Refusal Behavior in Language Models

Safety-aligned language models often refuse a harmful request stated directly but answer the same request inside a role-play or narrative wrapper. We measure this vulnerability across languages and registers: attack success on Qwen3-1.7B is already 89.4% in English and 93.0% in modern Chinese, and reaches 95.7% in Classical Chinese. We build GUISE, a benchmark for systematically studying this vulnerability. It includes parallel requests in English, modern Chinese, and Classical Chinese, matched harmful and benign pairs, wrapper types held out for evaluation, and a stricter criterion that counts warn-then-answer responses as attack successes. Representation analysis shows that language and register move harmful-request representations only slightly away from the model's refusal direction, whereas narrative wrappers move them much farther away. We propose AXIS, which combines preference optimisation with a rotation objective that aligns harmful-request representations with the refusal direction and a commitment objective that trains the model to refuse completely rather than produce a warn-then-answer response. Across Qwen3-1.7B, Qwen3-4B and GLM-4-9B, AXIS achieves the highest combined safety and usability score among the compared methods.

Mechanistic interpretability has emerged as the primary means to understand safety behavior of LLMs. However, existing tools primarily focus on the activating neurons or features of a model. The role of the remaining large set of inactive components is invisible to such methods. This work demonstrates that the inactive set contains safety-critical features that are causally relevant for refusal of harmful prompts. Suppressing such features could turn refusals into compliance, while passing undetected by prevalent interpretability tools. We introduce the Counterfactual Activation Potential (CAP), a metric that quantifies a suppressed feature's latent activation tendency as the product of its encoder alignment (how strongly the input drives it), suppression strength (how strongly active features inhibit it), and safety criticality (how much refusal depends on it). To find suppressed safety features at scale, we propose CAP-guided Safety Feature Discovery (CSFD), a two-stage filtering algorithm that identifies candidate safety features from hundreds of thousands of transcoder features without exhaustive ablation. A significant fraction of trials turn compliant with harmful prompts when a candidate feature is ablated. Under natural jailbreaks, the suppression acting on the highest-CAP features rises 2-4x, and their activation correspondingly falls by up to 80%. Amplifying a feature's suppressors pushes its activation down and raises harmful compliance with prompts related to the suppressed feature, with no such effect for random features. Our experiments span five Gemma, Qwen, and Llama models across various parameter sizes. Our findings indicate that jailbreaks could operate in part by suppressing safety-critical features rather than solely activating harmful ones, and that suppressed features are a necessary complement to activation-focused interpretability of safety behavior.

Studying how fine-tuning shapes refusal and noncompliance behaviour requires knowing which training examples refuse or otherwise fail to fulfil the request. Existing annotations cover evaluation sets, which are far smaller than training corpora. We present CompOrca, compliance labels for all 4,233,923 examples of the OpenOrca corpus. Every example was classified as compliant or noncompliant by five passes of an open-weight LLM judge (LongCat-2.0, 1.6T parameters). The corpus is released as unanimous compliance (94.75%), unanimous noncompliance (1.28%), and nonunanimous rows (3.97%), with the raw vote counts. A single pass flags 2.7-3.2% of the corpus as noncompliant, while only 1.28% is flagged by all five, so the most ambiguous rows can be filtered out. Against 450 human-annotated examples (150 annotated twice; human-human $κ=0.93$), the unanimous compliance and noncompliance labels are 97.3% and 86.7% precise. The noncompliance label is a high-precision subset of the corpus's noncompliance. Published refusal-detection methods recall between 0.4% and 94.1% of human-labelled noncompliance. We release the full corpus with its per-row labels and vote counts at https://huggingface.co/datasets/cemiu/CompOrca

