Models & learning
Inference-time guidance
Inference-time guidance steers a generative model's sampling process towards an additional condition or objective without updating the model's trained parameters. The guidance may modify a score, noise estimate, vector field, token probabilities, or candidate selection during generation; its guarantees and cost depend on the objective and sampler.
Also known as: test-time guidance, sampling guidance, inference guidance
Updated
Guidance changes sampling rather than training
A trained generative model defines how samples are produced from noise, masks, or another prior. Inference-time guidance adds information while that production process is running. The base model weights remain fixed, but each sampling step can be biased towards a class, text condition, geometric constraint, reward, or other differentiable objective.
Classifier guidance adds a classifier gradient to a diffusion model's score so that samples move towards a requested class. Classifier-free guidance combines conditional and unconditional model predictions instead of using a separate classifier. These are influential mechanisms, not the only forms of guidance.
The guided quantity depends on the generator
Guidance has to match the sampler's representation. A diffusion sampler may alter a noise or score prediction. A flow-matching model can modify its velocity field. A token generator can reweight next-token probabilities. Candidate ranking after generation is sometimes described as test-time optimisation, but it does not change how an individual sample forms.
This term is broader than prompt conditioning. A prompt is normally part of the model's learned input. Guidance adds a sampling-time signal, often with a tunable strength and extra computation.
Robot constraints can guide a motion generator
EmbodyHOI decodes wrist trajectories while a frozen hand–object motion generator is sampling. It runs a differentiable inverse-kinematics solve on the target humanoid and adds the IK-energy gradient to the hand-token velocity field. Later hand and object tokens can then form around the corrected wrists.
That is different from repairing the wrist after the whole interaction has been generated. Post-hoc repair can satisfy the geometric objective while breaking the relationship between the wrist, fingers, and object.
Guidance does not create a general guarantee
The guidance objective defines what is improved. Class guidance can strengthen class features without ensuring factual or physical correctness. Kinematic guidance can make a wrist more reachable without modelling contact force, stability, collision along the path, or actuator limits.
Strong guidance can also reduce diversity, distort samples, or increase compute. Validation should compare the guided output with an unguided baseline on the intended downstream task, preserve the sampling seed and guidance settings, and report failures rather than treating a lower guidance loss as sufficient evidence.
Sources
Related terms
Models & learning
Flow matching
Flow matching is a generative-model training method that learns a time-dependent vector field transporting samples from a simple prior distribution to a data distribution. Training regresses the vector field along chosen probability paths without solving the flow during each update; generation starts from the prior and numerically integrates the learned ordinary differential equation.
Models & learning
Diffusion policy
A diffusion policy is a learned robot policy that generates an action or action sequence through iterative denoising conditioned on observations and, where applicable, a task goal. This formulation can represent several plausible behaviours for the same scene and is commonly trained by imitation on time-aligned trajectories of observations and expert actions.
Models & learning
Reward model
A reward model is a learned function that estimates task progress, success, preference, or another reward signal from robot observations or trajectories. It can replace or supplement a hand-written reward when training or evaluating a policy, but its output is only as reliable as its labels, coverage, and resistance to shortcuts.
Hardware & control
Inverse kinematics
Inverse kinematics (IK) finds robot configurations that satisfy a desired position, orientation or other geometric constraints. It reverses the question asked by forward kinematics: instead of computing where a hand or foot is from the joint values, it searches for joint values that place it at a target. A target can have multiple solutions or no feasible solution.
Models & learning
Robot foundation model
A robot foundation model is a broadly pretrained model intended to provide a reusable starting point for multiple robot tasks, environments or embodiments. It learns from diverse robotics and sometimes web or human data, then acts directly or is adapted with target-domain data. The term describes a training and reuse strategy, not one fixed architecture.