finding

active

finding:at-layer-12-the-layer-analyzed-by-burger-et-al-2024-tp-and-tg-explain-similar-fractions-of-truth-related-variance-0-33-each

At layer 12 (the layer analyzed by Burger et al. 2024), tP and tG explain similar fractions of truth-related variance (~0.33 each).

Shows that Burger et al.'s layer choice corresponds to a transitional phase, not a universal property.

Source paper

extracted_from

Testing the Limits of Truth Directions in LLMs

(2026) · Angelos Poulis · Mark Crovella · Evimaria Terzi

Neighborhood — ranked by edge-count

Claims (1)

claim

The two-dimensional subspace reported by Burger et al. reflects a transitional phase in model processing rather than a universal property of truth directions.
supports
Reinterpretation of Burger et al.'s finding as layer-specific rather than universal.

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

In early layers, the polarity-dependent direction tP explains ~0.38 of truth-related variance at layer 7 vs ~0.09 for tG; by middle layers tG takes over and tP decays.finding0.848
Variance decomposition showing the disentanglement of polarity from truth across model depth.
The model converges to a more stable truth direction in middle-to-late layers, as evidenced by increasing cosine similarity between layer-wise probes.claim0.781
Supported by the geometric transition visible in cosine similarity heatmaps for F0-F3.
Truth-related directions reliably emerge at 60–75% of normalized layer depth in Qwen and Gemma modelsfinding0.771
Experiment 1 finding localizing where truth can be causally mediated
Single-layer analyses can be misleading because early-layer truth directions may reflect surface features with limited cross-task generalization.claim0.768
Methodological critique of prior work that fixed a single layer for truth probing.
The middle layer residual stream features are causally implicated in multi-step reasoning.claim0.760
Features for Kobe Bryant, California, Lakers participate in computing the capital answer.
F0-trained probes in layers 4-10 show inverted separation on F1 (AUROC ≈ 0), systematically misclassifying true statements as false.finding0.758
Demonstrates that early-layer probes capture sentence polarity rather than truth.
No single layer is universally optimal for probing truth directions; different tasks peak at different layers.claim0.756
Argues against the single-layer analysis approach of prior work.
Layer 29 (indexed at 10) of LLaMA3.1-8B on Strange Stories (2 scores) satisfies Criteria 1 and 2 under IIT 4.0 (temporal permutation).finding0.755
Third promising case from temporal permutation analysis.