ADVICE: Answer-Dependent Verbalized Confidence Estimation
Pith reviewed 2026-05-18 07:40 UTC · model grok-4.3
The pith
Making verbalized confidence depend on the model's answer improves calibration in large language models.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
ADVICE is a fine-tuning framework that promotes answer-grounded confidence estimation in LLMs. By addressing answer-independence, it substantially improves confidence calibration and generalizes to unseen settings without degrading task performance. The gains stem from enhanced answer dependence.
What carries the argument
ADVICE, a fine-tuning framework that promotes answer-grounded confidence estimation by conditioning confidence on the model's answer.
If this is right
- Confidence estimates become substantially better calibrated on the tasks used during fine-tuning.
- The improved calibration generalizes to settings and tasks the model has not seen.
- Accuracy on the original tasks remains unchanged.
- Overconfidence decreases because confidence statements now depend more closely on the answer provided.
Where Pith is reading between the lines
- The same answer-dependence principle could be tested on other uncertainty signals such as hedging phrases or probability distributions.
- Combining ADVICE-style training with methods that improve answer correctness might produce systems whose confidence and accuracy both rise together.
- In deployed chat systems this could let users adjust their trust level more precisely based on the model's stated confidence.
Load-bearing premise
Answer-independence drives overconfidence in verbalized estimates and fine-tuning can enforce dependence on the answer without creating new biases or hurting performance.
What would settle it
Measuring confidence scores after ADVICE fine-tuning and finding they remain independent of the chosen answer or show no gain in standard calibration metrics such as expected calibration error.
read the original abstract
Recent progress in large language models (LLMs) has enabled them to communicate their confidence in natural language, improving transparency and reliability. However, this expressiveness is often accompanied by systematic overconfidence, whose underlying causes remain poorly understood. In this work, we analyze the dynamics of verbalized confidence estimation and identify answer-independence -- the failure to condition confidence on the model's own answer -- as a primary driver of this behavior. To address this, we introduce ADVICE (Answer-Dependent Verbalized Confidence Estimation), a fine-tuning framework that promotes answer-grounded confidence estimation. Extensive experiments show that ADVICE substantially improves confidence calibration, while exhibiting strong generalization to unseen settings without degrading task performance. We further demonstrate that these gains stem from enhanced answer dependence, shedding light on the origins of overconfidence and enabling trustworthy confidence verbalization.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper identifies answer-independence as a primary cause of overconfidence in LLMs' verbalized confidence estimates. It introduces ADVICE, a fine-tuning framework intended to enforce answer-grounded confidence estimation, and reports that this yields substantially better calibration, strong generalization to unseen settings, no degradation in task performance, and that the gains are attributable to enhanced answer dependence.
Significance. If the causal attribution and experimental results hold, the work would provide both a practical fine-tuning recipe for more reliable natural-language confidence and a mechanistic explanation for overconfidence. The focus on answer dependence as a load-bearing factor distinguishes it from generic calibration methods and could guide future interventions.
major comments (3)
- [Abstract] Abstract: The abstract asserts 'extensive experiments' demonstrating substantial calibration improvements, strong generalization, and that 'these gains stem from enhanced answer dependence,' yet supplies no information on datasets, baselines, metrics, or controls. This absence prevents assessment of whether the data actually support the central claims.
- [§3] §3 (ADVICE framework): The training objective and loss terms used to promote answer dependence are not specified. Without an explicit formulation of how the fine-tuning objective differs from standard supervised confidence training, it is impossible to verify that the method targets answer conditioning rather than generic verbalized-confidence supervision.
- [§4] §4 (Experiments): No ablation is described that compares ADVICE against a baseline fine-tuned to output verbalized confidence without explicit answer conditioning. Such a control is required to establish that observed ECE reductions and generalization are caused by enhanced answer dependence rather than by supervised fine-tuning on confidence labels in general; absent this, the causal claim remains unsupported.
minor comments (2)
- [§2] The paper would benefit from a concise mathematical definition of 'answer-independence' (e.g., a conditional probability or dependence measure) to make the diagnostic claim precise.
