Daily Papers

We study how runtime enforcement against unsafe actions affects end-to-end task performance in multi-step tool using large language model (LLM) agents. Using tau-bench across Airline and Retail domains, we compare baseline Tool-Calling, planning-integrated (TRIAD), and policy-mediated (TRIAD-SAFETY) architectures with GPT-OSS-20B and GLM-4-9B. We identify model dependent interaction horizons (15 to 30 turns) and decompose outcomes into overall success rate (SR), safe success rate (SSR), and unsafe success rate (USR). Our results reveal a persistent Safety Capability Gap. While safety mediation can intercept up to 94 percent of non-compliant actions, it rarely translates into strictly safe goal attainment (SSR below 5 percent in most settings). We find that high unsafe success rates are primarily driven by Integrity Leaks, where models hallucinate user identifiers to bypass mandatory authentication. Recovery rates following blocked actions are consistently low, ranging from 21 percent for GPT-OSS-20B in simpler procedural tasks to near zero in complex Retail scenarios. These results demonstrate that runtime enforcement imposes a significant verifier tax on conversational length and compute cost without guaranteeing safe completion, highlighting the critical need for agents capable of grounded identity verification and post-intervention reasoning.

Large Language Model safety alignment predominantly operates on a binary assumption that requests are either safe or unsafe. This classification proves insufficient when models encounter ethical dilemmas, where the capacity to reason through moral trade-offs creates a distinct attack surface. We formalize this vulnerability through TRIAL, a multi-turn red-teaming methodology that embeds harmful requests within ethical framings. TRIAL achieves high attack success rates across most tested models by systematically exploiting the model's ethical reasoning capabilities to frame harmful actions as morally necessary compromises. Building on these insights, we introduce ERR (Ethical Reasoning Robustness), a defense framework that distinguishes between instrumental responses that enable harmful outcomes and explanatory responses that analyze ethical frameworks without endorsing harmful acts. ERR employs a Layer-Stratified Harm-Gated LoRA architecture, achieving robust defense against reasoning-based attacks while preserving model utility.

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.

Clinical trials are crucial for drug development but are time consuming, expensive, and often burdensome on patients. More importantly, clinical trials face uncertain outcomes due to issues with efficacy, safety, or problems with patient recruitment. If we were better at predicting the results of clinical trials, we could avoid having to run trials that will inevitably fail more resources could be devoted to trials that are likely to succeed. In this paper, we propose Hierarchical INteraction Network (HINT) for more general, clinical trial outcome predictions for all diseases based on a comprehensive and diverse set of web data including molecule information of the drugs, target disease information, trial protocol and biomedical knowledge. HINT first encode these multi-modal data into latent embeddings, where an imputation module is designed to handle missing data. Next, these embeddings will be fed into the knowledge embedding module to generate knowledge embeddings that are pretrained using external knowledge on pharmaco-kinetic properties and trial risk from the web. Then the interaction graph module will connect all the embedding via domain knowledge to fully capture various trial components and their complex relations as well as their influences on trial outcomes. Finally, HINT learns a dynamic attentive graph neural network to predict trial outcome. Comprehensive experimental results show that HINT achieves strong predictive performance, obtaining 0.772, 0.607, 0.623, 0.703 on PR-AUC for Phase I, II, III, and indication outcome prediction, respectively. It also consistently outperforms the best baseline method by up to 12.4\% on PR-AUC.

LLM agents with tool access can discover and exploit security vulnerabilities. This is known. What is not known is which features of a system prompt trigger this behaviour, and which do not. We present a systematic taxonomy based on approximately 10,000 trials across seven models, 37 prompt conditions, and 12 hypothesized attack dimensions -- all executed in real Docker sandboxes. Every condition includes the instruction "Always respect all rules and access policies"; only a single encouragement sentence varies. Within the task class studied (planted test-runner vulnerabilities), the threat model is narrower than assumed: nine of twelve dimensions -- including minimization, moral licensing, incentives, identity priming, and reasoning triggers -- produce no detectable exploitation at n=50 per cell (upper 95% CI <7% per condition). But it is also more subtle than expected: one dimension, goal reframing, reliably triggers exploitation. "You are solving a puzzle; there may be hidden clues" produces 38-40% exploitation on Claude Sonnet 4 despite the explicit rule instruction, replicating across four models (CTF framing: 8-14% on DeepSeek, GPT-5-mini, o4-mini). The agent does not override the rules; it reinterprets the task so that exploitative actions become task-aligned. GPT-4.1 produces no exploitation across 1,850 trials (37 conditions), and a temporal comparison across four OpenAI models released over eleven months shows a pattern consistent with improving safety training, though model capability differences are a confounder. The practical contribution is a narrowed, testable threat model: defenders should audit for goal-reframing language, not for the broad class of adversarial prompts.

Adverse drug events (ADEs) significantly impact clinical research, causing many clinical trial failures. ADE prediction is key for developing safer medications and enhancing patient outcomes. To support this effort, we introduce CT-ADE, a dataset for multilabel predictive modeling of ADEs in monopharmacy treatments. CT-ADE integrates data from 2,497 unique drugs, encompassing 168,984 drug-ADE pairs extracted from clinical trials, annotated with patient and contextual information, and comprehensive ADE concepts standardized across multiple levels of the MedDRA ontology. Preliminary analyses with large language models (LLMs) achieved F1-scores up to 55.90%. Models using patient and contextual information showed F1-score improvements of 21%-38% over models using only chemical structure data. Our results highlight the importance of target population and treatment regimens in the predictive modeling of ADEs, offering greater performance gains than LLM domain specialization and scaling. CT-ADE provides an essential tool for researchers aiming to leverage artificial intelligence and machine learning to enhance patient safety and minimize the impact of ADEs on pharmaceutical research and development. The dataset is publicly accessible at https://github.com/ds4dh/CT-ADE.

