[2606.07559] Phantom Transitions in Language Mannequin Superb-Tuning: A Density-Matrix Evaluation

View a PDF of the paper titled Phantom Transitions in Language Mannequin Superb-Tuning: A Density-Matrix Evaluation, by Vaibhav Prakash and 1 different authors

View PDF
HTML (experimental)

Summary:Language fashions fine-tuned the place the proper completion should outrank a near-synonym competitor typically fail silently. The cross-entropy loss falls monotonically whereas the proper token by no means overtakes the competitor within the mannequin’s rating. We examine this throughout 5 transformer architectures from two households spanning a sixfold parameter vary, on ten contexts whose appropriate and competing completions share substantial embedding overlap. We construct an order parameter combining the expected distribution with embedding overlap, as a density matrix as a result of that distribution lives over a non-orthogonal foundation. It decomposes additively right into a sign time period monitoring dedication to the proper token and a drag time period set by how the embedding bulk leaks likelihood into the rating. This isolates two failure modes. In kinematic failure the sign stays too small and the mannequin by no means commits. In structural failure the drag worsens throughout fine-tuning, so the mannequin degrades geometrically as its loss falls. The order parameter additionally reveals sharp jumps resembling part transitions. We take a look at the spontaneous-symmetry-breaking studying by monitoring it after each gradient step, and rule it out. The jumps persist below LoRA though the token embedding matrix by no means modifications. No geometric part transition is feasible when that geometry can’t transfer, so the discontinuity lies completely within the softmax readout. A number of dimensionless portions arrange the trajectory throughout architectures. One is constant throughout all 5 fashions below full fine-tuning. A second types architectures into two courses by their bulk embedding distribution and predicts whether or not LoRA alone could make a sentence commit. As a blind take a look at, the framework predicts a held-out structure’s essential studying charge to inside 2.1% of a later sweep. These outcomes characterize this near-synonym mechanism and wish recalibration earlier than extrapolation.