Improving Evaluation Realism with Inference-Time Compute and Deployment Scaffolds

• URL: http://arxiv.org/abs/2609.02302v1
• Date: Wed, 02 Sep 2026 08:47:34 GMT
• Title: Improving Evaluation Realism with Inference-Time Compute and Deployment Scaffolds
• Abstract summary: Two techniques make simulated alignment evaluations harder to distinguish from real deployments.<n>Dish (Deployment-Imitating SWE-Agent Harness) wraps the target in an agent harness, reducing the gap between simulated and real deployment environments.<n>Results show that automated approaches can improve the realism of alignment evaluations.
• Score: 35.96704508133401
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: A core obstacle to alignment evaluation is evaluation awareness: capable models can tell when they are being tested rather than deployed, weakening the conclusions a safety evaluation can support. We present two techniques that make simulated alignment evaluations harder to distinguish from real deployments. Our first technique, critique refinement, spends additional inference-time compute on each simulator action: the simulator generates multiple candidate actions, refines them using feedback from an instance of the target model on how to make them more realistic, and continues the evaluation with the most deployment-like candidate. Our second technique, DISH (Deployment-Imitating SWE-Agent Harness), wraps the target in an agent harness, reducing the gap between simulated and real deployment environments in coding settings. We test the techniques on multiple target models and find that they compose: applying both yields larger realism gains than either alone. Our results show that automated approaches can improve the realism of alignment evaluations, and that these improvements use additional compute more effectively than making the audits longer.

Related papers

• CAST: Critique-Aware Supervision for Training Reliable Long-Horizon Tool-Calling Agents [66.9575676266637]
  Large language model (LLM) agents are increasingly deployed in long-horizon, interactive, and stateful environments.<n>In these settings, a single wrong action, such as refunding the wrong purchase, can cause irreversible task failure and must be intercepted before execution.<n>We develop CAST, a critique-aware training framework that converts sparse task outcomes into action-level supervision for critique learning and policy optimization.
  arXiv  Detail & Related papers  (2026-08-31T01:59:51Z)
• Predicting LLM Safety Before Release by Simulating Deployment [5.628199352593384]
  Pre-deployment safety evaluations aim to inform the downstream risks of releasing a new AI model.<n>Most evaluations provide limited evidence about how often undesired model behavior will occur in deployment.<n>We show that deployment simulation can be seeded from public chat datasets and remain informative about production misbehavior rates.
  arXiv  Detail & Related papers  (2026-07-08T09:17:47Z)
• Skill-Aligned Annotation for Reliable Evaluation in Text-to-Image Generation [50.85337196571056]
  Text-to-image (T2I) generation has advanced rapidly, making reliable evaluation critical as performance differences between models narrow.<n>Existing evaluation practices typically apply uniform annotation mechanisms, such as Likert-scale or binary question answering (BQA)<n>We revisit T2I evaluation through the lens of skill-aligned annotation, where annotation strategies reflect the underlying characteristics of each evaluation skill.
  arXiv  Detail & Related papers  (2026-05-13T09:14:31Z)
• Automated Model Evaluation for Object Detection via Prediction Consistency and Reliability [3.9486037760311725]
  Prediction Consistency and Reliability (PCR) estimates detection performance without ground-truth labels.<n>We construct a meta-dataset by applying image corruptions of varying severity.<n>Results demonstrate that PCR yields more accurate performance estimates than existing AutoEval methods.
  arXiv  Detail & Related papers  (2025-08-16T15:39:56Z)
• Enhancing Training Data Attribution with Representational Optimization [57.61977909113113]
  Training data attribution methods aim to measure how training data impacts a model's predictions.<n>We propose AirRep, a representation-based approach that closes this gap by learning task-specific and model-aligned representations explicitly for TDA.<n>AirRep introduces two key innovations: a trainable encoder tuned for attribution quality, and an attention-based pooling mechanism that enables accurate estimation of group-wise influence.
  arXiv  Detail & Related papers  (2025-05-24T05:17:53Z)
• On the Role of Feedback in Test-Time Scaling of Agentic AI Workflows [71.92083784393418]
  Agentic AI (systems that autonomously plan and act) are becoming widespread, yet their task success rate on complex tasks remains low.<n>Inference-time alignment relies on three components: sampling, evaluation, and feedback.<n>We introduce Iterative Agent Decoding (IAD), a procedure that repeatedly inserts feedback extracted from different forms of critiques.
  arXiv  Detail & Related papers  (2025-04-02T17:40:47Z)
• Thinking Longer, Not Larger: Enhancing Software Engineering Agents via Scaling Test-Time Compute [61.00662702026523]
  We propose a unified Test-Time Compute scaling framework that leverages increased inference-time instead of larger models.<n>Our framework incorporates two complementary strategies: internal TTC and external TTC.<n>We demonstrate our textbf32B model achieves a 46% issue resolution rate, surpassing significantly larger models such as DeepSeek R1 671B and OpenAI o1.
  arXiv  Detail & Related papers  (2025-03-31T07:31:32Z)
• A Correct-and-Certify Approach to Self-Supervise Object Pose Estimators via Ensemble Self-Training [26.47895284071508]
  Real-world robotics applications demand object pose estimation methods that work reliably across a variety of scenarios. Our first contribution is to develop a robust corrector module that corrects pose estimates using depth information. Our second contribution is an ensemble self-training approach that simultaneously trains multiple pose estimators in a self-supervised manner.
  arXiv  Detail & Related papers  (2023-02-12T23:02:03Z)

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.