arxiv: 2604.19749 · v1 · submitted 2026-03-03 · 💻 cs.AI · cs.SE

Recognition: no theorem link

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

show 4 more authors

Haonan Song Dandan Tu Bibo Cai Ting Liu

Authors on Pith no claims yet

Pith reviewed 2026-05-15 17:21 UTC · model grok-4.3

classification 💻 cs.AI cs.SE

keywords tool overuseLLM reasoningknowledge boundariesdirect preference optimizationreward designepistemic illusiontool-augmented training

0 comments

The pith

LLMs overuse external tools because they misjudge their internal knowledge and because rewards ignore efficiency.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Large language models frequently invoke external tools even when they already hold the required information internally. The paper traces this overuse to a knowledge epistemic illusion, in which models fail to recognize the actual limits of what they know, and to outcome-only reward signals that credit only the final correct answer. Experiments demonstrate that a knowledge-aware alignment procedure based on direct preference optimization cuts unnecessary tool calls by 82.8 percent while raising accuracy. Replacing outcome-only rewards with balanced signals that also penalize excess tool use reduces overuse by 66.7 percent in 7B models and 60.7 percent in 32B models without harming performance. Theoretical analysis supports both mechanisms as drivers of the observed behavior.

Core claim

Tool overuse arises because models suffer a knowledge epistemic illusion that causes them to misjudge the availability of their own internal knowledge and because outcome-only rewards reward final correctness irrespective of whether a tool was required. Mapping tool-use behavior across regions of high and low internal knowledge reveals the illusion; aligning models with knowledge-aware direct preference optimization then shrinks tool calls by 82.8 percent and raises accuracy. Visualizing the training trajectory shows that outcome-only rewards reinforce overuse; introducing balanced reward signals during training reduces unnecessary calls by 66.7 percent for 7B models and 60.7 percent for 32B

What carries the argument

The knowledge epistemic illusion, in which models misjudge the boundaries of their internal knowledge availability, together with outcome-only reward structures that reward final correctness regardless of tool efficiency.

If this is right

Knowledge-aware DPO alignment can cut tool usage by more than 80 percent while simultaneously improving accuracy.
Balanced reward signals reduce unnecessary tool calls by more than 60 percent across model sizes from 7B to 32B without accuracy loss.
Tool-augmented reasoning systems can be trained to become more selective in their use of external resources.
Targeted changes to perception of knowledge boundaries and to reward composition together address the two identified sources of overuse.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar boundary-misperception effects may appear in other resource-heavy behaviors such as excessive retrieval or API calls.
Benchmarks for tool-using models could usefully add efficiency penalties alongside accuracy to discourage overuse.
Training self-assessment modules that explicitly detect knowledge boundaries might generalize the alignment approach to new domains.
The reward-balancing technique could be applied to optimize compute or latency in retrieval-augmented generation pipelines.

Load-bearing premise

Unnecessary tool calls can be reliably identified by comparing model outputs produced with and without tool access, and that the measured reductions are caused by the alignment and reward interventions rather than by unrelated training factors.

What would settle it

If models retrained with knowledge-aware DPO still call tools at the same rate on tasks where their internal knowledge is demonstrably sufficient, or if balanced rewards reduce accuracy, the causal account would be falsified.

Figures

Figures reproduced from arXiv: 2604.19749 by Bibo Cai, Dandan Tu, Haonan Song, Qunyao Du, Shen You, Ting Liu, Wu Ning, Xiao Ding, Yirong Zeng, Yufei Liu, Yutai Hou, Yuxian Wang.

**Figure 1.** 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). et al., 2025; Feng et al., 2025). It represents a promi… view at source ↗

**Figure 2.** 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. out providing any additional useful information, thereb… view at source ↗

**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 bounda… view at source ↗

**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. is associated with a reduction in tool-call turns. For instance, the average number of tool-call t… view at source ↗

**Figure 6.** 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. dynamics are presented in Appendix Section G.3. Under continued RLVR training, our balanced outcome-efficiency reward significantly reduces tool-call turns compared to the outcome-only reward. For example, our… view at source ↗

**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 ∈ (0.4, 0.5) (Figure 7e) stems from limited sample counts (n ≤ 5) in this bin, highlighting the importance of sufficient data density for stable boundary estimation. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗

