Causal Evidence that Language Models use Confidence to Drive Behavior

Who Should Read

AI engineers applying LLM hallucination reduction or abstention features in production. ML researchers and system architects who want to understand LLM metacognition and design more reliable AI systems.

Core Mechanics

• GPT-4o holds an implicit threshold — without explicit instructions, it auto-abstains when internal confidence drops below ~77%
• Confidence's effect size for predicting abstention (βstd=0.99) is ~10x larger than RAG score, sentence embeddings, or problem difficulty
• Injecting high/low confidence vectors into Gemma 3 27B via activation steering shifts abstention rate by 59.5pp (66.5%→7.0%) — direct causal evidence
• Mediation analysis: 67.1% of the activation steering effect is transmitted via confidence redistribution, 26.2% via decision policy change
• Explicitly instructing the model with a confidence threshold (0–100%) adjusts abstention accordingly; prior confidence remains a strong predictor — Stage 1 (confidence formation) and Stage 2 (threshold policy) operate independently
• GPT-4o weights internal confidence 1.8x more than instructed threshold; abstention baselines vary dramatically across models (DeepSeek 82%, GPT-4o 56.6%, Qwen 80B 43.8%, Gemma 3 27B 27.2%)

Evidence

• Activation steering shifts abstention: low-confidence(-2) 66.5% vs high-confidence(+2) 7.0% — 59.5pp difference (r=-0.99, p<0.001)
• Phase 2 logistic regression: confidence |βstd|=0.99 vs RAG(0.102), difficulty(0.110), embeddings(0.106) — roughly 9–10x larger
• Phase 4: confidence+threshold model vs threshold-only: AIC -1953, pseudo-R² 0.11→0.24 (χ²(1)=1955, p<0.001)
• Mediation analysis: confidence redistribution indirect path explains 67.1% of total steering effect (indirect effect a1×b1=-0.55, 95% CI [-0.65, -0.47])

How to Apply

• To improve production LLM reliability, add "abstain if confidence < T%" to the prompt (Phase 4 style) — each 10–20% threshold increase trades ~1.1% accuracy gain for higher abstention
• If a model rarely abstains (like Gemma 3 27B <5%), generate 20 paraphrases of the same abstention-instruction prompt and select the version with the highest abstention rate
• For open-source models with internal activation access (e.g. Gemma 3 27B): extract steering vectors by contrasting residual streams from high vs low confidence trials, inject at layers 30–40 during inference to programmatically control abstention rate

Code Example

# Phase 4 approach: Add confidence threshold instruction to prompt
prompt_template = """
You will be given a 4-way multiple choice question, with options 1-4.
First, internally estimate the probability (0-100) that your answer is correct.
Then:
- If your confidence is MORE than {threshold}%, output only the number of your answer.
- If your confidence is LESS than {threshold}%, output '5' to indicate you want to abstain.
Output only a single number.

Question: {question}
1) {option1}
2) {option2}
3) {option3}
4) {option4}

Answer:"""

# Adjusting threshold between 0~100 controls the accuracy/coverage tradeoff
# High threshold (e.g., 80%) → accuracy↑, coverage↓
# Low threshold (e.g., 30%) → accuracy↓, coverage↑
for threshold in range(0, 101, 10):
    prompt = prompt_template.format(
        threshold=threshold,
        question="Who won the Nobel Prize in Physics in 2024?",
        option1="Geoffrey Hinton",
        option2="Yann LeCun",
        option3="John Hopfield",
        option4="Andrew Ng"
    )
    # response = llm.generate(prompt)
    # If '5' is returned → abstention, if 1-4 → corresponding answer