Is Evaluation Awareness Just Format Sensitivity? Limitations of Probe-Based Evidence under Controlled Prompt Structure

• URL: http://arxiv.org/abs/2603.19426v1
• Date: Thu, 19 Mar 2026 19:39:35 GMT
• Title: Is Evaluation Awareness Just Format Sensitivity? Limitations of Probe-Based Evidence under Controlled Prompt Structure
• Abstract summary: We find that probes primarily track benchmark-canonical structure and fail to generalize to free-form prompts.<n>Standard probe-based methodologies do not reliably disentangle evaluation context from structural artifacts.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Prior work uses linear probes on benchmark prompts as evidence of evaluation awareness in large language models. Because evaluation context is typically entangled with benchmark format and genre, it is unclear whether probe-based signals reflect context or surface structure. We test whether these signals persist under partial control of prompt format using a controlled 2x2 dataset and diagnostic rewrites. We find that probes primarily track benchmark-canonical structure and fail to generalize to free-form prompts independent of linguistic style. Thus, standard probe-based methodologies do not reliably disentangle evaluation context from structural artifacts, limiting the evidential strength of existing results.

Related papers

• Structuring Semantic Embeddings for Principle Evaluation: A Prototype-Guided Contrastive Learning Approach [58.15427952222424]
  General-purpose text embeddings are widely deployed for such tasks, but broad semantic similarity can place semantically similar yet task-distinct examples.<n>We introduce Prototype-Guided Contrastive Learning (PGCL), a prototype-guided geometric regularization module built on top of frozen text embeddings.
  arXiv  Detail & Related papers  (2026-08-15T13:19:25Z)
• On the Non-Specificity of Statistical Measures Used in Script Decipherment [0.0]
  SIGIL is a purpose-built generative emblem system whose 3,000-text core corpus carries explicit compositional meanings.<n>A literature registry compiled in advance of evaluation records 54 methods and admits a method to exact scoring when both the published Indus outcome and a source-defined decision rule can be reproduced.<n> SIGIL receives the same category as the Indus corpus on every criterion scored this way, across repetition, directional-asymmetry, and lexical-distribution tests.
  arXiv  Detail & Related papers  (2026-08-04T01:33:02Z)
• From Signals to Transfer: A Factorised Study of Probe-Based Uncertainty Estimation in Large Language Models [51.470637128942336]
  We propose a factorised study of probe-based uncertainty estimation (UE) under matched conditions.<n>Our results show that raw hidden states and attention features are difficult to outperform in-domain.<n>We train benchmark-based pretrained probes that transfer reasonably well to open-ended factual generation.
  arXiv  Detail & Related papers  (2026-06-26T03:23:15Z)
• Localizing Prompt Ambiguity in Large Language Models with Probe-Targeted Attribution [9.857349182313092]
  We introduce PRIG, a gradient attribution method that uses a probe logit to attribute latent ambiguity to token positions.<n>Specifically, PRIG trains a linear probe to distinguish clear prompts from ambiguous prompts and attributes the probe score to earlier token representations in the residual stream.<n>In this setting, PRIG localizes ambiguous spans substantially better than gradient attribution baselines, achieving 0.840 AUROC on the combined synthetic benchmark and 0.891 AUROC on the gold set.
  arXiv  Detail & Related papers  (2026-06-03T22:15:43Z)
• When Negation Is a Geometry Problem in Vision-Language Models [32.51815690470519]
  Joint Vision-Language Embedding models such as CLIP typically fail at understanding negation in text queries.<n>We investigate an alternative evaluation framework based on Multimodal LLMs-as-a-judge, which excel at understanding simple yes/no questions about image content.
  arXiv  Detail & Related papers  (2026-03-20T23:06:23Z)
• Who Guards the Guardians? The Challenges of Evaluating Identifiability of Learned Representations [25.62598569130843]
  We show that each metric implicitly encodes assumptions about both the data-generating process (DGP) and the encoder.<n>When these assumptions are violated, metrics become misspecified and can produce systematic false positives and false negatives.<n>We introduce a taxonomy separating DGP assumptions from encoder geometry, use it to characterise the validity domains of existing metrics, and release an evaluation suite for reproducible stress testing and comparison.
  arXiv  Detail & Related papers  (2026-02-27T18:50:13Z)
• Alignment Verifiability in Large Language Models: Normative Indistinguishability under Behavioral Evaluation [0.0]
  We study alignment evaluation through the lens of statistical identifiability under partial observability.<n>We formalize the Alignment Verifiability Problem and introduce Normative Indistinguishability.<n>Our results show that behavioral benchmarks provide necessary but insufficient evidence for latent alignment under evaluation awareness.
  arXiv  Detail & Related papers  (2026-02-05T13:40:56Z)
• Probing Syntax in Large Language Models: Successes and Remaining Challenges [7.9494253785082405]
  It remains unclear whether structural and/or statistical factors systematically affect these syntactic representations.<n>We conduct an in-depth analysis of structural probes on three controlled benchmarks.
  arXiv  Detail & Related papers  (2025-08-05T08:41:14Z)
• SEOE: A Scalable and Reliable Semantic Evaluation Framework for Open Domain Event Detection [70.23196257213829]
  We propose a scalable and reliable Semantic-level Evaluation framework for Open domain Event detection.<n>Our proposed framework first constructs a scalable evaluation benchmark that currently includes 564 event types covering 7 major domains.<n>We then leverage large language models (LLMs) as automatic evaluation agents to compute a semantic F1-score, incorporating fine-grained definitions of semantically similar labels.
  arXiv  Detail & Related papers  (2025-03-05T09:37:05Z)
• StructTest: Benchmarking LLMs' Reasoning through Compositional Structured Outputs [78.84060166851805]
  StructTest is a novel benchmark that evaluates large language models (LLMs) on their ability to follow compositional instructions and generate structured outputs.<n> Assessments are conducted deterministically using a rule-based evaluator, which can be easily extended to new tasks and datasets.<n>We demonstrate that StructTest remains challenging even for top-performing models like Deepseek-V3/R1 and GPT-4o.
  arXiv  Detail & Related papers  (2024-12-23T22:08:40Z)
• Mitigating Bias in Text Classification via Prompt-Based Text Transformation [0.2864713389096699]
  We investigate whether prompting ChatGPT to rewrite text can reduce demographic signals while preserving meaning.<n>Results suggest that prompt-based rewriting offers a practical and generalisable approach for mitigating bias in text classification.
  arXiv  Detail & Related papers  (2023-05-09T13:10:23Z)
• On the Blind Spots of Model-Based Evaluation Metrics for Text Generation [79.01422521024834]
  We explore a useful but often neglected methodology for robustness analysis of text generation evaluation metrics. We design and synthesize a wide range of potential errors and check whether they result in a commensurate drop in the metric scores. Our experiments reveal interesting insensitivities, biases, or even loopholes in existing metrics.
  arXiv  Detail & Related papers  (2022-12-20T06:24:25Z)
• Not another Negation Benchmark: The NaN-NLI Test Suite for Sub-clausal Negation [59.307534363825816]
  Negation is poorly captured by current language models, although the extent of this problem is not widely understood. We introduce a natural language inference (NLI) test suite to enable probing the capabilities of NLP methods.
  arXiv  Detail & Related papers  (2022-10-06T23:39:01Z)
• Infusing Finetuning with Semantic Dependencies [62.37697048781823]
  We show that, unlike syntax, semantics is not brought to the surface by today's pretrained models. We then use convolutional graph encoders to explicitly incorporate semantic parses into task-specific finetuning.
  arXiv  Detail & Related papers  (2020-12-10T01:27:24Z)

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.