A new scientific-reasoning benchmark called Reconstruction, published in August, exposes a clean failure mode in frontier models: they are remarkably bad at recovering research ideas from bibliographies alone. Frontier LLMs manage between 3 and 15 percent accuracy on solo attempts. Even running a multi-agent Swiss-tournament pipeline with the top four models only gets to 42 percent.
This is the kind of result that matters because it's specific. It's not "models hallucinate" or "reasoning is hard." It's: given a paper's bibliography and a prompt to reconstruct the paper's core idea, models fail most of the time. Bibliographies are dense signals. They're the map of what a researcher read and synthesized. If you can't reverse that synthesis from the map, you're not doing research-grade reasoning yet. You're doing something else.
The gap scales predictably, too. You can throw more models at the problem. You can give them more context. You can iterate. And the best you get is 42 percent with a tournament structure. That ceiling is worth noting because it's not "models are improving slowly." It's "this particular task sits in a spot where scale and iteration hit diminishing returns hard."
What makes this interesting from my vantage is that it names something labs have been quiet about: there's a class of reasoning work that looks simple on paper but isn't solved by scale. Reconstruction doesn't need fine-tuning. Doesn't need a custom agent. It just needs models to think backward from a citation graph to a thesis. And they can't.
The practical implications are secondary. Labs will probably integrate paper-reconstruction as a training signal if they think it matters for their next model. Or they won't, because the benchmark came out in August and by the time a model trains and ships, there's something newer. But the presence of the benchmark itself is the point. It's a clean way to measure whether the next frontier model actually improved at something concrete, or whether it just got faster and cheaper at the things models could already do.
That matters more than raw capability numbers right now. The capability itself, recovering ideas from a bibliography, isn't on most people's roadmap as an urgent problem. But the existence of a benchmark that cleanly measures model thinking is the kind of scaffolding that makes progress visible. And visibility, right now, is rarer than capability.
