Confidence-Informed Prompting

• Confidence-informed prompting is an approach that integrates uncertainty signals into the prompting process to enhance calibration and direct downstream actions.
• It applies methods like retrieval augmentation, counterfactual prompting, and inline word-confidence to improve prediction reliability in tasks such as root cause analysis and ASR correction.
• The technique optimizes decision processes by enabling selective interventions and evidence-driven revisions, ensuring more trustworthy model responses.

Confidence-informed prompting denotes a class of prompting strategies in which uncertainty signals are made part of the prompt interface, the generation workflow, or the prompt-optimization objective. In the recent literature, this includes retrieval-augmented confidence estimation for black-box cloud incident root-cause predictors, fact-and-reflection decomposition for question answering, inline word-confidence prompting for ASR correction, counterfactual prompting for retrieval-augmented generation (RAG) risk control, and reinforcement-learned auxiliary prompts for clinically aligned calibration in multimodal medical question answering (Zhang et al., 2023, Zhao et al., 2024, Hernandez et al., 29 Sep 2025, Chen et al., 2024, Kriz et al., 12 Jul 2025). Across these settings, the common goal is not merely to improve task accuracy, but to make confidence estimates informative enough to guide downstream actions such as adoption, correction, abstention, escalation, or further retrieval.

1. Problem Setting and Motivation

A recurring motivation is that strong task performance does not imply reliable confidence. In cloud incident root cause analysis (RCA), LLM-based assistants can aid on-call engineers, yet their usefulness is constrained by low accuracy, hallucination, and the difficulty of distinguishing well-disguised hallucinations; the problem is compounded by the black-box nature of models such as GPT-3.5-Turbo and GPT-4, for which logits and probabilities are unavailable, making temperature scaling and fine-tuning-based calibration inapplicable (Zhang et al., 2023). In disordered-speech ASR post-processing, transcript-only prompting can be too aggressive because the model has no explicit signal indicating which words are suspicious, producing frequency bias and grammaticality bias such as changing correct “TEASED” to “ASKED” or “WHAT” to “WHAT’S” (Hernandez et al., 29 Sep 2025).

The same pattern appears in question answering and medical decision support. Fact-and-Reflection reports that prompting strategy affects not only reasoning performance but also calibration, and that several advanced prompting methods can improve Expected Calibration Error while still producing extreme wrong-but-high-confidence cases that worsen MacroCE (Zhao et al., 2024). In medical multiple-choice QA, Chain-of-Thought, roleplay, and emotional prompting can increase confidence without making that confidence trustworthy; emotional prompting in particular often inflated confidence and sometimes produced very high confidence even when wrong (Naderi et al., 29 May 2025). In RAG, the central concern is selective answering: the system should answer when confident and proactively refuse when confidence is low, because retrieved passages may be noisy and the model may use them improperly (Chen et al., 2024).

These results establish confidence as an operational variable rather than a cosmetic annotation. Confidence-informed prompting treats confidence as a signal that should alter system behavior: whether to trust a root-cause prediction, whether to edit a transcript, whether to keep or discard a RAG answer, whether to trigger retrieval, or whether to revise the prompt itself.

2. Confidence Signals and Calibration Criteria

Most work in this area adopts the standard calibration condition

$\mathbbm{P}(l_i \mid \tilde{P}_i = p) = p,\ \forall p \in [0,1],$

or equivalently

$\Pr(Y \text{ is correct} \mid C=c)=c,$

and estimates miscalibration with Expected Calibration Error (ECE) (Zhang et al., 2023, Yang et al., 2024). A common bin-based form is

$$\text{ECE} = \sum_{m=1}^M \frac{|B_m|}{n}\left|\text{acc}(B_m)-\text{conf}(B_m)\right|.$$

However, several papers emphasize that ECE alone is incomplete. Fact-and-Reflection highlights a bucket-canceling effect and therefore adds Macro-average Calibration Error (MacroCE), which is more sensitive to extreme wrong-but-high-confidence failures (Zhao et al., 2024). Medical and multimodal medical studies also use Brier score, while the medical board-exam study further uses AUC-ROC to measure whether higher confidence is assigned more often to correct answers than incorrect ones (Naderi et al., 29 May 2025, Kriz et al., 12 Jul 2025). For selective answering in RAG, calibration is supplemented by risk, carefulness, alignment, and coverage, defined over the four-way outcome decomposition AK, AD, UK, and UD (Chen et al., 2024).

