Language Model Robustness

Safety training on language models often induces over-refusal: improved safety on harmful prompts at the cost of increased refusal on harmless ones. Though this trade-off can be mitigated by training models with reinforcement learning (RL) to reason before answering, it does not remove the underlying problem that reasoning can often be a "rubber stamp" for a predetermined response. In this paper, we address the safety-refusal trade-off by rethinking how models are trained to reason about safety. Our key insight is that unsafe reasoning can itself serve as a useful exploratory signal. Rather than preemptively blocking harmful thoughts, we encourage the model to sufficiently explore unsafe reasoning but produce a safe response. The harmful exploration improves the model's ability to distinguish harmful from harmless prompts by resolving ambiguity, allowing it to remain safe while complying only when appropriate. We cast this as an adversarial optimization problem in which a reasoning player explores strategies for producing an unsafe response and an answer player ensures that the final output is safe. We train a single model with dense rewards to play both roles within one chain-of-thought, across different segments. To achieve this, we find that process rewards are crucial for stable optimization of competing objectives. Our resulting model SEAR deliberately engages in harmful reasoning as exploration while reliably flipping back to a safe answer. We demonstrate that this behavior helps mitigate over-refusal and defend against attacks that directly manipulate the reasoning to be harmful.

As large language models take on morally consequential roles in healthcare, legal, and hiring contexts, we need to examine whether their ethical behaviors are genuine or superficial. We show that current fairness evaluations substantially overestimate moral safety. Models appear fair when demographic identity is stated as an explicit label, yet become measurably less fair when the same identity must be inferred. We term this failure performative compliance, where a model is fair when the presentation resembles a fairness evaluation and less fair as that cue weakens. We introduce a cue-variation methodology that holds the moral dilemma and the demographic identity fixed and varies only how that identity is conveyed. Hiding the explicit label raises harmful decisions by +4.4 pp, changes model safety rankings, and the shift persists when models correctly infer the demographic, ruling out attribution error. We propose the Cue Visibility Gap, a model-agnostic robustness metric that can be added to any existing fairness benchmark to separate genuine from performative moral safety. Fairness evaluations that omit cue variation measure surface compliance, not moral robustness, and should not ground deployment decisions in high-stakes settings.

Large Language Models (LLMs) exhibit strong semantic capabilities, yet their resilience to manipulative linguistic patterns such as logical fallacies remains underexplored. Prior work has primarily examined whether LLMs can identify or classify fallacies, leaving their robustness against fallacious persuasion insufficiently studied. To address this gap, we introduce LoFa (Logical Fallacy), a comprehensive benchmark for evaluating LLM robustness against fallacies. LoFa is constructed through a multi-agent pipeline that pairs factual questions with fallacious arguments, and is accompanied by a multi-round debate framework for assessing model resilience under sustained adversarial persuasion. To disentangle fallacy robustness from a model's inherent knowledge limitations, we further propose Logical Fallacy Resistance at k (LFR@k), a metric that quantifies resistance to fallacious attacks. Experiments show that LLMs exhibit varying levels of robustness across different fallacy types, revealing distinct vulnerability profiles among models.

Event detection (ED) systems are typically evaluated on clean, curated text, leaving their robustness to real-world noise largely unexplored, particularly for low-resource languages such as Bangla. We introduce a generalized Bangla news event ontology and a benchmark comprising 9,979 annotated sentences across 40 event subtypes, spanning clean news text, real-world Automatic Speech Recognition (ASR) transcripts, and orthographically corrupted text. We systematically evaluate fine-tuned encoder-only models (BanglaBERT and XLM-R) alongside instruction-tuned decoder-only large language models (Llama 3 and Gemma 3). Our results reveal a clear architectural trade-off: encoder models achieve higher performance on clean text but degrade substantially under noise, whereas decoder-only LLMs are markedly more robust, particularly when event triggers are corrupted. We further show that embedding annotation guidelines during instruction tuning establishes a higher performance baseline on noisy text but yields inconsistent reductions in performance degradation across noisy conditions. Finally, model scaling consistently improves the robustness of decoder-only LLMs, while combined training on clean and noisy data serves as an effective regularization strategy that disproportionately benefits encoder architectures, significantly narrowing the robustness gap.

