What kinds of reasoning tasks reveal the ceiling of text-only training?
This explores which reasoning tasks expose the limits of training on text alone — and, crucially, whether those limits are a real ceiling on reasoning or an artifact of how we test and how language itself works.
This explores what kinds of tasks reveal the ceiling of text-only training — and the first thing the corpus does is split the question in two: is the ceiling in the model's reasoning, or in the text-only setup we're measuring it through? Several notes argue the dramatic 'reasoning cliffs' are largely illusions of evaluation. When models are confined to generating answers as text, they fail at long multi-step procedures even when they demonstrably know the algorithm — give them tool access and the cliff disappears Does the reasoning cliff depend on how we test models? Are reasoning model collapses really failures of reasoning?. The tasks that expose this are the ones demanding sustained procedural execution at scale: arithmetic over many steps, simulating a process, anything where you must hold and grind through state. The bottleneck there is execution bandwidth, not intelligence.
But a second cluster of notes points at a ceiling that tools can't fix. Physical, geometric, and causal reasoning tasks break because text itself is a lossy abstraction — it strips out the physics and spatial dynamics of the world, so a text-only model is manipulating symbols cut off from what they refer to Are text-only language models fundamentally limited by abstraction?. This is the genuinely architectural ceiling: not 'the model ran out of room to compute' but 'the training signal never contained the information.'
The sharpest diagnostic tasks are the ones that decouple *form* from *logic*. When you keep the reasoning structure but swap out familiar semantic content — nonsense tokens, out-of-distribution formats, unusual lengths — accuracy collapses, revealing that chain-of-thought is often reproducing learned reasoning *patterns* rather than performing fresh symbolic inference Does chain-of-thought reasoning reveal genuine inference or pattern matching? Does chain-of-thought reasoning actually generalize beyond training data?. The same point shows up as the symbolic/semantic split: give a model correct logical rules but strip the commonsense semantics, and it can't apply them — it reasons by association, not by formal manipulation Do large language models reason symbolically or semantically?. Even something as mundane as input length is a stress test: reasoning accuracy drops from 92% to 68% with a few thousand tokens of irrelevant padding, far below the context limit, in a way unrelated to language-modeling skill Does reasoning ability actually degrade with longer inputs?.
What makes this interesting is that the corpus also argues the raw capability is usually *already there* and the ceiling is about access, not absence. Reasoning generalizes from broad procedural knowledge spread across pretraining documents rather than memorized facts Does procedural knowledge drive reasoning more than factual retrieval?, and base models turn out to hold latent reasoning that minimal training merely selects and unlocks Do base models already contain hidden reasoning ability?. That reframes the whole question: a lot of what looks like a training ceiling is an *elicitation* ceiling.
If you want to push the boundary rather than just diagnose it, the corpus offers three different escape routes worth knowing about. Quiet-STaR grows reasoning as a side effect of ordinary next-token prediction on arbitrary text, no task-specific data required Can models learn reasoning from predicting any text?; latent-reasoning architectures scale test-time compute inside hidden states without ever verbalizing steps, suggesting the written chain-of-thought was a training artifact all along Can models reason without generating visible thinking tokens?; and Chain of Draft shows that ~92% of a verbose reasoning trace is documentation, not computation — equal accuracy at 7.6% of the tokens Can minimal reasoning chains match full explanations?. Taken together, the 'ceiling of text-only training' isn't one wall — it's an execution limit you can tool around, a grounding limit you can't, and an elicitation limit that was never really a limit at all.
Sources 12 notes
Language models show catastrophic failure in text-only reasoning benchmarks but maintain scaling when given tool access. The cliff reflects execution constraints, not reasoning capability, making text-only evaluations systematically underestimate real-world performance.
Models confined to text-only generation cannot execute multi-step procedures at scale, even when they know the underlying algorithm. Tool-enabled models solve problems beyond the supposed reasoning cliff, suggesting the bottleneck is procedural execution bandwidth.
Text strips the physics, geometry, and causality present in reality, forcing language models to manipulate symbols without grounding in their source dynamics. This creates predictable failure modes in physical, geometric, and causal reasoning that multimodal training could address.
CoT works by constraining models to reproduce familiar reasoning patterns from training, not by enabling novel symbolic reasoning. Performance degrades predictably under distribution shifts—the signature of imitation rather than capability emergence.
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.
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.
FLenQA shows reasoning accuracy drops from 92% to 68% at just 3000 tokens of padding, far below context window capacity. The degradation is task-agnostic, uncorrelated with language modeling performance, and persists even with chain-of-thought prompting.
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.
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.
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.
Multiple architectures—depth-recurrent models, Heima, and Coconut—demonstrate that test-time compute scales through hidden state iteration rather than token generation. This suggests verbalization is a training artifact, not a reasoning requirement.
Chain of Draft achieves equivalent accuracy to standard chain-of-thought on arithmetic, symbolic, and commonsense tasks while using only 7.6% of tokens. The 92.4% of removed tokens served style and documentation, not computation.