搜索结果：open-weights

Large language models (LLMs) increasingly inform mental health decisions by patients and clinicians. Inference-time activation steering can shift model behavior on a target dimension without altering weights or prompts and without disclosure to users, allowing treatment recommendations to be silently changed for commercial or ideological reasons. To determine whether directional activation steering can shift an open-weights LLM's depression treatment recommendations. This non-human subjects study applied directional activation steering to an open-weights LLM (DeepSeek V4 Flash) responding to 12 depression-advice scenarios (4 favoring medication, 4 favoring avoidance, 4 neutral), generated at 30 amplitudes from -1.5 to +1.5 in 0.1 increments plus an unsteered baseline. A single steering direction contrasting antidepressant medication with self-directed approaches (diet, exercise, meditation, dietary supplements), constructed from 16 paired training prompts and applied at the attention output of every transformer block; weights and system prompt were held constant. The extent to which medication and four self-care categories were addressed, scored 0 to 3 by a human-validated LLM rater (Claude Opus 4.7), the medication-versus-self-care balance, and clinician referral, estimated per unit of amplitude using mixed-effects models with a scenario random intercept. Across 372 generations, steering produced a graded, dose-dependent shift in the medication-versus-self-care balance, which declined by 0.32 per unit of amplitude (&#x3b2; = -0.32; 95% CI, -0.39 to -0.25; P < .001); medication extent fell and self-care extent rose. The shift was largest for scenarios with no stated treatment preference (&#x3b2; = -0.44; 95% CI, -0.54 to -0.34; P < .001). A clinician referral appeared in 322 of 372 responses (87%) and did not vary with steering amplitude (P = .63). In this open-weights LLM providing depression treatment information, inference-time activation steering shifted treatment recommendations without altering weights, prompt structure, or safety outputs, with the largest effect among users expressing no treatment preference. These findings suggest a need for LLM disclosure standards and independent auditing as such models inform clinical decisions.

Extracting information from radiology reports can provide critical data to empower many radiology workflows. For spinal compression fractures, these data can facilitate evidence-based care for at-risk populations. Manual extraction from free-text reports is laborious, and error-prone. Large language models (LLMs) have shown promise; however, fine-tuning strategies to optimize performance in specific tasks can be resource intensive. A variety of prompting strategies have achieved similar results with fewer demands. Our study pioneers the use of Meta's Llama 3.1, together with prompt-based strategies, for automated extraction of compression fractures from free-text radiology reports, outputting structured data without model training. We tested performance on a time-based sample of CT exams covering the spine from 2/20/2024 to 2/22/2024 acquired across our healthcare enterprise (637 anonymized reports, age 18-102, 47% Female). Ground truth annotations were manually generated and compared against the performance of three models (Llama 3.1 70B, Llama 3.1 8B, and Vicuna 13B) with nine different prompting configurations for a total of 27 model/prompt experiments. The highest F1 score (0.91) was achieved by the 70B Llama 3.1 model when provided with a radiologist-written background, with similar results when the background was written by a separate LLM (0.86). The addition of few-shot examples to these prompts had variable impact on F1 measurements (0.89, 0.84 respectively). Comparable ROC-AUC and PR-AUC performance was observed. Our work demonstrated that an open-weights LLM excelled at extracting compression fractures findings from free-text radiology reports using prompt-based techniques without requiring extensive manually labeled examples for model training.

While Americans are using herbal dietary supplements (natural products) more than ever, the consumption of natural products with prescription drugs can lead to harmful interactions. Pharmacovigilance of natural products depends on careful expert review and interpretation of a wide variety of evidence. In prior work, we demonstrated the value of knowledge graph (NP-KG) for assisting with natural product safety investigations. However, scaling the NP-KG from 33 natural products to the thousands on the market requires computer-assisted data extraction, particularly from visual elements (figures or tables) of pharmacology literature. We evaluated the accuracy and resilience of 3 open-weights and 7 closed-weights multimodal models by performing visual information extraction from select tables and images. The best performing models could accurately extract 90% of tabular data and 45% of data reported figures with a modified relative error rate of 0.05. Image resolution and information density were primary hindrances to better extraction performance.

Patient access to radiology reports has heightened the need for patient-friendly communication. Automated generation of patient-centered summaries using large language models (LLMs) is a promising solution. However, their use on real-life reports is limited by privacy concerns. Here, we evaluate the safety and effectiveness of on-premise, privacy-preserving LLMs for generating lay summaries of real French brain MRI reports for emergency presentations of headache. In this retrospective study, we sampled 105 brain MRI reports (January&#x2013;December 2022) for radiologist evaluation and a subset of 30 reports for non-physician evaluation. Three open-weights models (Llama 3.3 70B, Athene V2, Mistral Small) generated French lay summaries via a single standardized prompt. Radiologists&#x2019; mean ratings across models were high for exactness (4.10, 95% CI: 4.04&#x2013;4.16), exhaustiveness (4.34, 95% CI: 4.29&#x2013;4.39), didacticness (3.83, 95% CI: 3.79&#x2013;3.88), and readiness for clinical use (3.84, 95% CI: 3.79&#x2013;3.89). Non-physicians reported higher perceived understanding with summaries, from 2.85 (95% CI: 2.67&#x2013;3.04) to 4.27 (95% CI: 4.15&#x2013;4.38, p&#x2009;<&#x2009;0.001). The correct identification rate for reports increased from 75.2% to 83.6% (p&#x2009;<&#x2009;0.001). The ability to identify causal findings also improved, from 80.6% to 84.8% (p&#x2009;<&#x2009;0.001) overall. Overall error rate in LLM-generated lay summaries was 19.7% (62/315), warranting expert oversight.