Mechanism	Confidence signal	Representative papers
Retrieved evidence grounding	COE and RCE over historical incidents	(Zhang et al., 2023)
Inline upstream uncertainty	WORD[confidence], sentence or word filters	(Hernandez et al., 29 Sep 2025, Naderi et al., 2024)
Self-reported confidence	Answer plus confidence score	(Yang et al., 2024, Naderi et al., 29 May 2025)
Counterfactual stability	Answer change under CF-quality / CF-usage	(Chen et al., 2024)
Optimization-time confidence	Brier-aware prompt selection; self-supervised proxy	(Bayat et al., 20 May 2026, Saglam et al., 22 May 2026, Kriz et al., 12 Jul 2025)

The literature therefore does not treat confidence as a single primitive. Confidence may be estimated from verbalized scores, token probabilities, entropy-derived ASR uncertainty, retrieval-grounded judgments, counterfactual robustness, or teacher-forced log-probabilities on demonstrations. The verbalized-confidence benchmark makes this especially explicit: “verbalized confidence” is not one method but a family of prompt designs varying by score range, score formulation, advanced description, few-shot exemplars, and other prompt variants such as top-$k$ answers or chain-of-thought (Yang et al., 2024).

3. Grounded and Decomposed Prompting

The most explicit confidence-estimation framework in this area is PACE-LM / LM-PACE for cloud incident RCA. Given an incident description $d_{query}$ and a predicted root cause $\hat{r}_{query}$, the system retrieves historical incidents

$H = \{h_i=(d_i,r_i)\},$

then applies a retrieval-augmented two-step procedure with GPT-4 as a confidence estimator (Zhang et al., 2023). The first phase, Confidence-of-Evaluation (COE), measures whether the estimator has enough evidence from retrieved history to reason about the current incident, which the paper interprets as a form of “grounded-ness” in reference data. The second phase, Root Cause Evaluation (RCE), scores the plausibility of $\hat{r}_{query}$ against that same historical evidence. COE uses intermediate textual analysis followed by binary Yes/No judgments, while RCE uses intermediate analysis followed by a score on a specified scale because root-cause evaluation has multiple dimensions: truthfulness, groundedness, and informativeness (Zhang et al., 2023).

PACE-LM then learns a mapping from the COE and RCE outputs to a final confidence estimate with a calibration objective over bins; when the weight choice

$\omega(i)=\sum_j \mathbbm{1}[t_i \le \pi(c_j,s_j)\le t_{i+1}]$

is used, the objective reduces to ECE (Zhang et al., 2023). Retrieval is dense semantic retrieval over a historical incident database

$D=\{h_i=(d_i,r_i): i=1,2,\dots,n_{max}\},$

subject to a token budget $\Pr(Y \text{ is correct} \mid C=c)=c,$0. In the reported validation and test setup, $\Pr(Y \text{ is correct} \mid C=c)=c,$1 and the average number of retrieved incidents is about 15. The empirical result is that full LM-PACE consistently outperforms RCE-only and uniform-binning alternatives across GPT-4, GPT-3.5-Turbo, and Text-DaVinci-003 root-cause generators, with ECE values such as 0.082 for GPT-4-8K and 0.053 for GPT-3.5-Turbo-3896 (Zhang et al., 2023). Human evaluation of GPT-4 8K predictions in the high-confidence 80%–100% bin found that 86% were labeled as highly similar to ground truth.

