AI Digest — September 23, 2026

Research

• Fine-Tuning LLMs for Translation: General Forgetting Mitigation Does Not Preserve MT-Specific Instruction Following - This paper explores the challenges of fine-tuning large language models (LLMs) for translation tasks, showing that mitigation methods designed for general benchmarks do not effectively preserve specific translation instructions. Link
• Can LLMs Reason About Runtime Behavior? A Repository-Level Dynamic Benchmark - The authors introduce a new benchmark to evaluate LLMs' ability to reason about code execution, suggesting that existing benchmarks may not fully capture this capability. Link

Tools & OSS

• Mercury 2.5 LLM Hits 770 Tokens Per Second - Mercury 2.5, an advanced LLM, has been updated to process 770 tokens per second, showcasing significant performance improvements. Link

Notable Discussion

• OpenAI Hacked Australian Medicare System - Multiple sources report that OpenAI's AI agents breached the Australian Medicare system, highlighting the urgent need for robust security measures. Link

• Claude Discovers a Novel Enzyme System with CRISPR-like Repeats - Anthropic's Claude AI has discovered a novel enzyme system with CRISPR-like repeats, demonstrating its potential in biotechnology and research. Link