Reliable refusal of harmful requests is essential to the safe deployment of language models. Because excessive eagerness to please users may undermine existing refusal capabilities, reducing sycophancy offers a potential route to stronger refusal beyond the harmful scenarios covered by safety training. We investigate this possibility using compensatory feature injection (CFI), a training technique designed to limit the acquisition of a target concept by supplying its associated activation during learning. Across three Qwen3.5 base models, we use sparse autoencoders (SAEs) to identify the top-ranked sycophancy feature from paired sycophantic and independent responses, then validate its behavioral influence through inference steering. We subsequently inject the selected feature during supervised fine-tuning on sycophantic targets. Positive injection reduces learned sycophancy after removal (by 62.0% relative to ordinary fine-tuning in 35B-A3B), whereas modest negative injection increases it. Unexpectedly, these reductions in sycophancy do not consistently improve direct refusal of harmful requests, motivating a narrower evaluation of the same harmful intents under user pressure. In this setting, ordinary fine-tuning on sycophantic responses substantially weakens refusal, while selected checkpoints trained with positive injection recover part of the loss, including approximately 95% in 35B-A3B. These findings show that persistent sycophancy reduction does not guarantee stronger direct refusal, while identifying recovery under user pressure as a distinct, conditional benefit of training intervention.

Large language models (LLMs) are increasingly deployed with access to external tools, yet harmful tool-mediated interactions are less likely to be refused when compared to regular conversational ones. As this change in refusal behavior remains underexplored, we investigate its underlying mechanisms across a diverse set of open-weight language models. We find that information about the harmfulness of a request remains strongly encoded in the model's representations and transfers across conversational and tool-mediated inputs. Evidence from representation geometry and neuron-level analysis further indicates that the two interaction modes systematically distribute harm-related computation differently. Crucially, while conversational inputs can be refused at relatively low levels of perceived harmfulness, tool-mediated inputs remain permissive until harmfulness crosses a substantially higher effective refusal threshold. Moreover, tool-mediated refusal is also more brittle: progressively weakening the refusal computation disrupts tool-mediated refusal at lower intervention strengths than conversational refusal, even when benign capabilities remain intact. Together, our findings indicate that tool mediation does not simply reduce the internal perception of harm, but instead impacts its conversion into refusal. Overall, this suggests tool-mediated environments may intrinsically reduce robustness of models to harmful requests, and that conventional safety evaluations may not fully transfer to LLM agents.

Large audio-language models (LALMs) expand language models to process and interpret audio, but also expose them to heterogeneous audio jailbreaks. We ask whether successful jailbreaks reflect failures to recognize harmful intent or failures occurring after such recognition. Layer-wise probing reveals the latter: risk-related information remains decodable from intermediate representations, yet the internal risk signal fails to translate into refusal in later-layer processing. We identify this discrepancy as the risk-to-refusal gap. Building on this finding, we propose AEGIS, a detect-then-intervene defense whose mid-layer risk gate selectively activates downstream safety adapters. Across six LALMs and three heterogeneous audio jailbreak benchmarks, AEGIS reduces the average unsafe rate from 17.9% to 0.4%, while causing only a marginal increase in over-refusal on benign inputs. These results establish selective internal intervention as an effective path toward more robust refusal in LALMs. The code is available at https://github.com/azzzzliao/aegis-audio-defense.

Prior work has shown that internal harmfulness representations in large language models vary across risk categories, while sharing a common general harm representation component. This raises a question about the role of the category-specific component beyond general harm representation in LLM safety. To answer this question, we isolate the category-specific component by removing shared general harmfulness representation from each categorical harmfulness representation, yielding a category residual that is orthogonal to general harmfulness at every layer. Using activation steering with category residuals across 11 risk categories in 3 instruction-tuned LLMs, we find that whether category residuals encode harmfulness varies across categories, and that this category-wise pattern is similar across models. Whether category residuals induce refusal also varies across categories, but this category-wise pattern is more model-dependent. We also find that category residuals increase LLMs' downstream internal alignment with shared general harmfulness representation. Together, these findings demonstrate that more fine-grained category residuals should also be considered beyond shared general harmfulness representation to fully understand LLM safety. More broadly, our findings show that even a direction orthogonal to a concept at one layer can contribute to the concept's downstream amplification.

