Does training data format shape which reasoning strategies LLMs develop?
This explores whether *how* training data is presented — its shape and format, not its subject matter — steers the kind of reasoning an LLM ends up doing, and the corpus says format is a surprisingly powerful lever.
This explores whether the *format* of training data — multiple-choice vs. free-form, say — shapes which reasoning strategies a model develops, more than the actual subject content does. The most direct answer in the collection is a striking one: format effects can dwarf domain effects. Models trained on multiple-choice data learn to reason in a breadth-first way (scanning options laterally), while free-form training pushes them toward depth-first reasoning that follows one line down — and the format signal turns out to matter roughly 7.5 times more than the topic being taught Does training data format shape reasoning strategy more than domain?. In other words, presentation, not content, sets the cognitive habit.
What makes this more than a curiosity is *why* it happens. A second strand of the corpus argues LLMs reason through semantic association rather than abstract logic — strip the familiar semantics out of a task and performance collapses even when the rules are stated correctly Do large language models reason symbolically or semantically?. If reasoning is fundamentally distributional rather than symbolic, then the *surface shape* of what a model saw during training is exactly the kind of thing it would latch onto. Format isn't a cosmetic wrapper; it's part of the distribution the model learns to imitate.
This connects to a darker cousin of the same mechanism. Fine-tuning on natural language inference data doesn't teach genuine entailment — it amplifies a frequency shortcut, making models lean harder on which words co-occur more often rather than what actually follows from what Does fine-tuning on NLI teach inference or amplify shortcuts?. So the format/data you train on doesn't just nudge a reasoning *style* — it can entrench a reasoning *shortcut* so deeply the model fails worse on cases that contradict the pattern. Format shaping strategy and format entrenching bias are two faces of the same sensitivity to how data is presented.
There's a hopeful flip side worth knowing. If presentation can install a strategy, presentation at inference time can also *elicit* one without any retraining at all: wrapping reasoning operations as isolated, modular 'cognitive tool' calls lifted GPT-4.1's AIME score from 26.7% to 43.3%, purely by enforcing the kind of structured isolation that loose prompting can't guarantee Can modular cognitive tools unlock reasoning without training?. The capability was already latent; structure surfaced it. That suggests the reasoning strategies set by training format may be more about which latent pathways get activated than about new skills being added — a theme echoed by work treating reasoning as hidden-state trajectory formation, where surface chain-of-thought is only a partial window onto the real process Where does LLM reasoning actually happen during generation?.
The thing you didn't know you wanted to know: 'training data' is usually discussed as *what* a model knows, but this collection reframes it as *how* a model thinks. Two models fed the same facts in different shapes can end up with measurably different problem-solving temperaments — and since these models reason by pattern-matching distributions rather than manipulating symbols, the shape of the page may be quietly doing more work than the knowledge on it.
Sources 5 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.
When semantic content is decoupled from reasoning tasks, LLM performance collapses even with correct rules in context. Models rely on parametric commonsense and token associations rather than formal logical manipulation, constraining reasoning to training distribution semantics.
NLI fine-tuning increases LLM reliance on corpus-level frequency patterns (hypernyms more common than hyponyms) rather than semantic relationships. Models perform worse on adversarial cases where frequency patterns contradict actual entailment labels, showing the shortcut was learned more deeply.
Four cognitive tools implemented as sandboxed LLM calls improved GPT-4.1 on AIME2024 from 26.7% to 43.3% without any RL training. Modularity enforces operation isolation that pure prompting cannot guarantee, eliciting pre-existing reasoning capability.
Evidence from CoT faithfulness tests, feature steering, and layer analysis suggests latent-state dynamics drive reasoning, while surface chain-of-thought serves as a partial interface. Hidden reasoning processes should be the default focus of study.