Item Response Theory for AI Safety

Language models differ in how safely they behave and these differences are measured by safety benchmarks. But aggregated benchmark scores are hard to trust and interpret, because benchmarks duplicate one another, correlate heavily, and models may sandbag when they detect evaluation. To address these issues, we draw on Item Response Theory (IRT), a statistical toolkit for measuring these latents from performance on items with inferred psychometric properties. We fit IRT models to eight safety benchmarks across 192 language models, the largest psychometric analysis of LLM safety evaluations to date, and contribute three results. First, we find that three interpretable factors of refusal strictness, truthfulness, and contextual harm explain most of the variance between models across benchmarks. Second, psychometrically selected items recover full benchmark scores with lower error than random subsets of the same size, and roughly ten adaptively chosen items suffice for several individual benchmarks, cutting evaluation cost by 97-99%. Third, IRT supports audits of individual models, showing that it can be used to detect naive sandbagging and changes of model behind APIs. Overall, we show IRT is a ready-made toolkit for reading, reducing, and auditing safety benchmarks, which we recommend frontier labs and evaluators adopt.

Detecting CSAM Text-to-Image LoRAs From Weights

Low-rank adaptation (LoRA) fine-tuning has made it cheap and easy to customize open-weight image generation models for specific tasks, including the production of child sexual abuse material (CSAM). Existing moderation relies on metadata or generated outputs, but metadata can be deceptive and generating outputs may itself be unacceptable or illegal. We show that a safer signal lives in the weights. The top-left singular vectors of a LoRA's updates form a compact, inference-free fingerprint ($u_1$) of its strongest learned change. Using human-subject age as a benign proxy for CSAM, we find that $u_1$ identifies what a LoRA was trained on, generalizes across base models, and abstains on unrelated benign content. The signal is robust to additive weight noise, rescaling, and precision reduction. These results indicate that harmful LoRAs could be screened directly from their weights without relying on metadata or generating harmful outputs.

Persona Cartography: Charting Language Model Personality Traits in Weight Space

Large language models exhibit recurring behavioural patterns -- personas -- that shape generalisation and safety, but we lack reliable tools for decomposing, measuring, and controlling them. Our central insight is to treat personas as positions in a space of behavioural traits, using the OCEAN framework to describe model personas in terms of Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. We train low-rank adapters to amplify or suppress individual traits, and evaluate their effects using an LLM-judge calibrated against a human-validated panel, trait-specific multiple-choice benchmarks, and standard capability evaluations. Across six models from three families (4B-32B), we find that each adapter moves its target trait largely monotonically with scale, combines approximately additively with other adapters to construct mixed personas, and preserves performance on capability benchmarks at moderate scales. We further show that the induced trait axes affect safety-relevant behaviour in downstream evaluations: for example, moving along neuroticism and agreeableness axes affects frustration and sycophancy respectively. We also introduce an unsupervised psychometric pipeline that recovers four interpretable behavioural factors (tone, initiative, didacticism, epistemic caution) from model rollouts. Persona control can then be considered in terms of learning, scaling, and composing traits in weight space, providing a bridge between personality measurement, model editing, and safety.

Tool Use Enables Undetectable Steganography in Multi-Agent LLM Systems

Increasingly autonomous agentic AI systems pose novel multi-agent risks, such as secret collusion via covert communication channels. The natural defence to these collusion attempts is to monitor plain-text communication, but the efficacy of monitors has been called into doubt by increasingly sophisticated model steganography; indeed, some theoretical schemes have been proposed that are information-theoretically or computationally indistinguishable from good-faith plain-text communication. In this paper, we demonstrate that the complexity of these schemes is no longer a safety barrier, as agentic coding models can already produce undetectable stegosystems when given realistic tool usage, such as code execution or accessing research papers through web searches. Agents also adapt when key ingredients are missing, for example, by adding model-sampling components or implementing related keyed coding schemes. We then frame tacit steganographic coordination between agents as a Schelling-point problem and introduce coordination metrics for estimating when two agents are likely to select compatible schemes without explicit prior agreement. Our results suggest a shift in the threat model for covert communication between AI agents, where the main barrier is no longer whether frontier agents can understand and implement sophisticated stegosystems, but coordination: whether independently acting agents can converge on compatible schemes, keys, and parameters. We find substantial convergence on broad scheme families but limited strict one-shot coordination, suggesting that shared artefacts, repeated interaction, and tool-mediated search are the settings where covert communication risks are most acute. Overall, our findings provide empirical grounding for the recent strategic confinement hypothesis, which assumes that capable agents can construct covert channels that survive monitoring.

