Mechanistic Measurements · Interactive Summary

Instruction-Conditioned Separation & Displacement Across Layers

Qwen2.5-7B-Instruct (3 runs) + Qwen2.5-7B base (1 run) · TriviaQA slice · Layerwise AUC separability + displacement vector norms
Instruct Model Qwen2.5-7B-Instruct
Base Model Qwen2.5-7B (base)
Total Runs 4 (Mar 01–04 2026)
Metrics AUC (linear separability) · ‖v*‖ (norm)

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.

§ I

Methodological Condition

Methodological condition
“A claim about a transformer is only explanatory to the extent it can be expressed in front-to-back transformer dynamics.”

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.

§ II

Experimental Summary

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.

Run _203703
Instruct · 4-bit NF4
BF16 compute · double quant
Run _225321
Instruct · 8-bit INT8
FP16 weights off · thresh 6.0
Run _005558
Instruct · 8-bit INT8+FP16
FP16 activation kernel
Run _072206
Base · 8-bit INT8+FP16
Null control (same architecture)

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.

§ III

Interactive Results

Instruct Model (Representative Run)

AUC by Layer — Instruct · canonical prompt · Run _225321
Highlight: Layer 6 (AUC = 1.0)
Highlight: Layer 9 (AUC = 1.0)
This visualization interpolates sampled layers to 28 bars.
‖v*‖ by Layer — Instruct · canonical prompt · Run _225321
Highlight: Layer 26 (peak norm)
This visualization interpolates sampled layers to 28 bars.

Base Model (Null Control)

AUC by Layer — Base · canonical prompt · Run _072206
Highlight: Layer 14 (best AUC in this run)
Highlight: Layer 11 (trough in this run)
No early focal point is highlighted in this base run.
‖v*‖ by Layer — Base · canonical prompt · Run _072206
Highlight: Layer 26 (peak norm)
Layer 26 remains the norm peak in this base run.
Instruct · Run _005558 (matched)
Highlighted: Layer 6
Base · Run _072206
Highlights: Layer 14 (best), Layer 11 (trough)

What Is Being Claimed

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.
§ IV

Norm Amplification Trajectory

‖v*‖ Growth — Instruct vs Base (canonical)

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.

§ V

The 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.