Next-Chunk Reasoning: Why RL Might Not Actually Beat SFT for no-CoT Data

The transition from standard Supervised Fine-Tuning (SFT) to Reinforcement Learning (RL) has been hailed as the primary catalyst for the reasoning capabilities observed in recent frontier models. The general consensus in the engineering community has been that RL, particularly through methods like DeepSeek's GRPO or OpenAI's internal o1-training recipes, is strictly superior for teaching models "how to think" before they answer. However, new research titled Is Next-Chunk Reasoning RL Really Better than SFT? challenges this assumption, specifically when dealing with "no-CoT" data—datasets that contain reasoning-rich content but lack explicit Chain-of-Thought (CoT) steps.

The Challenge of no-CoT Data

Large-scale pre-training data is abundant, but high-quality reasoning data—where every logical step is articulated—is surprisingly scarce. Most technical manuals, scientific papers, and mathematical proofs (the "no-CoT" data) provide the problem statement and the final answer or a high-level derivation, skipping the granular "internal monologue" that modern reasoning models rely on.

Historically, the solution has been either to generate synthetic CoT using stronger models or to use RL to reward the model for reaching the correct final answer. The hypothesis was that RL would naturally incentivize the model to develop its own internal reasoning path. The recent work by researchers investigating "Next-Chunk Reasoning" suggests that the gap between RL and SFT in these scenarios may be narrower than previously thought, and in some cases, SFT actually maintains better data efficiency.

How Next-Chunk Reasoning Works

In a standard autoregressive setup, the model predicts the next token. In "Next-Chunk Reasoning," the model is trained to predict a sequence of tokens (a "chunk") that represents a logical leap. When training on no-CoT data, the goal is to bridge the gap between the premise and the conclusion.

The paper compares two primary training strategies:

SFT-based Next-Chunk: Predicting the next logical chunk directly from the input using cross-entropy loss.
RL-based Next-Chunk: Treating the chunk generation as a policy and rewarding the model based on the correctness or utility of the generated chunk in reaching the final answer.

The critical finding is that while RL is theoretically more flexible—allowing the model to explore arbitrary reasoning paths—it suffers from high variance and sparse rewards when the "chunks" are complex. SFT, despite being more rigid, provides a denser signal that helps the model learn the underlying structure of the data more effectively in the early stages of training.

The Trade-off: Exploration vs. Signal Density

In RL-based reasoning training (like the approach used in DeepSeek-R1), the model is rewarded for finding any path to the correct answer. This discovery process is what leads to emergent reasoning traits. However, for "no-CoT" data, the model effectively has to invent the missing links. If the reward signal is only "correct/incorrect" at the very end of the sequence, the credit assignment problem becomes intractable for long-horizon tasks.

The researchers found that SFT can act as a powerful regularization. By mimicking the "chunks" present in high-quality technical datasets, the model learns a distribution of "logical leaps" that are more likely to be correct. This is similar to how Quiet-STaR encourages models to think at every token, but at a more macro, chunk-based level.

Test-Time Compute and Inference Implications

One of the most significant implications of this study involves test-time scaling. Models like OpenAI's o1 utilize extra compute during inference to refine their reasoning. The effectiveness of this test-time compute is directly tied to how the model was trained.

If a model is trained exclusively via RL on sparse rewards, its reasoning traces can become idiosyncratic or "hacky," optimizing for the reward function rather than logical soundness. SFT-trained models tend to produce more conventional, human-like reasoning paths, which are often more robust when scaled with search algorithms like Monte Carlo Tree Search (MCTS) or Best-of-N sampling at test time.

Conclusion for Machine Learning Engineers

For engineers building specialized agents or fine-tuning models for niche technical domains, the takeaway is clear: do not discard SFT in favor of RL too early. While RL is essential for final alignment and maximizing performance on benchmarks with clear reward functions, SFT remains the bedrock for learning the fundamental architecture of reasoning, especially when your data doesn't have the luxury of explicit step-by-step labels.

Next-chunk reasoning highlights that we are still in the early days of understanding the optimal curriculum for intelligence. The debate isn't just about RL vs. SFT; it's about how we can best extract the latent logic trapped in the world's technical literature.

Sources:

Tags: ai, machinelearning, deeplearning, llm