A new paper argues today's top reasoning models struggle because they only know how to think in one direction.
Researchers studied why models like GPT-o1, GPT-o3, and DeepSeek-R1 still botch math problems despite generating long chains of thought. Analyzing accuracy across multiple problem sets, they identified five recurring failure patterns: models don't cover enough of the solution space, they make computational mistakes, they run with unverified assumptions, they ignore stated constraints, and they simply hit the response-length limit before finishing. Their proposed fix is a backward reasoning pattern, trained using a two-stage easy-to-hard math dataset that pairs each problem with both a forward and a backward solution path. A fine-grained reward system with smoothed signals is meant to stop the model from gaming the reward, while a gradually shifting sampling schedule keeps forward and backward training examples balanced.
Chain-of-thought models have mostly been improved by making them think longer, not differently - more tokens, more steps, same direction every time. This work treats reasoning direction as a choice the model should make for itself, which is a cheaper lever to pull than just throwing more test-time compute at a problem.
The paper reports improved accuracy and efficiency on math proofs, but it doesn't publish head-to-head numbers against the models it critiques, so the real test is whether this holds up outside the benchmark built to showcase it.