurately perceive their actual knowledge availability. To mitigate this, we propose a knowledge-aware epistemic boundary alignment strategy based on direct preference optimization, which reduces tool usage in by 82.8% while yielding an accuracy improvement. (2) Second, we establish a causal link between reward structures and tool-use behavior by visualizing the tool-augmented training process. It reveals that outcome-only rewards inadvertently encourage tool overuse by rewarding only final correctness, regardless of tool efficiency. To verify this, we balance reward signals during training rather than relying on outcome-only rewards, cutting unnecessary tool calls by 66.7% (7B) and 60.7% (32B) without sacrificing accuracy. Finally, we provide theoretical justification in this two lenses to understand tool overuse.

Figures & tables

Figure 1: An overview of tool overuse in LLMs. We provide a systematic investigation into why models invoke external tools unnecessarily. Our analysis reveals that this widespread phenomenon is driven by two primary mechanisms: knowledge epistemic illusion (a miscalibration of internal knowledge) and outcome-only rewards (optimization biases during training).

Model	Simple Samples (%)			Complex Samples (%)
Model	avg@8	avg@8 w/ tool	Tool Freq.	avg@8	avg@8 w/ tool	Tool Freq.
$\spadesuit$ GPT-5.2	99.28	95.01 ↓4.27	0.10	11.22	17.03 ↑5.81	0.20 ↑0.10
$\spadesuit$ Gemini-3	97.97	86.48 ↓11.49	0.86	11.91	62.86 ↑50.95	0.88 ↑0.02
$\spadesuit$ Claude-4.5	70.98	72.84 ↑1.86	0.44	23.38	68.54 ↑45.16	0.33 ↓0.11
$\spadesuit$ Deepseek-v3.2	99.29	94.18 ↓5.11	0.98	9.01	20.06 ↑11.05	1.26 ↑0.28
$\spadesuit$ Deepseek-R1	96.44	79.66 ↓16.78	0.68	16.92	52.84 ↑35.92	0.69 ↑0.01

Table 1: Evaluation of reasoning capabilities and tool-integration efficiency on GSM8K test benchmark. We report the average@8 accuracy and tool-use frequency across simple and complex samples of each LLM.

\spadesuit

denotes frontier models accessing by API,

\heartsuit

denotes RLVR-Trained models in TIR,

\clubsuit

denotes OSS foundation models. Tool Freq. denotes frequency of tool use for each example. In simple tasks, tool use slightly degrades performance, whereas in complex tasks, it substantially boosts accuracy. This finding highlights that tools are crucial for solving complex problems but should be avoided in simple ones.

Figure 2: The result of quantifying overuse in irrelevant tools. Higher scores indicate greater precision in tool selection, with fewer irrelevant calls. Top-tier models achieve only 80.2% performance on average, while open-source models get just 62.5% on average. It highlights that irrelevant tool overuse remains a pervasive issue, even among frontier models.

Panel of Figure 3 — Figure 3: Tool-use behavior and performance across the model’s internal knowledge availability. Contrary to the intuition that higher internal knowledge availability should correlate with fewer tool calls, the model exhibits a knowledge epistemic illusion. Notably, performance degrades when tools are invoked in high-availability regions ( $avg@1024>0.9$ ). This suggests that models often misjudge their internal boundaries, leading to an over-reliance on external tools even when internal knowledge is sufficient.

Panel of Figure 4 — Figure 4: Evaluation results comparing the base model and our K-DPO trained model. We report (1) the Avg@8 score with tools and (2) the knowledge–behavior correlation curve (measured by tool-call turns). Our approach reduces tool-call turns in higher $avg@1024$ ranges while improving overall tool-augmented performance.

Figure 5: Training dynamics of tool-call turns across model sizes. As RLVR training progresses, models increasingly overuse tools. The dashed line denotes the reasonable upper bound (tool@1024), which is consistently exceeded.

