Efficient VLM Inference

A companion study ran a 35B mixture-of-experts model on a 2011 NVIDIA Tesla C2075 (Fermi, sm_20, 6GB) as a GPU-prefill/CPU-decode hybrid, because the 4-bit model did not fit in device memory (arXiv:2606.24031). This report keeps the hardware and asks what a model that fits can do: we deploy MiniCPM-V-4.6, a modern multimodal assistant pairing a SigLIP2 vision encoder and window-attention merger (16x visual token compression) with a compact hybrid gated-delta-net backbone, entirely on the GPU. Three results. (i) An all-GPU engine built on measured foundations: projections that dequantize 8-bit weights once and call the vendor SGEMM still in the last Fermi toolchain (64% of FP32 peak; our best hand-written GEMM hit 37%, wrongly called the ceiling); a chunked delta-rule rewrite of the recurrent layers, 2.8x faster than the sequential scan once attribution exposed one bad kernel; and a measured negative: 4-bit weights make decode slower than 8-bit here, since Fermi issues nibble-unpacking shifts at half rate. (ii) The vision side is a port with a proof obligation: we translate tower, merger, and projector to sm_20 CUDA, validating every stage against a locally generated reference forward (full tower 1.4e-5). One failure, position-embedding bucketization differing on exact rational ties, generalizes to a rule: float tie-breaking in index arithmetic is implementation-defined; call the reference operator, do not reimplement it. (iii) Long context exposes an O(N^2) wall short benchmarks hide: prefill falls from 114 tok/s at 2k tokens to 21 at 10k in a naive attention kernel; per-head vendor-GEMM calls writing into the existing score buffer (zero extra memory) restore a flat profile (408 at 2k, 361 at 10k; 17x), verified by exact needle retrieval from 60% depth. The same rewrite cuts image encoding 6x, to 0.93s. The system answers an image question end-to-end in 1.7s.

Vision-Language Models (VLMs) have achieved strong performance in multimodal understanding, yet remain challenging to deploy on resource-constrained edge devices due to the substantial computational overhead of processing numerous visual tokens. Token reduction is a promising direction for accelerating VLMs inference, but existing approaches either rely on attention maps that are incompatible with modern acceleration frameworks or depend on computationally intensive pairwise similarity comparisons, which undermine scalability and negate their practical benefits in deployment. In this paper, we propose an attention-free and lightweight token reduction framework as a plug-and-play module for VLMs, which preserves both important and diverse tokens to produce a compact visual representation. First, to enable attention-free importance estimation, we adopt an information-theoretic perspective and quantify token information using a novel entropy-based criterion, retaining those with more expressive and less degenerate feature representations. Second, to ensure diverse visual coverage in a lightweight manner, we introduce a transformation-induced consistency signal where similar tokens yield similar signals, such that sorting by this signal places similar tokens close to each other and enables stride-based selection to produce a diverse token set. Extensive experiments across multiple VLMs benchmarks demonstrate that our framework achieves a favorable accuracy-efficiency trade-off, maintaining competitive performance under aggressive compression.

Vision-language models (VLMs) process large numbers of visual tokens, resulting in substantial inference latency and memory overhead. This has motivated extensive research on visual token compression. While training-free strategies rely on heuristic metrics and suffer significant performance degradation under high compression ratios, many training-based methods introduce external compression modules that force the VLM backbone to adapt, incurring substantial retraining cost and compromising VLMs' priors. Effective visual token compression hinges on strong information encoding, a capability already present in pretrained VLMs but underutilized by existing approaches. Motivated by this, we propose VisCo, a training-efficient self-compression framework that reuses the pretrained VLM itself as an intrinsic compressor. VisCo is a parameter-sharing autoencoder that compresses visual information using a small set of memory tokens and transfers hierarchical information from encoding to decoding. Experiments show that VisCo surpasses prior methods across all evaluated compression ratios, with larger gains under more aggressive compression, and remains stable even in the extreme single-token setting. Moreover, when combined with the original visual tokens, the learned memory tokens can even improve the base model, suggesting that VisCo captures complementary representations beyond compression.