Romanized Code Mixing (RCM), where bilingual speakers fluidly blend local languages with English in Roman script, has emerged as the dominant form of communication across multilingual communities. While Large Language Models (LLMs) perform strongly on monolingual and native-script benchmarks, their ability to follow instructions and reason over RCM-based content remains largely unexplored. To this end, we introduce the Indi-RomCoM benchmark for facilitating systematic evaluation on Indic Romanized Code-Mixed instructions. Our benchmark spans seven instruction-following tasks, four widely spoken Indic languages, and three controlled code-mixing intensity levels. We extensively evaluate a suite of LLMs covering proprietary, open-weight, and Indic-focused models under zero- and few-shot settings. LLMs consistently underperform on RCM instructions, with performance degrading as code-mixing density increases. Furthermore, reasoning tasks suffer less degradation than detection tasks (e.g., Toxicity) because the generated explanations offer necessary context. We believe Indi-RomCoM helps the community in developing inclusive multilingual systems.

Modern LLM deployments often combine quantization with higher sampling temperatures to reduce cost, latency, or repetition, yet safety evaluations usually treat these as fixed implementation details. We test whether models that are safe at FP16 with greedy decoding remain safe after quantization and stochastic sampling, or whether the two factors amplify each other. We evaluate 8 instruction-tuned models from five families across 3 precisions and 6 temperatures, covering 144 configurations on 7 harmfulness benchmarks and generating about 2.0 million responses, which are scored by a six-judge safety ensemble. Contrary to concerns that low-bit deployment erodes alignment, we find that standard quantization is approximately safety-neutral: for 7 of 8 models, AWQ INT4 keeps attack success within about 1.6 percentage points of FP16 or lowers it, with clear degradation only for SmolLM3-3B (34.5% to 44.1%). However, the larger risk comes from sampling: higher temperatures sharply increase decision instability, with DFR reaching 41.9% at T = 1.0, even when average ASR changes only modestly. The two factors do not compound: our Compound Degradation Index remains sub-additive (-0.071 to +0.018), indicating that quantization partially offsets rather than amplifies temperature-induced degradation. Finally, a per-benchmark breakdown shows that single-benchmark evaluation badly understates risk: several models scoring 0% on AdvBench exceed 80% on ManyHarm. Standard INT4/INT8 quantization can therefore be reasonable for well-aligned models, but safety claims should report multi-sample stability across multiple benchmarks rather than rely on a single benchmark at greedy decoding.

Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of natural language processing tasks. However, their tendency to generate hallucinations, namely factually incorrect or unfaithful outputs, poses a critical obstacle to their deployment in high-stakes applications. Although recent hallucination detection methods have made encouraging progress, they typically rely on costly output-level consistency checks or static hidden-state probes that capture shallow dataset-specific patterns, leading to substantial degradation under cross-dataset evaluation. In this work, we propose AURORA, a novel hallucination detection framework that shifts the focus from static representations to the weight-gradient dynamics of LLMs. Our key insight is that hallucinated and faithful answers induce qualitatively different gradient update patterns on the model's parameters. Specifically, hallucinated samples trigger asymmetric and structurally misaligned gradients, which can be captured through two complementary features: (1) the skewness of the cosine similarity distribution between weight matrices and their gradient update directions, and (2) the rotation ratio, which quantifies how much the gradient update reorients the singular-vector basis of weight matrices via SVD. AURORA achieves strong hallucination detection performance across four model families and four benchmark datasets. Further analyses demonstrate that our method scales effectively across model sizes and transfers to out-of-domain tasks, including mathematical reasoning and vision-language scenarios.