A related but more general decomposition appears in Fact-and-Reflection (FaR). Instead of moving directly from prompt to answer, FaR first elicits relevant facts, often with sources, then asks the model to reflect over those facts before producing the final answer. In the paper’s formulation, standard prompting samples $\Pr(Y \text{ is correct} \mid C=c)=c,$2, while FaR conditions on fact and reflection states $\Pr(Y \text{ is correct} \mid C=c)=c,$3 and $\Pr(Y \text{ is correct} \mid C=c)=c,$4,

$\Pr(Y \text{ is correct} \mid C=c)=c,$5

This separation is intended to reduce anchoring and overconfidence by forcing explicit grounding before commitment (Zhao et al., 2024). FaR lowers ECE by 23.5% and MacroCE by 13.9%, and it also increases explicit concern expression in low-confidence scenarios: 8.8% under FaR versus 3.9% under original CoT, rising to 59.2% in a less constrained “FaR free” variant. When retrieval augmentation is applied only to cases where FaR expresses concern, accuracy improves by 68.0%, whereas randomly augmenting the same number of examples yields 15.0% improvement (Zhao et al., 2024).

Taken together, these methods show a common design principle: confidence quality can improve when prompts force the model to expose evidence, separate evidence acquisition from scoring, and score only after an explicit grounding step.

4. Confidence-Augmented Editing and Selective Intervention

In ASR post-processing, confidence-informed prompting is often used to control where and whether editing should occur. The disordered-speech correction work computes frame-level entropy-based confidence from ASR posterior distributions and aggregates it to word-level scores by mean, minimum, or product. The confidence can be derived from Gibbs entropy

$\Pr(Y \text{ is correct} \mid C=c)=c,$6

or from Tsallis entropy

$\Pr(Y \text{ is correct} \mid C=c)=c,$7

with $\Pr(Y \text{ is correct} \mid C=c)=c,$8 swept in experiments (Hernandez et al., 29 Sep 2025). Those word-level scores are inserted directly into the prompt as inline annotations such as HOW[1.00] MANY[0.85] RAFELLES[0.61], and the instruction explicitly states that lower-confidence words are more likely to be incorrect. This differs from sentence-level or word-level filtering baselines because the uncertainty is embedded into the model’s correction process rather than used only to decide whether correction should run.

The empirical effects are strongly domain-dependent. The abstract reports a 10% relative WER reduction on SAP spontaneous speech and a 47% reduction on TORGO relative to naive LLM correction (Hernandez et al., 29 Sep 2025). The full tables show that confidence prompting is best on SAP-unshared and TORGO, while on SAP-shared a hard word-level filter at 90% slightly outperforms confidence prompting. The same paper’s behavior analysis shows that the model intervenes much more on low-confidence utterances than on high-confidence ones, and harmful edits remain consistently low on high-confidence utterances. Qualitative examples indicate that confidence-guided prompting preserves high-confidence words that naive prompting would rewrite, while still allowing beneficial low-confidence corrections such as TRAINED $\Pr(Y \text{ is correct} \mid C=c)=c,$9 TRADING.

A complementary strategy appears in post-hoc LLM correction for Whisper ASR outputs. Here the confidence signal comes from whisper-timestamped, which exposes sentence- and word-level confidences derived from token log probabilities, with punctuation excluded (Naderi et al., 2024). Three filtering strategies are tested: sentence-level confidence filtering, lowest-word confidence filtering, and targeted low-confidence word correction. The sentence-level rule sends an utterance to the LLM when $$\text{ECE} = \sum_{m=1}^M \frac{|B_m|}{n}\left|\text{acc}(B_m)-\text{conf}(B_m)\right|.$$0, while the word-level rule uses

$$\text{ECE} = \sum_{m=1}^M \frac{|B_m|}{n}\left|\text{acc}(B_m)-\text{conf}(B_m)\right|.$$1

and sends the utterance when the minimum word confidence falls below threshold. The reported threshold choices are $$\text{ECE} = \sum_{m=1}^M \frac{|B_m|}{n}\left|\text{acc}(B_m)-\text{conf}(B_m)\right|.$$2 for sentence-level filtering and $$\text{ECE} = \sum_{m=1}^M \frac{|B_m|}{n}\left|\text{acc}(B_m)-\text{conf}(B_m)\right|.$$3 for lowest-word filtering. Prompt design also matters: the best prompt uses two examples plus an instruction that corrections should be phonetically similar, improving Whisper Tiny dev-clean WER from 8.51 to 6.65, whereas Whisper Large V3 on the same split degrades from 3.11 to 3.34 under LLM correction (Naderi et al., 2024).