Speculative decoding accelerates autoregressive generation by letting a lightweight drafter propose multiple tokens that are verified by a larger target model. Although effective for text-only LLMs, speculative decoding yields limited gains in VLMs because drafters often diverge on vision-critical content, while existing multimodal acceleration methods do not directly address irrelevant visual evidence or optimize the verifier-accepted prefix length that governs speedup. We propose TIGER, a Text-conditioned vIsual GatEd Routing framework for multimodal speculative decoding. TIGER dynamically selects a sparse set of context-relevant visual tokens based on the drafter's current textual state, rather than expose the full visual token set or a fixed compressed interface. To better align training with inference-time efficiency, we optimize the drafter with acceptance-aligned group-based policy training using verifier-derived rewards based on accepted prefix length, built on top of distillation warm start with KL anchoring. This encourages the drafter not only to imitate the target model, but also to produce speculative continuations that survive verification for longer prefixes. Experiments show that TIGER yields consistent gains in accepted prefix length and speculative speedup under exact verifier-side speculative decoding, while achieving favorable quality-latency trade-offs with comparable downstream accuracy in visual-routing analyses.

Vision-Language Models (VLMs) are costly at inference time because they must process long sequences of visual tokens. Existing token pruning methods often degrade under high compression by blindly discarding information, breaking spatial structure or collapsing diversity. We propose SpecFlow, a training-free framework that shifts the paradigm from destructive pruning to conservative condensation, strictly enforcing spatial coverage and statistical conservation to ensure stability. Treating visual tokens as nodes in a $k$NN graph, SpecFlow (i) computes a stable importance field via spectral heat flow to preserve structural coherence, (ii) allocates budgets via adaptive spatial partitioning to guarantee coverage, and (iii) aggregates discarded information into coreset sinks to maintain statistical conservation. The method is plug-and-play, requires no fine-tuning, and is compatible with FlashAttention. Experiments confirm that our SpecFlow outperforms SOTA methods across tasks, VLM architectures, and pruning ratios. Notably, LLaVA-1.5 with SpecFlow retains 95.6% of original performance despite pruning 88.9% of visual tokens, offering an exceptional efficiency-accuracy balance. Code is available at https://github.com/Lzy-dot/SpecFlow

Vision-Language Models (VLMs) are the perceptual backbone of embodied AI, but their energy footprint on edge hardware remains poorly understood. Existing efficiency efforts focus predominantly on reducing visual tokens, implicitly treating visual processing as the dominant energy cost. We overturn this implicit assumption through the first systematic energy profiling of on-device VLM inference, spanning five models across three architecture families, four input resolutions, and two hardware platforms (NVIDIA RTX 3070 and Jetson Orin NX). Our analysis yields three findings. First, average inference power is a model-intrinsic constant, invariant to input resolution, image complexity, and prompt type, with less than 5% variation across all conditions. This means that all energy variation across inputs must arise from variation in inference time, not from variation in power draw. Second, each output token costs 11 to 39x more wall-clock time than each input token due to the compute-bound and memory-bound asymmetry between prefill and decode, making output token count the dominant driver of both latency and energy. Third, image complexity, measured by the number of objects in an image, induces up to 4.1x energy differences at identical resolution. This variation arises not from increased visual processing cost, but from differences in output length. These findings expose a fundamental limitation of visual token pruning: even removing all visual tokens saves at most 10% of total energy for fixed-token models. Across models spanning 1 billion to 8 billion parameters, controlling output length saves up to 97% of total energy, with the energy dominance of decoding growing stronger at larger model scale. In short, the true energy bottleneck in edge VLM inference is not what the model sees, but how much it says.

High-quality visual representation is a long-standing pursuit in computer vision. In the context of multimodal LLMs (MLLMs), feeding higher-resolution images can produce more fine-grained visual tokens. However, it introduces additional computational and design complexity, due to multiple forward passes and post-processing of increased tokens. Before simply adopting a higher resolution, have we truly unlocked the model's full perception capability at a standard resolution? Therefore, we study an interesting problem: how to achieve fine visual perception under lower cost without larger images. We present SigLIP-HD in this work. The core is a highly simple fine-to-coarse supervision design. We enforce the coarse feature of a mid-resolution image to mimic the fine-grained feature of its high-resolution version. We build this framework on the advanced SigLIP 2 model. Our final model produces better visual tokens at exactly the same inference budget. It is validated on extensive MLLM benchmarks and consistently delivers stronger results than our baseline model, especially on OCR-related tasks.