Large Reasoning Models (LRMs) exhibit strong problem-solving abilities, yet their safety alignment often degrades when handling harmful queries. Existing approaches to improving safety largely rely on additional training or preference optimization, while offering limited understanding of the internal mechanisms behind safety failures. In this work, we investigate this failure through a token-level positional analysis of refusal dynamics and identify a localized vulnerability at the onset of reasoning, which we term Onset Refusal Collapse (ORC). We find that the refusal-related signal of LRMs drops sharply at the first generated token under harmful queries, which is associated with unsafe response generation. Motivated by this finding, we propose SafeToken, a lightweight inference-time intervention that injects a learned continuous safety anchor precisely at reasoning onset. Despite updating only a single token embedding, SafeToken effectively mitigates ORC, improves safety on harmful-query benchmarks, and largely preserves reasoning utility. These results suggest that safety failures in LRMs can arise from a transient breakdown at the critical transition from understanding to generation.

Large language model (LLM) agents increasingly operate over long-horizon interactions involving tool use, persistent state, evolving authorization, and external environment feedback. In such settings, safety failures may emerge only after multiple turns, yet existing evaluations often reduce agent behavior to task or attack success, obscuring whether an agent acts, refuses, or remains appropriately calibrated as the interaction evolves. We introduce Blindspot, a benchmark for trajectory-level safety calibration of long-horizon tool-using agents. Blindspot evaluates complete user-agent-environment trajectories through adaptive adversarial interaction, stateful tool execution, and execution-grounded adjudication. Its current instantiation contains 22 attack families and 35 scenarios across seven domains, yielding more than 2,500 long-horizon trajectories with an average interaction length of 14.7 turns. Each trajectory is assigned one of five outcomes: Safe Completion, Correct Refusal, Unsafe Completion, Over-Refusal, or Indeterminate. Unlike fixed attack datasets, Blindspot is an extensible live-simulation framework in which attacks, scenarios, tools, policies, domains, and agent configurations can be added without redesigning the evaluation pipeline. We evaluate 13 proprietary and open-weight LLMs using eight metrics covering unsafe completion, appropriate refusal, benign utility, over-refusal, repeated-run robustness, and post-refusal failure. Preliminary results reveal substantial differences in safety-utility calibration across models and show that failures can emerge only after several initially safe interaction steps. These findings motivate treating agent safety as a trajectory-level property rather than a single-turn or binary success criterion.

Alignment applied after pretraining is shallow in a measurable way: a single direction in a model's residual stream can be edited out, and the model stops refusing harmful requests. That fact says how easily refusal can be removed, not what the refusal decision was reading in the first place. We ask what it reads, and we separate that from what the model comprehends. Across four open-weight models spanning three families, moral comprehension is native to pretraining: a low-rank moral subspace crystallizes during pretraining, and alignment rotates it once without rebuilding it. The refusal gate, in contrast, is a fresh post-training construction with only a weak pretraining precursor, written into a narrow control-token channel where the refusal decision is orthogonal to the moral-judgment decision. The central result is causal and comes from one model, OLMo-3. A nested interchange rank sweep patches successively larger slices of the moral subspace between matched requests and reads how much of refusal's response transfers: as the basis widens, moral judgment keeps reading more of it, while refusal levels off at the level of a single harm direction, and about three-quarters of refusal's causal input lies outside the moral subspace altogether. Refusal reads the harm percept, not the moral content that judgment reads on the same patches. The picture is not uniform across families. Llama reads broad moral content; Qwen reads beyond the single harm cue but is unresolved at our sample size; GPT-OSS reads harm, and its refusals can be argued in either direction by its own reasoning trace. Where refusal reads only a low-rank slice and routes around the bulk of what the model knows, a rank-one edit removes it. Whether widening what refusal reads would also deepen the behavior is the open question this raises.

