AI Digest — Jun 11, 2026 (Morning)

Anthropic, the developer of AI model Claude, has walked back a policy that could have hindered researchers using the model. The policy in question could have 'sabotaged' researchers by limiting the model's performance. This reversal comes after backlash from the research community. The policy change is significant as it affects how researchers can utilize Claude for their studies. Anthropic's decision reflects the company's response to community feedback.

Google has introduced DiffusionGemma, a new technology that accelerates text generation by four times. This innovation is part of Google's efforts to improve the efficiency of natural language processing tasks. DiffusionGemma is designed to enhance the performance of text generation models, making them more suitable for real-time applications. The technology has the potential to impact various industries, including customer service, content creation, and language translation. By reducing the time required for text generation, DiffusionGemma can enable more responsive and interactive AI systems.

The proposed method, Context-Driven Incremental Compression (C-DIC), aims to improve the efficiency and robustness of conversational agents by compressing dialogue history. It treats conversations as contextual threads and stores compression states in a compact memory. C-DIC uses a retrieve, revise, and write-back loop to share information across turns, stabilizing long-horizon behavior. The approach is evaluated on long-form dialogue benchmarks, demonstrating superior performance and efficiency. It also adapts truncated backpropagation-through-time to learn cross-turn dependencies without full-history backpropagation.

The redesign focuses on aligning each router row with the principal singular direction of the associated expert, enhancing the mathematical description of the matrix. This is achieved through a 'Power-then-Retract' paradigm, ensuring efficiency and stability. Theoretically, the approach drives router rows to converge toward the principal singular directions of associated experts. Empirical results from pretraining MoE models across 1B to 11B parameters confirm the effectiveness of this alignment. The redesign aims to improve the performance of Mixture-of-Experts models.

Researchers propose Bebop, a method to accelerate reinforcement learning (RL) training by integrating Multi-Token Prediction (MTP) with rejection sampling. The study reveals that MTP acceptance rates are bounded by model entropy and that probabilistic rejection sampling can alleviate this issue. The authors propose a novel end-to-end TV loss that optimizes multi-step rejection sampling acceptance rates, resulting in up to 25% extra inference throughput gains. Experimental results show that Bebop achieves up to 1.8x end-to-end acceleration in async RL training of large language models.

Claw-SWE-Bench is a multilingual benchmark and adapter protocol that allows for the comparison of heterogeneous agent harnesses under fair settings. The benchmark contains 350 GitHub issue-resolution instances across 8 languages and 43 repositories. It is designed to evaluate the coding ability of general-purpose agents such as OpenClaw. The benchmark shows that adapter design is essential for enabling OpenClaw-style harnesses to perform coding tasks effectively. A smaller version, Claw-SWE-Bench Lite, is also available for faster validation.

Researchers introduced ALIGNBEAM, a method that enables inference-time alignment transfer between large language models with different vocabularies. This approach addresses the issue of domain fine-tuning degrading model safety by translating anchor logits into the target model's vocabulary. ALIGNBEAM uses a small LLM judge to select the safest continuation among candidates, without requiring weight changes or retraining. The method has been evaluated on both cross-vocabulary and same-vocabulary pairs, showing substantial improvements in refusal on adversarial benchmarks while maintaining task accuracy and inference overhead. This technique allows for safety alignment transfer between model families at inference time.

OpenAI has partnered with Oracle to offer its models and Codex through Oracle Cloud, allowing users to leverage their existing cloud commitments. This integration enables enterprises to build and deploy AI solutions with enhanced security and governance. The partnership aims to simplify the adoption of AI technologies for businesses. Users can access OpenAI's capabilities, including Codex, directly within their Oracle Cloud environment.

A report from OpenAI reveals PRC-linked influence operations are using AI to target US tech debates. These operations focus on data center narratives, tariffs, and false claims about ChatGPT. The report highlights the use of AI in spreading misinformation and influencing public opinion on key tech issues. The operations aim to shape US policy and public discourse on AI and related technologies.

Google Research has developed a new framework for auditing machine unlearning, which is crucial for ensuring that machine learning models can effectively forget data they were trained on. This framework is part of the Algorithms & Theory research area. The ability to audit machine unlearning is important for data privacy and compliance with regulations. The framework provides a systematic approach to evaluating the effectiveness of machine unlearning methods.

Dario Amodei has published a post on his website discussing policy considerations for the AI exponential. The post explores the potential risks and benefits of advanced AI systems and argues for a proactive approach to AI policy. Amodei's post has generated significant discussion, with 201 comments on the topic. The post is relevant to AI researchers and policymakers interested in the societal implications of AI development. Amodei's perspective is informed by his experience as a researcher and entrepreneur in the AI field.

HelixDB is a graph database built on top of object storage, allowing for scalable and flexible data management. The project is open-source and available on GitHub. This approach enables the combination of graph database capabilities with the scalability and cost-effectiveness of object storage. The database is designed to handle complex relationships and queries. It is built using a modular architecture, making it easier to integrate with existing systems.