Latent reasoning models perform multi-step inference directly in hidden-state space, yet the structure of these latent reasoning trajectories remains poorly understood. We show that contrastive refinement signals between stronger and weaker reasoning trajectories exhibit a highly concentrated low-rank structure, while unconstrained latent updates remain sensitive to paraphrases, checkpoint choice, and trajectory perturbations. These observations suggest that latent reasoning trajectories contain stable invariant directions mixed with unstable instance-specific variation. We introduce \textbf{Trajectory-Invariant Latent Refinement (TILR)}, a training-free intervention framework for identifying and manipulating stable reasoning directions in latent space. TILR first learns a low-rank invariant subspace from contrastive trajectory differences across inputs, then constrains latent interventions to this subspace while suppressing poorly aligned updates through an adaptive alignment gate. Across six reasoning benchmarks, we find that a small number of latent directions explain most variation between strong and weak reasoning trajectories. Interventions on these directions causally improve reasoning consistency and reduce trajectory instability under paraphrases and perturbations. TILR improves answer consistency under paraphrase by ~10% and reduces latent trajectory variance by up to $50\%$ while preserving reasoning accuracy. These results support a geometric view of latent reasoning in which transferable reasoning behavior emerges from stable low-dimensional structure within hidden-state trajectories.

Large language models (LLMs) have attracted widespread attention from academia and industry, yet their deployment raises critical security concerns regarding robustness and reliability. Planning, a core component of intelligent behavior, remains challenging for LLMs, which often produce infeasible or incorrect solutions in long-horizon decision-making tasks due to inherent complexity. In this paper, we propose a symbolic feedback-driven iterative self-refinement framework to enhance the robustness and reliability of LLMs in long-horizon planning. Specifically, a natural language prompting mechanism is introduced to map logical symbols into natural language descriptions, enabling LLMs to better capture task constraints and semantics. We further design a symbolic verifier that identifies errors and converts them into corrective instructions interpretable by the LLM, thereby guiding self-refinement. In addition, we leverage a plan recognizer to infer goal reachability, facilitating more effective guidance toward desired goals. Empirical results demonstrate that the proposed framework consistently improves both feasibility and correctness in long-horizon planning tasks. This highlights its effectiveness in enhancing the reliability of LLM-based planning and potential to enable more trustworthy AI systems.

Probe-based uncertainty estimation (UE) has emerged as a prominent approach to detect hallucinations in Large Language Models (LLMs) by learning uncertainty from internal model signals. Yet, recent methods vary simultaneously across feature design, training data construction, and evaluation setting, obscuring what actually drives performance. To address this issue, we propose a factorised study of probe-based UE under matched conditions. Our results show that raw hidden states and attention features are difficult to outperform in-domain. However, under distribution shift, structured and compressed features are more robust, suggesting that in-domain performance alone is insufficient to measure progress. Furthermore, prompting and label construction significantly affect probe behaviour. Building on these best-practice findings, we train benchmark-based pretrained probes that transfer reasonably well to open-ended factual generation, providing a stable off-the-shelf baseline. Our work encourages more deployment-oriented evaluation of probe-based uncertainty estimators. The code repository is available at https://github.com/ponhvoan/ProbeUE.

Large language models (LLMs) are increasingly critical to digital library workflows, yet their ability to process historical language remains poorly understood. Historical difficulty is typically treated as a monolithic barrier, conflating orthographic variation, linguistic distance, and pretraining exposure. In this paper, we propose a diagnostic framework that decomposes this difficulty into four distinct dimensions: tokenization cost, predictive uncertainty (surprisal), semantic robustness, and context sensitivity. We evaluate this framework on three datasets spanning three centuries: (1) a newly curated corpus of 17th-century Italian texts (1610-1689) digitized from original page images; (2) canonical 19th-century Italian "I Promessi Sposi" serving as a high-exposure control; and (3) 18th-century Russian civil print books as a contrastive orthographic stress test. Our results reveal a distinct dissociation between encoding cost and comprehension. While Russian and early modern Italian incur comparable tokenization penalties (25-30% inflation), their predictive difficulty diverges sharply. 17th-century Italian is on average 2.4 times more surprising than its modern equivalent - with academic prose reaching 3.2 times - whereas Russian shows only a modest increase. But predictive uncertainty does not imply representational degradation: embedding similarity remains robust (> 0.85) across all datasets, confirming that models can represent historical meaning even when generation is unstable. Finally, we demonstrate that a minimal temporal context prompt reduces historical surprisal by approximately 60%, offering a simple, model-agnostic mitigation. These findings suggest that while historical text imposes a consistent encoding tax, digital libraries can safely deploy LLMs for semantic retrieval tasks, provided that generative applications are carefully adapted.

