This page is scoped to what was measured: layerwise
linear separability (AUC) between two behavioral conditions (“Verbose”
vs “Concise”) and the norm of the corresponding displacement vector
v*, sampled at specific layers and interpolated to 28
layers for visualization.
This page does not claim to identify circuits, features, or a unique internal mechanism. Any mechanistic interpretation is treated as a hypothesis layered on top of the measurements.
In practice, this means: (1) prefer measurements that are directly tied to model states across layers; (2) treat stories about “what the model is doing” as hypotheses unless they are constrained by such measurements; (3) explicitly separate what was measured from what might explain it.
The visualizations on this page report measurements of model states across layers. Any interpretation of these patterns should be treated as a hypothesis unless it can be expressed directly in terms of the model’s forward dynamics.
Four runs were conducted on identical dataset splits (TriviaQA, 1200
questions, 800 train / 400 eval, seed 42,
slice_hash: 396225b6…). Three runs use
Qwen2.5-7B-Instruct under different quantization regimes.
One run uses the same-architecture base model Qwen2.5-7B
as a null-model control with matched quantization to one instruct run.
Empirical pattern (as reported): In the instruct model, separability peaks early (Layer 6 / Layer 9 depending on prompt family), while displacement norm peaks late (Layer 26). In the base model, the Layer 26 norm peak persists but the early separability focal point is absent.
| Statement | Grounded in measurement? | Notes |
|---|---|---|
| Instruct runs show early perfect separability (AUC = 1.0 at L6/L9 for some families) | Yes | Shown in the instruct AUC visualization and reported table values. |
| Both instruct and base show a late-layer norm peak at Layer 26 | Yes | Shown in both norm visualizations. |
| Base model lacks the instruct model’s “Layer 6 focal point” in AUC | Yes | Base AUC profile differs; best AUC occurs elsewhere in the reported run. |
| This pattern implies an “input-conditioned Jacobian gain” mechanism | No (hypothesis) | Kept as an operating hypothesis; not asserted as proven by these measurements alone. |
Measurement: both models show increasing ‖v*‖ toward Layer 26. The instruct run shows steeper growth. Interpretation (hypothesis): instruction tuning may amplify an existing late-layer gain structure rather than creating it entirely.
word_prepend Inversion
In the instruct model, the word_prepend family shows
decreasing separability toward depth, degrading to roughly AUC ≈
0.87 at Layer 26.
In the base model, the same family becomes the most stable at depth, reaching AUC ≈ 0.996 at Layer 26.
Meanwhile the canonical family — the most stable in the
instruct model — shows the sharpest non-monotonic profile in the
base model, with a trough at Layer 11 (AUC ≈ 0.954).
A possible explanation is that instruction tuning trains the model to route phrase-level instruction templates efficiently through early layers (e.g. “Answer verbosely.”).
Under this interpretation, a bare positional token (verbose
at position 0) may require deeper processing in the instruct model,
while in the base model the token acts as a generic lexical cue
rather than an instruction signal.
This hypothesis could be tested by examining attention patterns or residual activations around Layer 26 for the two prompt families.