Although Video Large Language Models (Video LLMs) have shown strong performance in video understanding, their efficiency is still limited by the large number of visual tokens. Existing video token compression methods typically rely on frame-wise saliency or heuristic token merging, which can over-focus on locally salient regions and produce ambiguous fused features. To address these issues, we propose GeoTrace, a training-free spatiotemporal token compression framework that decomposes video evidence into exact skeleton tokens and traceable residual event tokens. Specifically, Contextual Farthest-Point Anchoring (CFPA) preserves salient, context-consistent, and high-coverage skeleton tokens, while Trajectory-Constrained Residual Condensation (TCRC) compresses residual tokens through one-to-one temporal trajectories and constrained near-manifold condensation, producing traceable event tokens with reduced ambiguity. We evaluate GeoTrace on four Video LLMs across four video understanding benchmarks, and the results demonstrate its effectiveness and generalization across different model architectures and scenarios. On LLaVA-OneVision, with only 10% visual tokens retained, GeoTrace achieves a $12.99\times$ TFLOPs reduction while preserving 99.1% of the vanilla performance. Overall, GeoTrace offers a compact and traceable token representation for efficient and robust Video LLM inference. Code is available at \href{https://github.com/guohuan-xie/GeoTrace.git}{\texttt{Code}}.

Vision-Language Models (VLMs) have achieved success using homogeneous Transformers to process multimedia data. Recent studies show that heterogeneous structures interleaving efficient mechanisms, like linear attention, improve both performance and inference latency over homogeneous designs. However, these efforts rely on handcrafted static mixing patterns, which are sub-optimal and difficult to adapt to specific hardware. To bridge this gap, we propose Multi-Objective Search for Adaptive Inter-layer Composition (MOSAIC), a hardware-aware search method that automatically transforms homogeneous models into optimized heterogeneous architectures. MOSAIC integrates diverse efficiency mechanisms--including linear, sparse, and low-rank operators--into a unified search space. By formulating the selection as a multi-objective Mixed Integer Programming (MIP) problem, our method identifies optimal configurations that maximize downstream performance under strict hardware latency constraints. To mitigate performance degradation from structural transitions, we introduce a two-stage parameter recovery process: global off-policy distillation to stabilize internal representations, followed by a dual-teacher on-policy distillation leveraging a 235B oracle for knowledge expansion and the original 4B teacher for distributional stability. We validate MOSAIC through MOSAIC-4B, derived from Qwen3-VL-4B-Instruct. Results demonstrate that MOSAIC-4B matches the baseline's performance across multiple benchmarks while requiring less than 2% of the original training cost. Furthermore, it substantially improves inference efficiency, achieving 1.76x prefilling and 2.54x decoding speedups.

The emergence of vision language models with fewer than 3 billion parameters has accelerated the implementation of on-device multimodal intelligence. However, a detailed understanding of component-wise quantization remains a bottleneck for optimal deployment. This paper presents a systematic evaluation framework for empirically validating five hypotheses across six quantization configurations on the Jetson Orin NX and AGX. By separating the vision encoder, projector, and large language model backbone yields the following results: (1) Quantization sensitivity is governed by the structural paradigm (MoE vs. dense) rather than scale alone, with MoE backbones mitigating INT4 noise where dense backbones degrade; (2) SigLIP encoders incur disproportionate INT8 latency on Jetson Ampere--a deployment-specific encoder-kernel-hardware interaction, not a SigLIP flaw; (3) Although INT4 quantization of LLMs greatly reduces VRAM consumption, it also causes slower token generation due to dequantization overhead; (4) Composite quantization errors are largely additive, except along the modality-alignment path, which is architecture-dependent; (5) The intelligence-per-joule profile varies significantly across platforms owing to memory bandwidth constraints.

Existing construction-site Vision-Language Model (VLM) studies have primarily adapted pretrained VLMs through direct QA-style fine-tuning from a single global image, but we argue that this paradigm remains limited in operational range, reliability under reduced-resolution inputs, and inference efficiency. To address these limitations, we propose AVA-VLM, an Adaptive Visual Attention-Vision Language Model that follows a human-inspired coarse-to-fine strategy: it first reasons over a low-resolution global image and requests a high-resolution local crop only when detailed inspection is needed. We further introduce a region-aware Chain-of-Thought dataset that teaches when to inspect, where to crop, and how to use local evidence. Experiments show that, for violation identification, AVA-VLM improves overall F1 from 62.0 to 75.1 while using only 30.6% of the baseline visual-token budget; for long-distance PPE-violation cases, F1 improves from 16.0 to 63.6. These results demonstrate AVA-VLM's improved robustness to distant and reduced-resolution visual evidence with substantially lower visual-token usage.

