A recent arXiv paper from Politecnico di Torino, "Dense Contexts Are Hard Contexts," pours cold water on the "million-token context" narrative. Researchers used three sets of "needle in a haystack" benchmarks with exactly the same length (~12K tokens), strictly controlled the position of the information, and only changed the information density — and found a long-overlooked phenomenon: even if the length doesn't change, the model's retrieval accuracy drops off a cliff as density rises. Open-source models that nearly maxed out on sparse text fell directly below 60% on the high-density scenario.
This overturns the implicit assumption that "context window = effective capacity." Over the past two years, the industry has equated "long context" with "long attention": pushing the window from 128K to 1M to 10M makes the benchmark numbers look good, the marketing copy sounds loud. But the new paper points out that the third axis that determines LLM performance is lexical item diversity (MATTR) — the same 12K tokens, novel-style prose (MATTR≈0.72) can be skip-read, while configuration-style / code-style / retrieval-assembled text (MATTR≈0.82) requires almost every token to be processed. Same length, different density, the difficulty is worlds apart.
For all the vendors and engineering teams piling on "million tokens," this is a sobering reminder: the actual usability of a 1M window may be even less than a "clean" 64K context for RAG-retrieved material stuffed into the prompt, multi-segment tool output stitched together by Agents, and long user-session history — as long as the content itself is high-density. Llama 4, Gemini 3.x, Qwen3.7-Max and other new flagships pile on high scores on NIAH, but that doesn't mean they will be stable in real Agent workflows.
What's more noteworthy is the next step: whether it is possible to train the model to be adaptive to density, or to do dynamic summarization/chunking on high-density segments at the inference end — this is the key to pulling "long context" from the benchmark back to production value.