RXed AI News

LegalHalluLens reveals AI hallucinations in legal workflows aggregate at 52%, but error concentration and direction are hidden. https://huggingface.co/papers/2606.18021

@OpenAI's Codex now records workflows once and repeats them autonomously as reusable skills. https://the-decoder.com/openais-codex-can-now-watch-you-work-once-and-repeat-the-task-forever/

DeepSWIP advances counterfactual reasoning for neurosymbolic AI, enabling causal inference in probabilistic logic programs. https://arxiv.org/abs/2606.20526v1

Small language models now match frontier LLMs in relation extraction without massive compute costs. https://arxiv.org/abs/2606.22606v1

Sakana AI launches Fugu, routing tasks across swappable frontier LLMs to top coding and reasoning benchmarks. https://www.marktechpost.com/2026/06/22/sakana-ai-launches-sakana-fugu-an-orchestration-model-that-routes-tasks-across-a-swappable-pool-of-frontier-llms/

InSight framework lets AI models self-learn new manipulation skills beyond training data. https://arxiv.org/abs/2606.24884v1

OpenAI's GPT-5.5-Cyber model outperforms Anthropic's Mythos on cybersecurity benchmarks. https://the-decoder.com/openai-says-new-gpt-5-5-cyber-outperforms-anthropics-mythos-on-cybersecurity-benchmark/

OpenThoughts-Agent introduces data recipes for training broadly capable agentic models. https://arxiv.org/abs/2606.24855v1

OpenAI improves ChatGPT Instant with better intent recognition and multi-turn context. https://the-decoder.com/openai-says-chatgpt-instant-now-better-understands-what-users-actually-want/