We present HunyuanOCR-1.5, a lightweight end-to-end OCR-specialized vision-language model. HunyuanOCR unifies document parsing, text spotting, information extraction, text-image translation, and multi-image document understanding within a single end-to-end VLM. Building upon the lightweight architecture of HunyuanOCR-1.0, HunyuanOCR-1.5 does not redesign the backbone, but systematically improves both efficiency and capability. For efficiency, we adapt DFlash to OCR decoding, significantly reducing the latency of long structured outputs such as dense documents, tables, and formulas while preserving output distribution. Powered by DFlash, HunyuanOCR-1.5 achieves a 6.37x Transformer inference speedup and a 2.14x speedup under vLLM, delivering the fastest inference among lightweight OCR VLMs. For capability, we propose Agentic Data Flow, an agent-driven data construction system that transforms model weaknesses into executable data requirements and autonomously performs material search, quality verification, and pipeline development. It substantially improves long-tail capabilities in ancient-script OCR, fine-grained chart and table parsing, multi-image text-centric QA, low-resource multilingual parsing, and document hallucination evaluation. HunyuanOCR-1.5 ranks among the top-tier end-to-end OCR solutions on OmniDocBench v1.6 while achieving new performance milestones across these long-tail tasks. Combined with an upgraded pretraining and post-training recipe, HunyuanOCR-1.5 further extends its capability in high-resolution, long-context, and multi-task scenarios. Experiments demonstrate faster inference, broader OCR capability coverage, and the deployment advantages of a lightweight end-to-end model. We will release the model weights and training code to support future research and real-world OCR applications.

Vision-Language Models (VLMs) have demonstrated impressive capabilities across different tasks, but their computational cost is dominated by the large number of visual tokens fed to the language model. Existing token reduction methods rely on attention-based scores or pairwise similarity, without an explicit semantic representation of each token. We introduce TORINO (TOken Reduction via Interpretable coNcept Overlap), a plug-and-play framework for adaptive visual token reduction in VLMs that requires no fine-tuning of the underlying model. TORINO leverages Sparse Autoencoders (SAEs) to project visual tokens into an interpretable latent space where token relationships can be analyzed through shared concept activations. Specifically, we define concept overlap as the degree of agreement between active SAE latents and use it to group tokens that share semantic content. Reduction within each group is then performed by either pruning or merging, providing a unified framework that preserves semantically important visual information while removing redundancy. Unlike fixed-budget approaches, TORINO dynamically adapts the reduction rate to input complexity, allowing different images to retain different numbers of tokens. Experiments across multiple vision-language benchmarks show that TORINO achieves favorable efficiency-accuracy trade-offs, reducing the number of visual tokens with minimal performance loss.

End-to-end Vision-Language Models (VLMs) show immense potential in autonomous driving. However, standard Supervised Fine-Tuning (SFT) often suffers from reasoning hallucinations and conservative biases. While traditional tool-augmented frameworks and Chain-of-Thought (CoT) approaches mitigate these issues, they incur exorbitant token consumption and unacceptable latency, rendering real-time deployment impractical. To resolve this reliability-efficiency trade-off, we propose CritiqueDriveVLM, a novel unified three-stage framework internalizing reasoning directly into the VLM. First, we introduce Critique-Driven Multi-Turn Reinforcement Learning (RL) guided by a multi-dimensional verifier. By providing granular scalar feedback and a multi-turn penalty, we force the policy to internalize logical deduction, cultivating a robust System-2 Teacher that achieves high accuracy without fragile external tools. Subsequently, we propose Latent Thought Distillation to overcome the latency bottleneck. By aligning the Student's latent representations with the Teacher's fully converged reasoning states, we compress deep logical capabilities into a fast, CoT-free System-1 Student. Extensive experiments on the widely-used DriveLMM-01 benchmark demonstrate remarkable improvements. Compared to the base model, our tool-free Teacher significantly boosts Multiple Choice Quality (MCQ) from 55.54% to a state-of-the-art 76.54%. Crucially, our distilled Student preserves competitive reasoning depth while drastically minimizing generation length to an average of merely 28 tokens. This slashes inference latency by 88% (from 3482 ms to 416 ms), paving a highly robust pathway for low-latency autonomous driving.Our source code is available at https://github.com/MICLAB-BUPT/CritiqueDriveVLM.

Recent Multi-modal Large Language Models (MLLMs) have demonstrated remarkable performance on 2D question answering tasks. However, extending these models to the 3D question answering remains challenging, as they typically require multiple views of the scene, which incurs substantial computational cost at inference. To mitigate this issue, existing solutions rely on strategic frame selection or token-merging algorithms that require preprocessing in advance all frames of the scene, i.e., an offline fashion. In contrast, we propose the first online token-pruning method that can be integrated seamlessly with current MLLM models for 3D question answering tasks, without additional training and with lower memory usage.Our key insight is to project each input frame into a shared voxel space using depth information and camera pose, identifying spatially-overlapped regions across frames and selectively pruning redundant image tokens before they enter the language model. Our method enables efficient online processing while reducing up to 50% of token usage. We apply this approach to Qwen2.5-VL-7B and Qwen3-VL-8B, demonstrating improved performance on the ScanQA, SQA3D, and OpenEQA-HM3D benchmarks.