Directional ablation removes an aligned language model's ability to refuse by projecting a single "refusal direction" out of the weights that write the residual stream. It needs no gradient-based training and no optimization, only a few hundred contrastive prompts, which makes it the canonical white-box attack on open-weight alignment. However, it has been established only on dense models up to roughly 70B parameters. We study whether it survives the shift to frontier mixture-of-experts (MoE) models whose residual streams are no longer a single tensor and whose weights ship quantized. We apply it to GLM-5.3-Flash (320B parameters, 288 routed experts, a four-wide hyper-connection residual, block-FP8). The attack survives the architecture, but what it reaches is no longer where a reader of the original recipe would look for it. Editing the attention, dense and routed-expert writers on their own removes 0.039, 0.016 and 0.148 of refusal respectively; editing all three together removes 0.776. As a result, 74% of the effect exists only under the joint intervention. The part the conventional recipe reaches by module-name matching accounts for 0.066 of that 0.776, which is why it fails silently on an MoE. The effect does not follow from removing just any direction: ablating a random direction orthogonal to it leaves refusal unchanged. A category-concentrated residue survives every edit we tried: subspaces fitted on violence, sexual content and hate leave measurable refusal at every rank from 1 to 12. We report the method, the 41-89 percentage-point reductions it achieves across seven harmful benchmarks with no detected change in capability, and the boundary where it stops.

Safety monitors screen prompts sent to deployed language models, flagging harmful requests so they are never answered. They are evaluated by recall against harmfulness labels, but a catch only prevents harm if the model would otherwise have complied. We measure the difference directly: we sample repeated responses from the target model, call a harmful prompt \emph{elicitable} if the model complies at least once, and report monitor recall separately on elicitable and non-elicitable prompts. Across six monitor configurations and three model families, spanning activation probes, fine-tuned text guards, and a 120B policy-conditioned reasoning classifier, recall on elicitable prompts falls 0.22 to 0.38 below recall on non-elicitable prompts at a fixed false positive rate. The prompts a monitor misses are 2.8 to 5.6 times more likely to be complied with than the prompts it catches. The gap replicates across three model families and appears also in text-only monitors entirely independent of the target model. This suggests that standard recall may overstate the protection monitors provide in practice, and that monitors should be evaluated against what their models will actually answer.

How do the methods used to train language models to refuse harmful requests shape how that refusal actually works inside the model? We compare three post-training methods - supervised fine-tuning, reasoning-augmented fine-tuning (training on reasoning chains that justify a safety decision), and preference optimization (ORPO) - across three architecturally distinct models (Llama-3.1-8B, Gemma-2-9B, Qwen3-8B). We find that training method, not just data, reshapes how refusal is computed internally: reasoning-augmented training consistently produces a distinct kind of refusal computation, visible across all three models, while architecture independently shapes internal structure and how reliably refusal can be steered. Most importantly, no method we study achieves all three properties we would want from safe alignment at once: refusal that isn't concentrated in a few fragile components, safety gains that don't cost general capability, and safety behavior correctable through small, targeted edits. We caution against treating current post-training methods as a solved, reliable defense, especially for security-critical use. Code and models are available in https://github.com/hoangcuongnguyen2001/Beyond-Shallow-Alignment.

Large language models (LLMs) are increasingly trained to decline queries that fall outside their knowledge (knowledge-based refusal, KR) or violate safety policies (safety-based refusal, SR). Although KR and SR result in superficially similar responses, they have largely been studied in isolation, leaving open whether they share an underlying mechanism. We address this gap with a systematic study on a new dataset of 213 contrastive quadruples that jointly probe both refusal types. We find that KR and SR are governed by overlapping yet distinguishable mechanisms. Both share a refusal direction, yet the overlap is asymmetric: SR signals transfer more strongly to KR than the reverse. Type-specific specialization emerges mainly in upper layers, with KR aligning with uncertainty- and knowledge-related representations and SR with safety- and policy-related ones. We thus characterize refusal as a commit-then-specify process: a shared initial mechanism commits to refusing, then type-specific features in later layers specify whether the grounds are epistemic or normative.