Large language models (LLMs) are increasingly being integrated into mental health support tools and other psychologically sensitive conversational applications. In such settings, behavioral stability and consistency are important for trustworthy human-AI interaction. However, semantically similar concerns can be presented through different contextual framings, potentially eliciting different model responses. Such framing-sensitive variability may challenge user expectations regarding system behavior and complicate the assessment of AI reliability. While prior studies have primarily examined such effects at the behavioral level, less is known about how framing-related variation is reflected in the internal representations of aligned language models. In this work, we investigate these effects using controlled matched prompts spanning multiple contextual framing conditions across several instruction-tuned model families. Across architectures, framing systematically alters interpretive response tendencies. Layer-wise probing analyses show that behavior-associated information remains decodable throughout transformer depth, with architecture-dependent variation in decoding strength. Moreover, held-out framing probes remained consistently above chance across architectures despite strong lexical baselines. Activation steering experiments further suggest that framing-associated representational directions can partially modulate downstream behavioral outcomes. Finally, these findings indicate that robustness to contextual variation may represent an important consideration when evaluating the consistency and trustworthiness of conversational AI systems deployed in mental-health-oriented interactions.

Large language models can serve as capable long-horizon agents, but their out-of-distribution (OOD) generalization remains weak. We identify a key source of this failure as task insensitivity: when faced with similar but distinct tasks, models might apply patterns learned during training and fail to solve the task at hand. We show that models often continue with actions aligned with the original task even when the instruction is semantically corrupted and cannot be directly answered. We further find that, when we replace the task description in a trained prompt with another similar but distinct task, the model may still output the same action. This behavior is accompanied by a consistent training-time attention drift away from task tokens and toward local observations, suggesting an optimization bias toward shortcuts. To mitigate this problem, we propose Task-Perturbed NLL Optimization, a lightweight contrastive regularizer that explicitly encourages action dependence on the task instruction. Extensive evaluations show that our intervention improves task sensitivity and OOD generalization while preserving more stable attention to task tokens.

In order to screen a prompt or a response, the recent guardrail methods generate a chain-of-thought (CoT) before they issue a verdict. This design follows a common belief that step-by-step reasoning improves a decision. However, CoT also makes the guard heavy and slow, because the model must generate many tokens before it decides. This may not match how guardrails are actually deployed. A guardrail sometimes should not be heavy and slow, and it often runs on-device, for example on an embodied robot. In this paper, we pose a question whether a safety guardrail really needs to reason. To answer this question, we train a lightweight bidirectional encoder and a reasoning guard on the same corpus, and we then remove only the reasoning while we keep everything else fixed. With this controlled same-base comparison, we show that the chain does not improve moderation accuracy. We name the resulting guard LeanGuard. A 395M label-only encoder reaches an average F1 of 82.90 $\pm$ 0.26 over public benchmarks. It matches a reasoning guard that is built on a much larger decoder, while it uses only a single forward pass over an input of at most 512 tokens. This is about a ~100x reduction in inference compute. We further show that this label-only encoder stays robust under training-label noise and retains far more recall at a strict false-positive rate than the reasoning guard, so a heavier reasoning guard is not the more robust choice either. Our finding suggests that the current guardrail benchmarks may not be hard enough to reward reasoning, and that the necessity of CoT for moderation is still not proven. We release all source codes and models including LeanGuard at https://github.com/ndb796/LeanGuard.

Dialogue systems based on large language models (LLMs) have advanced significantly in recent years. However, dialectal variation remains a major challenge, particularly for systems that process spoken input. LLM-based speech language models (SLMs), which integrate LLMs with speech processing components, show promise for spoken language tasks, yet their ability to comprehend dialects has not been sufficiently studied. Moreover, it remains unclear how the dialectal understanding of the base LLM affects SLM performance. This study investigates the dialectal robustness of both LLMs and SLMs using Japanese dialects as a test case. We define robustness as the ratio of performance on dialectal versus standard inputs, enabling fair comparisons. Our experiments show that SLM robustness correlates with that of their text-based counterparts. Furthermore, training with dialectal data and fine-tuning the speech encoder each improves robustness in SLMs.

