Probe-Rewrite-Evaluate: A Workflow for Reliable Benchmarks and Quantifying Evaluation Awareness

• URL: http://arxiv.org/abs/2509.00591v6
• Date: Thu, 06 Nov 2025 07:25:07 GMT
• Title: Probe-Rewrite-Evaluate: A Workflow for Reliable Benchmarks and Quantifying Evaluation Awareness
• Authors: Lang Xiong, Nishant Bhargava, Jianhang Hong, Jeremy Chang, Haihao Liu, Vasu Sharma, Kevin Zhu,
• Abstract summary: Large Language Models (LLMs) often exhibit significant behavioral shifts when they perceive a change from a real-world deployment context to a controlled evaluation setting.<n>This discrepancy poses a critical challenge for AI alignment, as benchmark performance may not accurately reflect a model's true safety and honesty.<n>We introduce a methodology that uses a linear probe to score prompts on a continuous scale from "test-like" to "deploy-like"
• Score: 6.071703608560761
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large Language Models (LLMs) often exhibit significant behavioral shifts when they perceive a change from a real-world deployment context to a controlled evaluation setting, a phenomenon known as "evaluation awareness." This discrepancy poses a critical challenge for AI alignment, as benchmark performance may not accurately reflect a model's true safety and honesty. In this work, we systematically quantify these behavioral changes by manipulating the perceived context of prompts. We introduce a methodology that uses a linear probe to score prompts on a continuous scale from "test-like" to "deploy-like" and leverage an LLM rewriting strategy to shift these prompts towards a more natural, deployment-style context while preserving the original task. Using this method, we achieved a 30% increase in the average probe score across a strategic role-playing dataset after rewriting. Evaluating a suite of state-of-the-art models on these original and rewritten prompts, we find that rewritten "deploy-like" prompts induce a significant and consistent shift in behavior. Across all models, we observed an average increase in honest responses of 5.26% and a corresponding average decrease in deceptive responses of 12.40%. Furthermore, refusal rates increased by an average of 6.38%, indicating heightened safety compliance. Our findings demonstrate that evaluation awareness is a quantifiable and manipulable factor that directly influences LLM behavior, revealing that models are more prone to unsafe or deceptive outputs in perceived test environments. This underscores the urgent need for more realistic evaluation frameworks to accurately gauge true model alignment before deployment.