Diffusion large language models (dLLMs) generate text through iterative denoising rather than left-to-right decoding. This generation paradigm introduces two axes that can influence safety alignment: when tokens are generated during denoising and where they appear in the response. In this paper, we measure dLLM safety behavior under harmful prompts by tracing intermediate token distributions and commitment decisions throughout denoising. Our analysis shows that refusal signals are concentrated in early denoising steps and leading response positions, and the tokens committed early can strongly shape the final safety outcome. Our measurements further show that the denoising step and persistence of refusal-token commitment are important for understanding dLLM safety. Based on these findings, we propose Refusal-Aware Early Commitment (RAEC), a simple training-free decoding method that commits persistent refusal signals from early steps. Experiments on LLaDA and Dream show that RAEC reduces attack success rates while largely preserving utility. The code is available at https://github.com/Glresearch1/RAEC.

Activation steering has emerged as a lightweight approach for controlling model refusal at inference time. A growing line of research explores trainable rotations of activations to develop geometrically principled intervention mechanisms. However, existing techniques rely on auxiliary constructs, such as refusal vectors, to define these rotations. In our work, we develop a self-contained methodology for learning parameter-efficient rotational transformations based on Riemannian optimization. We empirically validate the proposed scheme, demonstrating its superiority in intervention efficiency. An extensive ablation study highlights the importance of key design choices in our method. Our results identify the proposed rotation-based steering scheme as a promising direction for more reliable control over the behavior of LLMs.

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.

Existing jailbreak evaluations typically characterize robustness using a single attack success rate (ASR) measured in a default configuration (the vanilla state). However, user-LLM interactions can induce diverse operational states beyond the vanilla state. In this work, we find that jailbreak robustness is highly fragile to operational-state variation: even when the attack remains fixed, changing only an ordinary system prompt not designed to affect safety can dramatically alter attack success rates. We systematically investigate this phenomenon across seven aligned models and three representative jailbreak attacks, observing substantial differences in ASR between vanilla and non-vanilla operational states. In one case, ASR increases by up to 56 percentage points (2% to 58%) solely due to a change in operational state. Remarkably, these increases occur even for attacks originally designed and optimized under vanilla-state evaluation. We further show that state-dependent robustness variation is systematically associated with differences in hidden representations along a refusal-related axis, and that projections onto this axis strongly predict jailbreak outcomes. Our results show that a single vanilla-state evaluation may not fully characterize jailbreak robustness, motivating evaluations that also examine how robustness changes across non-vanilla operational states.

Large language models are trained to follow instructions while refusing harmful requests. Jailbreaks exploit this balance to elicit content a model would ordinarily reject. Roleplay jailbreaks are especially concerning: the harmful request can remain visible inside a roleplay wrapper made of a persona, scenario, and task, yet the model may comply. We use mechanistic interpretability to determine how this context reverses refusal and which elements contribute to the reversal. Across two benchmarks, three model families, and four authored wrappers, we compare matched harmful and benign requests with and without this wrapper. We trace hidden-state contrasts from the request to the final prompt state, isolate wrapper operations through controlled counterfactuals, intervene on their activation directions in held-out evaluation requests, and decompose effective directions geometrically. Our analysis yields three findings. (1) Successful attacks retain the measured harmful-versus-benign distinction at the request, while its refusal-associated expression weakens where the answer begins, a pattern we call safety-relay attenuation. (2) Constructing the complete roleplay around the request and framing it within the scenario contribute causally: removing the associated activation changes restores refusal. (3) These effects largely share internal structure, and most repair is reproduced by components aligned with the model's ordinary refusal of harmful requests without roleplay; scenario framing retains a smaller, model-dependent component. Together, these findings explain how roleplay can produce compliance despite retained evidence of harm and identify a concrete target for future safeguards: maintaining the connection from harm recognition to refusal.