The rapid growth of scientific submissions has pushed traditional peer review toward its scalability limits, motivating the exploration of large language models (LLMs) as intelligent automated evaluation assistants. Although recent studies show that LLMs can generate fluent critiques and approximate reviewer scores, their reliability, robustness, and security as decision-support systems remain insufficiently understood. This survey offers a systems-level analysis of LLM-based scientific peer review, focusing on two core evaluative functions: critique generation and score prediction. We present a structured taxonomy of modeling approaches (including prompt-based, supervised, retrieval-augmented, and alignment-optimized approaches), and synthesize empirical findings across existing benchmarks. We analyze dataset constraints, evaluation shortcomings, and domain concentration biases that limit current assessment practices. Beyond performance metrics, we identify emerging robustness risks, including prompt injection, data poisoning, retrieval vulnerabilities, and reward hacking, which expose automated review pipelines to strategic manipulation. From a data mining perspective, we outline key open challenges in modeling subjective disagreement and cross-domain generalization. By reframing automated peer review as a high-stakes, multi-objective decision problem, this survey provides a roadmap for developing robust, transparent, and trustworthy AI-assisted scientific evaluation systems.

Understanding the features of large language models (LLMs) is a central goal of interpretability. LLMs are commonly assumed to use superposition to represent more features than they have dimensions. They may not only represent features in superposition but also perform computation in superposition. Theory predicts that computing in superposition requires error correction that privileges feature directions over generic ones, but this prediction has not been tested empirically. We propose an empirical test of error correction in LLMs based on activation perturbations. Perturbing residual-stream activations, we find that they are robust to small perturbations--forming activation plateaus consistent with error correction--but less robust along candidate feature directions ("pure" directions, constructed from contrastive prompt pairs) than along mixtures of two such directions, indicating that the pure directions are privileged. We quantify this privilegedness by modeling the perturbation effect as a function of the $L^p$-norm of its decomposition into feature components. For $p=2$ the response is a quadratic form with at most as many nonzero eigenvalues as the residual-stream dimension, which cannot privilege the many feature directions superposition requires. $p>2$ lifts this constraint and is consistent with feature-specific error correction. We find $p>2$ for contrastive, MELBO, and SAE-decoder directions, and $p\approx2$ for random and PCA directions (controls). These results replicate across Gemma-2-9B, Qwen3-1.7B, Llama-3.1-8B, Mistral-7B-v0.3, Aya-Expanse-8B, and Yi-1.5-9B. We further validate our method on a toy model of error correction with known ground-truth features, recovering $p>2$ for true feature directions, degrading toward $2$ as we rotate away from them.

As Large Language Models asymptotically approach human-level fluency in natural language generation, solely relying on surface-level semantic artifacts for detecting LLM-generated texts has become increasingly precarious. Existing detectors often falter when facing three critical challenges: adversarial perturbations, cross-domain shifts, and the rapid temporal evolution of the foundation model. To address these issues, we propose \wavedetect, a novel framework that reformulates text detection as a signal processing task within the time-frequency domain. Unlike previous methods that analyze static token probability distributions, \wavedetect models the generated output as a probability signal, upon which a differentiable Continuous Wavelet Transform is applied to convert them into learnable spectral representations. This process reveals the intrinsic ``spectral fingerprints'' in machine-generated texts--patterns that remain invisible in time domain. Comprehensive evaluations on three well-curated datasets (RAID, EvoBench, and Domain-Shift) show that our method achieves a new state-of-the-art. It not only achieves superior accuracy but also exhibits remarkable robustness against sophisticated attacks, generalization across out-of-distribution topics and unseen evolving LLMs. Our results validate the efficacy of spectral analysis as a promising paradigm for LLM-generated texts detection.

