How do compact latent dynamics enable planning without explicit chain of thought?
This explores how reasoning that happens inside a model's hidden states — rather than spelled out as text — can still produce planning and problem-solving, and what the corpus knows about why latent reasoning works at all.
This explores how reasoning that lives in a model's internal representations, rather than in written-out chain-of-thought tokens, can still carry planning — and the corpus turns out to have a surprisingly direct answer: the reasoning was already in there, and chain-of-thought is just one way to surface it. The most striking evidence is that you can trigger reasoning by steering a single internal feature. Can we trigger reasoning without explicit chain-of-thought prompts? shows that nudging one sparse-autoencoder-identified direction matches or beats explicit chain-of-thought across six model families — and it activates early, overriding surface instructions. That reframes the whole question: latent reasoning isn't a workaround for missing CoT, it's the underlying capability that CoT happens to invoke.
If reasoning is a direction in activation space, it's also compressible and controllable. Can we steer reasoning toward brevity without retraining? finds that the difference between rambling and terse reasoning is a single linear vector extracted from 50 examples — cut verbosity by 67% with no accuracy loss. So the 'compact' in compact latent dynamics is literal: the same plan can occupy far fewer tokens, or none, because the trajectory through hidden space is what's doing the work. Can energy minimization unlock reasoning without domain-specific training? pushes this further, replacing token-by-token generation with gradient descent on an energy landscape — deliberation becomes an optimization over internal states rather than a written transcript, and it generalizes better out of distribution.
Where planning specifically comes in, two complementary moves matter: making latent dynamics stochastic, and conditioning them on the future. Can stochastic latent reasoning help models explore multiple solutions? replaces deterministic latent updates with sampling, so a recursive reasoner can hold a distribution over solutions instead of committing to one path — exactly what you need for ambiguous problems with several valid strategies. Its sibling result Can reasoning systems scale wider instead of only deeper? shows those stochastic latents let you scale reasoning 'wider' — sampling many parallel internal trajectories — instead of only deeper, sidestepping the latency cost of long serial chains. Planning here looks like exploring a branching space of latent futures, none of it written down.
The other route to planning-without-CoT is to teach the model to anticipate. Can embedding future information in training data improve planning? bakes future-information tokens into training data so a standard model learns goal-conditioned generation — better planning and algorithmic reasoning with no architectural change. And Can modular cognitive tools unlock reasoning without training? makes explicit what ties all of this together: the modular-tools paper lifted GPT-4.1 on AIME from 27% to 43% with no training at all, because the reasoning was latent and just needed structured isolation to come out. The thread across the corpus is consistent — capable models already contain the plans; compact latent dynamics are about steering, sampling, and conditioning those internal states rather than transcribing them into text.
Sources 7 notes
SAE-identified reasoning features can be directly steered to match or exceed chain-of-thought performance across six model families. This reasoning mode activates early in generation and overrides surface-level instructions, suggesting latent reasoning is a fundamental capability independent of explicit prompting.
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.
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.
GRAM replaces deterministic latent updates with stochastic sampling, enabling models to represent distributions over solutions rather than single predictions. This allows handling of ambiguous problems and multiple valid strategies that deterministic designs cannot represent.
GRAM shows that stochastic latent transitions enabling parallel trajectory sampling sidestep the serial latency cost of depth-only scaling. Width matches token-level parallelism benefits: independent paths sample the solution space without variance inflation.
TRELAWNEY augments training data with special tokens encapsulating future information, allowing models to learn goal-conditioned generation using standard infrastructure. Results show improved planning, algorithmic reasoning, and story generation without modifying architecture or training procedures.
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.