Tool Use Can Backfire in LLMs When Internal Reasoning Is Already Strong

Integrating external tools and evidence into large language models is widely seen as a straightforward way to enhance reasoning. But a new study challenges this assumption, demonstrating that the relationship is far more nuanced—and potentially counterproductive in certain scenarios.

The Problem with Blind Deference#

Researchers working with Python and Wikipedia tools discovered a striking pattern: external evidence reliably benefits reasoning when a model's initial beliefs are uncertain or weak. However, when the model already holds strong internal support for a correct answer, adding external evidence can flip that answer to an incorrect one. The core issue lies in how models allocate decision-making authority between their internal reasoning and external inputs.

The team frames this as a misallocation of revision authority—the model defers to external evidence even when its own internal reasoning is actually more reliable than what the tool provides. This happens because many systems default to treating tool outputs as authoritative, rather than weighing them against the model's existing confidence level.

Where the Harm Concentrates#

The researchers predicted that damage from misleading external evidence should cluster on high-confidence scenarios—cases where the model would have gotten the answer right without tool assistance. Testing this hypothesis through threshold localization and auditing methods, they confirmed that errors concentrate precisely where the model's unaided reasoning was strongest.

A separate intervention experiment proved particularly revealing: simply changing how the evidence is framed—altering the revision context—measurably changed how much damage misleading information caused.

A Practical Solution#

To address this problem, the team introduced CASE, a label-free controller that selects between unaided reasoning and tool-assisted pathways based on the model's answer-state certainty. Rather than always trusting external tools, CASE arbitrates which approach should take the lead. Early results show it outperforms existing confidence-based methods for deciding when to rely on tools.

These findings suggest a broader principle: effective tool integration requires mixed inference strategies that evaluate whether external evidence genuinely improves upon internal reasoning, rather than assuming it always will.