How does cross-domain reasoning transfer differ from domain-specific knowledge transfer?
This explores the difference between transferring *reasoning skills* that carry across domains (the how-to-think layer) and transferring *domain-specific knowledge* that's tied to particular facts and fields (the what-you-know layer) — and why the corpus treats these as two different machines with two different failure modes.
This explores the difference between transferring reasoning skills that travel across domains and transferring domain-specific knowledge bound to particular facts. The corpus draws a surprisingly clean line between the two — and the cleanest place to see it is a finding that they physically live in different parts of the network. One analysis locates knowledge retrieval in the *lower* layers of an LLM and reasoning adjustment in the *higher* layers Why does reasoning training help math but hurt medical tasks?. That separation isn't just architectural trivia: it explains why training a model harder on reasoning improves math but can actively *degrade* knowledge-heavy fields like medicine. Reasoning and knowledge are not the same substance, so improving one can erode the other.
The deeper reason they behave differently shows up in what each one depends on. Reasoning generalizes because it rides on *procedural* knowledge — broad, transferable patterns drawn from many diverse documents — whereas factual recall depends on narrow, document-specific memorization of the exact target fact Does procedural knowledge drive reasoning more than factual retrieval?. That's the mechanism behind cross-domain transfer: a procedure learned in one place is reusable elsewhere, but a memorized fact is not. The same intuition powers the finding that reconstructing experts' *hidden* thought processes — the self-talk, recall, and verification beneath a finished expert text — produces reasoning that transfers across domains, while training only on the polished surface text does not Can reconstructing expert thinking improve reasoning transfer?. You transfer reasoning by capturing process; you transfer knowledge by capturing content.
But cross-domain reasoning transfer has a hard limit the corpus is blunt about: it's bounded by the training distribution. Chain-of-thought reasoning degrades *predictably* under shifts in task, length, or format — models keep producing fluent reasoning that has quietly stopped being valid Does chain-of-thought reasoning actually generalize beyond training data?. So 'transferable' reasoning is really 'transferable within a neighborhood,' not unbounded. And on the knowledge side there's an even harder wall: prompting can only *activate* knowledge already in the model — no prompt strategy injects facts the model never learned Can prompt optimization teach models knowledge they lack?. Knowledge transfer therefore demands actually changing the weights; reasoning transfer can sometimes be coaxed out of what's latent.
This is why the two need different *methods* to install them. For domain knowledge, the corpus favors approaches that restructure what the model holds: knowledge-graph curricula that compose primitives into domain expertise Can knowledge graphs teach models deep domain expertise?, and RL-from-augmented-generation that internalizes coherent knowledge structures better than supervised fine-tuning because it rewards rational explanation, not just token-correct answers Can reinforcement learning embed domain knowledge more effectively than supervised fine-tuning?. For reasoning, the corpus leans on emergence: complex domain reasoning arises from RL on hard problems with only simple accuracy rewards Can simple rewards alone teach complex domain reasoning?, and reasoning competence can even emerge as a side effect of ordinary language modeling when models learn to generate rationales token-by-token on arbitrary text Can models learn reasoning from predicting any text?.
The payoff for a curious reader is the catch the corpus keeps circling back to: these two kinds of transfer *interfere* with each other. Every domain-adaptation method has a domain-conditional sweet spot, and the visible win — a benchmark bump — often hides degradation in reasoning faithfulness, capability transfer, or format flexibility How do domain training techniques actually reshape model behavior?. So the real question isn't 'reasoning transfer *or* knowledge transfer' — it's that pushing hard on either one tends to quietly tax the other, and most training recipes are silent about which side they're spending down.
Sources 10 notes
Two-phase inference model shows knowledge retrieval operates in lower network layers while reasoning adjustment happens in higher layers. This separation explains why reasoning training improves math but can degrade knowledge-intensive domains like medicine.
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.
Training on expert texts augmented with reconstructed thought processes (self-talk, knowledge recall, verification) produces reasoning skills that transfer across domains and adapt depth to problem difficulty, outperforming standard continual pretraining by up to 8 points on hard problems.
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.
Prompting works entirely within a model's pre-existing training distribution and cannot supply domain knowledge absent from training data. This creates a hard ceiling: no prompt strategy can compensate for missing foundational knowledge, only reorganize what already exists.
Fine-tuning a 32B model on 24,000 reasoning tasks derived from medical knowledge graph paths produces state-of-the-art performance across 15 medical domains, demonstrating that structured knowledge composition matters more than scale.
RLAG rewards both answer accuracy and explanation rationality by cycling between augmented and unaugmented generation, progressively internalizing coherent knowledge structures. This outperforms SFT because it prioritizes reasoning quality over token-level correctness.
Medical AI systems and o3 demonstrate that sophisticated domain reasoning emerges naturally from RL training on difficult problems with only basic accuracy signals, without requiring explicit chain-of-thought distillation from teacher models.
Quiet-STaR trains language models to generate rationales at every token position during pretraining on arbitrary internet text, enabling general reasoning without task-specific datasets. Rationale quality is judged by predictive accuracy rather than labeled correctness, allowing reasoning competence to emerge as a side effect of improved language modeling.
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.