Probing Routing-Conditional Calibration in Attention-Residual Transformers
This research paper investigates whether internal routing traces in Attention-Residual (AR) transformers provide stable evidence for post-hoc calibration beyond standard softmax confidence. As routing-augmented architectures increasingly claim to offer calibration-relevant uncertainty, the authors question if these traces genuinely improve model reliability. Using a matched-confidence diagnostic suite on AR transformers developed by the Kimi Team, the study stratifies examples by routing-derived states and compares subgroup gaps against null models. The results indicate that scalar routing summaries do not offer stable evidence of routing-conditional miscalibration, with weighted gaps remaining small or sensitive to random seeds. Furthermore, a minimal calibration probe using routing-depth variance failed to reliably improve Expected Calibration Error (ECE) compared to confidence-only controls. Even complex MLP models appeared to benefit from routing data only until capacity-matched controls were applied, revealing that shuffled routing profiles performed similarly. The authors conclude that apparent gains from routing-aware calibration are likely confounded by capacity and bandwidth factors, urging rigorous controls before interpreting such improvements as genuine internal-state calibration.
Wire timeline
Probing Routing-Conditional Calibration in Attention-Residual Transformers
This research paper investigates whether internal routing traces in Attention-Residual (AR) transformers provide stable evidence for post-hoc calibration beyond standard softmax confidence. As routing-augmented architectures increasingly claim to offer calibration-relevant uncertainty, the authors question if these traces genuinely improve model reliability. Using a matched-confidence diagnostic suite on AR transformers developed by the Kimi Team, the study stratifies examples by routing-derived states and compares subgroup gaps against null models. The results indicate that scalar routing summaries do not offer stable evidence of routing-conditional miscalibration, with weighted gaps remaining small or sensitive to random seeds. Furthermore, a minimal calibration probe using routing-depth variance failed to reliably improve Expected Calibration Error (ECE) compared to confidence-only controls. Even complex MLP models appeared to benefit from routing data only until capacity-matched controls were applied, revealing that shuffled routing profiles performed similarly. The authors conclude that apparent gains from routing-aware calibration are likely confounded by capacity and bandwidth factors, urging rigorous controls before interpreting such improvements as genuine internal-state calibration.
cs.AI updates on arXiv.org