- [Figures/Tables] Figure and table captions should explicitly state the evaluation metrics (e.g., ECE, Brier score) and the exact generalization settings tested.
Simulated Author's Rebuttal
We thank the referee for the constructive and detailed feedback. The comments identify important opportunities to improve clarity, explicitness, and support for causal claims. We address each major comment below and commit to revisions that strengthen the manuscript without altering its core contributions.
read point-by-point responses
-
Referee: [Abstract] Abstract: The abstract asserts 'extensive experiments' demonstrating substantial calibration improvements, strong generalization, and that 'these gains stem from enhanced answer dependence,' yet supplies no information on datasets, baselines, metrics, or controls. This absence prevents assessment of whether the data actually support the central claims.
Authors: We agree that the abstract would benefit from greater specificity to allow immediate assessment of the experimental scope. In the revised version we will concisely incorporate the primary datasets, key baselines, main metrics (including ECE), and reference to the answer-dependence controls while remaining within length constraints. revision: yes
-
Referee: [§3] §3 (ADVICE framework): The training objective and loss terms used to promote answer dependence are not specified. Without an explicit formulation of how the fine-tuning objective differs from standard supervised confidence training, it is impossible to verify that the method targets answer conditioning rather than generic verbalized-confidence supervision.
Authors: We acknowledge that an explicit mathematical statement of the objective would improve verifiability. Section 3 currently describes the framework in prose; we will add the precise loss formulation in the revision, highlighting the terms that enforce conditioning on the generated answer and how they differ from standard supervised confidence fine-tuning. revision: yes
-
Referee: [§4] §4 (Experiments): No ablation is described that compares ADVICE against a baseline fine-tuned to output verbalized confidence without explicit answer conditioning. Such a control is required to establish that observed ECE reductions and generalization are caused by enhanced answer dependence rather than by supervised fine-tuning on confidence labels in general; absent this, the causal claim remains unsupported.
Authors: We agree that an explicit ablation isolating the contribution of answer conditioning is necessary to substantiate the causal attribution. We will add this control experiment to the revised Section 4, comparing ADVICE against a baseline that receives standard supervised fine-tuning on verbalized confidence without the answer-dependence components, and report the resulting differences in calibration and generalization. revision: yes
Circularity Check
No significant circularity; empirical framework is self-contained
full rationale
The paper identifies answer-independence as a driver of overconfidence through analysis, then introduces ADVICE as a fine-tuning intervention to promote answer-grounded estimation, with results from experiments on calibration and generalization. No equations, fitted parameters renamed as predictions, or self-citation chains appear in the provided text that would reduce the central claims to inputs by construction. The approach is presented as an independent empirical response to an observed behavior rather than a tautological renaming or load-bearing self-reference, keeping the derivation chain non-circular.
Axiom & Free-Parameter Ledger
axioms (1)
- domain assumption Fine-tuning can promote answer dependence in verbalized confidence without degrading task performance or introducing new failure modes.
Forward citations
Cited by 2 Pith papers
-
Verbal Confidence Saturation in 3-9B Open-Weight Instruction-Tuned LLMs: A Pre-Registered Psychometric Validity Screen
Seven 3-9B instruction-tuned LLMs produce verbal confidence that saturates at high values and fails psychometric validity criteria for Type-2 discrimination under minimal elicitation.
-
Distilling Self-Consistency into Verbal Confidence: A Pre-Registered Negative Result and Post-Hoc Rescue on Gemma 3 4B
Fine-tuning Gemma 3 4B on unfiltered self-consistency targets produces a binary verbal correctness discriminator with AUROC 0.774 on TriviaQA, outperforming logit entropy after a modal-filtered pre-registration failed.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.