Also Worth Noting — 2026-04-10

Also Worth Noting

02 [Image Gen] DISCO AI Designs Enzymes, Catalytic Sites Included A new AI model, DISCO, can co-design protein sequences and structures, even generating the catalytic parts of enzymes from scratch. This is impressive because prior generative models could design proteins to bind ligands, but none could create enzymes without pre-specifying their catalytic residues. This breakthrough could unlock the creation of novel enzymes, expanding the range of chemistry available for drugs, industrial processes, and new materials. link

03 [RAG] Benchmarking Realistic LLM Agent Skill Usage A benchmark has been created to evaluate how effectively AI agents select and use tools from a library in complex, real-world situations. This is challenging because LLMs often struggle to identify and apply the most relevant skills from a large pool, unlike idealized tests where perfect skills are directly provided. This will help developers build more robust and reliable AI agents that can independently choose and utilize the right tools to solve practical problems. link

04 [RAG] Optimizing Retrieval for AI Agents Information retrieval systems can now be trained using the actual interactions and paths (trajectories) of AI agents. This method moves beyond traditional human click data, optimizing retrieval specifically for how AI agents process and use found information. Such specialized training will make AI agents more efficient and effective at finding and utilizing information in complex tasks. link

05 [RAG] Why Model Pruning Fails Generative AI Insights explain why shrinking AI models through pruning improves simple tasks but often breaks complex generative AI like RAG. Understanding this discrepancy is challenging because pruning aims to remove "less important" parameters, which are often surprisingly critical for complex generative outputs. This knowledge allows for more targeted and effective pruning strategies, creating efficient RAG and other generative AI models that retain high-quality output. link

06 [Video Gen] Video-MME-v2: Next-Gen Benchmark for Real Video Understanding Video-MME-v2 is a new benchmark designed to accurately test the true capabilities of video understanding AI models. It rigorously evaluates robustness and faithfulness, addressing the problem of existing benchmarks showing inflated scores that don't reflect real-world performance. This allows researchers to better understand and develop more reliable video AI, leading to practical advancements beyond current leaderboard results. link