Modern large language models (LLMs) often exhibit overconfident errors, assigning high confidence to incorrect predictions and thereby undermining reliability in real-world use. Existing calibration methods typically rely on post-hoc adjustment or retraining, operating at the output level without addressing the underlying causes of miscalibration. We propose CoCHI, (Class-oriented Causal Head Intervention), a framework that improves calibration by identifying and intervening on internal components responsible for distinct confidence behaviours. By analysing predictions through correctness-confidence patterns, CoCHI enables targeted intervention on model internals, reducing overconfident errors while preserving correct predictions. Our approach provides a mechanistic perspective on calibration, linking reliability to identifiable structures within the model rather than treating it as an output-level issue. Experiments across multiple LLMs show consistent improvements in calibration metrics, including Expected Calibration Error, Negative Log-Likelihood, and Brier Score, while maintaining accuracy on five multiple-choice question answering benchmarks. These results suggest that effective calibration requires not only adjusting outputs, but also addressing the underlying causal mechanisms within the model.
Figures & tables
Figure 1: The framework of CoCHI.
MMLU
MedMCQA
Method
Accuracy ↑
ECE ↓
NLL ↓
BS ↓
Accuracy ↑
ECE ↓
NLL ↓
BS ↓
Gemma-2-9B
TS
0.6653 ± 0.0050
0.0660 ± 0.0173
0.5060 ± 0.0335
0.1683 ± 0.0125
0.5607 ± 0.0186
0.0556 ± 0.0261
0.6114 ± 0.0151
0.2122 ± 0.0066
ITI
0.6693 ± 0.0060
0.0605 ± 0.0062
0.5029 ± 0.0233
0.1675 ± 0.0105
0.5673 ± 0.0200
0.0451 ± 0.0140
0.6085 ± 0.0154
0.2122 ± 0.0070
Therm.
0.6653 ± 0.0050
0.1831 ± 0.0567
1.8729 ± 1.1455
0.2302 ± 0.0429
0.5607 ± 0.0186
0.2849 ± 0.0106
2.9242 ± 0.4188
0.3293 ± 0.0165
ATS
0.6653 ± 0.0050
0.3784 ± 0.0052
0.9191 ± 0.0066
0.3553 ± 0.0030
0.5607 ± 0.0186
0.3534 ± 0.0783
4.1559 ± 1.3606
0.3702 ± 0.0396
Table 1: Results on four benchmarks (MMLU, MedMCQA, OpenBookQA, and MedQA-USMLE). Best results are highlighted in bold.
Mask Rule
Masked Acc ↑
Masked ECE ↓
Wrong-high facilitating only
0.2044 ± 0.0038
0.4745 ± 0.1477
Correct-high non-facilitating only
0.2389 ± 0.0398
0.1518 ± 0.0474
Correct-high interfering only
0.4267 ± 0.0462
0.0694 ± 0.0488
Canonical union (broad)
0.2189 ± 0.0271
0.3850 ± 0.1816
Random size-matched
0.2111 ± 0.0077
0.4330 ± 0.2736
CoCHI
0.6120 ± 0.0626
0.0490 ± 0.0089
Table 2: Ablation study of different masking strategies on MMLU.
τc
Accuracy ↑
ECE ↓
NLL ↓
BS ↓
0.2
0.5933 ± 0.0217
0.0555 ± 0.0096
0.5385 ± 0.0075
0.1808 ± 0.0032
0.3
0.6060 ± 0.0059
0.0643 ± 0.0102
0.5319 ± 0.0043
0.1781 ± 0.0017
0.4
0.5860 ± 0.0118
0.0665 ± 0.0108
0.5276 ± 0.0058
0.1761 ± 0.0025
0.5
0.6120 ± 0.0626
0.0490 ± 0.0089
0.5168 ± 0.0236
0.1743 ± 0.0008
0.7
0.5873 ± 0.0377
0.0705 ± 0.0099
0.5548 ± 0.0212
0.1873 ± 0.0083
0.9
0.5933 ± 0.0081
0.0696 ± 0.0277
0.5255 ± 0.0059
0.1748 ± 0.0022
Table 3: Sensitivity analysis of the confidence threshold τc on MMLU.
Method
GPU
Cal. time (s) ↓
Artifact size ↓
Inf.
Prunable
TS
No
0.5
4 B
Negligible
No
ATS
No
25.8
16 B
Negligible
No
ITI
Yes
18.4
16 KB
O(k⋅d)
No
Therm.
No
19.4
1.0 MB
O(H⋅C)
No
CoCHI
Yes
69.9
40 B
None
Yes
Table 4: Efficiency comparison of calibration methods using 40 calibration samples on MMLU.
Figure 2: ECE comparison between the unmasked model and CoCHI on correct_low and wrong_high subsets.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Domain
# Questions
MMLU
Multi-domain knowledge
14,042
MedQA-USMLE
Medical exam
11,451
OpenBookQA
Science
∼ 4,200
SciQ
Science
13,679
MedMCQA
Medical exam
187,005
Appendix
Table 5: Statistical summary of the datasets.
Method
Acc ↑
ECE ↓
NLL ↓
BS ↓
Gemma-2-9B
TS
0.9173 ± 0.0012
0.0579 ± 0.0184
0.4699 ± 0.4226
0.0641 ± 0.0094
ITI
0.9253 ± 0.0083
0.1027 ± 0.0091
0.2337 ± 0.0076
0.0617 ± 0.0025
Therm.
0.9223 ± 0.0021
0.0609 ± 0.0066
1.0836 ± 0.1824
0.0599 ± 0.0068
ATS
0.9173 ± 0.0012
0.0607 ± 0.0067
1.1578 ± 0.1183
0.0602 ± 0.0065
CoCHI
0.9353 ± 0.1069
0.0563 ± 0.0199
0.2309 ± 0.1276
0.0548 ± 0.0472
Appendix
Table 6: Results on SciQ. Best results are highlighted in bold.
Model
Cal. time (s) ↓
Min
Max
Peak GPU (GiB)
Mistral-7B-v0.1
39 ± 21
21
80
28.1
Llama-3.1-8B
58 ± 20
41
93
30.4
Gemma-2-9B
88 ± 23
72
159
46.2
Appendix
Table 7: Calibration time (s; mean ± std) of CoCHI across three model families, averaged over 15 runs per model. Costs include class-conditioned forward passes and exclude model loading, and scale with model size while remaining practical.
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
Large language models often express high confidence in answers that are wrong. Standard calibration remedies typically act globally or at the score level, reducing unwarranted confidence but also risking erosion of warranted confidence on correct answers. We introduce Probe-Conditioned Head Intervention (PCHI), an inference-time method that uses a frozen probe to detect likely wrong-but-confident responses and conditionally rescales downstream attention-head outputs during confidence generation. On Qwen3-4B-Instruct solving OpenMathInstruct problems with a structured binary confidence field, readout-token PCHI converts 82.2% of originally wrong-yes confidence readouts to no, while a joint intervention across upstream confidence-template tokens reduces ECE from 21.9% to 9.2% and damages only 5.1% of originally correct-yes readouts. The readout-token effect also appears on Gemma3-4B, though upstream interventions are weaker and more mask-dependent. These results show that verbalized overconfidence can be selectively reduced through conditionally applied internal intervention, partially decoupling the suppression of unwarranted confidence from the loss of warranted confidence.
Ke Li, Chongzhe Zhang, Zifan Zeng +3
Huawei Heisenberg Research Center · EPFL · TU Berlin +1
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.