Different Facets of Verbalised Overconfidence: an Interpretability Study
Large language models exhibit a systematic tendency toward overconfidence, producing assertive answers even when evidence supports hedging or abstention Overconfidence is most pronounced when models are prompted to output numeric confidence scores rather than using verbal markers or abstention A transcoder-based interpretability method identifies distinct feature sets: certainty relies on a broad coalition of shared features, while uncertainty is implemented as a sparse override via a small set
Analysis
TL;DR
- Large language models exhibit a systematic tendency toward overconfidence, producing assertive answers even when evidence supports hedging or abstention
- Overconfidence is most pronounced when models are prompted to output numeric confidence scores rather than using verbal markers or abstention
- A transcoder-based interpretability method identifies distinct feature sets: certainty relies on a broad coalition of shared features, while uncertainty is implemented as a sparse override via a small set of dedicated features
- Causal intervention on uncertainty features both proves the architectural imbalance underlying overconfidence and effectively mitigates overconfident errors
- The identified uncertainty features generalize across three expression settings, multiple languages, and an out-of-distribution modality task
Why It Matters
This research provides one of the first mechanistic explanations for LLM overconfidence, moving beyond behavioral observation to causal feature-level analysis. For AI practitioners building safety-critical systems, understanding that overconfidence stems from an architectural imbalance—rather than mere training noise—enables targeted interventions. The cross-setting generalization of findings also suggests these mechanisms are fundamental to current model architectures.
Technical Details
- Model and scope: Study focuses on Qwen3-4B, evaluating overconfidence across three uncertainty-expression modalities: verbal epistemic markers, abstention, and numeric confidence scores
- Experimental design: Uses controlled reasoning scenarios that manipulate logical necessity and possibility to create conditions where hedging or abstention is the correct response
- Transcoder-based interpretability: Proposes a novel method using transcoders to differentially identify features responsible for uncertainty versus certainty representations
- Causal intervention: Demonstrates that intervening on identified uncertainty features causally proves the imbalance and reduces overconfident errors
- Generalization: The same feature set transfers across verbal/numeric/abstention settings, multiple languages, and an out-of-distribution modality task
Industry Insight
- Safety-critical LLM deployments should prioritize uncertainty calibration, particularly avoiding numeric confidence score prompts that exacerbate overconfidence
- Feature-level interventions on transcoder-identified uncertainty mechanisms offer a promising path toward mitigating overconfidence without retraining
- The cross-lingual and cross-modal generalization of uncertainty features suggests interpretability-based safety interventions could be architecture-agnostic rather than model-specific
Disclaimer: The above content is generated by AI and is for reference only.