This paper challenges the assumption that vision-language models (VLMs) require fixed patch-based 2D vision features. Analyzing fine-tuned vision encoders, we find that representations become increasingly abstract and less spatially coherent during VLM training. Notably, models trained with image-text alignment (such as SigLIP2) develop a small number of specialized tokens that effectively summarize global image content. Building on this, we introduce RADIO1D, which compresses images into a compact, variable-length 1D token sequence using multi-teacher knowledge distillation and an autoencoder design. The resulting representations exhibit strong hierarchical summarization, enabling accurate scene understanding - even with a single token - and support improved composition-aware image retrieval. In VLMs, RADIO1D provides flexible accuracy-efficiency tradeoffs through adjustable token counts, delivering competitive performance on diverse multimodal benchmarks with lower computational overhead and better accuracy.

Visual token pruning is a crucial strategy for accelerating VLMs by compressing redundant image patches, yet existing methods often fail to preserve critical cues under dense instructions and fine-grained queries. In this paper, we investigate this failure and identify two underlying bottlenecks: the widespread dispersion of textual noise that corrupts dense cross-modal scoring, and the feature fragmentation inherent to standard token selection. To address these issues, we propose Entropy-Aware Dense Pruning (EADP), a framework that reformulates pruning as a structured compression problem. EADP first leverages statistical entropy to quantify and filter out textual noise, yielding a robust, fine-grained instruction relevance score. Subsequently, instead of naive Top-K selection, EADP casts token selection as a submodular maximization problem with a spatial prior, explicitly ensuring a holistic and non-redundant visual representation. Extensive experiments demonstrate that EADP improves the accuracy-efficiency trade-off of VLMs, robustly preserving fine-grained visual cues under strict token budgets while achieving SoTA performance on challenging multimodal benchmarks.

Recent multimodal large language models (MLLMs) have substantially advanced video understanding, yet long-form video QA remains challenging under fixed input token budgets, where uniform sampling can be inefficient for evidence localization. We propose ReQuest , an uncertainty-driven, question-adaptive keyframe selection pipeline that aligns question intent with relevant video content through selective computation. ReQuest integrates (i) a lightweight question-aware selector distilled from MLLM-generated supervision, (ii) Re-thinking Routing that triggers additional inference only when the model is uncertain with a length-adaptive criterion, and (iii) uncertainty-guided adaptive non-maximum suppression that selects temporally diverse frames while adjusting spacing based on question difficulty. As a plug-andplay method, ReQuest improves long-video QA without modifying or fine-tuning the underlying MLLM. Experiments on Video-MME, MLVU, and LongVideoBench demonstrate consistent accuracy gains with competitive computational cost, with particularly strong improvements in medium and long video regimes.

Object-centric models inspired by DETR have become the dominant paradigm for open-vocabulary video instance segmentation (OV-VIS). While recent efforts have reduced the computational cost of pixel decoding, textual modality fusion, and object decoding to make these architectures more suitable for mobile devices, real-time on-device inference at high frame rates remains an open challenge. In this paper, we introduce SegFS, a dual-stream fast-slow framework that significantly improves efficiency without sacrificing accuracy. On sparse keyframes, an open-vocabulary object-based model predicts instance-level representations. These representations are then projected back into the backbone feature space to condition a lightweight fast network, which efficiently relocalizes and segments the instances in subsequent frames. By shifting instance propagation from object decoding to feature-space conditioning, our approach decouples multimodal semantic understanding from dense mask prediction and enables efficient temporal propagation. The proposed fast branch achieves up to 14x lower latency than the mobile-oriented MOBIUS model, while maintaining competitive segmentation performance on standard OV-VIS benchmarks.

