Does training data format shape model reasoning more than domain content?
This explores whether *how* training data is presented (its format — multiple-choice, free-form, structured) shapes the way a model reasons more than *what* the data is about (its domain or subject matter), and the corpus comes down firmly on the side of format.
This explores whether the shape of training data — its format — does more to determine how a model reasons than the actual subject matter it's trained on. The most direct answer in the collection is striking: one study finds that format shapes a model's reasoning *strategy* about 7.5 times more strongly than domain does. Models trained on multiple-choice data learn to explore broadly (breadth-first), while models trained on free-form data learn to dig deep along one path (depth-first) — and that difference holds regardless of whether the content is math, law, or science Does training data format shape reasoning strategy more than domain?. Presentation, not topic, sets the reasoning habit.
What makes this more than a one-paper curiosity is that several other notes, coming at it from totally different angles, point the same direction. One line of work shows that reinforcement-learning post-training doesn't invent a reasoning style — it *amplifies one format that was already latent in pretraining* and quietly suppresses the alternatives, often within the first epoch, with the winner determined by model scale rather than by which format performs best Does RL training collapse format diversity in pretrained models?. So format isn't just an input you choose; it's a deep structural attractor the model collapses onto. Relatedly, chain-of-thought reasoning turns out to be *distribution-bound*: it degrades predictably when you shift the task, the length, or the format away from what the model saw in training — producing fluent text that imitates the *form* of reasoning without the underlying logic Does chain-of-thought reasoning actually generalize beyond training data?. Reasoning, in other words, is bolted to the formats it was practiced in.
But the corpus also resists the tidy headline, and that's where it gets interesting. A separate strand argues that what really transfers across domains is *procedural knowledge* — the broad, reusable "how to do this kind of problem" patterns drawn from diverse documents — as opposed to factual recall, which depends on narrow, document-specific memorization Does procedural knowledge drive reasoning more than factual retrieval?. That suggests the deeper axis isn't quite "format vs. domain" but "transferable procedure vs. pinned-down content," and format may matter so much precisely because it encodes procedure. Pushing further, StructTuning shows that *organizing* domain knowledge into a taxonomy — teaching the model where a fact sits in a conceptual structure — reaches 50% of full-corpus performance using just 0.3% of the data Can organizing knowledge structures beat raw training data volume?. Structure, a cousin of format, beats raw volume of content.
There's a practical edge to all this. Work on small models shows that DPO training — which explicitly shows the model correct *and* incorrect examples — outperforms ordinary fine-tuning specifically because it targets *output-format* failures that plain supervised training keeps getting wrong Can small models match large models on function calling?. And the survey of domain-adaptation methods warns that the visible wins from domain training often hide costs in exactly the format-flexibility dimension — models get better at the target content while losing their ability to shift formats How do domain training techniques actually reshape model behavior?. So format isn't only what shapes reasoning; it's also what's quietly damaged when you over-optimize for content.
The thing worth walking away with: if reasoning style is something a model *selects* from pretraining rather than acquires from content Do base models already contain hidden reasoning ability?, then the lever that actually moves how a model thinks may be the shape of what you show it — not the subject. Which reframes a lot of 'we need more domain data' instincts as possibly 'we need better-formatted data.'
Sources 8 notes
Models trained on multiple-choice data adopt breadth-first exploration (Cohen's d up to 1.5), while free-form training produces depth-first reasoning. Format effect dwarfs domain effect, meaning presentation matters far more than content type.
Controlled experiments show RL consistently amplifies one format distribution from pretraining within the first epoch while collapsing alternatives. The winning format depends on model scale, not necessarily performance, and is largely hidden when starting from proprietary pretrained models.
DataAlchemy experiments show CoT fails systematically under distributional shifts in task, length, and format. Models produce fluent but logically inconsistent reasoning — imitating reasoning form without valid underlying logic.
Analysis of 5 million pretraining documents shows reasoning relies on broad, transferable procedural knowledge from diverse sources, unlike factual recall which depends on narrow, document-specific memorization of target facts.
StructTuning achieves 50% of full-corpus performance using only 0.3% of training data by organizing chunks into auto-generated domain taxonomies. The model learns knowledge position within conceptual structures rather than raw text patterns, matching how students learn from textbooks.
Small models fine-tuned via DPO on correct and incorrect function-calling examples from a large teacher model achieve high accuracy on logical and mathematical tasks. DPO's explicit negative examples directly target the rigid output format failures where SFT alone underperforms.
Research shows every adaptation method—from parameter-efficient tuning to knowledge graph curricula—has optimal conditions tied to specific domains. The key finding: visible benefits like performance gains often come with hidden degradation in reasoning faithfulness, capability transfer, and format flexibility.
Five independent mechanisms—RL steering, critique fine-tuning, decoding changes, SAE feature steering, and RLVR—all elicit reasoning already present in base model activations. Post-training selects rather than creates reasoning; the bottleneck is elicitation, not capability acquisition.