Ayoub Belfatmi

Freezing the input factor A in LoRA, especially on top singular directions, unlocks substantial performance gains over freezing the output factor B, even at rank 2.

Targeted re-decoding can boost accuracy by up to 4.5 percentage points in quantized small language models, revealing a selective repair mechanism for problematic inference trajectories.

Centered token log-probability fails as a reliable monitor for reasoning models, but a new calibrated decoding controller enhances detection of unreliable outputs without extensive retraining.