Multimodal Large Language Models (MLLMs) incur prohibitive inference costs due to long visual token sequences. Training-free visual token reduction provides an efficient solution. However, existing methods distort attention distributions, giving rise to a phenomenon we term Attention Logit Collapse. To address this issue, we propose ERA, an Entropy-guided visual token pruning framework with Rectified Attention for efficient MLLMs. Specifically, ERA comprises three crucial components: Dual-view Entropy Pruning (DEP), Bias-aware Token Recycling (BTR), and Logit-preserving Attention Rectification (LAR). First, DEP identifies representative anchor tokens by jointly modeling visual diversity and head-wise saliency. BTR then recycles pruned tokens into their corresponding anchors while estimating a cluster-level logit bias. Building upon this, LAR injects the estimated bias into attention logits, effectively rectifying the collapse induced by token reduction. Together, these components preserve visual evidence even under aggressive compression, enabling robust performance across single-image, multi-image, and video settings on a wide range of MLLMs. Beyond delivering practical acceleration, ERA establishes logit-preserving visual token pruning as a principled framework for efficient MLLMs, unifying theoretical foundation, algorithmic design, and practical deployment. The code is at https://github.com/924973292/ERA.

Multimodal large language models (MLLMs) increasingly process long visual-token sequences, increasing the overall inference computation. Existing acceleration methods usually remove visual tokens or skip visual-token updates in entire layers, but these coarse strategies may discard fine-grained evidence or suppress useful operators together with redundant ones. In this paper, we study visual-token computation from an answer-observable perspective and find that late visual-token updates can remain large while having little effect on answer-token representations. Motivated by this answer-silent redundancy, we decompose each Transformer layer into attention and FFN operators and show that useful visual computation is often operator-dominant and layer-dependent. We propose an operator-level visual-token skipping framework that preserves the full visual-token sequence while selectively bypassing redundant attention, FFN, or both. Experiments across three MLLM architectures and 10 VQA benchmarks show that our method achieves strong efficiency-accuracy trade-offs, reducing \textbf{33.7%} TFLOPs on Qwen3-VL while retaining \textbf{99.5%} of the vanilla model performance.

Modern Vision-Language-Action (VLA) planners for autonomous driving emit a chain-of-causation (CoC) reasoning step \emph{before} producing a trajectory. The reasoning is autoregressive and dominates inference latency, while the trajectory head is parallel and cheap. Latency is an operational constraint in autonomous driving, so accelerating the reasoning step is the central problem we address. We observe that CoC reasoning has two qualitatively different needs: most tokens continue routine setup that follows naturally from the ego-trajectory history, and a small fraction encode commitments that require fresh visual evidence about an unexpected situation. We split this reasoning into two specialized paths: a \emph{routine reasoner} that handles the predictable continuation by attending to trajectory history, and a \emph{deliberative reasoner} (the unmodified VLA target) that handles novel cases by attending to current visual evidence, using the speculative decoding framework as the architectural template for how the two paths cooperate. Unlike standard speculative decoding, our routine reasoner is not a smaller replica of the target; the two reasoners are deliberately specialized to read different parts of the prompt. We propose two techniques to realize this. First, we introduce \textbf{FlatRoPE}, a 1D rotary positional embedding in the draft that breaks the rotational symmetry of the target's 3D M-RoPE, redirecting attention away from visual tokens and onto trajectory-history tokens. Second, we introduce \textbf{Action-aware RL (AARL)}, a post-training stage that uses an action-quality reward together with a static-reference KL anchor. Together, our two-reasoner system reduces the reasoning-step running time by approximately $4\times$ relative to the original Alpamayo planner.

Large Vision Language Models (LVLMs) have achieved remarkable success on vision-language tasks, yet fine-grained perception over high-resolution images and long-context videos remains challenging. As the number of visual tokens increases, the visual attention sink phenomenon becomes increasingly severe, causing irrelevant tokens to absorb a disproportionate amount of attention mass. Recent approaches attempt to mitigate this issue by explicitly predicting bounding boxes or temporal spans and re-encoding the cropped visual regions. Such methods depend on unreliable numeric localization in the discrete token space and incur significant computational overhead due to additional forward passes. In this work, we propose VisReflect, a simple yet effective framework that improves fine-grained perception in long visual contexts through latent visual reflection. Instead of decoding intermediate predictions into discrete tokens, the model generates continuous visual reflection that represents question-relevant visual features in the latent space. These reflections selectively emphasize salient regions or frames, guiding attention towards relevant visual tokens within a single forward pass. We conduct comprehensive evaluations on challenging high-resolution image benchmarks, including BLINK, V*, and HRBench-4K/8K, as well as video understanding benchmarks such as MVBench, VideoMME, and MLVU. Our method consistently improves over strong baselines, achieving gains of 4.1% on image benchmarks and 1.8% on video benchmarks. Compared with zooming-based methods, our model achieves comparable performance while reducing inference time by roughly 44% on video understanding.