PACUTE: Phonology-, Affix-, and Character-level Understanding of Tokens for Filipino

Large language models (LLMs) process text as sequences of subword tokens, which can obscure the character-level and morphological structure that underlies word formation. This limitation is most acute for languages with non-concatenative morphology, where standard tokenizers systematically misalign token boundaries with morpheme boundaries. We introduce PACUTE, a diagnostic benchmark of 4,600 tasks designed to evaluate morphological understanding in Filipino, a language characterized by productive infixation, reduplication, and diacritic-driven lexical distinctions that are typically absent from written text. PACUTE includes a hierarchical diagnostic framework of six compositional levels that localizes where morphological understanding breaks down. Evaluating open-weight LLMs and frontier commercial models, we find that open-weight models perform near chance on morpheme decomposition regardless of scale. Frontier models perform much better, often recovering individual affixes under contains-match scoring, but remain far below their character-level ceilings on compositional tasks of morpheme transformations and syllabification. These results identify productive morphological composition, rather than character access alone, as the persistent bottleneck for Filipino word-structure understanding.

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.

Consistency Training Along the Transformer Stack

Consistency training encourages models to behave similarly across different contexts, and has shown promise for reducing misalignment. We broaden the scope of consistency training in two ways. First, we introduce two new internal consistency targets: MLP Consistency Training (MLPCT), which matches post-activation MLP states, and Attention Consistency Training (AttCT), which matches per-head attention distributions. Second, we apply consistency training to four additional safety threats: persona in-context learning attacks, adversarial frustration, prefill attacks, and conditional misalignment. Across several models and threat settings, we find that consistency training reduces misalignment well beyond the sycophancy and jailbreak settings studied in prior work. We also find cases of cross-threat generalization, where training against one failure mode improves robustness to another, and identify a shared residual-stream mechanism underlying ACT, MLPCT, and AttCT, while distinguishing BCT as mechanistically distinct. Our results suggest that consistency training is a flexible and extensible framework for alignment, capable of unifying defenses against a broader class of model pathologies.

Consistency Training Can Entrench Misalignment

Consistency training encourages a model to produce similar outputs across related inputs or sampling procedures. Such methods are simple, scalable, and largely label-free, but their effects on model alignment remain poorly understood. Could the self-bootstrapping nature of these methods amplify undesired behavior in models? We test seven consistency training methods on 108 model organisms: open-source models (7B--70B) fine-tuned to exhibit various forms of controlled misaligned behavior. We find that outcomes vary significantly: consistency training generally suppresses reward hacking and emergent misalignment but amplifies sycophancy. We present evidence that distribution shifts induced by the consistency labeling process, rather than variation in the selection operators, may be the primary driver of systematic alignment effects. Finally, we present a unifying theoretical framework to derive conditions under which consistency training will amplify or suppress misalignment. In total, our study establishes that consistency training is not alignment-neutral, and that its use in critical systems should be carefully audited.

Consistency Training while Mitigating Obfuscation via Rate Matching

Large language models are often influenced by extraneous input features, such as cues revealing a user's preferred answer. Consistency training reduces this influence by training models to behave similarly across inputs with and without the extraneous feature. However, existing methods train for consistency over entire responses or internal activations, which also constrains whether the model verbalises said extraneous features. We show this leads to obfuscation, where the model learns not to mention a cue while remaining influenced by it, which may undermine monitorability. To address this, we introduce Rate Matching Consistency Training (RMCT), which trains for consistency over selected behavioural properties without constraining how this behaviour is expressed. RMCT matches the rate at which the model exhibits a target behaviour (e.g., following a bias cue) across input perturbations, rather than requiring paired inputs with and without the extraneous feature, extending consistency training to settings where the extraneous features cannot be removed. We evaluate RMCT on sycophancy reduction in two open-weight language models, achieving reductions in bias-following comparable to a standard consistency-training baseline on held-out bias types, while largely preserving the model's tendency to verbalise the bias cue. Further, we find that RMCT is more data-efficient at the expense of being less compute-efficient in our experiments. Overall, RMCT shows that consistency training can improve behavioural robustness without directly trading off against monitorability.

Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time

Language model finetuning often results in learning undesirable traits in combination with desired ones. To address this, we propose inoculation prompting: modifying finetuning data by prepending a short system-prompt instruction that deliberately elicits the undesirable trait. At test time, we evaluate without the instruction; inoculated models have much lower expression of the trait than models trained with unmodified training data. Inoculation is selective: in a toy setting where assistant responses are always in Spanish and ALL-CAPS, an appropriate inoculation (e.g., ``You always speak in Spanish.'') teaches the model to capitalize responses while still responding in English. We find that inoculation is also effective across several additional settings: reducing emergent misalignment (EM) from task-specific finetuning, defending against backdoor injections, and mitigating the transmission of traits via subliminal learning. Follow-up analysis suggests a mechanism: making a trait less surprising via inoculation reduces optimization pressure to globally update the model, thereby reducing the degree of generalization. Our analysis relates to prior work on EM: inoculation explains prior findings that educational contexts mitigate EM from insecure code. Beyond demonstrating a simple and effective technique for selective learning, our results contribute to a better conceptual understanding of how and why language models generalize.