A new benchmark called Reconstruction, published this month, exposes a hard limit in how frontier language models handle research-level reasoning. The test is simple in concept but ruthless in execution: given only a bibliography from a published research paper, can a model recover the core ideas?
The results are blunt. Frontier LLMs managed 3-15% accuracy on this task. A multi-agent setup running a Swiss-tournament pipeline, agents competing and querying each other, barely budged that needle.
This matters because recovering ideas from citations is a core research skill. Researchers read a reference list and infer what problems someone solved, what method they used, why it mattered. It's the difference between reading papers mechanically and understanding them. If frontier models can't do this, they're not ready for the kinds of reasoning tasks that actual research requires.
The benchmark doesn't just measure raw knowledge. It measures something closer to causal reasoning, the ability to infer intent and contribution from indirect signals. A model that knows every paper ever written might still fail this task. You need to understand the relationship between citations and claims.
The most interesting detail is what happens when you throw multi-agent architecture at the problem. Usually, having multiple models reason in parallel and compete produces better outputs. Here it didn't move the dial much. That suggests the bottleneck isn't coordination or debate. It's something deeper: the models lack the reasoning structure needed to make those inferences in the first place.
This is the kind of benchmark that matters precisely because it's not flashy. GPQA scores and SWE-Bench rankings get attention. A test that says "your model is mediocre at understanding what researchers actually do" doesn't make headlines. But it tells you something real about how much work remains.
