Agentic reinforcement learning (RL) trains large language models to autonomously call tools during reasoning, with search as the most common application. These models excel at multi-step reasoning tasks, but their safety properties are not well understood. In this study, we show that RL-trained search models inherit refusal from instruction tuning and often deflect harmful requests by turning them into safe queries. However, this safety is fragile. Two simple attacks, one that forces the model to begin response with search (Search attack), another that encourages models to repeatedly search (Multi-search attack), trigger cascades of harmful searches and answers. The attacks succeed by triggering models to generate harmful, request-mirroring search queries before they can generate the inherited refusal tokens. This exposes a core weakness of current RL training: it rewards continued generation of effective queries without accounting for their harmfulness. As a result, RL search models have vulnerabilities that users can easily exploit, making it urgent to develop safety-aware agentic RL pipelines optimising for safe search. Instruction tuning (IT) is the standard method to align large language models (LLMs) with human preferences and teach them to refuse harmful requests (Schulman et al., 2017; Shao et al., 2024). However, IT only shapes static responses and is insufficient in agentic settings, where models must also decide when and how to call external tools, capabilities not explicitly learned during pre-training (Zhang et al., 2025). Agentic reinforcement learning (RL) for tool-use (Zhang et al., 2025) tackles this by fine-tuning models to interleave reasoning with tool use (Dong et al., 2025). In practice, search is the most common tool: agentic RL rewards effective, well-timed search queries and achieves strong gains on multi-hop reasoning tasks (Song et al., 2025a;b; Jin et al., 2025). Despite the progress, effect of agentic RL on safety of IT models remains unclear. While prior work reported safety degradation of retrieval-augmented agents (Y u et al., 2025), little is known about whether agentic RL for search preserves refusal of harmful requests. As agentic RL is now deployed in closed-source systems such as OpenAI's DeepSearch (OpenAI, 2025), this evaluation gap can create real deployment risks.

Traditional Chinese Medicine (TCM), with a history spanning over two millennia, plays a role in global healthcare. However, applying large language models (LLMs) to TCM remains challenging due to its reliance on holistic reasoning, implicit logic, and multimodal diagnostic cues. Existing TCM-domain LLMs have made progress in text-based understanding but lack multimodal integration, interpretability, and clinical applicability. To address these limitations, we developed BenCao, a ChatGPT-based multimodal assistant for TCM, integrating structured knowledge bases, diagnostic data, and expert feedback refinement. BenCao was trained through natural language instruction tuning rather than parameter retraining, aligning with expert-level reasoning and ethical norms specific to TCM. The system incorporates a comprehensive knowledge base of over 1,000 classical and modern texts, a scenario-based instruction framework for diverse interactions, a chain-of-thought simulation mechanism for interpretable reasoning, and a feedback refinement process involving licensed TCM practitioners. BenCao connects to external APIs for tongue-image classification and multimodal database retrieval, enabling dynamic access to diagnostic resources. In evaluations across single-choice question benchmarks and multimodal classification tasks, BenCao achieved superior accuracy to general-domain and TCM-domain models, particularly in diagnostics, herb recognition, and constitution classification. The model was deployed as an interactive application on the OpenAI GPTs Store, accessed by nearly 1,000 users globally as of October 2025. This study demonstrates the feasibility of developing a TCM-domain LLM through natural language-based instruction tuning and multimodal integration, offering a practical framework for aligning generative AI with traditional medical reasoning and a scalable pathway for real-world deployment.

ABSTRACT We propose Schr odinger Bridge Mamba (SBM), a new concept of training-inference framework motivated by the inherent compatibility between Schr odinger Bridge (SB) training paradigm and selective state-space model Mamba. Experiments on a joint denoising and dereverberation task using four benchmark datasets demonstrate that SBM, with only 1-step inference, outperforms strong baselines with 1-step or iterative inference and achieves the best real-time factor (RTF). Beyond speech enhancement, we discuss the integration of SB paradigm and selective state-space model architecture based on their underlying alignment, which indicates a promising direction for exploring new deep generative models potentially applicable to a broad range of generative tasks. Index T erms-- Schr odinger Bridge, Mamba, Deep generative model, Speech enhancement 1. INTRODUCTION Deep generative models have been increasingly employed for speech enhancement (SE) tasks. By learning the underlying distribution of clean audio given its degraded counterpart, generative models are capable of generating high-quality speech from low-quality inputs that include noise, reverberation, clipping, bandwidth limitation or a mixture of these artifacts.

The rise of generative AI (GenAI) has impacted many aspects of human life. As these systems become embedded in everyday practices, understanding public trust in them also becomes essential for responsible adoption and governance. Prior work on trust in AI has largely drawn from psychology and human-computer interaction, but there is a lack of computational, large-scale, and longitudinal approaches to measuring trust and distrust in GenAI and large language models (LLMs). This paper presents the first computational study of Trust and Distrust in GenAI, using a multi-year Reddit dataset (2022--2025) spanning 39 subreddits and 197,618 posts. Crowd-sourced annotations of a representative sample were combined with classification models to scale analysis. We find that Trust and Distrust are nearly balanced over time, with shifts around major model releases. Technical performance and usability dominate as dimensions, while personal experience is the most frequent reason shaping attitudes. Distinct patterns also emerge across trustors (e.g., experts, ethicists, general users). Our results provide a methodological framework for large-scale Trust analysis and insights into evolving public perceptions of GenAI.

