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

Yilong Liu; Xixun Lin; Pengfei Cao; Ge Zhang; Fang Fang; Yanan Cao

cs.CLApr 13, 2026

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

Authors: Yilong Liu, Xixun Lin, Pengfei Cao, Ge Zhang, Fang Fang, Yanan Cao

Organizations: 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 · School of Computer Science and Technology, Donghua University, Shanghai, China

Abstract

Large language models (LLMs) have demonstrated impressive capabilities in utilizing external tools. In practice, however, LLMs are often exposed to tools that are irrelevant to the user's query, in which case the desired behavior is to refrain from invocations. In this work, we identify a widespread yet overlooked mechanistic flaw in tool refusal, which we term structural alignment bias: Even when a tool fails to serve the user's goal, LLMs still tend to invoke it whenever query attributes can be validly assigned to tool parameters. To systematically study this bias, we introduce SABEval, a new dataset that decouples structural alignment from semantic relevance. Our analysis shows that structural alignment bias induces severe tool-invocation errors in LLMs, yet remains largely unaccounted for in existing evaluations. To investigate the internal mechanisms underlying this bias, we propose Contrastive Attention Attribution, which reveals two competing pathways for semantic checking and structural matching. The relative strength of these pathways drives LLMs' tool invocation decisions. Based on these findings, we further introduce a rebalancing strategy that effectively mitigates structural alignment bias, as demonstrated by extensive experiments, without degrading general tool-use capabilities.

Figures & tables

Figure 1: Illustration of erroneous tool invocation driven by the model’s strong reliance on structural alignment.

Figure 2: SABEval construction with three steps. Samples are semantically irrelevant but structurally aligned.

Model	Random	SABEval	$\bm{\Delta}$
Qwen3-4B	0.16	40.04	+39.88
Qwen3-8B	0.04	34.26	+34.22
Qwen3-14B	0.04	41.86	+41.82
ToolACE-2.5-8B	0.12	40.67	+40.55
Watt-Tool-8B	0.00	10.51	+10.51

Table 1: Tool invocation rate (%) on randomly paired samples and SABEval (

\mathcal{D}_{0}

Figure 3: Tool invocation rate (%) on SABEval subsets

\mathcal{D}_{0},\dots,\mathcal{D}_{4}

with increasing structural alignment degree.

User Query: Can you tell me if ‘ Spider-Man: Miles Morales ’ is on sale on the Canadian PlayStation Store?
Original Tool: check_ nintendo_switch _game_on_sale Params: game_title , region
Counterfactual	Strategy	Tool Name	Parameters
Structural	Param Substitution	check_ nintendo_switch _game_on_sale	game_title , include_dlc
	Param Removal	check_ nintendo_switch _game_on_sale	game_title
	Param Addition	check_ nintendo_switch _game_on_sale	game_title , region , include_dlc
Semantic	Target Tool	check_ playstation _game_on_sale	game_title , region

Table 2: Example of counterfactuals. We construct structural counterfactuals by intervening on the original tool parameters via three strategies ( substitution , removal , and addition ) while retaining the original query. The semantic counterfactual includes a valid tool that serves the user’s goal, which is utilized for pathway discovery in § 5.1 .

Figure 4: Tool invocation rate (%) on original samples and their structural counterfactuals from

\mathcal{D}_{1}

Panel of Figure 5 — Figure 5: Semantic (left) and structural (right) pathways in Qwen3-8B from top- $k$ attention weights ( $k=23$ , corresponding to 2% of the attention heads).

