LLM Refusal Behavior

Military Large Language Models (LLMs) must provide accurate information to the warfighter in time-critical and dangerous situations. However, today's LLMs are imbued with safety behaviors that cause the LLM to refuse many legitimate queries in the military domain, particularly those related to violence, terrorism, or military technology. Our gold benchmark for assessing refusal rates, which was developed by veterans of the US Army and special forces, is to our knowledge the first dataset of its kind. We present results for refusal and deflection rates on 31 public models and 3 military models. We observe hard rejection rates as high as 98.2% and soft deflection rates ranging from 0% to 21.3%. We also present results on two additional synthetic datasets and show their correlations with the gold dataset. Finally, we perform abliteration using the Heretic library on a military-tuned gpt-oss-20b model, showing an absolute increase in answer rate of 66.5 points but an average relative decrease of 2% on other military tasks. In our concluding remarks, we argue for deeper specialization, including with mid-training and end-to-end post-training, to achieve zero refusals and maximum military task accuracy for closed military models.

Large language models and LLM-based agents are increasingly used for cybersecurity tasks that are inherently dual-use. Existing approaches to refusal, spanning academic policy frameworks and commercially deployed systems, often rely on broad topic-based bans or offensive-focused taxonomies. As a result, they can yield inconsistent decisions, over-restrict legitimate defenders, and behave brittlely under obfuscation or request segmentation. We argue that effective refusal requires explicitly modeling the trade-off between offensive risk and defensive benefit, rather than relying solely on intent or offensive classification. In this paper, we introduce a content-based framework for designing and auditing cyber refusal policies that makes offense-defense tradeoffs explicit. The framework characterizes requests along five dimensions: Offensive Action Contribution, Offensive Risk, Technical Complexity, Defensive Benefit, and Expected Frequency for Legitimate Users, grounded in the technical substance of the request rather than stated intent. We demonstrate that this content-grounded approach resolves inconsistencies in current frontier model behavior and allows organizations to construct tunable, risk-aware refusal policies.

Modern deployments require LLMs to enforce safety policies at scale, yet many controls rely on inference-time interventions that add recurring compute cost and serving complexity. Activation steering is widely used, but it requires runtime hooks and scales cost with the number of generations; conditional variants improve selectivity by gating when steering is applied but still retain an inference-time control path. We ask whether selective refusal can be moved entirely offline: can a mechanistic understanding of category-specific refusal be distilled into a circuit-restricted weight update that deploys as a standard checkpoint? We propose C-Δθ Circuit Restricted Weight Arithmetic}, which (i) localizes refusal-causal computation as a sparse circuit using EAP-IG and (ii) computes a constrained weight update ΔθC supported only on that circuit (typically <5% of parameters). Applying ΔθC yields a drop-in edited checkpoint with no inference-time hooks, shifting cost from per request intervention to a one-time offline update. We evaluate category-targeted selectivity and capability retention on refusal and utility benchmarks.

Prior work argues that refusal in large language models is mediated by a single direction, enabling steering and abliteration. We show that this account is incomplete: across diverse refusal and non-compliance categories, refusal behaviors correspond to geometrically distinct directions in activation space. Yet activation steering along any refusal-related direction produces nearly identical refusal--over-refusal trade-offs, acting as a shared one-dimensional control knob. Thus, different directions primarily affect not whether the model refuses, but how it refuses. Using sparse autoencoders, we uncover a structured internal representation of refusal: a reusable core of shared refusal latents supplemented by style- and domain-specific latents. Linear interventions collapse this structure into uniform behavioral control, flattening mechanistic differences across refusal types. Our results reconcile the apparent simplicity of refusal steering with the diversity of refusal behaviors, and clarify the limits of linear interpretability for aligned model behavior.

