Technical Summary: The Computational Basis of Confidence in Large Language Models

Problem Statement

Reliable confidence—the probability that a model's own answer is correct—is a prerequisite for the trustworthy deployment of Large Language Models (LLMs). While existing literature has extensively evaluated confidence signals based on their predictive accuracy (how well they correlate with correctness) and calibration (how well reported probabilities match empirical frequencies), a fundamental question remains unanswered: What does the confidence signal itself represent computationally?

Current signals, such as answer-token probabilities or verbal reports, are often treated as empirical predictors of correctness. However, it is unclear whether these signals are merely heuristic preference scores that combine evidence in a non-Bayesian manner, or if they function as monotonic readouts of a latent decision variable. In computational neuroscience, a latent decision variable is a noisy internal representation of evidence that, under a normative process model, is sufficient to compute the posterior probability of correctness (statistical decision confidence, or SDC). Distinguishing between a heuristic score and a normative readout is critical for understanding the internal mechanics of model confidence.

Methodology

The authors employ the framework of Statistical Decision Confidence (SDC), a normative model from computational neuroscience, to test whether answer logits in multimodal LLMs behave as readouts of a latent decision variable. The study focuses on the answer-logit difference (LD = $\ell_A - \ell_B$) as the candidate readout.

To test this, the authors evaluate four qualitative signatures predicted by the SDC framework across three distinct decision regimes:

1. Psychometric Function: Choice probability must increase monotonically with signed stimulus strength.
2. Signed Evidence Encoding: The signed LD must vary monotonically with the signed stimulus strength.
3. Trial-Specific Confidence (Fixed-Strength Prediction): At a fixed objective stimulus strength, the magnitude of LD ($|LD|$) must predict correctness. This tests whether trial-to-trial fluctuations in LD reflect internal noise rather than just changes in objective difficulty.
4. Folded-X Pattern: As stimulus strength increases, $|LD|$ should increase on correct trials but decrease on error trials. This geometric signature arises because high-strength errors occur only when noise pushes the latent decision variable just across the decision boundary.

Experimental Setup:

• Models: Three multimodal non-reasoning models (Qwen2.5-VL 7B, Gemma 3 12B, Gemma 4 12B) and one multimodal reasoning model (Gemini Flash 3).
• Tasks:
  • Perceptual Discrimination: Three synthetic visual tasks (contrast, object size, shape categorization) where stimulus strength is experimentally controlled. To induce trial-to-trial variation in deterministic models, the authors generated multiple stimulus exemplars at each strength level by randomizing task-irrelevant ("nuisance") features (e.g., spatial jitter, dot realization).
  • Memory-Based Decision: A city-population comparison task requiring retrieval of world knowledge.
  • Complex Visual Reasoning: CLEVR tasks (count equality, object existence) with parametric Gaussian blur.
• Behavioral Validation: An abstention task was conducted where models were instructed to withhold answers if their internal confidence (derived from LD) suggested a probability of correctness below 60%.

Key Results

1. Evidence Encoding and Latent Variable Signatures in Perceptual Tasks

Across the three perceptual tasks, the models satisfied all four SDC signatures:

• Psychometric & Signed Encoding: Choice probability and signed LD varied monotonically with stimulus strength, confirming that models encode task-relevant evidence.
• Fixed-Strength Prediction: Even when objective stimulus difficulty was held constant, larger $|LD|$ significantly predicted higher accuracy (e.g., in the contrast task, adding $|LD|$ to a stimulus-strength model improved AUROC from 0.864 to 0.907). This demonstrates that LD carries trial-specific decision evidence beyond objective difficulty.
• Folded-X Pattern: The interaction between correctness and stimulus strength produced the characteristic folded-X geometry. On correct trials, $|LD|$ increased with stimulus strength; on error trials, $|LD|$ decreased. This was observed in Qwen, Gemma 3, and Gemma 4, with a single minor exception (Gemma 3 on the contrast task showed a flat rather than inverted error branch, though within-strength prediction remained strong).

2. Generalization to Memory and Reasoning

• Memory-Based Decisions: In the city-population task, the psychometric function was shallow due to high noise in retrieved knowledge. However, the latent decision variable signatures held: $|LD|$ predicted correctness within fixed population-evidence bins, and the folded-X pattern was present (and even sharper in Gemma 3). This indicates that LD acts as a readout of latent decision evidence even when the objective stimulus axis is a noisy proxy for the model's internal evidence.
• Complex Visual Reasoning (CLEVR): In CLEVR tasks with Gaussian blur, the first two signatures (relying on a signed evidence axis) were not directly applicable as blur removes information rather than favoring a specific answer. However, the third signature held: $|LD|$ predicted correctness beyond blur strength and scene controls. The folded-X pattern was not observed; error branches were flat or positive rather than negative. The authors attribute this to the lack of a known normative process model for complex reasoning, preventing the derivation of the specific geometric prediction.

3. Behavioral Consequences

In the abstention task, models selectively withheld low-confidence answers based on the LD signal.

• Models abstained on 87.7% of trials, committing only to high-confidence responses.
• Accuracy on committed trials rose from a baseline of 92.2% to 99.95%.
• The abstention policy tracked the internal LD signal better than objective stimulus strength, and the resulting behavior closely approximated an idealized threshold on $|LD|$.

Significance and Claims

The paper provides a computational account of confidence in multimodal LLMs, moving beyond the empirical observation that confidence signals predict correctness.

• Latent Decision Variable Readout: The primary claim is that in settings where a normative process model can be specified (simple perceptual and memory-based tasks), answer logits behave not merely as heuristic preference scores, but as monotonic readouts of a latent decision variable. This variable is, in principle, sufficient to compute the posterior probability that the current choice is correct.
• Unifying Framework: The study establishes Statistical Decision Confidence (SDC) as a unifying computational framework for studying confidence across biological and artificial intelligence. It delineates the conditions under which answer logits function as normative readouts versus heuristic scores.
• Limitations and Boundaries: The authors modestly note that the absence of the folded-X pattern in complex reasoning tasks (CLEVR) does not necessarily imply a failure of SDC. Instead, it highlights the current boundary of the framework: without an explicit normative process model for complex reasoning, the specific geometric signatures cannot be derived or tested. The results suggest that while LD carries trial-specific information about correctness in complex tasks, its status as a normative latent variable remains unresolved until appropriate process models are developed.

In summary, the work demonstrates that for simple and memory-based decisions, the confidence signal in LLMs is structurally isomorphic to the latent decision variables found in biological systems, providing a rigorous foundation for interpreting model confidence.