OpenAI Renegotiates Pentagon Deal After Backlash, Admits 'Opportunistic and Sloppy' Handling In a swift reversal, OpenAI is renegotiating its recently announced agreement to provide AI models to the U.S. Department of Defense (DoD), also known as the Department of War (DoW). The move comes after the deal faced significant backlash, with critics accusing the company of prioritizing profits over ethical concerns. OpenAI CEO Sam Altman acknowledged the company "shouldn't have rushed" to finalize the agreement, which he admitted "looked opportunistic and sloppy." The original deal, announced on Friday, allowed the DoD to deploy OpenAI's models, including the popular ChatGPT, on classified military networks. The renegotiated terms will now explicitly prohibit the use of OpenAI's AI for domestic surveillance of U.S. citizens, a key concern raised by critics. Altman stated the new contract language will ensure OpenAI's systems are not "intentionally used for domestic surveillance of U.S. persons and nationals," in line with constitutional and legal protections. The agreement will also bar the DoD's intelligence components, such as the National Security Agency and Defense Intelligence Agency, from using OpenAI's services without a separate contract modification. Altman's comments come after Anthropic, a rival AI company, faced pressure from the Trump administration for refusing to remove safeguards against the use of its technology for mass surveillance and autonomous weapons. The administration labeled Anthropic a "supply-chain risk," leading the company to reject a similar Pentagon deal. In contrast, OpenAI has been more willing to work with the government, with Altman stating the company was "genuinely trying to de-escalate things and avoid a much worse outcome." However, the optics of the initial deal struck a nerve, with some users launching a "delete ChatGPT" campaign and Anthropic's Claude chatbot surging in popularity. Legal experts have raised concerns about the broader implications of AI companies partnering with the military, particularly regarding the potential for misuse and the erosion of public trust. The renegotiated terms aim to address these concerns, but the debate over the ethical boundaries of AI development and deployment is likely to continue.
Meta's AI Ambitions: Integrating Rivals, Expanding Capabilities, and Navigating Copyright Challenges In a move to bolster its AI capabilities, Meta (formerly Facebook) is reportedly integrating models from rival labs, including Gemini, ChatGPT, and Claude, into its own Meta AI platform. This comes as the tech giant prepares to release new in-house models, such as "Avocado," and expand the functionality of its AI assistant, including the addition of scheduled tasks and voice agent support. Meta's acquisition of Manus AI also appears to be a key part of its AI strategy, with the company testing Manus-style agents and exploring ways to integrate the technology into Meta AI. Additionally, the platform is said to be testing an experimental AI shopping tool, allowing users to receive product recommendations and links to e-commerce sites. However, Meta's AI ambitions have not been without controversy. The company is facing a significant copyright infringement lawsuit from adult film producers Strike 3 Holdings and Counterlife Media, who allege that Meta downloaded at least 2,396 of their films without authorization to use in AI training. The potential damages in this case could exceed $350 million. In a related case, Meta has secured a partial fair use victory in a lawsuit filed by several book authors, who accused the company of using pirated books to train its large language model, LLaMA. While the court granted Meta summary judgment on specific claims, it also outlined how copyright challenges against AI developers might succeed in the future, emphasizing the importance of proving potential market harm from AI-generated content. These legal battles highlight the complex and evolving landscape of AI development, where tech giants like Meta must navigate the delicate balance between innovation and respecting intellectual property rights. As the company continues to expand its AI capabilities, it will likely face increased scrutiny and potential legal challenges from content creators and rightsholders seeking to protect their work.
Distilling the tool-using capabilities of large language models (LLMs) into smaller, more efficient small language models (SLMs) is a key challenge for their practical application. The predominant approach, supervised fine-tuning (SFT), suffers from poor generalization as it trains models to imitate a static set of teacher trajectories rather than learn a robust methodology. While reinforcement learning (RL) offers an alternative, the standard RL using sparse rewards fails to effectively guide SLMs, causing them to struggle with inefficient exploration and adopt suboptimal strategies. To address these distinct challenges, we propose MENTOR, a framework that synergistically combines RL with teacher-guided distillation. Instead of simple imitation, MENTOR employs an RL-based process to learn a more generalizable policy through exploration. In addition, to solve the problem of reward sparsity, it uses a teacher's reference trajectory to construct a dense, composite teacher-guided reward that provides fine-grained guidance. Extensive experiments demonstrate that MENTOR significantly improves the cross-domain generalization and strategic competence of SLMs compared to both SFT and standard sparse-reward RL baselines.
Gradient-based data attribution methods, such as influence functions, are critical for understanding the impact of individual training samples without requiring repeated model retraining. However, their scalability is often limited by the high computational and memory costs associated with per-sample gradient computation. In this work, we propose GraSS, a novel gradient compression algorithm and its variants FactGraSS for linear layers specifically, that explicitly leverage the inherent sparsity of per-sample gradients to achieve sub-linear space and time complexity. Extensive experiments demonstrate the effectiveness of our approach, achieving substantial speedups while preserving data influence fidelity. In particular, FactGraSS achieves up to 165% faster throughput on billion-scale models compared to the previous state-of-the-art baselines. Our code is publicly available at https://github.com/TRAIS-Lab/GraSS.
Large language models (LLMs) have demonstrated promising performance in generating diagnostic conclusions from imaging findings, thereby supporting radiology reporting, trainee education, and quality control. However, systematic guidance on how to optimize prompt design across different clinical contexts remains underexplored. Moreover, a comprehensive and standardized framework for assessing the trustworthiness of LLM-generated radiology reports is yet to be established. This study aims to enhance the trustworthiness of LLM-generated liver MRI reports by introducing a Multi-Dimensional Credibility Assessment (MDCA) framework and providing guidance on institution-specific prompt optimization. The proposed framework is applied to evaluate and compare the performance of several advanced LLMs, including Kimi-K2-Instruct-0905, Qwen3-235B-A22B-Instruct-2507, DeepSeek-V3, and ByteDance-Seed-OSS-36B-Instruct, using the SiliconFlow platform.