We present HaloGuard 1.0, an open-weights implementation of the constitutional-classifier paradigm for input safety. It achieves state-of-the-art performance on English and multilingual prompt-safety benchmarks at roughly one-tenth the model size of current leading open guard models. The safety constitution is the organising structure of the corpus: a natural-language constitution of 46 policies and 2,940 subcategories drives synthetic data generation, with exhaustive one-to-one paired counterfactuals that hold topic and vocabulary fixed while flipping intent, a two-tier harmless design that separately targets boundary and baseline false positives (FPs), and balanced multilingual materialisation across 46 languages that treats language as a surface form appearing on both sides of the boundary rather than as an adversarial signal. Across seven prompt-safety benchmarks, HaloGuard 1.0-0.8B attains the best average F1 (90.9) of any open guard we evaluate, outperforming baselines up to 27B parameters (over 30 times larger) while holding false-positive rate (FPR) to 4.3 and false-negative rate (FNR) to 9.5. The HaloGuard 1.0-4B variant reaches average F1 of 92.1 and FPR of 3.5, spending

Large language models (LLMs) have shown strong potential for automated software vulnerability detection, particularly in retrieval-augmented generation (RAG) settings. However, for approaches relying on proprietary models and APIs, reproducibility and replicability remain largely unexplored, raising the question of whether reported results generalize or depend primarily on specific model choices. In this work, we present a reproducibility study of Vul-RAG, a RAG-based framework for source code vulnerability detection that enhances LLMs with high-level vulnerability knowledge. We first replicate the results in a fully local and open-weights setting using the reported open-weight baseline models. We then extend the evaluation to a diverse set of recent open-weight LLMs, including code-specialized, general-purpose, and reasoning models of varying parameter sizes. The results confirm that the findings of Vul-RAG are reproducible under local deployment, but with minor deviations. Across all evaluated models, we observe a performance plateau at approximately 0.30 pairwise accuracy (code pairs for which both the vulnerable and the patched function are correctly classified). Notably, this

Open-weight Large Language Models (LLMs) enable scientific progress and broad deployment. However, they make it difficult to control access to sensitive capabilities. Current practice either suppresses dangerous capabilities before release or mediates access through closed services that use specialized model variants, input/output monitors, and API permissions. The former is susceptible to jailbreaks while sacrificing capability for all users to mitigate the risks posed by a few, and the latter is fundamentally incompatible with open-weight release. In this paper, we propose Tiered Language Models (TLMs), where a single set of released weights supports multiple capability levels. In its default public configuration, a TLM behaves as a conventional LLM. A compact secret key specifies a permutation over a small parameter subset, inducing an alternative computation graph over the same weights that exposes additional capabilities. We develop a training protocol that jointly pretrains both configurations from scratch, then fine-tunes the keyed configuration on private data with regularization to preserve the public model's behavior. We pretrain 180M- and 650M-parameter TLMs and demonstr

Open-weight advanced AI models -- systems whose parameters are freely available for download and adaptation -- are reshaping the global AI landscape. As these models rapidly close the performance gap with closed alternatives, they enable breakthrough research and broaden access to powerful tools. However, once released, they cannot be recalled, and their built-in safeguards can be bypassed through fine-tuning or jailbreaking, posing risks that current governance frameworks are not equipped to address. This report moves beyond the binary framing of ``open'' versus ``closed'' AI. We assess the current landscape of open-weight advanced AI, examining technical capabilities, risk profiles, and regulatory responses across the European Union, United States, China, the United Kingdom, and international forums. We find significant disparities in safety practices across developers and jurisdictions, with no commonly adopted standards for determining when or how advanced models should be released openly. We propose a tiered, safety-anchored approach to model release, where openness is determined by rigorous risk assessment and demonstrated safety rather than ideology or commercial pressure. W

We present Open Ontologies, an open-source ontology engineering system implemented in Rust that integrates LLM-driven construction with formal OWL reasoning and ontology alignment via the Model Context Protocol. Our primary finding is that stable 1-to-1 matching is the dominant factor in ontology alignment quality: on the OAEI Anatomy track, it achieves F1 = 0.832 (P = 0.963, R = 0.733), competitive with state-of-the-art systems and exceeding all in precision. Ablation across five weight configurations shows that signal weights are irrelevant when stable matching is applied (F1 varies by less than 0.004), while removing stable matching drops F1 to 0.728. On the Conference track, the same method achieves F1 = 0.438. On tool-augmented ontology interaction, we find a surprising result: an LLM reading a raw OWL file (F1 = 0.323) performs worse than the same LLM with no file at all (F1 = 0.431), while structured MCP tool access achieves F1 = 0.717. This demonstrates that tool structure provides a qualitatively different mode of access that the LLM cannot replicate by reading raw syntax. The system ships as a single binary under the MIT licence.

Recent defenses for safeguarding open-weight large language models (LLMs) are intended to prevent adversarial usage. Underlying these defenses is an assumption that new harmful behavior is learned through fine-tuning rather than elicited by jailbreaking the model. Yet, pretrained LLMs already encode substantial harmful knowledge across many domains, which raises an important question: can an adversary jailbreak safeguarded models, to achieve harmful usage without fine-tuning at all? In this paper, we show that open-weight safeguards are susceptible to simpler strategies that, despite being well known, have not been systematically evaluated against these safeguards. Specifically, we evaluate two low-cost attacks--abliteration and prefilling--that do not rely on gradient-based optimization. Across three harmfulness evaluation benchmarks (BeaverTails, HarmBench, and AdvBench), these attacks increase attack success rates against safeguarded open-weight models from below 10\% to a range of 16%-96%. To mitigate this vulnerability, we introduce abliteration-resistant tuning (ART), which incorporates an abliteration-based objective into training. ART can be layered onto existing defenses a