Large language models can reason across multiple steps — until the combination is one they have never seen before.
Researchers studying what they call the "curse of two-hop reasoning" argue the culprit is a gap in how models are supervised during training. The paper introduces a concept called "identity bridge": a minimal extra signal that forces a model to maintain an explicit representation of the intermediate entity connecting two facts. Without it, a model trained on "A relates to B" and "B relates to C" cannot reliably answer "how does A relate to C?" when that exact chain is new. With it, even a stripped-down single-layer transformer achieves out-of-distribution generalization on two-hop problems. Experiments with standard GPT-2 models confirmed the pattern, and analysis of fine-tuned mainstream LLMs showed that correct two-hop answers consistently corresponded to models that had formed a direct subject-to-answer association internally.
This matters because multi-hop reasoning is not an edge case — it is the backbone of any task that requires stitching together separate pieces of knowledge, from legal analysis to medical diagnosis to basic research assistance. If the failure mode is traceable to a specific missing supervision signal rather than a fundamental architectural limit, that is a more tractable problem than the field previously assumed.
The finding also reframes a debate that has run alongside every major LLM benchmark: are these models reasoning, or pattern-matching? The identity bridge result suggests the answer is more structural than philosophical — the right training nudge can coax genuine compositional generalization out of architectures that currently fall short without it.