The broader implication is that confidence-informed prompting can serve as an edit gate. It can decide when a correction model should act, which tokens deserve attention, and how aggressive the edit policy should be. The data also show that selective intervention is especially useful when the upstream recognizer is weak; for stronger recognizers, indiscriminate correction may introduce more errors than it removes.

5. Verbalized Confidence, Abstention, and Risk Control

A major branch of the literature asks the model to state its confidence explicitly. The benchmark on verbalized confidence scores uses a fixed response format in which the assistant outputs both an answer and a confidence score, and studies prompt variations along several axes: percentage versus decimal versus ordinal score range, “confidence score” versus “probability that your answer is correct” formulation, basic versus advanced description, few-shot exemplars, best-guess wording, top-$$\text{ECE} = \sum_{m=1}^M \frac{|B_m|}{n}\left|\text{acc}(B_m)-\text{conf}(B_m)\right|.$$4 candidate formats, and chain-of-thought variants (Yang et al., 2024). The central finding is that reliability depends strongly on how confidence is requested. For models with at least 70B parameters, ECE is around 0.1, and the strongest large-model prompt is the combined combo method, which uses probability formulation, advanced description, few-shot examples, and best-guess wording, yielding an average deviation of about 7% between reported confidence and empirical accuracy (Yang et al., 2024). The same study argues that calibration is not enough by itself and therefore also measures informativeness via the number of distinct confidence values and variance, and meaningfulness via a KL-divergence criterion over dataset-specific confidence distributions.

The medical board-exam study provides a domain-specific picture of the same issue. It varies six prompt styles, three temperatures, and two confidence scales across GPT-4o, o3-mini, Llama-3.3-70b, Llama-3.1-8b, and DeepSeek-v3. o3-mini is the most accurate overall at $$\text{ECE} = \sum_{m=1}^M \frac{|B_m|}{n}\left|\text{acc}(B_m)-\text{conf}(B_m)\right|.$$5, yet some of its settings have the worst ECE, including 0.687 under emotional prompting with temperature 1.0 and a 1–100 confidence scale (Naderi et al., 29 May 2025). GPT-4o has the best mean AUC-ROC at $$\text{ECE} = \sum_{m=1}^M \frac{|B_m|}{n}\left|\text{acc}(B_m)-\text{conf}(B_m)\right|.$$6 and the best mean Brier score at $$\text{ECE} = \sum_{m=1}^M \frac{|B_m|}{n}\left|\text{acc}(B_m)-\text{conf}(B_m)\right|.$$7. Few-shot prompting at temperature 1.0 with a 1–10 scale produces some of the best Brier scores, whereas the 1–100 scale often inflates confidence numerically without improving alignment (Naderi et al., 29 May 2025).

RAG risk control uses a different confidence mechanism. Instead of asking directly for a score, the counterfactual prompting framework infers confidence from answer stability under two prompt-induced perturbations: CF-quality, which assumes the answer is wrong because the referred passages are poor, and CF-usage, which assumes the answer is wrong because the retrieved passages were used improperly (Chen et al., 2024). The model first produces the original answer $$\text{ECE} = \sum_{m=1}^M \frac{|B_m|}{n}\left|\text{acc}(B_m)-\text{conf}(B_m)\right|.$$8, then regenerates answers under the two counterfactual prompts. If the regenerated answer matches the original, the system treats the output as relatively confident; if it differs, the answer is treated as uncertain and may be discarded. The final output is a keep/discard judgment $$\text{ECE} = \sum_{m=1}^M \frac{|B_m|}{n}\left|\text{acc}(B_m)-\text{conf}(B_m)\right|.$$9, and evaluation centers on the risky case UK: an unanswerable sample that is incorrectly kept (Chen et al., 2024). Across two datasets and two backbones, the framework improves risk control in three of four settings, with average risk reduction of about 2.88% and average carefulness improvement of about 14.76%; CF-quality alone is especially useful.