Safety alignment in Large Language Models is critical for healthcare; however, reliance on binary refusal boundaries often results in over-refusal of benign queries or unsafe compliance with harmful ones. While existing benchmarks measure these extremes, they fail to evaluate Safe Completion: the model's ability to maximise helpfulness on dual-use or borderline queries by providing safe, high-level guidance without crossing into actionable harm. We introduce Health-ORSC-Bench, the first large-scale benchmark designed to systematically measure Over-Refusal and Safe Completion quality in healthcare. Comprising 31,920 benign boundary prompts across seven health categories (e.g., self-harm, medical misinformation), our framework uses an automated pipeline with human validation to test models at varying levels of intent ambiguity. We evaluate 30 state-of-the-art LLMs, including GPT-5 and Claude-4, revealing a significant tension: safety-optimised models frequently refuse up to 80% of "Hard" benign prompts, while domain-specific models often sacrifice safety for utility. Our findings demonstrate that model family and size significantly influence calibration: larger frontier models (e.g., GPT-5, Llama-4) exhibit "safety-pessimism" and higher over-refusal than smaller or MoE-based counterparts (e.g., Qwen-3-Next), highlighting that current LLMs struggle to balance refusal and compliance. Health-ORSC-Bench provides a rigorous standard for calibrating the next generation of medical AI assistants toward nuanced, safe, and helpful completions. Furthermore, our benchmark facilitates reproducible evaluation, encourages safety calibration, and supports development of clinically reliable, context-aware, human-aligned medical AI systems. Our code and data are available at: https://github.com/ZhihaoZhang97/Health-ORSC-Bench. Warning: Some contents may include toxic or undesired contents.

Current safety evaluations of large language models rely on single-shot testing, implicitly assuming that model responses are deterministic and representative of the model's safety alignment. We challenge this assumption by investigating the stability of safety refusal decisions across random seeds and temperature settings. Testing four instruction-tuned models from three families (Llama 3.1 8B, Qwen 2.5 7B, Qwen 3 8B, Gemma 3 12B) on 876 harmful prompts across 20 sampling configurations (4 temperatures x 5 seeds), we find that 18-28% of prompts exhibit decision flips (the REFUSE/PARTIAL/COMPLY label changes across configurations), depending on the model. Our Safety Stability Index (SSI) reveals that higher temperatures significantly reduce decision stability (Friedman chi-squared = 396.81, p < 0.001), with mean within-temperature SSI dropping from 0.977 at temperature 0.0 to 0.942 at temperature 1.0. We validate findings across all model families using Claude 3.5 Haiku as a unified external judge, achieving 89.1% inter-judge agreement with the Llama 70B judge on the two models both judges labeled (Cohen's kappa = 0.62). Within each model, prompts with higher compliance rates exhibit lower stability (Spearman rho = -0.47 to -0.70, all p < 0.001), indicating that models "waver" more on borderline requests. These findings demonstrate that single-shot safety evaluations are insufficient for reliable safety assessment and that evaluation protocols must account for stochastic variation in model behavior. For Llama 3.1 8B, single-shot evaluation agrees with multi-sample ground truth only 92.5% of the time when pooling across temperatures (98.7% at greedy to 90.3% at temperature 1.0), and we recommend scaling samples with temperature--one at greedy, three at low temperature, more at higher temperatures (where three reach only ~95%), and ten when pooling--rather than a single flat threshold.

LLMs are trained to refuse harmful instructions, but do they truly understand harmfulness beyond just refusing? Prior work has shown that LLMs' refusal behaviors can be mediated by a one-dimensional subspace, i.e., a refusal direction. In this work, we identify a new dimension to analyze safety mechanisms in LLMs, i.e., harmfulness, which is encoded internally as a separate concept from refusal. There exists a harmfulness direction that is distinct from the refusal direction. As causal evidence, steering along the harmfulness direction can lead LLMs to interpret harmless instructions as harmful, but steering along the refusal direction tends to elicit refusal responses directly without reversing the model's judgment on harmfulness. Furthermore, using our identified harmfulness concept, we find that certain jailbreak methods work by reducing the refusal signals without reversing the model's internal belief of harmfulness. We also find that adversarially finetuning models to accept harmful instructions has minimal impact on the model's internal belief of harmfulness. These insights lead to a practical safety application: The model's latent harmfulness representation can serve as an intrinsic safeguard (Latent Guard) for detecting unsafe inputs and reducing over-refusals that is robust to finetuning attacks. For instance, our Latent Guard achieves performance comparable to or better than Llama Guard 3 8B, a dedicated finetuned safeguard model, across different jailbreak methods. Our findings suggest that LLMs' internal understanding of harmfulness is more robust than their refusal decision to diverse input instructions, offering a new perspective to study AI safety.