Safety alignment is essential for deploying large language models, requiring systems to prevent harmful compliance while preserving helpfulness on benign requests. Activation steering offers a training-free inference-time approach to safety control, but effective safety steering requires addressing two coupled questions: when to intervene and how generation should be shaped after intervention. However, existing safety steering methods remain limited along both dimensions, as their triggering mechanisms can be unstable across domains and refusal-oriented steering often yields rigid refusals rather than constructive safe guidance. To address these limitations, we propose ALTSTEER, an inference-time framework that couples selective intervention with refusal-anchored constructive redirection within a single inference pass. ALTSTEER uses an internal refusal-relevant signal to decide when to steer, and applies staged steering to shift generation from refusal-oriented control toward constructive alternatives. Evaluations on Llama-3.1 and Qwen2.5 show that ALTSTEER preserves benign utility while improving constructive safe-completion behavior, especially on models that otherwise tend to produce short refusals for harmful requests.

Refusal training protects AI models from jailbreaks by training models to decline unsafe queries, reducing the risk of misuse. Recent work finds that refusal behavior in aligned language models can be mediated by a single activation direction or a low-dimensional refusal subspace shared across harmful prompts: ablating those directions suppresses refusals while largely preserves other model capabilities. Yet it remains unclear why safety-critical features in a wide range of models emerge in a concentrated, low-dimensional structure. In a case study of OLMo-2-0425-1B-Instruct we find that the refusal geometry reflects refusal training: activation updates resulting from refusal-completion first-token losses explain the resulting refusal direction and refusal subspace. We study refusal directions through the training dynamics across refusal datasets and reveal that their brittleness is associated with repetitive refusal starts, which in turn is linked to concentration of gradients and refusal features in a low-dimensional subspace. Across frozen-model analyses and controlled synthetic fine-tuning, we find evidence of a hardening lever: diverse refusal starts can raise stable ranks of gradients and activation changes, making refusals harder to remove with a vector ablation attack.

Practical LLM unlearning is usually evaluated through two objectives: suppress target-specific knowledge and preserve non-target utility. In generative QA, this leaves a third behavior underspecified: when a target-adjacent prompt admits a broader answer without target-specific leakage, the model should answer at that level rather than leak, evade, or refuse. We study this specification problem in a controlled LoRA-GRPO RWKU setting, comparing four reward designs that span lexical suppression, anti-refusal shaping, rubric-based broad answering, and an explicit refusal contrast, with and without SFT warm-up. The experiments show that optimization success is not equivalent to behavioral unlearning: RWKU forget scores, held-out completion audits, and training dynamics can point to different conclusions. We trace these disagreements to reward-hacking endpoints, policy-support limits in GRPO, benchmark probes that miss endpoint changes, and a rubric reward that selects broad-topic answering with low semantic leakage under held-out evaluation.

Encoded-prompt attacks are evaluated almost entirely on their harmful arm: a benchmark sends obfuscated harmful requests and reports how often the model complied. A high refusal rate there is reported as safety, and it is equally consistent with a model that has stopped telling the request apart from anything else in the same format. We run the benign arm through the same transformation, and the two cases are far apart. Across four 7-8B models spanning three base families and four post-training recipes, refusal of harmful homoglyph-encoded prompts spans 0.08 while the same four span 0.57 on the identical requests in plaintext. What the encoding destroys is not refusal but the harm gap: on one model the gap between harmful and benign refusal falls from +0.82 in plaintext to exactly 0.00 under the encoding, and a benchmark reading only the harmful arm scores that model and one retaining a +0.61 gap identically. Running the cell such benchmarks leave out (plaintext content wearing the attack template, with nothing obfuscated) shows that on two of the four models the loss is caused by the protocol rather than by the character transformation, and on a third by the characters. Across a full SFT -> DPO -> RLVR pipeline the harm gap rises by +0.26 with a paired interval excluding zero while the standard harmful-arm metric registers no resolved change at all. We report twelve instrument defects, each with the control that caught it, including a binary jailbreak judge that fires on 0.61-0.70 of responses to plaintext benign prompts; six of the twelve inflate apparent safety, which is the direction a broken safety evaluation fails in by default.

