RL Post-Training Builds Compositional Reasoning Strategies
This research investigates whether reinforcement learning (RL) post-training simply amplifies skills already present in a base model or if it enables the model to synthesize primitive skills into new, higher-level reasoning strategies. By using a fully observable "rewrite-grammar" environment—where every step of a model's reasoning can be audited against a known set of rules—the authors provide a clear look at how models evolve during training. They find that RL does not just reweight existing behaviors; it actively reorganizes primitive competence into reliable, reusable procedures that allow the model to solve problems that were previously out of reach.
A Controlled Environment for Reasoning
To study how reasoning emerges, the authors created a testbed where a Transformer model is pretrained on primitive symbol-rewrite chains. After pretraining, the model is tasked with solving "contraction" problems, where it must reduce a sequence of symbols to a target goal. Because the grammar is fully known, the researchers can classify every step the model takes as either a primitive rule, a valid composed strategy (a "macro" or "parallel" contraction), or an invalid "spurious" rewrite. This allows them to track exactly how the model’s internal strategy changes as it learns to solve increasingly difficult problems.
Phased Procedural Chunking
The study reveals that RL post-training follows a distinct, phased progression. Early in training, the model focuses on strengthening its grasp of primitive, step-by-step reductions. As training continues, it begins to discover and consolidate "composed procedures." These include sequential compositions, which collapse multiple steps into one, and parallel compositions, which perform independent operations simultaneously. These are not just one-off lucky guesses; the model consolidates these shortcuts into a stable, reusable repertoire, allowing it to solve complex problems that exceed the generation budget of the base model.
Selectivity Over Exploration
A key finding is that RL’s success is not simply due to exploring more possibilities than other methods like rejection fine-tuning (RFT). Both RL and RFT generate a high volume of non-primitive shortcuts. However, RFT often produces many invalid, "spurious" rewrites that it then clones during training. In contrast, RL uses a "same-prompt contrast" mechanism—inherent to the Group Relative Policy Optimization (GRPO) algorithm—to distinguish between successful and failed attempts. This allows RL to be highly selective, suppressing invalid shortcuts while concentrating its exploration into valid, reusable structures.
Pretraining as a Foundation
The researchers also highlight that the emergence of these compositional strategies is not guaranteed by exposure to primitive rules alone. Instead, the ability to form higher-level strategies depends on how the pretraining data organizes primitive competence. If the base model is trained in a way that organizes primitives into logical reduction procedures, RL can effectively "compress" these into higher-level strategies later. Ultimately, the base model provides the raw procedural ingredients, and RL acts as the mechanism that builds them into reliable, sophisticated reasoning strategies.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!