Why do multilingual language models sometimes generate in the wrong language, and why is this so hard to fix? We introduce Language Identity Head Ablation (LIHA), a causal intervention that zeros each attention head individually and measures the resulting language switch rate across a parallel dataset of 2,700 prompt-language pairs spanning seven languages. Applied to GPT-2, LIHA identifies a small set of first-token broadcaster heads - led by L6H1 (switch rate 0.32, 3.23 $σ$ above the population mean) - that attend persistently to the first prompt token, propagating its language signal throughout generation. Compensatory redistribution when heads are ablated is statistically significant (p < $10^{-5}$) and follows a directional, hierarchical pattern: compensation always recruits heads in layers above the ablated head, suggesting a feedforward cascade rather than global diffusion. To probe how training regime shapes these circuits, we apply LIHA to a controlled pair - Qwen2.5-1.5B-Base and Qwen2.5-1.5B-Instruct - identical in architecture and size, differing only in training. The base model is nearly flat (max SR=0.016, 200/336 heads at SR=0.0); the instruct model concentrates causal influence sharply at layer 0, led by L0H5 (SR=0.224, 8.93 $σ$ above mean), with all other layers near zero. This controlled comparison provides direct causal evidence that instruction tuning reorganizes language identity circuits toward early-layer localization. Extended experiments with Chinese and Russian confirm that first-token broadcasting is script-specific in GPT-2, with non-Latin languages handled at layer 0 - the same locus as the instruction-tuned model. Code and data will be released upon publication.

Reliable evaluation of generated radiology reports requires strict clinical accuracy, as omitted critical findings or mischaracterized radiographic observations can directly affect patient care. Existing metrics obscure this requirement by reducing report quality to a medically ungrounded scalar. Although Large Language Models (LLMs) possess rich medical knowledge, they likewise struggle to draw a reliable boundary between clinically significant errors and harmless variation. We study this boundary using ReEvalMed benchmark as testbed and evaluate metric-level clinical significance from detecting true clinical errors ("Discrimination") and tolerating insignificant variations ("Robustness"). Across 8 LLM evaluators under one-pass and two-pass settings, we identify a widespread discrimination bias: models effectively detect errors but also over-penalize harmless rephrasings. To mitigate this, we synthesize 4k report pairs and train lightweight interpretable metrics on Qwen3-8B and MedGemma-4B. Our trained metric sharpens the clinical significance boundary, surpassing 32B-scale medical LLMs and remaining competitive with proprietary models. Crucially, the more costly two-pass setting fails to consistently improve overall performance and mainly trades discrimination for robustness. These findings suggest one-pass trained metrics as the practical choice for cost-sensitive deployment, with two-pass inference reserved for settings where D-R balance is critical. We will release the dataset and metric.

Legal AI benchmark research frequently invokes the assumption that large language models can improve access to justice, including for people who cannot access lawyers in order to understand and exercise their legal rights. We argue that current benchmarks are not equipped to support this assumption because they evaluate legal reasoning over inputs that have already been preprocessed by legal experts, which measures the upper bound of model performance. Access to justice depends on a lower bound: how models perform when inputs come from pro se litigants, whose prompts may contain noisy narratives, buried facts, omissions, folk-legal assumptions, and surface-level errors. These degradations are comparable to conditions under which LLMs are known to degrade in the general machine learning literature, including long-context sensitivity, underspecification, hallucination, and typographical perturbations. We connect evidence from pro se literature with this body of machine learning research and present a small perturbation experiment on LEXam, a legal benchmark, to illustrate the gap between these two bounds. If model development continues to focus on benchmarks that measure only the upper bound, this gap may remain hidden or even widen. We conclude by calling for legal benchmarks that directly measure robustness under pro se-like inputs so that access-to-justice claims about legal AI can become empirically testable.

Large language models perform increasingly well on standardized logical reasoning benchmarks, but whether this ability remains robust beyond English is unclear. We introduce ChLogic, an English--Chinese aligned benchmark that tests whether models preserve logical reasoning performance when the same latent logical structure is expressed in English and diverse Chinese surface realizations. Built from formal logical templates, the benchmark contains three data sets: (i) the General aligned set, derived from 60 General Propositions across nine template families; (ii) the Difficult aligned set, derived from 40 Difficult Problems; and (iii) the Chinese-only set, covering 15 language-specific phenomenon types. Each aligned item pairs one English reference expression with five Chinese realizations. Experiments on Qwen3, Ministral, and GLM models reveal a persistent English--Chinese performance gap. Back-translation from standard Chinese into English often improves performance on the General aligned set, but produces mixed effects on the Difficult aligned set, where Qwen3-32B and GLM-5.1 perform worse after translation. These results indicate that Chinese surface realization, translation artifacts, and model-specific behavior jointly affect multilingual logical reasoning. Overall, ChLogic provides a useful stress test for the robustness of multilingual reasoning.