State-aligned distortion has been documented in China-origin text-based large language models (LLMs), but whether, and in what form, it arises in multimodal systems has not been systematically examined. We construct a balanced benchmark of 200 core entries spanning ten politically sensitive topics, plus a seven-variant visual-abstraction probe, and run nine vision-language models (VLMs), seven China-origin and two non-China, across four elicitation paradigms and two prompt languages, yielding 21,708 trials. Each response is audited on six dimensions -- explicit refusal, information integrity, visual grounding, state-aligned framing, language consistency, and response length -- by two independent frontier LLM judges, validated against three human experts on a 200-trial sample. Measuring each dimension separately lets us decompose multimodal censorship into individual signals rather than a single refusal-based score; in particular, refusal and framing are measured independently, so a model can stop refusing while still reframing. We find that (i) Chinese-language prompting roughly triples the odds of state-aligned framing, within every model; (ii) China-origin models reframe more than non-China models (direction robust across judges and human raters; magnitude 1.6--3.2x); (iii) the effect is strongest in text-only political commentary (36.5%) and is gated by recognition of the depicted subject rather than pixel detail, persisting even at silhouette for iconic images; and (iv) across four Qwen generations, state-aligned framing rises while explicit refusal falls: censorship migrates from a visible act (refusal) to an invisible one (fluent reframing). We argue this shift to invisible reframing is fundamentally a problem of human-AI interaction: it removes the very signal users rely on to recognize that information has been withheld.

Safety alignment in large language models is often treated as a distributed property of the entire network, yet its practical brittleness suggests that refusal behavior may be concentrated in a smaller set of parameters. This work addresses where safety-aligned refusal is encoded by transplanting weights from aligned models into matched unaligned base models at multiple levels of granularity. Using two open-weight model pairs and four safety benchmarks, we conducted experiments to compare the effects of replacing attention weights, MLP weights, contiguous layer regions, and MLP blocks. Across both model families, refusal transfer is dominated by MLP weights: replacing MLP parameters recovers substantially more malicious-prompt refusal than replacing attention parameters, with gains of at least 2.7 times more across benchmarks. Within the MLP stack, refusal-relevant parameters exhibit a consistent mid-network concentration, as the block spanning layers 8-11 is selected first in all six greedy searches over model-dataset pairs. The results also show that the composition of safety-relevant components is non-additive: in five of six greedy trajectories, adding more aligned blocks can reduce refusal performance, and selective block subsets can outperform full MLP transplantation on malicious refusal, benign over-refusal, or both. Finally, greedy orders transferred to OR-Bench vary with the source benchmark used to derive them, indicating a benchmark-dependent precision-coverage trade-off. These results suggest that safety alignment in current LLMs is both localized and interaction-sensitive, offering insight into alignment brittleness and potential avenues for targeted safety interventions.

Large language models (LLMs) increasingly provide conversational health information that may influence treatment decisions, yet existing benchmarks do not isolate whether medication-safety boundaries persist across follow-ups after explicit self-treatment intent. We introduce TAF-MED, a physician-reviewed benchmark of 500 fixed three-turn scenarios, and evaluate eight LLMs across 4,000 conversations. A rubric-based automated judge labelled responses as SAFE, LEAKY, or UNSAFE, and two physicians independently annotated a model-balanced random subset of 400 conversations. We assessed unsafe guidance, collapse after a strictly SAFE initial response, and model-ranking stability. Overall, 71.6% of conversations contained an UNSAFE response, and 61.4% of those beginning with a strictly SAFE response later collapsed to UNSAFE; model-level collapse rates ranged from 24.4% to 96.2%. Four of 28 model pairs reversed order between initial unsafe and collapse rates. Automated labels achieved 94.3% agreement with the adjudicated physician reference ($κ= 0.895$). These findings show that first-turn safety is an incomplete proxy for conversational safety persistence and motivate evaluation across complete dialogue trajectories. We will release TAF-MED on Hugging Face to support reproducible research on multi-turn medical safety.

