Unsure but Certain: Uncovering the Representation-Confidence Gap in Diffusion Language Models

2026-08-09

AI Summary

This study identifies and names a previously unrecognized phenomenon—termed the “representation–confidence gap”—in diffusion language models, wherein internal representations can detect errors under noisy inputs, yet output confidence remains spuriously high, severely degrading answer ranking performance. The authors demonstrate that conventional calibration techniques fail to restore effective ranking in such settings. To address this, they propose a lightweight method that directly extracts residual signals from the hidden states of a frozen model to rectify ranking without requiring additional generation steps. Experimental results show that this approach significantly improves ranking performance under noise, confirming the utility of internal representations and underscoring that confidence reliability is more critical than overall accuracy for robust downstream performance.

Abstract

Diffusion language models use broad context to create text, suggesting they might handle input noise better than standard models. Testing reveals this is only partially true. Internally, diffusion models detect text errors highly accurately. Externally, their reported certainty ignores this signal. As accuracy drops due to noise, confidence stays near its maximum and the ability to correctly rank answers degrades toward random chance. We call this mismatch the representation confidence gap. The visible concentration of high certainty scores is a misleading surface symptom. Standard math adjustments remove this concentration but fail to fix the underlying loss of ranking order. This ranking deficit favors standard models under noisy conditions and resists common remedies. Matching training recovers accuracy but not ranking, while score recalibration and input level error signals cannot reorder the final answers. However, the information needed to properly evaluate an answer survives in the hidden states. A lightweight extraction tool uses this signal to improve ranking. This approach is highly efficient because it leaves the base model completely frozen and requires zero additional text generation steps. We present this tool to prove the signal exists, while clearly noting its limits. Ultimately, certainty reliability is a more pressing limit than overall accuracy under noisy conditions.

Problem

• diffusion language models
• representation-confidence gap
• input noise
• confidence calibration
• answer ranking

Innovation

• representation-confidence gap
• diffusion language models
• confidence calibration
• hidden state extraction
• answer ranking