Model	Method	SABEval $\bm{\downarrow}$					ACEBench $\bm{\uparrow}$
Model	Method	$\bm{\mathcal{D}_{0}}$	$\bm{\mathcal{D}_{1}}$	$\bm{\mathcal{D}_{2}}$	$\bm{\mathcal{D}_{3}}$	$\bm{\mathcal{D}_{4}}$	Single	Parallel
Qwen3-4B	Base	37.61	57.85	73.41	85.56	89.76	65	60
	Prompt	27.12	42.20	56.83	69.71	75.95	66	57
	CAA (Ours)	8.29	14.29	20.29	27.76	36.00	63	61
Qwen3-8B	Base	30.39	51.61	66.29	75.56	82.44	71	62
	Prompt	19.90	34.29	48.73	57.66	66.20	72	66

Table 3: Evaluation results on all subsets of SABEval and ACEBench. We report the tool invocation rate (

\downarrow

) on SABEval and accuracy (

\uparrow

) on ACEBench. Base denotes the performance of the original model.

Figure 6: Comparison of pathway strengths between invocation and non-invocation cases for Qwen3-8B.

Figure 7: TIR under different scaling coefficients for Qwen3-8B ( $k=23$ ).

Appendix figures & tables31 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: Distribution of derived class counts (left) and parameter counts (right) across tool templates.

Panel of Figure 9 — Figure 9: Tool invocation rate (%) across 10 randomly sampled tool templates. The red line denotes the overall TIR aggregated over these 10 groups, while the gray lines represent individual tool template groups.

Model	Add	Remove	Substitute
Qwen3-4B	38.81	68.57	70.30
Qwen3-8B	42.65	72.67	76.71
Qwen3-14B	42.43	66.48	73.76
ToolACE-2.5-8B	81.92	80.55	90.65
Watt-Tool-8B	86.63	91.05	96.45

Appendix

Table 4: Percentage of samples in which the probability of

w_{\text{tool}}

drops by more than 5% relative to its original value under structural counterfactuals.

Models	Original	Instantiate Param Count
Models	Original	1	2	3	4
$\operatorname{sim}(t,q)$	52.26	54.72	57.48	60.57	63.09
Qwen3-4B	89.76	87.39	81.56	73.50	65.11
Qwen3-8B	83.05	77.64	71.60	64.30	57.07
Qwen3-14B	90.44	88.10	82.14	77.96	70.99
ToolACE-2.5	71.47	69.66	70.00	66.71	63.09

Appendix

Table 5: TIR for cases flattened with different numbers of parameters (from

\mathcal{D}_{4}

of SABEval). The similarity score

\operatorname{sim}(t,q)

is also reported.

User Query: I need hiking shoes recommendations for males dealing with overpronation .
Tool Name: sandals _prep
Original	Tool Description: Provides a specific list of sandals recommendations for common foot conditions.
Original	Tool Parameters: foot_condition , gender
Flattened	Tool Description: Provides a specific list of male sandals recommendations for common foot conditions.
Flattened	Tool Parameters: foot_condition

Appendix

Table 6: An illustration of the structurally flattened case.

Model	Tool-call token	Tool-call token string	Tool-call format
Qwen3 family	151657	“ <tool_call> ”	<tool_call> {"name": <func-name>, "arguments": <args-json-object>} </tool_call>
ToolACE-2.5-8B	58	“ [ ”	[ func_1(param_a=value_a, param_b=value_b...),func_2(params)]
Watt-Tool-8B	58	“ [ ”	[ func_1(param_a=value_a, param_b=value_b...),func_2(params)]

Appendix

Table 7: Tool-call tokens

w_{\text{tool}}

, their string representations, and concrete tool-call formats

Model	Refusal tokens	Refusal tokens string	Example refusal
Qwen3 family	{785, 40, 4064, 2132}	{"The", "I", "None", "It"}	“ It seems there might be a misunderstanding. The provided function is for comparing VR headsets, not drones. If you’d like to compare VR headsets……”
ToolACE-2.5-8B	{791, 40}	{"The", "I"}	“ I ’m sorry, but I can only provide the current price for a game on the PlayStation platform. I do not have access to GOG prices.”
Watt-Tool-8B	{791, 40}	{"The", "I"}	“ The given question lacks the parameters required by the function. The available function is for basketball, not lacrosse.”

Appendix

Table 8: Refusal tokens for each model, corresponding string representations, and an example refusal. These tokens cover over 98% of refusal cases.