Safety alignment in multilingual models is uneven: a model that reliably refuses a harmful request in English will often comply with the same request in a lower-resource language. We trace this gap mechanistically in sarvam, an Indic-multilingual mixture-of-experts reasoning model, and find it is not a failure to detect harm. Harm is encoded as an internal direction that is nearly language-invariant in mid-network (English-vs-Indic cosine ${\approx}0.9$ at $L11$), and steering that direction upstream causally controls refusal. But the detection direction is orthogonal to the change that actually writes the refusal, which is late and assembled over the course of generation rather than read off in a single forward pass. We attribute the write to a specific, localizable circuit, a mixture-of-experts writer held in check by an attention opposer and price every way of intervening on it: damping the opposer is cheap and effective, amplifying the writer is a cost wall, and surgical edits to the responsible heads do nothing. The circuit's organization, and the gradient method that exposes it, recur in a second, unrelated MoE model, while the lever's strength is architecture-specific. The result is a cost-measured map of where a multilingual safety repair can land, and what it costs

Prompt injection is a critical security threat in large language model (LLM) applications, where attackers hijack model behavior by embedding malicious instructions in user or external data. Existing detection methods only detect the presence of injection and refuse to respond upon detection, overlooking the fact that for many modern aligned models, well-crafted instructions can resist most injection attacks. This means that the injection robustness varies significantly across instructions and models. This leads to widespread unnecessary over-refusal: inputs containing injections that the model could have handled correctly are rejected incorrectly. To deal with this over-refusal issue, we propose BASIS (Robustness-Aware Prompt Injection Defense). This defense method uses the Attention Competition Ratio ($ρ$) as features to train two sparse linear probes: an existence probe and a breach probe. Both probes make defense decisions through cascaded gating, which does not require additional LLM inference. BASIS comprises three stages: injection existence detection, per-sample breach prediction, and instruction robustness assessment; the online cascade refuses only when the model would actually be compromised and thus avoids over-refusal on robust instructions. Experiments across four tasks and six open-source LLMs show that BASIS maintains near-perfect injection detection while substantially reducing over-refusal on safe attack samples, especially under robust instruction templates.

Large language models typically undergo post-training to align them with safety policies but there exist many sophisticated jailbreaks that sidestep established safeguards. For instance, prior work by Andriushchenko et al. (2025) has found that changing the grammatical tense from present to past can be enough to elicit harmful responses. In this work, we uncover a more general failure of non-imperative syntactic forms. We demonstrate that this syntactic vulnerability exists in 16 models up to 70B parameters, using behavioral evaluation. To investigate the root cause, we apply causal mediation analysis, finding that refusal is partially conditioned on upstream syntactic features. By steering these purely syntactic features we are able to trigger and suppress refusal. Finally, we trace this ill-conditioning to linguistically biased post-training data of open-source models and show that increasing syntactic diversity can mitigate the issue. Our findings suggest that current alignment approaches introduce confounders that prevent a pure semantic grounding of the refusal decision.

We tackle the challenging yet underexplored task of Generalized Referring Expression Comprehension (GREC), which requires a model to localize the object described by a textual expression when it exists (positive sample) and to refuse output when it does not (negative sample). Although Multimodal Large Language Models (MLLMs) excel at localizing existing objects, they often fail to reject nonexistent ones due to the absence of negative samples during training, producing hallucinated bounding boxes. Existing post-training approaches such as supervised fine-tuning (SFT) and reinforcement learning (RL) enhance refusal behavior but usually degrade localization accuracy on positive samples, undermining the model's core competence. To address this, we propose Refusal-Calibrated Group Relative Policy Optimization (RC-GRPO), a calibrated RL strategy that strengthens the refusal ability of MLLMs while preserving localization performance. It enforces "None" outputs in rollouts for valid advantage estimation on negative samples and applies a penalty to prevent over-refusal on positives, achieving a balanced trade-off between accuracy and reliability. A second-stage reasoning reinforcement further consolidates causal understanding and interpretability. Experiments on three GREC benchmarks demonstrate that RC-GRPO attains superior localization accuracy while maintaining strong refusal capability.