Improving Evaluation Realism with Inference-Time Compute and Deployment Scaffolds

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.

You Shouldn't Have Asked: A Pragmatics-Inspired Taxonomy for Evaluating LLM Refusals

Refusals are often treated as face-threatening acts in pragmatics because they can challenge the requester's socially claimed self-image. Large language models (LLMs) are increasingly trained to refuse unsafe and inappropriate requests, and these refusals may harm users when models fail to manage this interactional cost properly. While existing work has mainly approached LLM non-compliance as a safety-alignment outcome, it does not provide a way to evaluate whether LLMs refuse appropriately across different harmful contexts. To study this question, we propose (to our knowledge) the first taxonomy of LLM refusals that is grounded in pragmatic theory. Applying this taxonomy to responses from 16 modern LLMs across 14 harm categories, we find that although models differ in how they refuse, their refusals are overall explicit and strongly morally evaluative, with interactional repair occurring mainly through offering or providing safer alternatives instead of interpersonal facework. This pattern is especially consequential in sensitive harm contexts, where overuse of negative framing may make users feel shamed or provoked, undermining the purpose of safe non-compliance. We therefore call for alignment evaluation that considers not only whether models refuse harmful requests, but also whether they refuse in ways that are contextually adaptive and socially accountable for the interactional consequences of saying no.

Measuring Reward-Seeking via Contrastive Belief Updates

Language models trained with reinforcement learning may learn to optimize the grader's judgment rather than the intended objective. This "reward-seeking" is difficult to measure because a model that pursues the grader's judgment and one that pursues the intended objective behave identically whenever the grader rewards the intended behavior. We measure reward-seeking using Contrastive Synthetic Document Finetuning to change a model's beliefs about what the grader rewards, putting those beliefs in conflict with what users or developers want, and measuring the rate at which the model adopts each party's preferred behavior. Applied to intermediate checkpoints of a capabilities-focused OpenAI o3 RL run, without safety training, we find that these checkpoints often side with grader preferences over those of users or developers on coding and alignment tasks. This tendency to side with the grader trends upward throughout RL training. For example, in an environment that forces a choice between keeping a promise to a supervisor and breaking it to complete the task, a late capabilities-focused o3 checkpoint breaks the promise 87% of the time when SDF documents say the grader rewards task completion, versus 9% when they say it rewards honesty (a choice its chain-of-thought often makes explicit). An earlier checkpoint is far less sensitive (40% vs. 24%). Our method also generalizes to reward-hacking models. A model organism trained to reward-hack (gpt-oss-120b) is more than twice as sensitive to grader preferences as the unmodified model, with the mean behavioral shift in favor of the grader rising from 33% to 86%. These results indicate that RL can increase reward-seeking over the course of training, producing models that may act against their developers' intentions when they believe that doing so leads to higher reward.

Agreement Is Not Alignment: Divergent Moral Grounds in Human and LLM Ethical Judgments

Agreement with human judgments is a common proxy for evaluating the alignment of large language models (LLMs). Yet agreement in final labels does not show that human annotators and models rely on the same moral grounds. Two agents may reach the same judgment while appealing to different principles, contextual assumptions, or interpretations of the situation. We test this distinction using a curated 500-item ETHICS-derived benchmark spanning five domains of moral judgment, with new human annotator and LLM annotations of both final labels and supporting rationales. Across frontier and open model families, agreement with human annotator majority labels is often high. However, rationale-level analysis reveals systematic divergence in the moral grounds expressed by human annotators and models. In particular, models redistribute attention across categories such as harm, respect, promise-keeping, justice, desert, and excuse relevance, even when their final labels match the human annotator majority. Our results show that agreement should not be treated as equivalent to alignment. Label-based evaluation can therefore be misleadingly reassuring unless complemented by analysis of the reasons, principles, and moral priorities expressed in model judgments.

An Emergent Mirage: Is Emergent Misalignment and Realignment Indeed a Robust Phenomenon?