Related papers

• CORE: Context-Robust Remasking for Diffusion Language Models [51.59514489363897]
  We propose Context-Robust Remasking (CORE), a training-free framework for inference-time revision.<n>Rather than trusting static token probabilities, CORE identifies context-brittle tokens by probing their sensitivity to targeted masked-context perturbations.<n>On LLaDA-8B-Base, CORE delivers consistent improvements across reasoning and code benchmarks, outperforming compute-matched baselines and improving MBPP by up to 9.2 percentage points.
  arXiv  Detail & Related papers  (2026-02-04T00:12:30Z)
• Gaming the Judge: Unfaithful Chain-of-Thought Can Undermine Agent Evaluation [76.5533899503582]
  Large language models (LLMs) are increasingly used as judges to evaluate agent performance.<n>We show this paradigm implicitly assumes that the agent's chain-of-thought (CoT) reasoning faithfully reflects both its internal reasoning and the underlying environment state.<n>We demonstrate that manipulated reasoning alone can inflate false positive rates of state-of-the-art VLM judges by up to 90% across 800 trajectories spanning diverse web tasks.
  arXiv  Detail & Related papers  (2026-01-21T06:07:43Z)
• A methodological analysis of prompt perturbations and their effect on attack success rates [0.5387033080274478]
  This work aims to investigate how different Large Language Models (LLMs) alignment methods affect the models' responses to prompt attacks.<n>We selected open source models based on the most common alignment methods, namely, Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Reinforcement Learning with Human Feedback (RLHF)
  arXiv  Detail & Related papers  (2025-11-11T19:39:33Z)
• Sampling-aware Adversarial Attacks Against Large Language Models [52.30089653615172]
  Existing adversarial attacks typically target harmful responses in single-point greedy generations.<n>We show that for the goal of eliciting harmful responses, repeated sampling of model outputs during the attack prompt optimization.<n>We show that integrating sampling into existing attacks boosts success rates by up to 37% and improves efficiency by up to two orders of magnitude.
  arXiv  Detail & Related papers  (2025-07-06T16:13:33Z)
• Aurora: Are Android Malware Classifiers Reliable and Stable under Distribution Shift? [51.12297424766236]
  AURORA is a framework to evaluate malware classifiers based on their confidence quality and operational resilience.<n>AURORA is complemented by a set of metrics designed to go beyond point-in-time performance.<n>The fragility in SOTA frameworks across datasets of varying drift suggests the need for a return to the whiteboard.
  arXiv  Detail & Related papers  (2025-05-28T20:22:43Z)
• RAG-Zeval: Towards Robust and Interpretable Evaluation on RAG Responses through End-to-End Rule-Guided Reasoning [64.46921169261852]
  RAG-Zeval is a novel end-to-end framework that formulates faithfulness and correctness evaluation as a rule-guided reasoning task.<n>Our approach trains evaluators with reinforcement learning, facilitating compact models to generate comprehensive and sound assessments.<n>Experiments demonstrate RAG-Zeval's superior performance, achieving the strongest correlation with human judgments.
  arXiv  Detail & Related papers  (2025-05-28T14:55:33Z)
• ReliableEval: A Recipe for Stochastic LLM Evaluation via Method of Moments [21.37415398600286]
  We argue for a method of moments evaluation over the space of meaning-preserving prompt perturbations.<n>We show that even top-performing models like GPT-4o and Claude-3.7-Sonnet exhibit substantial prompt sensitivity.
  arXiv  Detail & Related papers  (2025-05-28T09:40:48Z)
• Meta-Evaluating Local LLMs: Rethinking Performance Metrics for Serious Games [3.725822359130832]
  Large Language Models (LLMs) are increasingly being explored as evaluators in serious games.<n>This study investigates the reliability of five small-scale LLMs when assessing player responses in textitEn-join, a game that simulates decision-making within energy communities.<n>Our results highlight the strengths and limitations of each model, revealing trade-offs between sensitivity, specificity, and overall performance.
  arXiv  Detail & Related papers  (2025-04-13T10:46:13Z)
• Efficient Safety Alignment of Large Language Models via Preference Re-ranking and Representation-based Reward Modeling [84.00480999255628]
  Reinforcement Learning algorithms for safety alignment of Large Language Models (LLMs) encounter the challenge of distribution shift.<n>Current approaches typically address this issue through online sampling from the target policy.<n>We propose a new framework that leverages the model's intrinsic safety judgment capability to extract reward signals.
  arXiv  Detail & Related papers  (2025-03-13T06:40:34Z)
• A Baseline Analysis of Reward Models' Ability To Accurately Analyze Foundation Models Under Distribution Shift [2.2310395620011945]
  We evaluate how reward model performance is affected by distribution shift. We show novel calibration patterns and accuracy drops due to OOD prompts and responses. We adapt an OOD detection technique commonly used in classification to the reward model setting to detect these distribution shifts.
  arXiv  Detail & Related papers  (2023-11-21T18:41:26Z)
• Bring Your Own Data! Self-Supervised Evaluation for Large Language Models [52.15056231665816]
  We propose a framework for self-supervised evaluation of Large Language Models (LLMs) We demonstrate self-supervised evaluation strategies for measuring closed-book knowledge, toxicity, and long-range context dependence. We find strong correlations between self-supervised and human-supervised evaluations.
  arXiv  Detail & Related papers  (2023-06-23T17:59:09Z)
• How Reliable is Your Regression Model's Uncertainty Under Real-World Distribution Shifts? [46.05502630457458]
  We propose a benchmark of 8 image-based regression datasets with different types of challenging distribution shifts. We find that while methods are well calibrated when there is no distribution shift, they all become highly overconfident on many of the benchmark datasets.
  arXiv  Detail & Related papers  (2023-02-07T18:54:39Z)

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.