Researchers ran a controlled bake-off of the three main ways to give large language models new information, and the results are messier than any single vendor's marketing would suggest.
The study compares representation-based methods (KV-cache approaches like Cartridges and Compaction) against parametric fine-tuning across five knowledge-heavy benchmarks. In ideal conditions, Cartridges win almost every storage budget, beating fine-tuning by 10 points. Compaction only keeps pace with Cartridges at low compression rates; push compression past 50x and it falls 10 points behind fine-tuning instead. In a more realistic test involving retrieval across multiple documents, Cartridges are the only method that matches plain in-context learning, outperforming fine-tuning by 29 points and Compaction by 15. But Cartridges have a catch: they are among the few methods, alongside full fine-tuning and large adapter layers, that suffer catastrophic forgetting, dropping 6 percent on unrelated benchmarks and 13 percent on coding tasks.
This matters because "how do I get new facts into a model" is no longer an academic question. Every RAG pipeline, every fine-tuned support bot, every long-context agent is implicitly betting on one of these three approaches, usually without benchmarking the alternatives. The finding that the most accurate method also degrades unrelated skills is the kind of tradeoff that gets discovered in production, not in a demo.
None of this settles the context-versus-weights debate that has run since retrieval-augmented generation first challenged pure fine-tuning. It just confirms there is no free lunch: better knowledge injection currently costs you something else, and teams need to pick their poison based on what they can afford to forget.