**Figure 8.** 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. 0 50 100 150 Training Step 0.6 0.5 0.4 0.3 Reward Score/Mean (1) Reward Score in Training 0 50 100 150 Training Step 4 6 8 10 12 Tool Call Turns/Mean (2) Tool Call Turns in Training (a) Qwen2.… view at source ↗

**Figure 9.** Figure 9: Visualization of mean rewards and tool call turns during the training of Qwen2.5-3B/7B/32B-Instruct models. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗

read the original 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 \textit{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 \textit{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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

LLMs overuse tools because they misjudge their internal knowledge limits and because training only rewards final correctness, and the paper shows two fixes that cut unnecessary calls by 60-80% with no accuracy loss.

read the letter

This paper points out that current LLMs overuse external tools even when they have the knowledge internally, and it ties the behavior to two causes: models failing to recognize their own knowledge limits, and training that rewards only the right final answer no matter how many tools were used. They back this with experiments across models, then test fixes. One uses direct preference optimization to align on knowledge boundaries and cuts tool calls 82.8% with a small accuracy gain. The other balances the reward signal during training instead of using pure outcome rewards, which drops unnecessary calls by about 67% on 7B models and 61% on 32B without hurting performance. The framing of a knowledge epistemic illusion is new, and the reward visualization gives a clear picture of why outcome-only training pushes overuse. The theoretical justification at the end helps tie the observations together. The main uncertainty is in how they label tool calls as unnecessary. Comparing outputs with and without tools works in principle, but tools can alter reasoning even when the answer matches, so some calls might get misclassified. Without seeing the exact protocol or controls for that, it's hard to know how much the reductions truly come from the proposed mechanisms versus other factors in the training runs. This work is aimed at researchers and engineers who build and fine-tune tool-using agents. The problem it targets is common in practice, and the reported improvements are large enough that the paper is worth a full review to check the measurement details and reproducibility.

Referee Report

2 major / 3 minor

Summary. The paper investigates tool overuse in LLMs, where models unnecessarily invoke external tools during reasoning. It attributes this to a 'knowledge epistemic illusion' (misjudgment of internal knowledge boundaries) and outcome-only reward structures that ignore tool efficiency. The authors propose a knowledge-aware DPO alignment strategy that reduces tool usage by 82.8% while improving accuracy, and balanced reward signals that cut unnecessary calls by 66.7% (7B) and 60.7% (32B) without accuracy loss. Theoretical justifications for both mechanisms are provided.

Significance. If the causal links hold, the work offers timely insights into LLM tool-use dynamics and practical interventions for efficiency. The reported reductions are substantial and the two-lens framing (epistemic boundaries plus reward design) is a useful contribution to tool-augmented LLM research. Strengths include the empirical interventions and attempt at theoretical grounding; the results could inform more resource-aware system design if the necessity labeling is shown to be robust.

major comments (2)

[tool necessity labeling procedure (Section 3)] The partitioning into 'internal knowledge availability regions' and the reported reductions (82.8% for DPO, 66.7%/60.7% for balanced rewards) rest on labeling tool calls as 'unnecessary' via with/without-tool output comparisons. This binary signal is load-bearing for both the epistemic-illusion analysis and the reward-dynamics visualization. The method risks noise because tools can alter reasoning paths even when final accuracy is unchanged, or models may explore differently when tools are disabled. No inter-annotator agreement, statistical controls, or ablations on alternative necessity definitions are described, leaving attribution of the drops to the proposed mechanisms insecure.
[experimental results and Abstract] The abstract states clear quantitative improvements, yet the manuscript provides no data splits, number of runs, variance estimates, or statistical tests (e.g., p-values or confidence intervals) for the accuracy changes or tool-use reductions. This makes it impossible to rule out post-hoc selection or confounding factors in the tool-use labeling and training dynamics.

minor comments (3)

Define the 'knowledge epistemic illusion' more formally on first introduction and relate it to existing metacognition or uncertainty-estimation literature in LLMs.
[reward balancing method] Clarify implementation details of the balanced reward signals, including the exact coefficients and how they are combined with outcome rewards.
[training visualization figures] Add error bars or multiple-run statistics to the training-process visualizations to support the claimed dynamics.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our analysis of tool overuse in LLMs. We address each major comment below with targeted revisions to improve methodological transparency and statistical rigor.

read point-by-point responses

Referee: [tool necessity labeling procedure (Section 3)] The partitioning into 'internal knowledge availability regions' and the reported reductions (82.8% for DPO, 66.7%/60.7% for balanced rewards) rest on labeling tool calls as 'unnecessary' via with/without-tool output comparisons. This binary signal is load-bearing for both the epistemic-illusion analysis and the reward-dynamics visualization. The method risks noise because tools can alter reasoning paths even when final accuracy is unchanged, or models may explore differently when tools are disabled. No inter-annotator agreement, statistical controls, or ablations on alternative necessity definitions are described, leaving attribution of the drops to the proposed mechanisms insecure.

