UniJev: Separating Accuracy from Calibration in Trimodal System One Decision Models

Key points

• System One models produce probabilistic judgments that software can consume directly, making reliable confidence as important as answer accuracy.1
• Yet joint training across text, images, and video remains underexplored, and the contributions of supervised fine-tuning (SFT) and reinforcement learning for calibrated decisions (RLCD) are difficult to separate.1
• We introduce UniJev, a trimodal framework that scores candidate answers with a shared head under joint supervised training across text, images, and video.1
• Supervised decision models improve accuracy over their backbone readouts across all three modalities, while the existing RLCD algorithms change accuracy little and have inconsistent effects on visual calibration across recipes.1

Sources (1)

• [1]UniJev: Separating Accuracy from Calibration in Trimodal System One Decision Models
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 04:49 AM

  We introduce Cora (Confidence--Outcome Reliability Alignment), which groups current predictions by confidence and optimizes each prediction's confidence toward the empirical accuracy of its bin.

  System One models produce probabilistic judgments that software can consume directly, making reliable confidence as important as answer accuracy.

Extractive summary: sentences quoted from the sources.