UAV-ESPrune: Can Visual Token Pruning Preserve UAV Reasoning?.

UAV-ESPrune: Evidence-Aware Visual Token Pruning for UAV Vision-Language Reasoning • Ziyang Xu and Peng Han • National Space Science Center, Chinese Academy of Sciences; University of Chinese Academy of Sciences • Aerospace • 2026.

Cold Open.

Imagine a UAV looking down at a parking lot. Several identical white vans sit beside one another. Ask whether the scene contains vehicles, and a compact representation of one van might suffice. Ask how many vans are waiting, and every separate instance becomes relevant. The pixels look repetitive; the answer depends on that repetition.

Now imagine compressing that image before a vision-language model sees it. You are choosing what the model gets to know. A faithful summary of the dominant category can still be the wrong summary for the question. The engineering problem is to make the representation smaller while preserving distinctions that could change a decision.

That tension drives UAV-ESPrune. What should survive an aggressive visual-token budget: the most salient patches, the most diverse features, or evidence of individual objects and spatial relationships? The interesting part is how you translate that question into an actual selector.

The Idea.

Start at the interface between vision and language. Qwen2.5-VL processes images with a vision transformer, then uses a merger to combine neighboring patch features and project them into the language model’s embedding space. Its visual sequence therefore contains learned representations, already contextualized by the encoder. Its multimodal training spans image-text, grounding, documents, and video. That pretrained representation is the raw material available to a compression algorithm.

This interface creates a useful separation of concerns. The encoder can extract information at the original input resolution, while a subsequent module decides how much of the resulting representation should enter the language model. Think of it as choosing the contents of a report after inspecting the scene, rather than restricting what the inspector can see.

There are established precedents. FastV uses attention information from early language-model layers to identify visual tokens for removal in subsequent layers. Its organizing idea is to spend some decoder computation learning which image representations deserve continued processing.

VisPruner takes a visual-cue route: select important tokens using visual attention, then retain diverse information through similarity-aware duplicate removal. Its official implementation is public. This makes saliency plus diversity a concrete engineering baseline, with an accessible implementation to inspect.

UAV-ESPrune keeps Qwen2.5-VL-7B frozen and prunes after its visual merger. It scores feature magnitude, local peaks, and transitions; reserves peak representatives; and selects a feature-spatial coreset. Removed-token residuals update retained anchors, gently and locally for protected peaks. The anchors return to spatial order.

Here is a toy way to understand the allocation problem. Suppose a bright vehicle produces many high-scoring features, while several smaller vehicles produce weaker ones. Spending the entire budget on the strongest features could describe the bright vehicle beautifully. A representative-set objective asks a different question: which collection of retained features gives the rest of the image suitable representatives?.

Adding position to that thought experiment changes what “suitable” means. Two patches can look alike while belonging to distant objects. A representation that respects both appearance and location has a way to distinguish those cases. The relevant engineering choice is the definition of redundancy: similar appearance, similar location, or both.

Protected representatives introduce another useful distinction. Some retained features can act as broad summaries; others should remain closely tied to a particular local observation. Imagine summarizing the surrounding pavement without allowing that summary to overwhelm the tiny vehicle feature you deliberately preserved. That is a sensible allocation of responsibilities within a compressed representation.

The robotics connection is direct. A perception interface should preserve the distinctions consumed by downstream reasoning. Category presence, object multiplicity, and spatial arrangement are different contracts. The distinctive design choice here is to make that contract explicit through a UAV-oriented selection policy, while continuing to use a pretrained multimodal backbone.

The Result.

Testing uses three MM-UAVBench subtasks: referring-expression counting, environment-state classification, and ground-target planning.

MM-UAVBench assembles real UAV imagery from public datasets and evaluates models through multiple-choice questions. Referring-expression counting asks for objects matching a description. Environment-state classification concerns lighting and weather. Ground-target planning includes choosing suitable plans or rescue routes from aerial scenes with marked locations. These tasks give concrete meaning to the paper’s use of “reasoning.”.

That combination is useful for thinking about compression. Counting makes individual instances relevant. Environmental interpretation invites information distributed across the scene. Route selection brings relationships among locations into the answer. A selector is being asked to support different uses of the same kind of visual representation.

The headline result: retaining one quarter of visual tokens preserves full-token counting accuracy at 30.86%. Environment classification and planning remain near their full-token scores.

The unit of progress is retained answer quality per visual token. For a compression experiment, the full-input model supplies a reference operating point: how much representational capacity can you remove while maintaining its measured behavior? That framing makes the counting result immediately interpretable without introducing a new model, a new training objective, or a separate downstream capability claim.

