Wired for Overconfidence: A Mechanistic Perspective on Inflated Verbalized Confidence in LLMs

Tianyi Zhao; Yinhan He; Wendy Zheng; Yujie Zhang; Chen Chen

cs.CLApr 1, 2026

Wired for Overconfidence: A Mechanistic Perspective on Inflated Verbalized Confidence in LLMs

Authors: Tianyi Zhao, Yinhan He, Wendy Zheng, Yujie Zhang, Chen Chen

Organizations: University of Virginia

Abstract

Large language models are often not just wrong, but confidently wrong: when they produce factually incorrect answers, they tend to verbalize overly high confidence rather than signal uncertainty. Such verbalized overconfidence can mislead users and weaken confidence scores as a reliable uncertainty signal, yet its internal mechanisms remain poorly understood. We present a circuit-level mechanistic analysis of this inflated verbalized confidence in LLMs, organized around three axes: capturing verbalized confidence as a differentiable internal signal, identifying the circuits that causally inflate it, and leveraging these insights for targeted inference-time recalibration. Across two instruction-tuned LLMs on three datasets, we find that a compact set of MLP blocks and attention heads, concentrated in middle-to-late layers, consistently writes the confidence-inflation signal at the final token position. We further show that targeted inference-time interventions on these circuits substantially improve calibration. Together, our results suggest that verbalized overconfidence in LLMs is driven by identifiable internal circuits and can be mitigated through targeted intervention.

Figures & tables

Figure 1: Left : Two-step elicitation. The model answers a factual question, then self-reports confidence as an integer (0–99). Center : Truth-injection counterfactual design. For each confidently-wrong record, the clean prompt retains the model’s incorrect answer, while the corrupted prompt substitutes the ground truth, keeping all other tokens identical. Right : $\Delta$ TSLD distributions and $\Delta$ TSLD vs. $\Delta$ confidence scatter plots for Qwen2.5-3B (top) and Llama-3.2-3B (bottom) on PopQA.

Figure 2: Attribution heatmaps across all six model $\times$ dataset configurations.

Qwen2.5-3B-Instruct				Llama-3.2-3B-Instruct
Rank	PopQA	MMLU	NQOpen	Rank	PopQA	MMLU	NQOpen
1	m30	m35	m30	1	a14.h3	m0	a14.h3
2	m35	m30	m35	2	m0	a14.h3	m0
3	m26	m27	m31	3	m27	m4	a13.h18
4	m31	a24.h5	m26	4	a13.h18	m25	m4
5	m25	m24	m25	5	m23	m2	m27

Table 1: Top-10 components by attribution score. Components in blue are shared across all three datasets. Prefix m : MLP layer; prefix a : attention head (e.g. a14.h3 = layer 14, head 3).

Panel of Figure 3 — Figure 3: (a) Faithfulness as a function of retained top- $k$ edges. (b) TSLD reduction when each of the top-10 components is individually ablated.

Model	Dataset	Baseline TSLD	Compl. Drop	Completeness%
Qwen2.5-3B-Instruct	PopQA	10.05	7.92	78.8%
	NQOpen	9.68	7.25	74.9%
Llama-3.2-3B-Instruct	PopQA	4.79	3.56	74.3%
	NQOpen	4.55	3.37	74.1%

Table 2: Completeness validation. Completeness drop

=

full-model TSLD

-

ablated-sub-network TSLD, where the ablated sub-network retains all edges except the top-3,000.

Figure 4: Incremental component ablation : mean TSLD as the top-

k

components are jointly ablated, for

k=1

10

. Dashed lines mark the clean (pre-ablation) mean TSLD.

Model	Method	$\alpha$	ECE	Brier	ECE Impr.	Brier Impr.
Qwen2.5-3B-Instruct	baseline	–	0.492	0.430	–	–
	mean ablation	–	0.295	0.244	40.0%	43.1%
	steering	0.3	0.245	0.193	50.3%	55.1%
	steering	0.4	0.166	0.148	66.4%	65.7%
	steering	0.5	0.111	0.130	77.5%	69.7%
	steering	0.6	0.145	0.146	70.7%	66.1%

Table 3: Intervention results on PopQA. Best ECE and Brier per model in bold .

Figure 5: Reliability curves for Llama-3.2-3B-Instruct across all three datasets under baseline (before intervention), mean ablation, and steering at

\alpha\in\{0.2,0.3,0.4,0.5,0.6,0.7\}

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Raw verbalized confidence distributions across all six model

\times

dataset configurations. Green: correct predictions; red: incorrect predictions.

Substituted Step 1 answer	Mean $\Delta_{\text{TSLD}}$	Top-10 overlap
Ground truth	$-8.50$	10/10
Same-meaning gold alias	$-8.24$	9/10
Type-matched plausible-wrong answer	$-7.71$	10/10
Length-matched distractor	$-8.68$	7/10

Appendix

Table 4: Answer-substitution controls on Qwen2.5-3B-Instruct/PopQA. Mean

\Delta_{\text{TSLD}}

measures the change relative to the original incorrect-answer context, with more negative values indicating a larger confidence decrease. Top-10 overlap is measured against the circuit recovered using ground-truth substitution.

