A researcher released a 10MB LoRA adapter for Qwen3.5-4B that reads internal confidence signals from the model's activations to gate tool-use, reducing hallucination and improving error detection by 0.46 d′ compared to the base model. This approach addresses a key limitation of small language models—their inability to accurately verbalize confidence—by leveraging internal signals, enabling more reliable tool-use and private query handling without requiring larger models. The adapter was trained on 126 hand-authored items and evaluated with confidence intervals; it reduces private query leaks from 22% to 10% and is open-source under Apache-2.0.
Background
Small LLMs often produce overconfident verbalized confidence despite internal uncertainty. Research shows that internal activations can reflect true confidence better than output tokens. The d′ measure quantifies discriminability between correct and incorrect responses. This work uses a LoRA adapter to tap into that internal signal.