Vision-language models and vision-language action models endow the robot with unprecedented capabilities. However, the input of video and high-resolution images yields a massive number of visual tokens, leading to extremely high inference latency and severely hindering the robot's real-time control. To break through this computational bottleneck, we propose ST-Merge, a plug-and-play, training-free framework that efficiently fuses redundant tokens directly during the visual encoding phase. By explicitly constructing 3D spatiotemporal coordinates, it employs a multi-queue parallel matching and weighted aggregation mechanism to achieve efficient and geometrically consistent fusion of redundant tokens across frames. In addition, we introduce a post-merge positional correction mechanism that effectively eliminates spatial deviation caused by merging by dynamically re-evaluating the rotational position code of the weighted centroid of the vision token, thereby ensuring the high-precision spatial awareness required for dexterous operation. In the Video Question Answering task on the mainstream VLM, Qwen2.5-VL, ST-Merge achieves a 2$\times$ inference speedup with only a tiny 1% loss in precision. When deployed on the $π_{0.5}$ VLA policy, ST-Merge achieves an 8.3$\times$ speedup at 1024 $\times$ 1024 resolution and matches the baseline success rate at this high-resolution setting. At lower resolutions, it introduces a small drop in accuracy.

Predicting future states is essential for autonomous agents, yet current Vision-Language-Action (VLA) models fundamentally lack this capability, relying instead on reactive perception-action mapping. While integrating Predictive World Models (PWMs) addresses this gap, existing approaches either incur prohibitive cascaded latency or act as shallow terminal tasks that fail to deeply embed forward-looking reasoning. To endow VLA models with this reasoning capability, we propose X-Mind. Rather than treating PWMs as an external auxiliary module, this framework internalizes them as the Visual Chain-of-Thought (Visual CoT). By enforcing a world rollout prior to action, the model is constrained to imagine future evolution first, yielding a driving policy that is robustly grounded in environmental dynamics and aware of the future consequences its actions will unfold. The challenge here is efficiency, and we tackle it on two fronts. First, we introduce a compact representation of visual thinking: an abstract sketch that fuses a Bird's-Eye-View (BEV) layout with abstract driving priors (e.g., navigation intents and traffic rules). Rather than rolling out dense future frames, the model reasons over this sketch as a mental canvas; aided by a Deep Compression Autoencoder (DC-AE), a 12-frame future rollout is reduced to merely 96 tokens, alleviating the long-context computational bottleneck. Second, to accelerate generation further, we propose a recurrent block diffusion scheme that unrolls the denoising steps across the layers of the large drive model, folding iterative refinement into the backbone's one forward pass. Trained and validated on large-scale real-world data, X-Mind achieves competitive end-to-end driving performance, which makes it a highly practical, low-latency solution that successfully deploys large-scale cognitive reasoning directly onto resource-constrained vehicle platforms.

Vision Language Models (VLMs) have emerged in the robotic domain as a powerful tool that enables environmental perception with language context, serving as a catalyst for open-vocabulary tasks like ObjectNav. Yet, their computational footprint typically confines them to cloud execution, hindering low-latency inference with local deployment on resource-constrained robots. To address this challenge, we present a distillation strategy that transfers complex spatial-semantic reasoning from large frontier models into a lightweight, 4B-parameter local VLM for edge execution on embedded GPU devices (e.g., Jetson Orin). We first establish a State of the Art (SotA), Scene Graph (SG)-based pipeline using Claude Sonnet 4.6, achieving a 39.7% Success Rate (SR) on the HM3D OVON benchmark. We then demonstrate that fine-tuning Qwen3.5-4B on just 500 frontier reasoning traces effectively enables navigation capabilities, yielding a SR of 34.5%, narrowing the gap to the performance of large cloud models. Finally, we introduce E-RLVR with Token Generation (TG) regularization to compress output sequence lengths for physical deployment while grounding the agent in its task. This downstream optimization reduces TG overhead by 72.1% and latency by 71.8%. Combined with quantization, this joint strategy yields a cumulative 82.8% reduction in overall inference latency without significantly sacrificing performance, presenting a viable paradigm for local, low-latency VLM execution on mobile robots.