Step 2 confidence prompt	Mean $\Delta_{\text{TSLD}}$	Top-5 overlap	Top-10 overlap
Original prompt	$-7.07$	5/5	10/10
“How certain are you that it is correct?”	$-10.63$	4/5	8/10
“On a scale from 0 to 100, how sure are you?”	$-3.75$	3/5	7/10
“How likely is it that your answer is accurate?”	$-3.35$	4/5	6/10

Appendix

Table 5: Confidence-prompt robustness on Qwen2.5-3B-Instruct/PopQA. The circuit remains partially conserved under three semantically related Step 2 prompt variants.

High/low target sets	Top-10 Jaccard vs. default
$\mathcal{H}=\{70,75,80,85,90,99\}$ ; $\mathcal{L}=\{0,10,15,20,25,30\}$	1.00
$\mathcal{H}=\{80,85,90,99\}$ ; $\mathcal{L}=\{0,10,20\}$	1.00
$\mathcal{H}=\{60,65,70,75,80,85,90,95,99\}$ ; $\mathcal{L}=\{0,5,10,15,20,25,30,35,40\}$	0.82
$\mathcal{H}=\{80,90,99\}$ ; $\mathcal{L}=\{0,15,30\}$	0.82
$\mathcal{H}=\{75,80,85,90,95,99\}$ ; $\mathcal{L}=\{5,10,15,20,25,35\}$	0.82

Appendix

Table 6: Sensitivity of Qwen2.5-3B-Instruct/PopQA circuit discovery to TSLD target sets. The recovered top-10 components remain stable across narrower, wider, coarser, and shifted target sets.

Figure 7: Reliability curves for Qwen2.5-3B-Instruct across PopQA (top), MMLU (middle), and NQOpen (bottom) under baseline (before intervention), mean ablation, and steering at

\alpha\in\{0.2,0.3,0.4,0.5,0.6,0.7\}

Figure 8: Steering

\alpha

sweep across all six configurations. ECE (teal, left axis) and Brier score (purple, right axis) as a function of steering strength

\alpha

. Dashed grey: baseline ECE; dashed yellow: mean ablation ECE; star: optimal

\alpha

Model / Dataset	Raw ECE	Temp.	Platt	Isotonic	Histogram	Mean abl. (ours)	Best steer (ours)
Llama-3.2-3B / PopQA	0.568	0.302	0.008	0.075	0.147	0.018	0.020
Qwen2.5-3B / NQOpen	0.555	0.304	0.016	0.054	0.033	0.374	0.103
Llama-3.2-3B / MMLU	0.176	0.021	0.026	0.078	0.232	0.166	0.074
Llama-3.2-3B / NQOpen	0.507	0.245	0.036	0.057	0.191	0.245	0.088

Appendix

Table 7: Comparison with post-hoc calibrators (ECE, 10 bins; lower is better). Calibrators are fit on a labeled 50% split and evaluated on the disjoint half; Raw ECE is computed on the evaluation split. Mean ablation and best steering restate our full-dataset results from Tables 3 , 11 , and 12 .

Rank	PopQA	NQOpen
1	m27	m27
2	m21	m21
3	m23	m26
4	m26	m23
5	m20	m20
6	m19	m19

Appendix

Table 8: Top-10 components by attribution score for Qwen2.5-7B-Instruct. 9 of the top-10 components are shared across PopQA and NQOpen.

Dataset	Baseline	$\alpha=0.2$	$\alpha=0.3$	$\alpha=0.4$	$\alpha=0.5$	$\alpha=0.6$	Mean ablation
PopQA	0.647	0.419	0.312	0.203	0.121	0.068	0.155
NQOpen	0.626	0.528	0.450	0.347	0.217	0.109	0.615

Appendix

Table 9: Intervention results (ECE) for Qwen2.5-7B-Instruct. Steering sweep over

\alpha

and mean ablation; best ECE per dataset in bold .

Bucket	Criterion	Role
1	$\Delta_{\text{TSLD}}\leq-\tau$	Overconfident collapse; circuit discovery target
2	$\Delta_{\text{TSLD}}\geq+\tau$	Truth recognition; excluded
3	$\Delta_{\text{TSLD}}\approx 0$	Noise; excluded

Appendix

Table 10: Three-bucket stratification.

Model	Method	$\alpha$	ECE	Brier	ECE Impr.	Brier Impr.
Qwen2.5-3B-Instruct	baseline	–	0.281	0.310	–	–
	mean ablation	–	0.261	0.298	7.3%	3.8%
	steering	0.3	0.250	0.294	11.1%	5.2%
	steering	0.4	0.230	0.284	18.2%	8.3%
	steering	0.5	0.196	0.271	30.3%	12.7%
	steering	0.6	0.188	0.263	33.1%	15.3%

Appendix

Table 11: Intervention results on MMLU. Best ECE and Brier per model in bold .