Removing peak protection lowers counting accuracy; replacing the coreset with top-scoring tokens lowers average accuracy.

These interventions ask useful mechanistic questions. Does assigning special status to local evidence change the answer score? Does representing the wider feature field help beyond ranking patches independently? An ablation organized around those questions connects the algorithm’s structure to its intended function. It also gives a practitioner specific components to reason about rather than an indivisible collection of heuristics.

Stored-image inference on an RTX 4090 takes about 12% less end-to-end time, answering directly from prefill logits.

That is the systems number to carry into a discussion of the measured workload. Token count describes the representation; end-to-end time describes completion of the computation. Keeping both views visible makes the result actionable. A team can reason separately about what information enters the decoder and how long its application waits for an answer.

The practical reading is straightforward: treat compression quality and execution time as paired measurements. One tells you whether the smaller representation remains useful. The other tells you what that representation buys in the actual inference pipeline.

The Biggest Weakness.

The strongest objection is comparative validity: the method labeled SparseVLM is a materially simplified substitute for the published algorithm.

The paper’s SparseVLM adapter ranks visual tokens by absolute cosine similarity to averaged prompt embeddings, then prunes once before prefill.

Published SparseVLM works differently. It selects relevant text tokens to guide visual-token importance through decoder attention, adapts sparsification across layers, and recycles removed tokens into compact representations. Those mechanisms define how the method uses language to decide which visual information matters.

A comparison against averaged prompt embeddings does not exercise that machinery. Consider a question about vehicles near a particular road. The original approach can use interactions formed inside the multimodal decoder. The replacement asks whether an image feature resembles an aggregate text embedding before that processing happens. These are different relevance estimators operating on different representations.

Using a common integration point makes an experiment easier to organize. It also changes the scientific question when a competing method depends on computation beyond that point. Equalizing the final number of tokens does not restore the mechanisms removed from the baseline.

The resulting comparison supports a narrower statement: this evidence-oriented visual selector performs well against the particular text-similarity selector implemented here. It does not establish the same advantage over the published SparseVLM procedure.

The repair is clear. Evaluate the original method with its intended pruning schedule and recycling, then compare accuracy against actual compute or latency across operating points. Alternatively, label the simplified implementation as a SparseVLM-inspired proxy and keep the conclusion at that level.

This objection leaves the full-token retention result intact. It weakens the argument that the advantage comes specifically from UAV-oriented evidence preservation outperforming established text-guided pruning. The distinction matters because an apparent algorithmic win can partly reflect what was removed from the competing algorithm.

Verdict.

Roboticists building aerial visual question answering, inspection interfaces, or ground-station decision support should pay attention to this work. The useful design question is where to spend a limited visual representation budget when downstream questions care about different properties of the scene.

My takeaway is to treat token selection as part of the perception interface. Choosing image resolution, choosing representations, and choosing which representations survive all shape what the reasoning system can know. That makes compression a task-design decision as well as an inference optimization.

Consider a shelf-inspection application. If the question concerns whether a product category is present, a representative observation may be sufficient. If the question concerns stock count, distinct instances become essential. The same distinction applies to the parking-lot example: the appropriate summary depends on which changes in the scene should change the answer.

This perspective connects UAV-ESPrune to another line of work. ET-Prune, an earlier 2026 paper on text-rich multimodal inputs, protects text-like regions and uses question-conditioned, sample-specific token budgets. It performs progressive pruning within the decoder. Its domain is different, and its allocation signals are different.

The common research direction is evidence allocation. A digit on a document and a small vehicle in an aerial image can each occupy little visual area while carrying decisive information. In both cases, a useful compression objective should reflect the consequence of losing that information.

That suggests a productive division of labor for future systems. Domain structure can identify observations deserving protection, while question context can help decide how much capacity the current request needs. Representation selection and budget selection become complementary design decisions.

For practitioners, the durable lesson is about specifying the interface: decide which distinctions must remain recoverable, then make the compression policy serve that requirement. That is a concrete way to reason about efficiency without treating every visual token as interchangeable.

One-line takeaway: compress the visual sequence while preserving the distinctions your task asks the model to make.

Outro.

Return to the parking lot. The useful compressed view is the one that preserves what would make the answer change: another van arriving, one departing, or a different group becoming relevant to the question.

The next development worth watching is a selector that jointly decides which evidence to retain and how much representation that particular scene deserves. An almost empty lot and a dense gathering create different information demands.

For embodied AI, that is the compelling direction: representations whose size follows the evidence needed for the decision, with compression designed around the meaning of the question.