Search for: All records

Emerging large reasoning models (LRMs), such as DeepSeek-R1 models, leverage long chain-of-thought (CoT) reasoning to generate structured intermediate steps, enhancing their reasoning capabilities. However, long CoT does not inherently guarantee safe outputs, potentially leading to harmful consequences such as the introduction of security vulnerabilities in code or the spread of misinformation. Current research on large language model (LLM) safety usually focuses on short-answer responses, overlooking the long CoT style outputs of LRMs. To bridge this gap, we conduct a systematic study of LRM safety. First, we investigate safety evaluators calibrated against human annotations. Using our newly developed metrics, we thoroughly assess the safety of 13 state-of-the-art LRMs on StrongReject and WildJailbreak datasets. Our results show that LRMs are not safe compared to their reasoning advance. Further, we perform a fine-grained analysis of the reasoning trace and final answer. We find that three decoding strategies-ZeroThink, LessThink, and MoreThink-can improve model safety without additional training. However, these strategies either use constrained reasoning traces or incur high inference costs. To better strengthen LRM safety, we introduce SafeChain, the first-of-its-kind safety training dataset in CoT style. We fine-tune two LRMs with SafeChain, showing that it not only enhances model safety but also preserves performance across 6 reasoning benchmarks. 

Clinical decision-making is inherently complex and fast-paced, particularly in emergency departments (EDs) where critical, rapid and high-stakes decisions are made. Clinical Decision Rules (CDRs) are standardized evidence-based tools that combine signs, symptoms, and clinical variables into decision trees to make consistent and accurate diagnoses. CDR usage is often hindered by the clinician's cognitive load, limiting their ability to quickly recall and apply the appropriate rules. We introduce CDR-Agent, a novel LLM-based system designed to enhance ED decision-making by autonomously identifying and applying the most appropriate CDRs based on unstructured clinical notes. To validate CDR-Agent, we curated two novel ED datasets: synthetic and CDR-Bench, although CDR-Agent is applicable to non ED clinics. CDR-Agent achieves a 56.3\% (synthetic) and 8.7\% (CDR-Bench) accuracy gain relative to the standalone LLM baseline in CDR selection. Moreover, CDR-Agent significantly reduces computational overhead. Using these datasets, we demonstrated that CDR-Agent not only selects relevant CDRs efficiently, but makes cautious yet effective imaging decisions by minimizing unnecessary interventions while successfully identifying most positively diagnosed cases, outperforming traditional LLM prompting approaches. 

Safety is critical to the usage of large language models (LLMs). Multiple techniques such as data filtering and supervised fine tuning have been developed to strengthen LLM safety. However, currently known techniques presume that corpora used for safety alignment of LLMs are solely interpreted by semantics. This assumption, however, does not hold in real-world applications, which leads to severe vulnerabilities in LLMs. For example, users of forums often use ASCII art, a form of text-based art, to convey image information. In this paper, we propose a novel ASCII art-based jailbreak attack and introduce a comprehensive benchmark Vision-in-Text Challenge (VITC) to evaluate the capabilities of LLMs in recognizing prompts that cannot be solely interpreted by semantics. We show that five SOTA LLMs (GPT-3.5, GPT-4, Gemini, Claude, and Llama2) struggle to recognize prompts provided in the form of ASCII art. Based on this observation, we develop the jailbreak attack ArtPrompt, which leverages the poor performance of LLMs in recognizing ASCII art to bypass safety measures and elicit undesired behaviors from LLMs. ArtPrompt only requires black-box access to the victim LLMs, making it a practical attack. We evaluate ArtPrompt on five SOTA LLMs, and show that ArtPrompt can effectively and efficiently induce undesired behaviors from all five LLMs. Our code is available at https: //github.com/uw-nsl/ArtPrompt. 