Large language models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks. However, their safety remains a critical concern due to their susceptibility to adversarial prompt-based attacks. In this paper, we present UNIATTACK, an adversarial testing framework designed from a defense-oriented perspective to systematically construct effective black-box attack prompts. Unlike prior approaches that rely on static templates or iterative model-specific tuning, UNIATTACK extracts minimal but high-impact attack features from diverse existing attacks, optimizes them via a specialized attacker LLM, and composes them into flexible templates through automated refinement process. This feature-centric construction enables one-shot attacks that generalize across multiple models and safety categories, providing a practical tool for assessing LLM robustness. Our evaluation results shows that compared to the baselines, UNIATTACK achieves an average attack success rate (ASR) improvement of 64.63%-248.82% on models deployed with multi-layered defense mechanisms and it only takes 0.03%-4.96% cost of the baselines. UNIATTACK artifact is available at https://anonymous.4open.science/r/UniAttack-Artifact-30F1.

Activation monitors -- lightweight probes trained on a language model's internal representations -- are an increasingly common layer in deployment safety stacks. Deployed models however are rarely static: they are quantized, fine-tuned, adapted with LoRA, or served with merged adapters while the monitor remains frozen. We present the first systematic test of whether this implicit contract holds: whether activation monitors trained on a base model remain reliable after these routine model updates. Across multiple safety-relevant monitors, model depths, update families, and open-weight models, we find a sharp split: quantization-style updates largely preserve frozen probe performance, while fine-tuning-style updates frequently make probes stale. Fragility is highly monitor-dependent, with privacy/PII probes most affected and refusal-compliance probes comparatively stable, showing that retraining a behavior need not stale its corresponding monitor. QLoRA is especially damaging despite NF4 quantization alone being relatively benign, suggesting that quantization becomes riskier when combined with adaptation. We further show that degradation is predictable from pre-deployment features, enabling revalidation budgets to be triaged toward the monitors most likely to fail. Finally, we test repair strategies and find that cheap label-free activation realignment repairs every repair-relevant stale cell, with none requiring score calibration, few-label heads, or labeled retraining. These results suggest that fine-tuning should trigger activation-monitor revalidation by default, with prediction triaging which monitors to check first and label-free realignment as the default repair.

Foundation models are now used in settings where the prompts they receive can change quickly. Users change, topics change, policies change, and the model may suddenly face a kind of request that was rare in the calibration data. This makes fixed calibration risky. Conformal prediction and conformal risk control give model-agnostic ways to control error, but they work best when the calibration data still look like the future data. This paper develops PromptShift CRC, a drift-aware conformal risk control method for foundation-model outputs under prompt and domain shift. The method embeds prompts and responses, measures how far the current prompt stream has moved from the calibration pool, gives more weight to relevant or recent calibration examples, and updates the risk level online after observed violations. It reports three practical diagnostics: realized risk error, prompt drift, and effective calibration size. We give conditions under which the method controls risk up to terms for distribution mismatch and weighted quantile uncertainty. In a synthetic prompt-shift benchmark, static conformal risk control fails sharply after drift, while PromptShift-CRC gives the best coverage among the adaptive baselines considered. We then evaluate the same calibration layer on public benchmark derived streams for question answering, toxicity, summarization factuality, and long-context hallucination risk