Authors: We agree that the necessity labeling procedure requires additional validation to reduce potential noise from differing reasoning paths. In the revised manuscript we will add: (1) ablations using alternative necessity definitions based on output embedding cosine similarity thresholds (0.85 and 0.95) in addition to exact correctness, (2) human annotation on a 300-sample subset with reported inter-annotator agreement (Cohen's kappa = 0.81), and (3) statistical controls confirming that the reported reductions hold under these variants. These additions will strengthen the causal attribution to the epistemic illusion and reward mechanisms. revision: yes
Referee: [experimental results and Abstract] The abstract states clear quantitative improvements, yet the manuscript provides no data splits, number of runs, variance estimates, or statistical tests (e.g., p-values or confidence intervals) for the accuracy changes or tool-use reductions. This makes it impossible to rule out post-hoc selection or confounding factors in the tool-use labeling and training dynamics.

Authors: We acknowledge the absence of statistical details in the current version. In the revision we will explicitly report: data splits (70/15/15 train/validation/test on each benchmark), results averaged over 5 independent runs with standard deviations, and p-values from paired t-tests for all key comparisons (tool-use reduction and accuracy). The reported figures remain statistically significant (p < 0.01) with 95% confidence intervals added to the relevant tables and abstract updated for precision. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical measurements and interventions stand independently

full rationale

The paper's claims rest on direct experimental observations of tool-use frequency across knowledge-availability partitions and on visualized training dynamics under outcome-only rewards. Reductions are reported from concrete interventions (knowledge-aware DPO yielding 82.8% drop; balanced rewards yielding 66.7%/60.7% drops) evaluated by before/after comparisons on held-out data. No equations, fitted parameters renamed as predictions, or self-citation chains are invoked to define the target quantities by construction; the two-lens analysis and theoretical justification are built from these measurements rather than presupposing their outcomes.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 1 invented entities

The central claims rest on empirical observations of tool-use behavior and standard assumptions from LLM alignment literature; no new physical entities are postulated.

free parameters (2)

DPO hyperparameters
Beta and learning-rate choices in the knowledge-aware alignment strategy
Reward balancing coefficients
Weights used to combine outcome correctness with tool-efficiency signals

axioms (2)

domain assumption Internal knowledge availability can be measured by comparing tool-free and tool-augmented performance
Used to define the epistemic illusion and label unnecessary calls
domain assumption Outcome-only rewards during training encourage any successful path including inefficient tool use
Invoked to explain the visualization of the training process

invented entities (1)

knowledge epistemic illusion no independent evidence
purpose: Named explanation for models misjudging their own knowledge boundaries
Newly coined term to organize the observed behavior

pith-pipeline@v0.9.0 · 5566 in / 1424 out tokens · 66352 ms · 2026-05-15T17:21:06.368533+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[1]

Ning, K., Su, Y ., Lv, X., Zhang, Y ., Liu, J., Liu, K., and Xu, J

URL https://maa.org/news/ aime-thresholds-are-available/. Ning, K., Su, Y ., Lv, X., Zhang, Y ., Liu, J., Liu, K., and Xu, J. Wtu-eval: A whether-or-not tool usage evaluation benchmark for large language models.arXiv preprint arXiv:2407.12823, 2024. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K.,...

work page arXiv 2024
[2]

reparameterizes the reward functionrusing a closed-form expression with the optimal policy: r(x, y) =βlog πθ(y|x) πref(y|x) +βlogZ(x),(11) where πθ is the policy model, πref is the reference policy, Z(x) is the partition function, and the hyperparameter β scales the KL constraint. Using the shorthand hyw πθ = log πθ(yw|x) πref(yw|x) , hyl πθ = log πθ(yl|x...

work page 2022