Figure 6: Training dynamics of tool-call turns under two reward schemes in 7B/32B-scale model. As RLVR training progresses, our balanced reward decrease the tool-call turns compared to the outcome-only reward.

Model	GSM8K test	AIME24	AIME25
Qwen2.5-7B	60.71 1.5	10.42 1.4	2.92 1.5
+ $\heartsuit$ Outcome	92.34 2.2	40.00 4.5	23.33 8.6
+ $\heartsuit$ Balanced	92.44 0.3	35.83 2.3	24.17 2.5
Qwen2.5-32B	91.39 1.0	23.33 1.4	15.0 1.4
+ $\heartsuit$ Outcome	94.14 0.5	56.89 3.2	54.67 4.1
+ $\heartsuit$ Balanced	94.90 0.1	59.17 1.5	53.33 1.7

Table 2: Performance comparison of tool-augmented models across three benchmarks. Bold indicates the highest accuracy (avg@8); subscripts denote tool-call turns, and its underlining marks the lowest value.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Table 9

$G$	$\tau$	GSM8K		AIME 24		AIME 25		Average
$G$	$\tau$	Acc. $\uparrow$	#Calls $\downarrow$	Acc. $\uparrow$	#Calls $\downarrow$	Acc. $\uparrow$	#Calls $\downarrow$	Acc. $\uparrow$	#Calls $\downarrow$
Qwen2.5-7B-inst		60.71	1.5	10.42	1.37	2.92	1.45	24.68	1.44
ReTool-7B		92.34	2.24	40.00	4.50	23.33	8.60	51.89	5.11
8	0.8	91.89	1.95	53.33	4.43	33.33	7.90	59.52	4.76
16	1.0	92.04	2.09	40.00	6.60	33.33	6.63	55.12	5.11
32	1.2	92.34	2.30	50.00	5.40	30.00	6.47	57.45	4.72

Appendix

Table 5: Ablation study on group size (

G

) and sampling temperature (

\tau

), compared with baseline models. We report Accuracy (%) and average Tool Calls (#Calls) across three benchmarks. Average represents the arithmetic mean of accuracy and calls across all datasets.

Model	GSM8K(1319)		AIME24/25 (60)
Model	Simple	Complex	Simple	Complex
$\spadesuit$ GPT-5.2	1260	59	50	4
$\spadesuit$ Gemini-3	1123	196	28	22
$\spadesuit$ Claude-4.5	360	959	22	37
$\spadesuit$ Deepseek-v3.2	1276	43	50	10
$\spadesuit$ Deepseek-R1	1137	182	13	38

Appendix

Table 6: Based on avg@8 metric, statistical distribution of the number after division for each LLM.

\spadesuit

denotes frontier models,

\heartsuit

denotes RLVR-Trained models in TIR,

\clubsuit

denotes OSS foundation models.

Model	Simple Samples (%)			Complex Samples (%)
Model	avg@8	avg@8 w/ tool	Tool Freq.	avg@8	avg@8 w/ tool	Tool Freq.
$\spadesuit$ GPT-5.2	99.75	52.25 ↓47.50	1.33	0.00	0.00 ↑0.00	0.47 ↓0.86
$\spadesuit$ Gemini-3	88.16	71.05 ↓17.11	1.92	23.86	39.20 ↑15.34	3.76 ↑1.84
$\spadesuit$ Claude-4.5	88.64	70.45 ↓18.19	0.76	9.80	16.89 ↑7.09	0.80 ↑0.04
$\spadesuit$ Deepseek-v3.2	95.25	73.75 ↓21.50	3.12	5.00	51.25 ↑46.25	5.50 ↑2.38
$\spadesuit$ Deepseek-R1	90.38	66.35 ↓24.03	0.13	8.55	12.83 ↑4.28	0.25 ↑0.12

Appendix

Table 7: Evaluation of reasoning capabilities and tool-integration efficiency on AIME 24/25 benchmark. We report the average@8 accuracy and tool invocation frequency across simple and complex task samples.

