A robot policy can identify the right objects or predict a plausible next frame and still lack enough information to act. Recognition does not guarantee precise geometry, a usable interaction pose or a representation of the destination named by the instruction. The GIFT paper, first submitted on September 3, 2026, calls this the action-sufficiency gap.
This is a useful debugging frame for both vision-language-action models and world-action models. Ask what the action head can recover from its intermediate features, rather than treating a correct scene description or realistic forecast as proof that the representation supports control. For the broader architecture distinction, see the VLA versus VLM guide.
Diagnose the missing information before changing the policy
Consider an illustrative instruction: “Put the test tube into the requested rack slot.” A model may recognize the tube and rack, and a world model may forecast a credible insertion. The action can still fail in three different ways:
- Geometry is missing. The features do not preserve the tube-slot alignment or approach direction accurately enough.
- The interaction is ambiguous. The policy confuses the manipulated object, reference object or suitable gripper configuration.
- The goal is underspecified. The instruction selects one slot, while the representation treats the whole rack as the destination.
This hypothetical example suggests three probes for an actual failure. Change the camera or object pose while keeping the instruction fixed to probe spatial sensitivity. Add a similar distractor to probe interaction roles. Keep the objects fixed while changing only the named destination to probe goal grounding. These tests can narrow a failure pattern; a failed rollout alone cannot prove which internal feature caused it.
The same probes help distinguish architecture changes from supervision changes. A larger vision encoder may retain more detail without preserving the detail the controller needs. An additional conditioning path may improve one benchmark while making the deployed policy depend on another predicted feature. The experiment should isolate whether the shared representation improved, whether the action interface changed, or both.
GIFT turns the three probes into training objectives
GIFT means Guided Intermediate Feature Training. Its method trains intermediate visual features against three targets: a frozen geometry teacher, object-centric interaction configurations, and instruction-conditioned goal regions. In the default no-injection design, auxiliary predictions shape features through training losses but are excluded from action inputs. The teacher and auxiliary heads are unnecessary at deployment.
The paper applies this supervision to a StarVLA-OFT-based VLA and two Fast-WAM variants. For the underlying distinction, StarVLA organizes visual-language features and action modules in a modular policy framework. Fast-WAM distinguishes directly producing actions from generating a future video and recovering actions through inverse dynamics. Training with video prediction does not necessarily require generating that video during deployment.
One paired result illustrates the evidence: on LIBERO-Plus, GIFT-WAM-Fast rises from its baseline’s 60.0% to 72.6% success. Both train on original LIBERO data without adapting to the perturbed benchmark; each perturbed instance is evaluated once. The VLA comparison also changes its action head, so the paper’s same-head, no-guidance control matters when attributing gains. These are author-run results, not independent reproduction.
Turn action sufficiency into an acceptance test
“Action-sufficient” should become a recorded test rather than a model label:
- Define the failed action decision, including the observation, instruction and required contact or motion.
- Identify whether geometry, interaction role or instruction-specific goal information was absent, unstable or ignored by the action head.
- Create one controlled perturbation for that factor while holding the others as constant as possible.
- Compare the original policy with a supervision-only variant using the same action route, data split and evaluation budget.
- If an auxiliary feature is injected at inference, test it separately and log failures in the auxiliary prediction as well as the final action.
- Review per-shift results and repeated trials. An average can hide a regression under the exact change that triggered the original failure.
Keep a record of what each probe actually changes. Moving a camera can alter visibility as well as geometry; adding a distractor can change occlusion as well as object selection. Inspect auxiliary predictions alongside the final action, and retain ambiguous cases. A cleaner error description is useful even when the available evidence does not yet identify a cause.
For the evaluation record itself, the robot VLA evaluation guide provides a broader framework for scenarios, intervention and repeated trials. The GIFT-specific addition is to bind each scenario to the information the action pathway needs, then keep the supervision change separate from an action-head or inference-interface change.
When this intervention is worth testing
GIFT’s supervision pattern is a reasonable candidate when a policy already has adequate semantic recognition, failures track viewpoint, distractors or target-region changes, and training can expose the required geometry and interaction labels. It is less informative when failures come from latency, calibration, action scaling, control frequency or hardware limits. Those problems can produce similar endpoint failures without being representation failures.
It also carries a labeling and teacher-dependence cost. Before adopting it, record how geometry targets, entity roles and goal masks are produced; how noisy they are; and whether their definitions match the target task. A clean auxiliary loss built from the wrong interaction role can make the representation more consistent and still make the robot choose the wrong action.
As checked on September 27, 2026, the official GIFT project page points to a website repository, whose inspected tree contains no training code or model weights. This finding is limited to that linked repository. No model, simulator or robot execution was performed for this article.
The practical lesson is the diagnostic split. A policy can know what is in view while failing to preserve where, how and which destination well enough for action. GIFT supplies one training intervention and paired evidence for that hypothesis. Adoption still requires released artifacts, independent execution and a test built around the failure that matters on the target system.