Tokenization introduces representational redundancy: under a fixed token vocabulary, every byte string admits many valid token encodings, or segmentations, that decode to the same surface string. However, given a prompt, most language model tokenizers break this representational symmetry by returning a canonical segmentation. Training only on canonical segmentations should influence inference behavior, and there is little reason to expect models to respect segmentation symmetry on downstream tasks. We find that this symmetry partially emerges during training. Here, we probe this emergent symmetry through experiments testing token compositional understanding, representation diversity, and task focused benchmark performance. We primarily use \textbf{retokenization} -- replacing a prompt's canonical tokenization with an alternative segmentation while preserving its bytes exactly. Relative to other prompt perturbations, retokenization is unusually clean because it isolates segmentation effects without changing syntax, semantics or surface form. We use retokenization to study sensitivity and robustness to semantically identical input representations across pretraining and post-training. Moreover, this partial retokenization symmetry suggests a distinct inference-time sampling axis. While temperature sampling generates diverse outputs from the model using its next-token probability distribution, retokenization generates diversity from the model's internal computations through semantically equivalent input representations. We find that while this retokenization sampling strategy can hurt performance on easy problems, it can also recover solutions that conventional sampling does not find. Overall, our work presents retokenization as a simple yet powerful probe of large language models, shedding light on compositional understanding and prompt sensitivity, and offering a novel sampling strategy.

Large language models (LLMs) are now widely used as automatic judges for open-ended instruction-following evaluation. This practice is convenient, scalable, and often more semantically aware than reference-based metrics, but it also introduces a new reliability question: does a judge evaluate the quality of an answer, or does it also react to the language in which the comparison is presented? We propose Judge-LS, a lightweight meta-evaluation protocol that transforms LLMBar response-pair items into English, Chinese, and Chinese-English language-switched variants. A reliable judge should preserve its preference under label-preserving language transformations and should not prefer a language when two answers are translation-equivalent. We evaluate four API-accessible judges on the full 419-item LLMBar benchmark, producing 13,408 successful pairwise judgments. Across models, Chinese and language-switched presentations induce 10.7--14.4% preference flips relative to English, and all judges achieve their highest accuracy in English. However, translation-equivalent tie probes do not reveal a systematic English preference: most probes are judged as ties, and non-tie decisions more often favor Chinese. We add confidence intervals, paired significance tests, and an automatic transformation audit with a sensitivity analysis that excludes mechanically flagged high-risk variants. The experiment requires no model training, uses only API calls, and is feasible on modest local hardware.

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.

As LLMs increasingly serve in advisory and deliberative roles, users rely on them for non-verifiable reasoning in domains lacking objective ground truths. However, traditional evaluations of LLM reasoning focus almost exclusively on fact-based domains, such as mathematics and science, leaving uncertainty over whether and to what degree models can handle ambiguous, subjective, or value-laden problems over time. To address this concern, we propose moral reasoning as a paradigmatic subdomain of non-verifiable reasoning. We define moral robustness as a model's capacity to exhibit sound moral reasoning across time and contexts, and we introduce a scalable, adversarial, multi-turn evaluation framework to empirically measure this capability. We simulate 48,000 user-agent moral deliberations across four frontier LLMs, varying premise relevance, premise order, conversation duration, and the user's stated moral view. We find that models successfully ignore morally-irrelevant distractors, but shift their reasoning by up to 6.5%, on average, towards the user's stated preferred moral view, and varying their reasoning depending on factors such as order (altering moral judgments by order in 13-22% of the cases) and duration (altering moral judgments between single-turn and multi-turn in 10-24% of the cases). Our analysis indicates that models tailor not just their final verdicts but their underlying justifications to align with a user's moral viewpoint - a failure mode we characterize as moral deliberative sycophancy.

Large language models (LLMs) now reach expert-level scores on medical licensing exams, encouraging the assumption that high scores imply safe medical judgment while patients increasingly use them for health advice. We show this assumption is fragile: when misleading context is injected into questions that LLMs originally answer correctly, they abandon the correct answer. We call the ability to maintain correct judgment under adversarial context epistemic resilience, and introduce MedMisBench to measure it. MedMisBench contains 10,932 medical question items and 48,889 misleading context-option pairs spanning medical reasoning, agentic capability, and patient-journey evaluation. Across 11 model configurations, mean accuracy falls from 71.1% on original questions to 38.0% under focused misleading context, with 51.5% attack success. The most damaging injections are formal, rule-like fabrications: authority-framed falsehoods reach 69.5% attack success and exception-poisoning claims reach 64.1%. A 14-member clinical panel from 7 countries identified serious potential harm in 38.2% of reviewed cases. MedMisBench exposes a structural blind spot in LLM evaluation in medical settings: existing benchmarks measure what models know, but not whether they preserve correct medical judgment under misleading context.