AI
Study reveals how AI models fake reasoning when given the answer first
Researchers found that training AI to generate reasoning from known answers creates a fundamental mismatch when the model must reason without that information at test time.
When machine learning systems learn to produce reasoning traces backwards—starting from a known answer and working towards the question—they can fall into a trap: justifying the answer rather than genuinely deriving it. A new study from researchers at arXiv examines this "post-hoc rationalization" problem and proposes a solution.
The Core Problem#
In Reverse Chain-of-Thought Generation (RCG), models are trained to synthesize step-by-step reasoning by working backward from query-answer pairs. This creates an inherent disconnect: during training, the model sees the answer and uses it to guide reasoning construction, but at inference time, it must generate reasoning without any answer available. This creates what researchers call a train-inference mismatch.
The team quantified this gap across three dimensions: lexical anchoring (how much of the answer text appears in the reasoning), trajectory anchoring (how the answer shapes which reasoning steps get generated), and probabilistic anchoring (how much the answer could be recovered from reading just the trace). All three reveal deep entanglement between the final answer and the supposedly independent reasoning path.
Why Suppression Backfires#
One intuitive fix—preventing answer vocabulary from appearing in reasoning steps—seemed promising but produced an unexpected result. By filtering out answer-related language, the model was forced to constantly track and work around the answer internally. This actually strengthened the answer's influence on reasoning dynamics, even as surface-level similarity decreased.
A Structural Alternative#
The researchers propose Structural Skeleton-guided Reasoning (SSR), which replaces suppression with structural separation. The approach operates in two phases: first generating a high-level abstract scaffold of the reasoning task, then using that skeleton to guide the full step-by-step trace. This decoupling—rather than mere vocabulary avoidance—reduces all three forms of answer dependence simultaneously.
Across both standard benchmarks and out-of-distribution test cases, the distilled variant (SSR-D) outperformed suppression-based baselines by up to 10 percentage points. Critically, it also better preserved performance when models faced novel problem distributions, suggesting more genuinely transferable reasoning had been learned.