When OpenAI released its AI video-generation app, Sora, in September, it promised that "you are in control of your likeness end-to-end." The app allows users to include themselves and their friends in videos through a feature called "cameos"--the app scans a user's face and performs a liveness check, providing data to generate a video of the user and to authenticate their consent for friends to use their likeness on the app. But Reality Defender, a company specializing in identifying deepfakes, says it was able to bypass Sora's anti-impersonation safeguards within 24 hours. Platforms such as Sora give a "plausible sense of security," says Reality Defender CEO Ben Colman, despite the fact that "anybody can use completely off-the-shelf tools" to pass authentication as someone else. Reality Defender's researchers used publicly available footage of notable individuals, including CEOs and entertainers, from earnings calls and media interviews.

When you purchase through links in our articles, we may earn a small commission. Developers and creatives looking for greater control and privacy with their AI are increasingly turning to locally run models like OpenAI's new gpt-oss family of models, which are both lightweight and incredibly functional on end-user hardware. Indeed, you can have it run on consumer GPUs with just 16GB of memory. That makes it possible to use a wide range of hardware - with NVIDIA GPUs emerging as the best way to run these sorts of open-weight models. While nations and companies rush to develop their own bespoke AI solutions to a range of tasks, open source and open-weight models like OpenAI's new gpt-oss-20b are finding much more adoption.

The strategic choice of model "openness" has become a defining issue for the foundation model (FM) ecosystem. While this choice is intensely debated, its underlying economic drivers remain underexplored. We construct a two-period game-theoretic model to analyze how openness shapes competition in an AI value chain, featuring an incumbent developer, a downstream deployer, and an entrant developer. Openness exerts a dual effect: it amplifies knowledge spillovers to the entrant, but it also enhances the incumbent's advantage through a "data flywheel effect," whereby greater user engagement today further lowers the deployer's future fine-tuning cost. Our analysis reveals that the incumbent's optimal first-period openness is surprisingly non-monotonic in the strength of the data flywheel effect. When the data flywheel effect is either weak or very strong, the incumbent prefers a higher level of openness; however, for an intermediate range, it strategically restricts openness to impair the entrant's learning. This dynamic gives rise to an "openness trap," a critical policy paradox where transparency mandates can backfire by removing firms' strategic flexibility, reducing investment, and lowering welfare. We extend the model to show that other common interventions can be similarly ineffective. Vertical integration, for instance, only benefits the ecosystem when the data flywheel effect is strong enough to overcome the loss of a potentially more efficient competitor. Likewise, government subsidies intended to spur adoption can be captured entirely by the incumbent through strategic price and openness adjustments, leaving the rest of the value chain worse off. By modeling the developer's strategic response to competitive and regulatory pressures, we provide a robust framework for analyzing competition and designing effective policy in the complex and rapidly evolving FM ecosystem.

The developments in Generative AI technologies have paved the way for numerous innovations in different fields. Recently, Generative AI has been proposed as a competitor to AES systems in evaluating student essays automatically. Considering the potential limitations of AI in processing idioms, this study assessed the scoring performances of Generative AI models for essays with and without idioms by incorporating insights from Corpus Linguistics and Computational Linguistics. Two equal essay lists were created from 348 student essays taken from a corpus: one with multiple idioms present in each essay and another with no idioms in essays. Three Generative AI models (ChatGPT, Gemini, and Deepseek) were asked to score all essays in both lists three times, using the same rubric used by human raters in assigning essay scores. The results revealed excellent consistency for all models, but Gemini outperformed its competitors in interrater reliability with human raters. There was also no detectable bias for any demographic group in AI assessment. For essays with multiple idioms, Gemini followed a the most similar pattern to human raters. While the models in the study demonstrated potential for a hybrid approach, Gemini was the best candidate for the task due to its ability to handle figurative language and showed promise for handling essay-scoring tasks alone in the future.

Monte Carlo (TSMC) to provide scalable stepwise supervision using small LMs. This yields more efficient training, improved stability, and higher accuracy. OpenAI's o1 series (OpenAI, 2024) introduce inference-time scaling by increasing the length of the Chain-of-Thought (CoT) (Wei et al., 2022) reasoning process. Despite their empirical success, RL approaches that generate the entire reasoning chain in a single forward pass face notable limitations, including CoT derailment, where the reasoning trajectory drifts off course due to accumulated errors, and the inherent challenges of long-horizon RL with sparse outcome rewards. This sequential scaling strategy, i.e., simply extending the CoT length, can therefore be insufficient (Y ang et al., 2025). To improve planning quality, we introduce Multi-Path Plan Aggregation (MPP A). For each planning step, the model generates multiple alternative plans and aggregates them into an improved plan before proceeding to the subsequent execution steps. Beyond enhancing planning, we identify a fundamental challenge in credit assignment for long-horizon policy learning (Kaelbling et al., 1996). Existing RL fine-tuning frameworks struggle to provide effective process-level supervision (Guo et al., 2025). First, evaluating the correctness of intermediate steps is inherently difficult. Automated annotation using LLM judges (Gu et al., 2024) often yield unreliable or noisy signals Second, introducing a separate process reward model (PRM) adds complexity. We then define the process preference between two candidate continuations at the same step by comparing their incremental log-weights. We repurpose Twisted Sequential Monte Carlo (TSMC) to provide process-level preferences for online Step-DPO training. Results show that our approach consistently outperforms both distillation-based long-CoT methods and RL methods that rely solely on outcome rewards. The Chain-of-Thought trajectories can be lengthy and the positions of the first error vary considerably, making outcome-based RL fine-tuning inefficient. Training long trajectories with outcome rewards is highly inefficient.

OpenAI patched the ShadowLeak attack in August 2025 after hackers exploited ChatGPT to exfiltrate Gmail data through invisible email prompts without user interaction.