Intention Distillation injects behavior intent into VLA action decoders

Intention Distillation (INDI) uses a frozen teacher VLM to extract behavior-level intent during training and recovers it at an intermediate decoder layer at inference, improving VLA success on SimplerEnv-Bridge from 64.3% to 84.7%, on RoboCasa Kitchen from 64.1% to 70.3%, and in real-world tasks from 62.0% to 68.7%.

Hugging Face Daily Papers — Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models

September 1, 20261 source60% confidence

78Strategic

Strategic score

Evidence strength76

Technical novelty82

Commercial potential64

INDI's new claim is that action decoders should model the semantic objective of forthcoming behavior, not only the demonstrated motor command or concrete future realizations. The constraint moved is the assumption that behavior-level intent can only be supervised from privileged future information: at inference, the policy is trained to recover a teacher-VLM-derived multimodal intent latent from standard inputs at an intermediate decoder layer, so the future-informed training signal does not impose an inference-time sensor or trajectory requirement.¹

Context

In VLA models, action decoders are conventionally trained by behavior cloning, which supervises the demonstrated motor command while leaving the local objective served by that behavior implicit. Future-based supervision adds frames, latent observations, trajectories, or motion representations, but these capture particular realizations rather than the shared semantic objective. INDI addresses this gap by having a frozen teacher VLM interpret a demonstrated segment and then distilling the resulting intent into the deployed policy's decoder.¹

Assessment

The reported evidence includes large absolute improvements on SimplerEnv-Bridge (64.3% to 84.7%) and more moderate gains on RoboCasa Kitchen (64.1% to 70.3%), with consistent gains on π0.5 in both benchmarks. Real-world average success improves from 62.0% to 68.7%, and the largest gains occur on longer-horizon tasks. The consistency across two VLA families and the real-world transfer strengthens the causal claim that decoder-organized intent, rather than a single benchmark quirk, drives the gains.¹

Caveats

The main evidence gap is the abstract-level reporting: no error bars, sample sizes, ablations, or statistical tests are provided, and the 'controlled' baselines are not fully specified. What would change my mind would be evidence that the recovered latent does not organize downstream predictions in an objective-dependent way, or that the observed gains disappear when baselines are given equivalent extra training-time signals without intent distillation. The paper also does not specify task-level real-world breakdown beyond the horizon effect.¹

Source text

Hugging Face Daily Papers — Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models · August 24, 2026

Computer Science > Robotics

Title:Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models

Abstract:Vision-Language-Action (VLA) models can turn multimodal context into robot actions, but their action decoders are still trained largely by behavior cloning. This supervises which motor command was demonstrated while leaving implicit the local objective served by the behavior under the instruction. Future-based supervision enriches action learning with frames, latent observations, trajectories, or motion representations, but these signals capture particular realizations of what may happen rather than the shared semantic objective of the forthcoming behavior. We propose Intention Distillation (INDI), which distills behavior-level intent into the action decoder. During training, a frozen teacher VLM interprets a demonstrated segment from the current observation, instruction, coarse action summary, and corresponding execution video. From its standard inputs, the deployed VLA recovers the resulting multimodal intent representation at an intermediate decoder layer and uses it to organize action prediction together with representations of how the behavior unfolds and what it achieves. On SimplerEnv-Bridge, INDI improves GR00T-N1.7 from 64.3% to 84.7%, and on RoboCasa Kitchen it improves the controlled GR00T-N1.7 baseline from 64.1% to 70.3%, with consistent gains on $\pi_{0.5}$ across both benchmarks. In real-world tasks, INDI improves average success from 62.0% to 68.7%, with gains of up to 12.0 pp on longer-horizon tasks. Further analyses show that the recovered latent is used by the decoder, captures behavior objective and execution progress, and organizes downstream predictions in an objective-dependent manner. These results show that action decoders benefit from explicitly modeling the semantic objective of the behavior they generate.

Current browse context:

Bookmark

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

Here 'behavior intent' is an unstated technical assumption: it is a multimodal latent representation produced by the frozen teacher VLM from coarse action summaries and execution video, not a natural-language goal. The design's force is that the decoder is trained to recover this latent from standard inputs alone, so the privileged future signal is compressed into policy representations.

The reported baselines are denoted 'controlled', which likely means the authors ran their own GR00T-N1.7 reproduction for fair comparison, a detail the abstract does not clarify. The π0.5 results are cited as consistent gains, but no absolute numbers are given in the abstract, so the effect size on that model family cannot be assessed.

'Coarse action summary' is a term of art that is not defined in the abstract. It appears as an input to the teacher VLM during training but not among the deployed policy's standard inputs, implying the student must infer intent without this information; the summary itself is one of the future-based realizations the method claims to go beyond.

The source text above remains the property of its publisher and is reproduced for commentary.

Hugging Face Daily Papers — Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models

August 24, 2026endistinct angle

Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models

↑ Back to top