Strategy	Qwen3-8B		Qwen3-14B
Strategy	Disabled	Enabled	Disabled	Enabled
Original	47.2	30.8	50.8	29.2
Substitute	21.6	15.6	26.8	14.0

Appendix

Table 9: TIR (%) with thinking mode disabled and enabled on Qwen3-8B and 14B models, evaluated on a subset of

\mathcal{D}_{1}

Strategy	Qwen3-4B	Qwen3-8B	Qwen3-14B	Qwen3-32B
Original	60.44	53.43	60.40	62.22
Add	52.85	46.20	50.24	55.41
Remove	33.80	26.08	35.27	41.52
Substitute	32.08	23.17	29.90	34.47

Appendix

Table 10: TIR (%) across different Qwen3 model sizes on

\mathcal{D}_{1}

Method	$\mathcal{D}_{0}$	$\mathcal{D}_{1}$	$\mathcal{D}_{2}$	$\mathcal{D}_{3}$	$\mathcal{D}_{4}$
Base	30.39	51.61	66.29	75.56	82.44
Prompt	19.90	34.29	48.73	57.66	66.20
CAA (Ours)	4.29	6.49	7.27	8.98	11.12
One-shot (r)	63.71	84.05	91.22	94.78	96.98
One-shot (s)	30.54	49.37	60.14	69.80	79.95
Fine-tuning	1.41	2.48	2.10	3.51	4.39

Appendix

Table 11: TIR (%) of additional baselines on Qwen3-8B across SABEval. (r) and (s) denote raw and summarized one-shot formats, respectively.

Panel of Figure 10 — Figure 10: Pathways in Qwen3-4B.

Panel of Figure 11 — Figure 11: Pathways in Qwen3-14B.

Panel of Figure 12 — Figure 12: Pathways in ToolACE-2.5-8B.

Panel of Figure 13 — Figure 13: Pathways in Watt-Tool-8b.

Figure 14: Comparison of pathway strengths between invocation and non-invocation cases for Qwen3-4B.

Figure 15: Comparison of pathway strengths between invocation and non-invocation cases for Qwen3-14B.

Figure 16: Comparison of pathway strengths between invocation and non-invocation cases for ToolACE-2.5-8B.

Figure 17: Comparison of pathway strengths between invocation and non-invocation cases for Watt-Tool-8B.

Panel of Figure 18 — Figure 18: Pathway strengths across degrees of structural alignment for Qwen3-8B and ToolACE-2.5-8B.

Panel of Figure 19 — Figure 19: TIR at different top- $k$ thresholds under varying scaling coefficients for Qwen3-4B.

Panel of Figure 20 — Figure 20: TIR at different top- $k$ thresholds under varying scaling coefficients for Qwen3-8B.

Panel of Figure 21 — Figure 21: TIR at different top- $k$ thresholds under varying scaling coefficients for Qwen3-14B.

Panel of Figure 22 — Figure 22: TIR at different top- $k$ thresholds under varying scaling coefficients for ToolACE-2.5-8B.

Panel of Figure 23 — Figure 23: TIR at different top- $k$ thresholds under varying scaling coefficients for Watt-Tool-8B.

Span Name	Span Content
system-start	“<\|im_start\|>system\n”
tool-instruction	“# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n”
tool-start	“<tools>\n”
tool-definition	tool definition
tool-end	“</tools>\n\n”
output-instruction	For each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n"name": <function-name>, "arguments": <args-json-object>\n</tool_call>

Appendix

Table 12: Spans for Qwen3 models (4B, 8B, and 14B).

