A new paper on arXiv proposes a self-improvement loop for reasoning models, built on a simple observation: even when a problem is too hard for a model to solve on its own, an additionally supplied solution might still enable learning. The authors frame this as a way to train models by looking back at problems they initially failed, rather than only on successes.
The method, described in the abstract as "Hindsight Hierarchies," appears to use these supplied solutions as a form of hindsight feedback. Because the source is only a brief abstract, the exact mechanism is not fully detailed, but the core idea is that the model can improve by revisiting difficult problems with the benefit of a provided answer.
The significance lies in extending training beyond a model's current ability. Instead of discarding problems that exceed its grasp, the loop turns them into teaching signals. The paper does not compare with other approaches in the available text, so this summary reflects only the authors' stated contribution.