DVTR-KD: Distilling Vision-Language Models via Intermediate Visual Token Reconstruction

Vision-language models (VLMs) have achieved remarkable multimodal understanding and reasoning capabilities, yet remain computationally expensive due to dense visual tokenization. Existing efficiency approaches either merge redundant visual tokens or drop them progressively in language backbone, often trading accuracy for speed. In this work, we propose DUET-VLM, a versatile plug-and-play dual compression framework that consists of (a) vision-only redundancy aware compression of vision encoder's output into information-preserving tokens, followed by (b) layer-wise, salient text-guided dropping of visual tokens within the language backbone to progressively prune less informative tokens. This coordinated token management enables aggressive compression while retaining critical semantics. On LLaVA-1.5-7B, our approach maintains over 99% of baseline accuracy with 67% fewer tokens, and still retains >97% even at 89% reduction. With this dual-stage compression during training, it achieves 99.7% accuracy at 67% and 97.6% at 89%, surpassing prior SoTA visual token reduction methods across multiple benchmarks. When integrated into Video-LLaVA-7B, it even surpasses the baseline -- achieving >100% accuracy with a substantial 53.1% token reduction and retaining 97.6% accuracy under an extreme 93.4% setting. These results highlight end-to-end training with DUET-VLM, enabling robust adaptation to reduced visual (image/video) input without sacrificing accuracy, producing compact yet semantically rich representations within the same computational budget. Our code is available at https://github.com/AMD-AGI/DUET-VLM.

Multimodal Large Language Models (MLLMs) encode images into dense visual token sequences, creating substantial attention and KV-cache overhead. Existing visual token compression methods reduce this cost by pruning low-saliency tokens in shallow layers, but this practice couples temporary computational deactivation with irreversible information deletion. We show that this coupling breaks an intrinsic inference pattern in MLLMs. Through layer-wise attention analysis, we identify Visual Backtracking: visual pivots remain spatially stable, yet in middle layers they actively retrieve context from tokens that appeared unimportant earlier. Thus, early low-saliency tokens can be inactive for computation while still being useful for later reasoning. Motivated by this finding, we propose ReTrace, a backtracking-aware compression framework that introduces a dormant token state. ReTrace caches non-pivot tokens outside expensive layer-wise computation, restores them during the backtracking window, and prunes only after their information has been absorbed. Experiments on Qwen2.5-VL and InternVL3 demonstrate nearly 90% reductions in FLOPs and KV cache, a 2.83$\times$ prefilling speedup, and strong performance retention across complex vision-language benchmarks.

While knowledge distillation (KD) is widely adopted for training lightweight models by leveraging supervision from larger teacher models, relying solely on output token distributions has proven insufficient for compressing Multimodal Large Language Models (MLLMs). Since output tokens are a byproduct of the model attending to visual inputs, prior works have explored explicitly distilling attention to provide a direct supervisory signal. While promising, the precise utility of which attention signals to distill remains under-explored. In this work, we challenge the conventional reliance on prompt-to-vision attention by revealing that downstream performance correlates strongly with response-to-vision attention similarity to the teacher, but negligibly with that of prompt-conditioned attention. Furthermore, we observe that attention distributions exhibit significant variance across individual tokens, indicating that a uniform distillation objective is suboptimal. To this end, we introduce Token-level Response-visual Attention Guidance (TRAG), a distillation objective that 1) shifts the focus to response-to-vision signals and 2) employs token-specific objectives by adaptively weighting the Kullback-Leibler divergence based on attention entropy, effectively guiding the student to mirror the teacher's precise visual focus. Extensive experimental results on multiple benchmarks demonstrate that TRAG significantly outperforms prior distillation baselines.