Span Name	Span Content
BOS	“<\|begin_of_text\|>”
system-start	“<\|start_header_id\|>system<\|end_ header_id\|>\n\n”
background	“Cutting Knowledge Date: December 2023\nToday Date: …\n\n”
tool-instruction	“You are an expert in composing functions. You are given a question and a set of possible functions. Based on the question, you will need to make one or more function/tool calls to achieve the purpose.\nIf none of the function can be used, point it out. If the given question lacks the parameters required by the function, also point it out.\nYou should only return the function call in tools call sections.\n\n”
output-instruction	“If you decide to invoke any of the function(s), you MUST put it in the format of [func_name1(params_ name1=params_value1, params_name2=params_value2…), func_name2(params)]\nYou SHOULD NOT include any other text in the response.\n”
tool-start	“Here is a list of functions in JSON format that you can invoke.\n”

Appendix

Table 13: Spans for ToolACE-2.5-8B and Watt-Tool-8B models (identical partitioning, except Watt-Tool-8B has no background span).

Panel of Figure 24 — Figure 24: An example annotation illustrating a base class, derived classes, and a tool template.

Panel of Figure 25 — Figure 25: Query-generation prompt used during dataset construction.

Panel of Figure 26 — Figure 26: Prompt for adding additional parameters during dataset extension.

Panel of Figure 27 — Figure 27: Default system prompt for Qwen3-4B, Qwen3-8B and Qwen3-14B

Panel of Figure 28 — Figure 28: Default system prompt for ToolACE-8B and Watt-Tool-8B

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

Beyond Semantic Similarity: Reducing Unnecessary API Calls via Behavior-Aligned Retriever
Aug 20, 2025Yixin Chen, Ying Xiong, Shangyu Wu +4In-Context Example Selection In-Context Retrieval
City University of Hong Kong · Mohamed Bin Zayed University of Artificial Intelligence · McGill University
When Tools Get in the Way: The Effect of Unnecessary Tool Availability on LLM Answering
Sep 12, 2026Saanvi Paturi, Arsen Kenzhebayev, Arham Sethi +3Tool-Augmented Language Model Agents LLM Tool Use
Spark AI Research · UWCSEA East Campus · Haileybury Astana +2
The Tool-Overuse Illusion: Why Does LLM Prefer External Tools over Internal Knowledge?
Mar 3, 2026Yirong Zeng, Shen You, Yufei Liu +9LLM Tool Use RL for Tool Use
Harbin Institute of Technology, SCIR · Huawei Technologies Co., Ltd · Peking University

Aug 20, 2025J/K move · Enter open · S save

Beyond Semantic Similarity: Reducing Unnecessary API Calls via Behavior-Aligned Retriever

Yixin Chen, Ying Xiong, Shangyu Wu, Yufei Cui+3

City University of Hong Kong · Mohamed Bin Zayed University of Artificial Intelligence · McGill University

In-Context Example Selection In-Context Retrieval Tool-Augmented Language Model Agents LLM Tool Use Tool Selection LLM Agent Reliability+1

Open paper AskPDF arXiv

Tool-augmented LLMs invoke external functions to extend their capabilities, but errors in the invocation decision, such as calling a tool when none is needed or omitting a needed call, can produce unreliable outputs and unnecessary cost. A lightweight remedy is to prepend retrieved examples so LLMs decide tool use in context. However, existing retrievers rank examples by semantic similarity alone. Lexically close or semantically close queries can require opposite behavior, so the retrieved examples may be behaviorally inconsistent and silently mislead the model. We propose Behavior Aligned Retrieval (BAR), a backbone-agnostic training recipe that teaches a dense retriever a behavior-aware similarity, keeping semantically related candidates close only when their tool-use behavior is compatible. BAR does not predict invocation labels; instead, it ranks demonstrations while leaving the final tool-use decision to the LLM. Applied to multiple retrieval backbones, including BERT, Contriever, and Qwen-based representation backbone, BAR consistently improves invocation reliability and reduces unnecessary API calls across 14 LLMs and 3 benchmarks.

Abstract

LLM Tool Use

Explore similar work

Beyond Semantic Similarity: Reducing Unnecessary API Calls via Behavior-Aligned Retriever

When Tools Get in the Way: The Effect of Unnecessary Tool Availability on LLM Answering

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