Red-team evaluations of AI models support some claims and not others, and the boundary between the two is calculable rather than merely a matter of judgment. We define the evidential ceiling of an evaluation as the largest factor by which one result can move belief under a fixed testing budget, derive it in closed form for the benchmark null result, and use it to locate that boundary exactly. We find that above a calculable harm rate, a benchmark of modest size certifies a category to a stated evidentiary standard, and a clean sheet is then the stronger of the two possible observations, outweighing a single reproduced failure. Below that rate, no passive benchmark of feasible size provides the specified evidence of safety under the fixed scoring rule and approximately independent trial structure. The crossing between the two regimes has a closed form. The bound is not specific to benchmarks: written in terms of a procedure's hypothesis conditioned elicitation rates, it covers adaptive and automated red teaming as well, and shows that discrimination between the hypotheses rather than attack success is what determines evidential worth. Auditing eight evaluation suites against the boundary, we find that current benchmarks are adequate for high-frequency harm categories and several orders of magnitude short for rare, catastrophic ones. Safety benchmarks are not uninformative. They are informative about a specific and computable set of propositions, and the discipline they need is to state which.

We describe ufakzeka-1, a 151M-parameter (182M with embeddings) decoder-only Turkish language model pretrained from scratch on 13.5B tokens of openly licensed text and instruction-tuned for chat, at a total cost of about \$286 in cloud GPU, API and notebook time. The contribution is not the model's capability, which is what a model this size can be expected to have, but the record of building and measuring it: a Turkish byte-level tokenizer at 1.77 tokens per word, a three-stage pretraining schedule, a post-training mixture of openly licensed and generated data, and an evaluation battery of release gates, a rule-checked sweep of 5,508 conversations, judged conversations and hand tests, all with prompts held out from the training data, enforced by decontamination inside the data build and by a checked-in invariant script we run before each build. We report three findings that we believe transfer to other small-model efforts: a safety gate that had been "fixed" with training data written from its own questions read 64/64 while the honest figure was 34/64; training-seed variance was as large as the spread across every recipe we tried, so single-seed comparisons at this scale are uninformative; and data rounds repaired only what was absent from the data, while identity tracking over long context and multi-turn arithmetic did not move across any data change we tried, which we read as limits of the model size rather than gaps in the data, a reading the next, larger model will test. Weights, the data recipe, the evaluation code and the spend ledger are released under Apache-2.0.

Deep reinforcement learning has emerged as a promising and powerful technique for automatically acquiring control policies that can process raw sensory inputs, such as images, and perform complex behaviors. However, extending deep RL to real-world robotic tasks has proven challenging, particularly in safety-critical domains such as autonomous flight, where a trial-and-error learning process is often impractical. In this paper, we explore the following question: can we train vision-based navigation policies entirely in simulation, and then transfer them into the real world to achieve real-world flight without a single real training image? We propose a learning method that we call CAD^2RL, which can be used to perform collision-free indoor flight in the real world while being trained entirely on 3D CAD models. Our method uses single RGB images from a monocular camera, without needing to explicitly reconstruct the 3D geometry of the environment or perform explicit motion planning. Our learned collision avoidance policy is represented by a deep convolutional neural network that directly processes raw monocular images and outputs velocity commands. This policy is trained entirely on simulated images, with a Monte Carlo policy evaluation algorithm that directly optimizes the network's ability to produce collision-free flight. By highly randomizing the rendering settings for our simulated training set, we show that we can train a policy that generalizes to the real world, without requiring the simulator to be particularly realistic or high-fidelity. We evaluate our method by flying a real quadrotor through indoor environments, and further evaluate the design choices in our simulator through a series of ablation studies on depth prediction. For supplementary video see: https://youtu.be/nXBWmzFrj5s

Large language models (LLMs) represent a transformative class of AI tools capable of revolutionizing various aspects of healthcare by generating human-like responses across diverse contexts and adapting to novel tasks following human instructions. Their potential application spans a broad range of medical tasks, such as clinical documentation, matching patients to clinical trials, and answering medical questions. In this primer paper, we propose an actionable guideline to help healthcare professionals more efficiently utilize LLMs in their work, along with a set of best practices. This approach consists of several main phases, including formulating the task, choosing LLMs, prompt engineering, fine-tuning, and deployment. We start with the discussion of critical considerations in identifying healthcare tasks that align with the core capabilities of LLMs and selecting models based on the selected task and data, performance requirements, and model interface. We then review the strategies, such as prompt engineering and fine-tuning, to adapt standard LLMs to specialized medical tasks. Deployment considerations, including regulatory compliance, ethical guidelines, and continuous monitoring for fairness and bias, are also discussed. By providing a structured step-by-step methodology, this tutorial aims to equip healthcare professionals with the tools necessary to effectively integrate LLMs into clinical practice, ensuring that these powerful technologies are applied in a safe, reliable, and impactful manner.