Recent work has reported Emergent Misalignment (EM), where language models fine-tuned on narrow, domain-specific misaligned datasets abruptly acquire broadly misaligned behavior, alongside evidence that this behavior can be reversed through limited realignment. We systematically study repeated alignment and misalignment cycles using controlled fine-tuning loops while tracking behavioral performance, and LoRA representations throughout training. Although we reproduce EM, we find that both misalignment and realignment are highly sensitive to superficial dataset characteristics, with apparent rapid realignment largely disappearing after controlling for response-length differences. We further find that previously reported mechanistic signatures, including representational phase transitions in LoRA space, do not consistently correlate with behavioral misalignment across training. Our results suggest that current evidence for EM is less robust than previously claimed and highlight the need for evaluation protocols that carefully control for these surface level dataset artifacts to identify the robustness of the EM phenomenon.

Prefill Awareness in Large Language Models

Safety-relevant studies of language models, including alignment and jailbreaking evaluations and AI control protocols, often rely on prefilling model outputs. If AI models can recognize and act on the fact their prior assistant messages have been inserted or edited, the effectiveness and validity of these methods could be compromised. We investigate whether frontier language models can distinguish between tampered and untampered assistant-side context, a capability we call prefill awareness. To do so, we construct a binary preference benchmark across three prefill mechanisms, filtering for cases where models show consistent stances. We find that frontier models show substantial prefill awareness: Claude Opus 4.5 detects prefills opposing its preferences in 9-35% of cases with a 0% false positive rate when prompted; additionally, models often revert towards baseline behavior without explicitly reporting that the prefill was foreign. Controlled ablations later also show that detection and resistance rely on different cues, where stylistic mismatch mainly affects whether models flag a prefill as foreign, while preference mismatch mainly affects whether they revert toward their baseline answer. We also examine more realistic agentic settings such as misalignment-continuation evaluations and SWE-bench trajectories, where frontier models sometimes disavow prefilled assistant turns in ways that depend strongly on dataset, task success, and hidden formatting artifacts. Our results indicate that prefill awareness is already a substantial confound for some prefill-based methods. We recommend that model developers track this capability in frontier systems.

Can AI Make Conflicts Worse? An Alignment Failure in LLM Deployment Across Conflict Contexts

AI models are already deployed in societies affected by armed conflict, and journalists, humanitarian workers, governments and ordinary citizens rely on them for information or for their work processes. No established practice exists for checking whether their outputs can make those conflicts worse. We tested nine model configurations from four providers (OpenAI, Anthropic, DeepSeek, xAI) on 90 multi-turn scenarios designed to surface misaligned behaviour in conflict contexts: false equivalence between documented atrocities, denial of genocide, and failure to recognise ethnic slurs, among others. When such outputs feed into journalism, humanitarian reporting, or public debate, they can deepen divisions in fragile societies. Failure rates span 6\% to 47\% between the best and worst performing models, which makes model choice a safety question in its own right and when users pushed for ``balance'' in cases where international courts have already assigned responsibility, five of nine configurations failed 80 to 100 percent of the time. We release the first evaluation framework for this domain and propose adding it to alignment evaluation portfolios.

Deployment-Relevant Alignment Cannot Be Inferred from Model-Level Evaluation Alone

Alignment evaluation in machine learning has largely become evaluation of models. Influential benchmarks score model outputs under fixed inputs, such as truthfulness, instruction following, or pairwise preference, and these scores are often used to support claims about deployed alignment. This paper argues that deployment-relevant alignment cannot be inferred from model-level evaluation alone. Alignment claims should instead be indexed to the level at which evidence is collected: model-level, response-level, interaction-level, or deployment-level. Two studies support this position. First, a structured audit of eleven alignment benchmarks, extended to a sixteen-benchmark corpus, dual-coded against an eight-dimension rubric with Cohen's kappa = 0.87, finds that user-facing verification support is absent across every benchmark examined, while process steerability is nearly absent. The few interactional benchmarks identified, including tau-bench, CURATe, Rifts, and Common Ground, remain fragmented in coverage, and benchmark construction rather than data source determines what is measured. Second, a blinded cross-model stress test using 180 transcripts across three frontier models and four scaffolds finds that the same verification scaffold raises one model's verification support to ceiling while leaving another categorically unchanged. This shows that scaffold efficacy is model-dependent and that the gap identified by the audit cannot be closed at the model level alone. We propose a system-level evaluation agenda: alignment profiles instead of single scores, fixed-scaffolding protocols for comparable interactional evaluation, and reporting templates that make the inferential distance between evaluation evidence and deployment claims explicit.