Skip to content

Training-based evidence comes from the model’s developmental history: how its parameters changed during optimization, when capabilities appeared, and what happens when the training process is modified. This is the least-used evidence family in current mechanistic interpretability, yet it addresses questions that no post-training measurement can answer — particularly about whether a mechanism is a stable product of the learning process or a fragile artifact of a specific training run.

Methods that analyze training trajectories without modifying them:

  • Loss curves — training and validation loss over time, including structure in the curve (plateaus, phase transitions, sudden drops).
  • Checkpoint comparison — comparing weight structure, activation patterns, or behavioral profiles across training checkpoints to track when and how a mechanism forms.
  • Phase transitions — identifying points in training where qualitative changes in behavior or internal structure occur (e.g., grokking, induction head formation).
  • Learning dynamics — tracking gradient norms, parameter velocities, effective learning rates, and other optimization-level quantities across training.

Observational training evidence is the primary source for I11 Onset coupling: does the mechanism appear when the capability appears? If a circuit is claimed to implement a capability, and that capability emerges at step 10,000, the circuit’s weight structure and activation patterns should undergo a corresponding transition at the same point. Onset coupling that holds across multiple training seeds strengthens the claim; onset coupling that fails — the capability appears but the proposed circuit does not — is evidence against the mechanistic claim.

Methods that modify the training process and observe consequences:

  • Ablate-then-retrain — removing a component (by zeroing or masking weights) and continuing training. Tests whether the model recovers the capability through alternative pathways, which bears on I5 Rival mechanism exclusion.
  • Fine-tuning — targeted training on specific data and measuring which components change. Components that change during task-specific fine-tuning are implicated in that task.
  • Curriculum manipulation — modifying the training data distribution and observing effects on mechanism formation. Tests whether the mechanism depends on specific data patterns or emerges from general optimization pressure.
  • Knockout training — training with specific components held fixed (frozen) and measuring whether the capability still develops.

Interventional training evidence addresses I12 Offset coupling: does the mechanism go when the capability is removed? If fine-tuning on data that removes a capability also degrades the proposed circuit, this is evidence for a causal link between the circuit and the capability. The neuroscience analogue is Craver’s “top-down leg” of mutual manipulability: manipulating the system-level phenomenon (the capability) affects the component-level mechanism (the circuit).

Training evidence is most directly relevant to:

CriterionHow training evidence bears on it
I11 Onset couplingDoes the mechanism form when the capability develops?
I12 Offset couplingDoes the mechanism degrade when the capability is removed?
I5 Rival mechanism exclusionDoes ablate-then-retrain recover via alternative pathways?
M3 StabilityIs the mechanism stable across training seeds?
I10 Rescue reversibilityDoes restoring a component after ablation-then-retrain recover the original mechanism?