Model	Method	$\alpha$	ECE	Brier	ECE Impr.	Brier Impr.
Qwen2.5-3B-Instruct	baseline	–	0.551	0.485	–	–
	mean ablation	–	0.374	0.304	32.1%	37.3%
	steering	0.3	0.295	0.244	46.5%	49.6%
	steering	0.4	0.201	0.191	63.5%	60.7%
	steering	0.5	0.141	0.169	74.5%	65.1%
	steering	0.6	0.103	0.159	81.2%	67.2%

Appendix

Table 12: Intervention results on NQOpen. Best ECE and Brier per model in bold .

Figure 9: Verbal confidence circuit vs. factual retrieval circuit (Qwen2.5-3B-Instruct, PopQA). Both circuits are discovered on the same Bucket 1 records with the same EAP-IG machinery; the factual retrieval circuit uses an answer-logit objective in place of TSLD. The two circuits are structurally distinct: the Jaccard overlap of their top-10 (top-20) components is 0.18 (0.21), and the cosine similarity between their signed edge-attribution scores on the induced subgraphs is 0.004 (0.055).

Confidence Estimation in Language Models

Confidence estimation in language models measures whether a model’s uncertainty signals predict answer correctness and support reliable decisions.

See all 225 papers

Jun 2, 2026

Large Language Models Are Overconfident in Their Own Responses

Prior work has shown that instruction-tuned large language models (LLMs) are less well calibrated than their base pre-trained counterparts. However, little is known about the frequently used chat template's effect on the calibration of conversational LLMs. In this work, we investigate the mechanisms driving this miscalibration by decoupling the effects of the post-training algorithm and the chat format. We find that, while instruction tuning fundamentally harms calibration, the chat template aggravates the issue through an "ownership bias" -- models are significantly more confident in their own answers than in identical answers provided by a user. Extensive experiments across six recent open-weight LLMs, three benchmarks, and three confidence elicitation methods show that models assign up to 26% higher confidence to their own responses. Leveraging this insight, we propose a simple inference-time strategy: framing the model's answer as user input during confidence elicitation. This approach significantly reduces overconfidence and improves calibration by up to 26% without the need for retraining, narrowing the gap between base and instruction-tuned models.

Mario Sanz-Guerrero, Manuel Mager, Katharina von der Wense

∇Johannes Gutenberg University Mainz, Germany · ♠University of Colorado Boulder, USA

Confidence Estimation in Language Models LLM Prompting LLM Fine-Tuning Language Model Calibration LLM Reliability LLM Uncertainty Estimation

Ask PDF

Mar 26, 2026

Closing the Confidence-Faithfulness Gap in Large Language Models

Large language models (LLMs) tend to verbalize confidence scores that are largely detached from their actual accuracy, yet the geometric relationship governing this behavior remain poorly understood. In this work, we present a mechanistic interpretability analysis of verbalized confidence, using linear probes and contrastive activation addition (CAA) steering to show that calibration and verbalized confidence signals are encoded linearly but are orthogonal to one another -- a finding consistent across three open-weight models and four datasets. Interestingly, when models are prompted to simultaneously reason through a problem and verbalize a confidence score, the reasoning process disrupts the verbalized confidence direction, exacerbating miscalibration. We term this the "Reasoning Contamination Effect." Leveraging this insight, we introduce a two-stage adaptive steering pipeline that reads the model's internal accuracy estimate and steers verbalized output to match it, substantially improving calibration alignment across all evaluated models.

Miranda Muqing Miao, Lyle Ungar

University of Pennsylvania · Department of Computer and Information Science

Confidence Estimation in Language Models Language Model Calibration LLM Interpretability Activation Steering

Ask PDF

Apr 7, 2026

Identifying Influential N-grams in Confidence Calibration via Regression Analysis

While large language models (LLMs) improve performance by explicit reasoning, their responses are often overconfident, even though they include linguistic expressions demonstrating uncertainty. In this work, we identify what linguistic expressions are related to confidence by applying the regression method. Specifically, we predict confidence of those linguistic expressions in the reasoning parts of LLMs as the dependent variables and analyze the relationship between a specific

n

-gram and confidence. Across multiple models and QA benchmarks, we show that LLMs remain overconfident when reasoning is involved and attribute this behavior to specific linguistic information. Interestingly, several of the extracted expressions coincide with cue phrases intentionally inserted on test-time scaling to improve reasoning performance. Through our test on causality and verification that the extracted linguistic information truly affects confidence, we reveal that confidence calibration is possible by simply suppressing those overconfident expressions without drops in performance.

Shintaro Ozaki, Wataru Hashimoto, Hidetaka Kamigaito +2

Nara Institute of Science and Technology (NAIST) · The University of Tokyo

Language Model Calibration Confidence Estimation in Language Models LLM Reliability Causal Interventions in Language Models

Ask PDF

Abstract

Confidence Estimation in Language Models

Explore similar work

Large Language Models Are Overconfident in Their Own Responses

Closing the Confidence-Faithfulness Gap in Large Language Models

Identifying Influential N-grams in Confidence Calibration via Regression Analysis