Representational Depth of Evaluation Awareness Shifts With Scale in Open-Weight Language Models

• URL: http://arxiv.org/abs/2606.29196v1
• Date: Sun, 28 Jun 2026 04:48:17 GMT
• Title: Representational Depth of Evaluation Awareness Shifts With Scale in Open-Weight Language Models
• Abstract summary: A model that recognises an evaluation context could alter its behaviour strategically, making downstream benchmarks harder to interpret.<n>Using 11 models spanning Qwen 2.5, Gemma 2, and Llama 3.2, we find a systematic size-dependent shift in representational depth.<n>This suggests that scale changes not only the strength of evaluation-awareness but also where it is most linearly recoverable in the network.
• Score: 0.0
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Do language models know when they are being tested? This question matters for AI safety: a model that recognises an evaluation context could alter its behaviour strategically, making downstream benchmarks harder to interpret. Using 11 models spanning Qwen 2.5, Gemma 2, and Llama 3.2, we find a systematic size-dependent shift in representational depth: in both Qwen 2.5 and Gemma 2, the layer at which evaluation-awareness is most linearly recoverable moves from late layers in smaller models to early layers in larger ones. This suggests that scale changes not only the strength of evaluation-awareness but also where it is most linearly recoverable in the network. This depth shift helps explain why within-family scaling trajectories are non-monotonic or inverse rather than smooth and family-general, showing that a simple universal power-law account is not supported under denser within-family sampling. Finally, white-box probe signals are consistently stronger than black-box behavioural expression, and the relationship between the two varies by family in ways not predicted by probe AUROC alone.

Related papers

• What Does Layer-Importance Reveal About Transformers and State-Space Models? [27.103840607246685]
  We decompose layer importance into two distinct notions:Necessity and Plasticity.<n>In every evaluated residual transformer up to $14$B parameters, Necessity and Plasticity anti-align across depth, whereas in the evaluated Mamba-style SSMs they point to overlapping regions.<n>In the evaluated transformers, concentrating updates in the most plastic layers increases catastrophic forgetting, while this tier-dependent effect disappears in the evaluated Mamba-style SSMs.
  arXiv  Detail & Related papers  (2026-09-15T02:34:13Z)
• When Is a Steerable Concept Representation Real? Measurement Confounds in a Cross-Family Audit of Neuroscience Parallels in LLMs [8.657685524975246]
  Large language models (LLMs) are increasingly reported to exhibit human-like neural and cognitive signatures.<n>Our main experiment examines the causal steerability of concept directions.<n>With raw activation units and a fixed layer and coefficient, steerability appears to increase with model scale.<n>With residual-norm-comparable interventions and held-out operating-point selection, concept steering remains significant at every scale.
  arXiv  Detail & Related papers  (2026-08-08T14:37:53Z)
• Will Scaling Improve Social Simulation with LLMs? [87.15677770205751]
  Large Language Model (LLM) social simulations are a promising research method, but they are not yet faithful enough to be adopted widely.<n>We use scaling laws to study the relationship between LLMs' compute scale, general capability benchmarks, and the fidelity of social simulation.
  arXiv  Detail & Related papers  (2026-07-02T17:30:38Z)
• LogitDynamics: Reliable ViT Error Detection from Layerwise Logit Trajectories [1.4610038284393168]
  We study error prediction in Vision Transformers (ViTs)<n>Motivated by internal-signal hallucination detection in large language models, we investigate whether similar depth-wise signals exist inViTs.<n>By attaching lightweight linear heads to intermediate layers, we extract features from the last L layers that capture both the logits of the predicted class and its top-K competitors.<n>A linear probe trained on these features predicts the error indicator.
  arXiv  Detail & Related papers  (2026-04-12T13:43:40Z)
• MAR-MAER: Metric-Aware and Ambiguity-Adaptive Autoregressive Image Generation [1.4552327135549117]
  We introduce MAR-MAER, an innovative hierarchical autoregressive framework.<n>It combines two main components: It is a metric-aware embedding regularization method, and it is a latent model used for handling ambiguous semantics.<n>Our method aligns the model's internal representations with human-preferred quality metrics, such as CLIPScore and HPSv2.<n>It achieves excellent performance in both metric consistency and semantic flexibility.
  arXiv  Detail & Related papers  (2026-04-02T10:19:46Z)
• From Early Encoding to Late Suppression: Interpreting LLMs on Character Counting Tasks [49.57538588967748]
  Large language models (LLMs) exhibit failures on elementary symbolic tasks such as character counting in a word, despite excelling on complex benchmarks.<n>We uncover a consistent phenomenon across modern architectures, including LLaMA, Qwen, and Gemma.<n>We show that symbolic reasoning failures in LLMs are not due to missing representations or insufficient scale, but arise from structured interference within the model's computation graph.
  arXiv  Detail & Related papers  (2026-04-01T11:40:12Z)
• Sparse Semantic Dimension as a Generalization Certificate for LLMs [53.681678236115836]
  We introduce the Sparse Semantic Dimension (SSD), a complexity measure derived from the active feature vocabulary of a Sparse Autoencoder (SAE) trained on the model's layers.<n>We validate this framework on GPT-2 Small and Gemma-2B, demonstrating that our bound provides non-vacuous certificates at realistic sample sizes.
  arXiv  Detail & Related papers  (2026-02-11T21:45:18Z)
• Same Answer, Different Representations: Hidden instability in VLMs [65.36933543377346]
  We introduce a representation-aware and frequency-aware evaluation framework that measures internal embedding drift, spectral sensitivity, and structural smoothness.<n>We apply this framework to modern Vision Language Models (VLMs) across the SEEDBench, MMMU, and POPE datasets.
  arXiv  Detail & Related papers  (2026-02-06T12:24:26Z)
• Superposition Yields Robust Neural Scaling [22.05527470952902]
  We show that representation superposition, meaning that LLMs represent more features than they have dimensions, can be a key contributor to loss and cause neural scaling.<n>Our results identify representation superposition as a central driver of neural scaling laws, providing insights into questions like when neural scaling laws can be improved and when they will break down.
  arXiv  Detail & Related papers  (2025-05-15T16:18:13Z)
• Interpretability at Scale: Identifying Causal Mechanisms in Alpaca [62.65877150123775]
  We use Boundless DAS to efficiently search for interpretable causal structure in large language models while they follow instructions. Our findings mark a first step toward faithfully understanding the inner-workings of our ever-growing and most widely deployed language models.
  arXiv  Detail & Related papers  (2023-05-15T17:15:40Z)
• Second-Moment Loss: A Novel Regression Objective for Improved Uncertainties [7.766663822644739]
  Quantification of uncertainty is one of the most promising approaches to establish safe machine learning. One of the most commonly used approaches so far is Monte Carlo dropout, which is computationally cheap and easy to apply in practice. We propose a new objective, referred to as second-moment loss ( UCI), to address this issue.
  arXiv  Detail & Related papers  (2020-12-23T14:17:33Z)
• How do Decisions Emerge across Layers in Neural Models? Interpretation with Differentiable Masking [70.92463223410225]
  DiffMask learns to mask-out subsets of the input while maintaining differentiability. Decision to include or disregard an input token is made with a simple model based on intermediate hidden layers. This lets us not only plot attribution heatmaps but also analyze how decisions are formed across network layers.
  arXiv  Detail & Related papers  (2020-04-30T17:36:14Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.