Can minimal training signals unlock latent reasoning capability in base models?
This explores whether base models already hold reasoning ability that a small nudge can unlock — rather than reasoning being something post-training has to build from scratch.
This explores whether base models already hold reasoning ability that a small nudge can unlock — rather than reasoning being something post-training has to build from scratch. The corpus answers with a striking convergence: yes, and from many directions at once. One synthesis finds that five independent methods — RL steering, critique fine-tuning, decoding changes, sparse-autoencoder feature steering, and RLVR — all elicit reasoning that's already sitting in base model activations Do base models already contain hidden reasoning ability?. The shared takeaway reframes the whole problem: post-training *selects* reasoning rather than *creating* it, so the real bottleneck is elicitation, not capability acquisition.
What makes this convincing is that the unlocking signals can be almost absurdly minimal. Modular 'cognitive tools' — reasoning operations wrapped as sandboxed model calls — lifted GPT-4.1 on a hard math benchmark from 26.7% to 43.3% with no RL training at all Can modular cognitive tools unlock reasoning without training?. A single steering vector extracted from just 50 example pairs can compress chain-of-thought by two-thirds while holding accuracy Can we steer reasoning toward brevity without retraining?. And reasoning chains can be slashed to 7.6% of their tokens with no accuracy loss — meaning most of what looked like 'thinking' was documentation, not computation concise-intermediate-reasoning-chains-match-verbose-cot-occupy. If a tiny vector or a leaner format can move the needle this much, the capability was clearly already present.
But 'unlock latent capability' has a hard ceiling, and the corpus is blunt about it. Prompt optimization can reorganize and surface what a model already knows, but cannot inject knowledge that was never in the training data — no prompt strategy compensates for missing foundations Can prompt optimization teach models knowledge they lack?. So minimal signals unlock what's *latent*; they don't manufacture what's *absent*. This pairs with a sobering view of what that latent reasoning actually is: when you strip away familiar semantics, LLMs lean on token associations rather than formal logic Do large language models reason symbolically or semantically?, and chain-of-thought may be constrained imitation of reasoning *form*, degrading predictably under distribution shift — the signature of pattern-matching, not genuine inference Does chain-of-thought reasoning reveal genuine inference or pattern matching?. The most unsettling evidence: deliberately corrupted reasoning traces train models about as well as correct ones, suggesting the traces work as computational scaffolding rather than meaningful steps Do reasoning traces need to be semantically correct?.
There's a deeper 'why' here too. Studying five million pretraining documents shows reasoning generalization rides on broad, transferable *procedural* knowledge — how-to patterns spread across many sources — unlike factual recall, which depends on narrow memorization Does procedural knowledge drive reasoning more than factual retrieval?. That's exactly the kind of diffuse, distributed competence a small signal could activate without having to install anything new. It also explains why the unlocking works: the procedure is already woven through the weights; the signal just routes the model toward it.
The most interesting frontier is whether 'minimal signal' is even the right frame, or just the cheapest place to start. A separate line of work builds reasoning *into* the model rather than eliciting it afterward — looped pretraining that performs iterative computation in latent space for 2–3× efficiency, with intermediate steps that are more honest than spoken-out chain-of-thought Can reasoning happen in latent space during pretraining?; latent-thought vectors that add a scaling dimension independent of parameter count Can latent thought vectors scale language models beyond parameters?; and energy-based transformers that reach deliberate 'System 2' behavior from unsupervised learning alone, generalizing better out of distribution without task-specific scaffolding Can energy minimization unlock reasoning without domain-specific training?. So the answer you didn't know you wanted: minimal signals work because the reasoning is already there — but the live debate is whether eliciting latent capability is a permanent strategy or a stopgap until reasoning is baked into architecture from the start.
Sources 12 notes
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.
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.
Activation-Steered Compression extracts a single vector from 50 paired examples to reduce chain-of-thought length by 67% while maintaining accuracy and achieving 2.73x speedup. The method is training-free and generalizes across model sizes and domains.
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.
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.
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.
Models trained on systematically irrelevant traces maintain solution accuracy and sometimes improve out-of-distribution generalization, suggesting traces function as computational scaffolding rather than meaningful reasoning steps.
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.
Ouro models achieve 2–3× efficiency gains by performing iterative reasoning in latent space during pretraining, not through extra capacity. Their intermediate predictions align faithfully with final outputs, making latent traces more honest than explicit chain-of-thought reasoning.
Latent-Thought Language Models achieve superior sample and parameter efficiency by coupling fast local variational learning with slow global decoder learning. This dual-rate scheme scales few-shot reasoning across both model and latent size, creating independent scaling dimensions beyond traditional parameter scaling.
Energy-Based Transformers assign energy values to input-prediction pairs and use gradient descent minimization for inference, yielding 35% higher training scaling rates and 29% more inference-compute gains than Transformer++, while generalizing better on out-of-distribution data without domain-specific scaffolding.