Unified Definition and Notation

Overview

Confidence is used broadly across LLM systems in many forms: probability-like estimates, uncertainty scores, log-probability quantities, verbalized confidence, sample-agreement metrics, semantic uncertainty, verifier/judge/reward-model scores, or hybrids combining multiple sources. Despite this diversity, these signals share a common operational feature: each must affect a downstream decision to be useful in control.

This framework abstracts away the surface-level differences and focuses on the functional role of confidence as a control signal that transforms decision states and shapes which actions are taken.

Decision State & Units

At any step t in a generative process, we have a decision state that captures the context:

Given this state, we must choose among discrete decision units U_t = {u₁, u₂, …, u_n}. The unit granularity varies by application:

• token — next word in decoding
• span — consecutive tokens or phrases
• claim — atomic fact or statement
• chunk — passage or document segment
• example — training or demonstration instance
• response — complete answer or turn
• model/tool — expert or module to invoke
• step — action in a trajectory or plan
• trajectory — complete sequence of actions
• vote — aggregate choice in multi-sample voting

Confidence Signal

A confidence signal is any reliability-relevant score or distribution produced by the model or auxiliary system. In its most general form:

This signal is a vector (possibly m = 1) that quantifies how much to trust, weight, or pursue each decision unit u given the current state ξ_t. Sources include:

• Self — token log-probabilities, hidden states, verbalized uncertainty
• Sample-based — agreement across multiple samples, semantic entropy, consistency scores
• Auxiliary — verifier scores, reward-model outputs, router predictions, judge decisions
• External — retrieval relevance, tool execution success, environment feedback
• Hybrid — combinations of the above

Decision Process

The pipeline from state to action flows through four stages:

Key steps:

• Transform T̃_t: Normalize, calibrate, or combine confidence signals. Examples: softmax (for selection), clipping (for filtering), weighted averaging (for fusion).
• Policy δ_t: Map transformed scores to a decision rule. Examples: argmax (select highest), threshold (filter below cutoff), softmax-sample (probabilistic), or defer (abstain).
• Action a_t: Execute the policy to commit to one unit or a set of units.
• Updated state ξ_t+1: Incorporates action outcome and feeds forward.

Three Axes of Variation

Confidence systems vary along three independent dimensions. Understanding these axes helps map papers and design new methods systematically.

Lifecycle: From Training to Deployment

This framework specializes across the six domains covered in the survey:

• Training (§3): Confidence selects training examples and weights gradients. Sources: self-confidence (answer likelihood), auxiliary (judge/verifier), hybrid (uncertainty + influence). Units: examples, tokens, preferences. Role: learning signal allocation and supervision quality control.
• Inference (§4): Confidence shapes next-token distribution, when to stop generation, and which candidate to select. Sources: self (logprobs, verbalized, semantic entropy), auxiliary (PRM). Units: tokens, answers, steps. Role: decoding control, stopping, output selection.
• Routing (§5): Confidence predicts query suitability (pre-call) or answer quality (post-hoc) for cascading or model selection. Sources: auxiliary (routers), self (calibrated), hybrid (score gap). Units: queries, answers. Role: deferral, escalation, portfolio orchestration.
• RAG (§6): Confidence determines whether to retrieve, what context to keep, and whether to abstain. Sources: self (low-token confidence triggers retrieval), auxiliary (retrieval quality), hybrid (consistency + relevance). Units: tokens/sentences (trigger), documents/passages (filter), queries (routing). Role: retrieval triggering, context filtering, groundedness checking.
• Risk (§7): Confidence detects hallucinations, calibrates predictions, and enables conformal guarantees and abstention. Sources: self (semantic entropy, self-check), auxiliary (probes, verifiers), hybrid (calibrated predictor). Units: claims, answers. Role: hallucination detection, coverage guarantees, reliability certification.
• Agentic (§8): Confidence guides search (MCTS with PRMs), triggers escalation and backtracking, aggregates multi-agent votes. Sources: auxiliary (verifiers, rewards), peer (voting), self (intrinsic verification). Units: steps, trajectories, responses. Role: search guidance, control flow, consensus and escalation.