A<sc>bstract</sc> This paper presents a study of the charmless three-body decays$$ {B}_{(s)}^0\to {K}_{\mathrm{S}}^0{h}^{+}h{\prime}^{-} $$ $B_{\left(s\right)}^0\to K_S^0{h}^{+}h{\prime }^{-}$ (whereh⁽′⁾=π, K), using a sample ofppcollision data collected by the LHCb experiment during Runs 1 and 2 of the LHC, corresponding to an integrated luminosity of 9 fb⁻¹. The decay$$ {B}_s^0\to {K}_{\mathrm{S}}^0{K}^{+}{K}^{-} $$ $B_s^0\to K_S^0{K}^{+}{K}^{-}$ is observed for the first time, and the following ratios of branching fractions are measured:$$ {\displaystyle \begin{array}{l}\frac{\mathcal{B}\left({B}^0\to {K}_{\mathrm{S}}^0{K}^{+}{K}^{-}\right)}{\mathcal{B}\left({B}^0\to {K}_{\mathrm{S}}^0{\pi}^{+}{\pi}^{-}\right)}=0.578\pm 0.007\pm 0.017,\\ {}\frac{\mathcal{B}\left({B}^0\to {K}_{\mathrm{S}}^0{K}^{\pm }{K}^{\mp}\right)}{\mathcal{B}\left({B}^0\to {K}_{\mathrm{S}}^0{\pi}^{+}{\pi}^{-}\right)}=0.1363\pm 0.0035\pm 0.051,\\ {}\begin{array}{l}\frac{\mathcal{B}\left({B}_s^0\to {K}_{\mathrm{S}}^0{\pi}^{+}{\pi}^{-}\right)}{\mathcal{B}\left({B}^0\to {K}_{\mathrm{S}}^0{\pi}^{+}{\pi}^{-}\right)}=0.269\pm 0.011\pm 0.015\pm 0.008,\\ {}\frac{\mathcal{B}\left({B}_s^0\to {K}_{\mathrm{S}}^0{K}^{+}{K}^{-}\right)}{\mathcal{B}\left({B}^0\to {K}_{\mathrm{S}}^0{\pi}^{+}{\pi}^{-}\right)}=0.0303\pm 0.0041\pm 0.0025\pm 0.0009,\\ {}\frac{\mathcal{B}\left({B}_s^0\to {K}_{\mathrm{S}}^0{K}^{\pm }{\pi}^{\mp}\right)}{\mathcal{B}\left({B}^0\to {K}_{\mathrm{S}}^0{\pi}^{+}{\pi}^{-}\right)}=1.818\pm 0.021\pm 0.031\pm 0.056,\end{array}\end{array}} $$ $\begin{array}{c}\frac{B\left(B^0\to K_S^0{K}^{+}{K}^{-}\right)}{B\left(B^0\to K_S^0{\pi }^{+}{\pi }^{-}\right)}=0.578\pm 0.007\pm 0.017,\\ \frac{B\left(B^0\to K_S^0{K}^{\pm }{K}^{\mp }\right)}{B\left(B^0\to K_S^0{\pi }^{+}{\pi }^{-}\right)}=0.1363\pm 0.0035\pm 0.051,\\ \begin{array}{c}\frac{B\left(B_s^0\to K_S^0{\pi }^{+}{\pi }^{-}\right)}{B\left(B^0\to K_S^0{\pi }^{+}{\pi }^{-}\right)}=0.269\pm 0.011\pm 0.015\pm 0.008,\\ \frac{B\left(B_s^0\to K_S^0{K}^{+}{K}^{-}\right)}{B\left(B^0\to K_S^0{\pi }^{+}{\pi }^{-}\right)}=0.0303\pm 0.0041\pm 0.0025\pm 0.0009,\\ \frac{B\left(B_s^0\to K_S^0{K}^{\pm }{\pi }^{\mp }\right)}{B\left(B^0\to K_S^0{\pi }^{+}{\pi }^{-}\right)}=1.818\pm 0.021\pm 0.031\pm 0.056,\end{array}\end{array}$ where the uncertainties are statistical, systematic, and due to knowledge of the ratio of hadronisation fractions of the$$ {B}_s^0 $$ $B_s^0$ andB⁰mesons, respectively. 

The time-dependent $CP$ asymmetry in $B^0\to J/\psi \rho (770{)}^0$ decays is measured using proton-proton collision data corresponding to an integrated luminosity of $6{\mathrm{fb}}^{-1}$ , collected with the LHCb detector at a center-of-mass energy of 13 TeV during the years 2015–2018. The $CP$ -violation parameters for this process are determined to be $2{\beta }_{c\overline{c}d}^{\mathrm{eff}}=0.710\pm 0.084\pm 0.051\mathrm{rad}$ and $\left|\lambda \right|=1.019\pm 0.034\pm 0.024$ , where the first uncertainty is statistical and the second systematic. This constitutes the first observation of time-dependent $CP$ violation in $B^0\to J/\psi \rho (770{)}^0$ decays. Assuming approximate SU(3) flavor symmetry, these results are combined with the previous consistent LHCb measurement to set the most stringent constraint on the penguin contribution, $\mathrm{\Delta }{\varphi }_s$ , to the $CP$ -violating phase ${\varphi }_s$ in $B_s^0\to J/\psi \varphi \left(1020\right)$ decays, yielding $\mathrm{\Delta }{\varphi }_s=5.0\pm 4.6\mathrm{mrad}$ . 

Muon-electron universality is tested in $B^0\to K^{*0}{\ell }^{+}{\ell }^{-}$ decays, in the dilepton-invariant-mass region above the $\psi \left(2S\right)$ resonance. The analysis uses beauty mesons produced in proton-proton collisions recorded by the LHCb detector at center-of-mass energies of 7, 8, and 13 TeV, corresponding to an integrated luminosity of $9{\mathrm{fb}}^{-1}$ . The ratio of branching fractions between the muon and electron channels $R_{K^{*0}}$ is measured to be ${1.08}_{-0.12}^{+0.14}\left(\mathrm{stat}\right)\pm 0.07\left(\mathrm{syst}\right)$ for a dilepton-invariant-mass squared above $14.0{\mathrm{GeV}}^2/c^4$ , consistent with the standard model prediction. This result represents the most precise measurement of $R_{K^{*0}}$ in this region and the first such measurement performed at a hadron collider.