MECHANISTIC INTERPRETABILITY · 2026

Why are LLMs
confidently wrong?

Better confidence starts inside the model.

Correcting overconfident errors through Class-oriented Causal Head Intervention. Find the heads that drive unreliable confidence. Intervene precisely.

Explore the researchDiscover CoCHI
Jing Ren†Bowen Li†Ziqi Xu*Xuechao YangFeng Xia

RMIT University † Equal contribution   * Corresponding author

CAUSAL HEAD EXPLORERILLUSTRATIVE
ATTENTION HEADSTARGETED INTERVENTION
Error pathway Preserved

Trace the mechanisms behind model confidence.

CALIBRATION, FROM THE INSIDE OUT.Scroll to explore ↓
3LLM families
5QA benchmarks
≤40Calibration examples / dataset
8–10Heads selected per model

01 / THE IDEA

Confidence is a signal.
It should be a reliable one.

A model can be certain and still be wrong. CoCHI asks which internal components make that happen.

Most calibration methods adjust output probabilities. CoCHI instead examines the model’s attention heads, linking their causal roles to distinct patterns of correctness and confidence.

Using a small calibration set, it identifies heads that facilitate overconfident errors and components that do not support confident, correct predictions. A compact, protected mask suppresses selected heads during inference while keeping the pretrained model parameters frozen.

FOUR CALIBRATION CLASSESτc = 0.5
01Correct / High

Protect the heads that support reliable answers.

02Correct / Low

Right answer. Uncertain confidence.

03 PRIMARY TARGETWrong / High

Find the mechanisms behind confident errors.

04Wrong / Low

Wrong answer. Uncertainty acknowledged.

Classes are defined on labeled calibration data, not by access to test-time correctness.

02 / UNDER THE HOOD

Small intervention.
Causal impact.

From behavioral patterns to a compact structural mask. Four steps, one frozen model.

01 / DECOMPOSE

Separate behaviors

Partition predictions by correctness and confidence to isolate overconfident errors.

correctness × confidence
02 / ATTRIBUTE

Discover causal roles

Optimize head gates on each class to identify facilitating and non-facilitating heads.

necessary + sufficient
03 / CONSTRUCT

Build a protected mask

Combine error and noise heads, protect correct-high facilitators, then rank and select top-k.

Hmask = Herr ∪ Hnoise
04 / INTERVENE

Gate at inference

Suppress the selected head outputs. Preserve the remaining computation and model weights.

z̃h(x) = gh · zh(x)
THE COMPLETE FRAMEWORKOpen figure ↗
CoCHI framework: calibration class decomposition, class-conditioned causal head attribution, canonical-union mask construction, and inference-time intervention.
FIG. 01 Class-oriented Causal Head Intervention. Calibration from the inside out.

Keep what makes the model right.
Heads that facilitate correct-high predictions are protected from masking, with an additional ranking penalty to preserve useful computation.

03 / THE EVIDENCE

More reliable confidence.
Across models and domains.

Explore reported results across general knowledge, science, and medical QA.

Reported benchmark results
MethodAccuracy ↑ECE ↓NLL ↓Brier Score ↓

Results shown as mean ± standard deviation. Best means in each metric are highlighted. ECE, NLL, and Brier Score: lower is better. Accuracy: higher is better.

01

Mechanisms matter

Class-conditioned causal selection targets specific sources of unreliable confidence.

02

Compact by design

CoCHI selects 10 heads for LLaMA and 8 for Gemma and Mistral, using a small calibration set.

03

A clear research scope

Evaluated on multiple-choice QA. Open-ended generation and multi-turn interactions remain future work.