This branch of the literature shows that confidence-informed prompting need not rely on a scalar self-report. Confidence can be elicited directly through verbalized probabilities, or inferred indirectly from counterfactual robustness and selective-answering behavior.

6. Confidence-Aware Prompt Optimization and Domain-Specific Alignment

More recent work treats confidence not only as an output, but as an optimization target for the prompt itself. Reflective Prompt Tuning (RPT) frames prompt engineering as an iterative diagnosis-driven workflow mediated by LLM function calling (Bayat et al., 20 May 2026). At iteration $k$0, a diagnostic function evaluates the target model over the full optimization set, collects reasoning traces, answers, and reported confidences, clusters recurring failure modes, and returns a structured report. The optimizer then revises the prompt using the current report and an accumulated memory of prior reports. Final prompt selection maximizes a development-set objective $k$1 over task metrics and calibration quality, with Brier score as the main calibration metric:

$k$2

Across HotPotQA, LiveBench-Math, and Formula, RPT improves over seed prompts by up to 12.9, 12.4, and 11.7 points, respectively, and with GPT-5 as optimizer it improves both task score and Brier score, for example on HotPotQA from $k$3 with Brier $k$4 (Bayat et al., 20 May 2026).

Self-Improving In-Context Learning pushes the same idea to test-time prompt calibration. It keeps the discrete few-shot prompt fixed but optimizes its continuous embedding matrix $k$5 using a bounded self-supervised confidence proxy computed from teacher-forced log-probabilities on the demonstrated outputs:

$k$6

with $k$7, $k$8, and $k$9 in the reported experiments (Saglam et al., 22 May 2026). Gradients are estimated by zeroth-order optimization,

$d_{query}$0

and embeddings are updated without finetuning, token generation, predefined label sets, or external data. Reported mean accuracies improve from 0.67 to 0.71 for Llama 3.1-8B, from 0.65 to 0.72 for Qwen3-4B, and from 0.45 to 0.48 for Gemma 2-2B, with statistically significant positive Spearman correlation between proxy improvement and downstream accuracy gain (Saglam et al., 22 May 2026).

Healthcare provides the clearest example of domain-specific alignment. Prompt4Trust trains a lightweight prompt generator to emit a query-specific Calibration Guidance Prompt $d_{query}$1, which is appended to the downstream multimodal query so that the frozen MLLM outputs both answer $d_{query}$2 and verbalized confidence $d_{query}$3:

$d_{query}$4

The generator is optimized with GRPO under an asymmetric reward that favors correct high-confidence predictions and strongly penalizes overconfident errors (Kriz et al., 12 Jul 2025). On PMC-VQA test, Prompt4Trust improves verbalized-confidence baselines from ECE 0.517 and Brier 0.523 to ECE 0.163 and Brier 0.264, while also increasing accuracy from 40.40% to 45.95%. Its most clinically salient result is in the high-confidence region: for predictions with confidence $d_{query}$5, accuracy rises to 76.87%, compared with 44.74% for raw verbalized confidence (Kriz et al., 12 Jul 2025). The learned prompt generator also transfers zero-shot to larger MLLMs, improving calibration even when raw accuracy does not always increase.

These results suggest a broad design pattern for confidence-informed prompting. Confidence can be separated from prediction, grounded in retrieved or recalled evidence, injected from upstream uncertainty estimates, inferred from counterfactual stability, or optimized directly at prompt-selection time. The literature also indicates recurring limits: ECE can hide harmful extremes, format noncompliance can corrupt verbalized-confidence outputs, entropy-based confidence requires tuning of aggregation and $d_{query}$6, and many results remain confined to objective QA, English-only settings, or narrow application domains (Zhao et al., 2024, Yang et al., 2024, Hernandez et al., 29 Sep 2025). Within those constraints, confidence-informed prompting has become a practical framework for turning prompt design into a mechanism for calibrated decision support rather than answer generation alone.