In real Qwen2.5-32B-Instruct and Qwen3-14B (and consistently across the Qwen-2.5, Qwen-3 and Gemma-3 families, 4B-32B), logical-validity judgments and content plausibility are both linearly represented and strongly geometrically aligned, and the degree of this alignment predicts the magnitude of behavioral content-effect biases

- Bertolazzi, Pezzelle & Bernardi study content effects (plausibility of a reasoning problem's content biasing judgments of its logical validity) in real Qwen-2.5, Qwen-3, and Gemma-3 family models (4B-32B parameters), primarily reporting results for Qwen2.5-32B-Instruct and Qwen3-14B as reference models with consistent patterns across all evaluated models. [bertolazzi-etal-2025-content-effects-validity-plausibility] - A 3D PCA projection of layer-50 hidden states from real Qwen2.5-32B-Instruct in the zero-shot setting shows four well-separated clusters corresponding to valid/invalid and true/false model predictions, with the plausibility clusters displaced along a similar direction to the validity clusters; the authors describe this as a qualitative visualization suggesting parallel representational structure, and quantify it with the average cosine similarity between difference-in-means validity and plausibility vectors (0.48-0.64 over highly steerable layers, vs 0.10-0.13 for hypernymy and -0.12 to -0.17 for harmlessness controls; Sec. 4.2, Fig. 3). [bertolazzi-etal-2025-content-effects-validity-plausibility] - Using steering vectors, they show that plausibility vectors can causally flip validity judgements in a subset of late layers and vice versa (Fig. 4), and a mixed-effects regression over 57 model-prompt points finds average plausibility-validity similarity to be a significant positive predictor of the behavioral content effect (beta=0.557, p=0.017; Sec. 4.3), associating the representational alignment with the reasoning bias. [bertolazzi-etal-2025-content-effects-validity-plausibility]