Layer-wise probing shows clearer middle-layer separability in the general model and a weaker, deeper separability profile in the coder model¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? The general model reaches peak linear separability at Layer 29 (FDR ≈ 0.0303) for difficulty and Layer 28 (FDR ≈ 0.0224) for Bloom targets, while the coder model's difficulty peak FDR remains low (≈ 0.0108 at Layer 28) and its Bloom peak shifts to Layer 36 of 48. → Yi Zhang & Julia Rayz 2026

Evidence¶

Yi Zhang & Julia Rayz 2026¶

Yi Zhang & Julia Rayz. (2026). From Execution to Education: A Bloom-Aligned Framework for Measuring Educational Control in LLMs. Preprint. https://arxiv.org/abs/2607.08009 doi:10.48550/arxiv.2607.08009

q2 · i? · causal · r2

Layer-by-layer Fisher's Discriminant Ratio probing of residual-stream activations at the last instruction token across all 48 layers, with 95% confidence intervals from bootstrap resampling with 1000 samples (Figure 1).

"Specifically, the model reaches peak linear separability at Layer 29 (FDR ≈ 0.0303) for difficulty and Layer 28 (FDR≈ 0.0224) for Bloom targets."

Yi Zhang & Julia Rayz 2026 (2)¶

Yi Zhang & Julia Rayz. (2026). From Execution to Education: A Bloom-Aligned Framework for Measuring Educational Control in LLMs. Preprint. https://arxiv.org/abs/2607.08009

q2 · i? · causal · r2

Coder-model FDR trajectory contrasted with the general model's; projection distributions (Figures 2 and 3) show the coder model with more overlap and higher variance along the mean-difference direction.

"For general difficulty instructions, its peak FDR remains low (≈ 0.0108 at Layer 28), indicating more overlapping activation distributions under this linear diagnostic. Furthermore, when processing targeted Bloom prompts, the layer of maximum separability shifts deeper into the network, peaking at Layer 36 of 48."

Discussion¶

Related Claims¶

• Both models target Higher Bloom levels far more accurately than Lower levels, with the general model outperforming the coder model upward — related
• LLMs show a directional asymmetry in educational control: they reliably increase cognitive demand but struggle to lower it — related