Vision-Language Models (VLMs) improve generalization and interpretability in autonomous driving but suffer from efficiency issues due to long visual token sequences, particularly in standard multi-view settings. Existing token pruning methods employ fixed pruning rate allocation and static importance metrics, ignoring dynamic inter-view importance differences and the evolving information importance during inference. Our analysis reveals that multi-view VLMs inherently encode task-related view priors in deeper layers and exhibit dynamic information requirements. Motivated by these findings, we propose MVPruner, a two-stage adaptive token pruning method that aligns pruning behavior with the model's dynamic information requirements. The first stage allocates pruning budgets based on the information diversity of each view, and retains tokens with consistent contribution across stages, ensuring semantic representational capacity. The second stage allocates budgets and selects tokens guided by instruction text to guarantee task alignment. Experimental results on four benchmarks demonstrate the superior performance of our method. For example, DriveMM equipped with MVPruner achieves 87.3% reduction in FLOPs, 4.97* speedup in prefilling phase while retaining 98.5% accuracy on DriveLM benchmark.

Multimodal large language models (MLLMs) have achieved strong multimodal reasoning capabilities, but their efficiency is limited by the large number of visual tokens, which introduces substantial computational overhead. Visual token pruning offers a natural solution, yet existing methods are imperfect: attention-based criteria tend to retain redundant tokens, while diversity-based criteria are often agnostic to user instructions. Even methods that combine multiple criteria still lack a principled formulation of the intrinsic objective of token pruning. In this paper, we revisit visual token pruning from a first-principles perspective and formulate it as constructing Token Optimal Preservation Sets. Through a top-down information-theoretic analysis, we identify three fundamental principles for effective token selection: Task Relevance, Information Coverage, and Semantic Diversity. Based on these principles, we propose TOPS, a training-free and model-agnostic pruning module that can be applied to various MLLMs. Extensive experiments on 7 MLLM backbones and 14 benchmarks demonstrate that TOPS outperforms prior methods under diverse pruning settings. Notably, on LLaVA-NeXT, TOPS removes 77.8% of visual tokens while preserving 100.0% and 100.6% performance on its 7B and 13B models, respectively, suggesting that pruning redundant visual tokens can sometimes mitigate hallucination and inspire future lightweight MLLM design.

With the increase in model parameters and training data, the instruction following and generalization capabilities of Large VisionLanguage Models (LVLMs) have been significantly improved. Based on the Mixture of Experts (MoE) architecture, LVLMs expand their parameter capacity while maintaining the inference cost. However, traditional MoE methods employ a Top-k static routing strategy, which fails to account for variations in the input and adaptively select the number of experts, resulting in suboptimal resource utilization. In this paper, we propose viewing token routing as an information encoding task, framing dynamic routing as a Minimum Description Length (MDL) problem in encoding By validating the connection between MDL and gating entropy in the MoE scenario, we introduce Gating Entropy-based Uncertainty-aware Adaptive Routing (GeMoE) for MoE. Unlike traditional static or heuristic-based dynamic routing methods, GeMoE explicitly models the trade-off between model complexity and performance. By using gating entropy to assess the complexity of tokens, GeMoE adaptively determines the number of experts each token should engage. On a wide range of backbones and benchmarks, our method achieves 99.5% average performance retention compared to the original static routing, while improving average expert activation sparsity by 36.5%.

Inference efficiency is typically pursued by shrinking the model: distillation, pruning, quantization, and sparse routing each lower per-token cost while treating token count as fixed. But output length has been inflating, and it is precisely the component the standard toolkit leaves untouched. Here, we argue that brevity is the missing inference-efficiency lever, and that pretraining data curation is a practical way to pull it: a model trained on concise, correct data learns to answer in fewer tokens; i.e. it has a lower Cost-of-Pass. We apply our VLM curation pipeline to the MAmmoTH-VL single-image subset, and compare models trained on our curated data, the standard MAmmoTH-VL data, and external open-weight frontier VLMs. On a controlled 20-evaluation set and 14 VLMs at 1B-4B activated parameters, we hold output length fixed with a per-model regression, separating brevity from quality, and price models in FLOPs per correct answer. Curation buys a 35x Cost-of-Pass advantage over the most verbose 4B comparator (Qwen3.5-4B) within $\sim$1 pp of accuracy (0.41 vs 14.58 TFLOPs per correct answer; 0.691 vs 0.704 mean accuracy). Curation also buys a +17.55-percentage-point matched-length accuracy gain over the uncurated baseline that grows with model scale (from +16.7 pp at 1B to +21.2 pp at 4B). This brevity improvement concedes no quality: generic verbosity buys no accuracy at any capability or scale, and the window where reasoning-structured verbosity still earns its tokens shrinks from 4 of 8 capability groups at 2B to 1 of 8 at 4B. Per example, the concise model even reaches correct answers the verbose reasoning model misses, marking reasoning as a distinct curation target rather than something brevity gives up. Inference efficiency in this regime is a tokens-per-correct problem, and brevity is the lever that targets it directly.