Competence Gate: Using Small Model Confidence for Tool Calls

Synthium- · reddit · 2026-07-05

The author created a 10MB LoRA adapter and orchestration layer for Qwen3.5-4B that decides per query whether to answer directly, search the web, or retrieve local documents, refusing to fabricate facts if unverified. The core idea: small instruction models struggle to verbalize their confidence (tested 7 models from 3–9B, all hitting a confidence ceiling), but their internal activations actually contain this signal. The adapter reads this directly to gate tool calls. Experiments show this gating is better at catching errors than the base model's tool calls (d′ increased by 0.46, 95% CI [0.01,0.89]), and in multi-token cases, 87% were indeed wrong. A dual-signal version routes privacy queries to local retrieval, reducing the proportion of privacy questions sent to public web search from 22% to 10%. It runs locally on Apple Silicon/MLX and offers a GGUF version.

Original post →

More from coding & agent

coding & agent channel →