\spadesuit

denotes frontier models,

\heartsuit

denotes RLVR-Trained models in TIR,

\clubsuit

denotes OSS foundation models. * this model specially suppress invalid tool-use in RLVR training.

Panel of Figure 7 — Figure 7: Tool-use behavior and performance across the model’s epistemic boundary for other models. An anomalous dip in Qwen2.5-32B-Instruct’s tool-augmented curve within $avg@1024\in(0.4,0.5)$ (Figure 7(e) ) stems from limited sample counts ( $n\leq 5$ ) in this bin, highlighting the importance of sufficient data density for stable boundary estimation.

Figure 8: Comparison of reward trajectories during the training process for Qwen-7B (left) and Qwen-32B (right). The solid lines represent the initial training phase, while the dashed lines denote the continual training stage.

Panel of Figure 9 — Figure 9: Visualization of mean rewards and tool call turns during the training of Qwen2.5-3B/7B/32B-Instruct models.

LLM Tool Use

Large language model tool use studies how models select, call, and coordinate external tools to extend their reasoning and take actions.

See all 295 papers

Spurious Tool Use: When RL Agents Learn the Wrong Reason to Act
Sep 14, 2026Yiwei Yang, Haoxiang Zhang, Bingbing Wen +6LLM Agent Reliability RL for Tool Use
University of Washington · University of California San Diego · Stanford University
Are Tools All We Need? Unveiling the Tool-Use Tax in LLM Agents
Apr 30, 2026Kaituo Zhang, Zhen Xiong, Mingyu Zhong +4Tool-Augmented Language Model Agents Agentic Reasoning
University of Houston · University of Southern California · 3New York University +2
Do LLMs Know Tool Irrelevance? Demystifying Structural Alignment Bias in Tool Invocations
Apr 13, 2026Yilong Liu, Xixun Lin, Pengfei Cao +3LLM Tool Use Tool-Use Evaluation
Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China · School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China · Institute of Automation, Chinese Academy of Sciences, Beijing, China +1

Sep 14, 2026J/K move · Enter open · S save

Spurious Tool Use: When RL Agents Learn the Wrong Reason to Act

Yiwei Yang, Haoxiang Zhang, Bingbing Wen, Yao Lu+5

University of Washington · University of California San Diego · Stanford University

LLM Agent Reliability RL for Tool Use

Open paper AskPDF arXiv

Large language model (LLM) agents increasingly interleave natural language reasoning with external tools such as web search and code execution. These tool-use policies are often optimized via reinforcement learning (RL), which can amplify spurious correlations in the training data. In this work, we study when and why RL-trained agents learn shortcut tool-selection policies: invoking tools based on superficial prompt cues rather than genuine task requirements. We construct controlled synthetic environments combining factual question answering and mathematical reasoning tasks, and inject cues that are strongly correlated with specific tools during training but causally irrelevant to tool necessity. Across counterfactual evaluations where cues are present but the associated tools are not required, agents exhibit substantial shortcut behavior, with spurious tool invocation rates increasing by up to 39 percent. However, shortcut formation is not universal: across the conditions we test, it arises only when the agent has already learned to use the target tool reliably, suggesting that task competence, rather than dataset imbalance alone, is a key factor in shortcut learning. A swapped-cue analysis further shows that semantic alignment between cues and tools substantially amplifies this effect. To mitigate these failures, we introduce a dense, decision-level reward in which an LLM judge evaluates the necessity of each tool call. This tool-necessity reward effectively suppresses cue-driven tool use while preserving task performance, providing a practical approach to improving the robustness of LLM agent tool-use policies.

Abstract

LLM Tool Use

Explore similar work

Spurious Tool Use: When RL Agents Learn the Wrong Reason to Act

Are Tools All We Need? Unveiling the Tool-Use Tax in LLM Agents

Do LLMs Know Tool Irrelevance? Demystifying Structural Alignment Bias in Tool Invocations