VLM Inference

Large Vision-Language Models (LVLMs) are rapidly evolving toward true multimodal reasoning, with visual search representing a concrete instantiation of the thinking-with-images paradigm. However, LVLM visual search faces two key challenges: incompatibility among intrinsic capabilities after post-training, and interference in long multi-step reasoning contexts. To address these, we identify two novel insights. First, self-regulation between pre- and post-training LVLMs leverages the intrinsic single-step capabilities of the pre-training model to mitigate capability deterioration and long-context interference. Second, probability-based prophetic sampling, replacing naive prompting, provides a probabilistic interface where the pre-training model acts as a prophet and the post-training model selectively accepts prophetic tokens under its output distribution, preserving coherent multi-step reasoning. Building on these insights, we introduce SeProD, a self-prophetic decoding framework that leverages intrinsic single-step capabilities to enable coherent multi-step reasoning in a training-free, plug-and-play manner. Experiments show that SeProD consistently improves multiple visual-search LVLMs across all 12 splits of 4 visual search benchmarks, as well as across general VQA benchmarks, without added computational overhead, thanks to its parallel prophetic acceptance mechanism.

Flowcharts are widely used in industrial requirements, but usually remain embedded as static images. Vision Language Models (VLMs) show promise in the conversion of these flowcharts into machine-readable models for RE activities, yet, when directly applied to flowchart conversion, they often fail on topology-critical visual details. To address this, we propose EdgeFlow that augments a VLM's original input with a deterministically extracted Canny edge map-acting as a structural prior-to improve flowchart-to-Mermaid conversion, without requiring annotated training data or domain-specific model fine-tuning. We evaluate EdgeFlow on IndusReqFlow, a dataset sourced from real-world requirements. Compared with off-the-shelf VLMs, EdgeFlow improves node-level F1 by 17.39 percentage points and edge-level F1 by 16.94 percentage points. At the path level, EdgeFlow improves path F1 by 11.06 percentage points, enabling better support for model-based testing. These results demonstrate that EdgeFlow provides a practical, training-free means to improve topology-preserving flowchart-to-Mermaid conversion for industrial RE. Cross-dataset evaluation results on a public synthetic benchmark show no significant improvement; this highlights the need for diverse benchmarks incorporating industrial data for the comprehensive evaluation of future VLM-based RE tools.

Deploying Vision-Language Models on resource-constrained hardware typically requires INT8 quantization, but in joint-embedding architectures such as CLIP this introduces a failure mode distinct from quantized CNN classifiers: activation noise accumulated across transformer blocks perturbs the direction of the multimodal embedding, eroding the cosine alignment on which zero-shot retrieval depends. We characterize this as Quantization-Induced Representation Collapse (QIRC) and quantify it on INT8 CLIP ViT-B/32, where the layer-wise noise-to-signal ratio grows from below 10% in shallow blocks to 52% at Layer 11. We propose LRA-EE (Layer-wise Representation-Aware Early Exit), which bypasses noise-saturated deep layers via Spatio-Semantic Aggregation (replacing the immature shallow [CLS] with a global patch-token average), a learned multi-feature gate (confidence, top-2 margin, spatial-activation variance), and Layer-adaptive Confidence Thresholding calibrated to each layer's Information-to-Noise Ratio. On ImageNet-1K zero-shot classification, LRA-EE reduces FLOPs by 13.4% and improves Top-1 accuracy by +2.44%p (58.72% -> 61.16%) over the INT8 baseline. A four-quadrant decomposition isolates the Rescue Effect: 9.5% of samples are correctly classified at shallow exits but lost to noise at full depth, against only 7.1% suffering the inverse.

Metaphorical videos are prevalent across various real-world scenarios to convey complex ideas, and understanding them typically requires high-order cognitive capabilities. The lack of systematic studies on metaphorical video understanding not only constrains the real-world applicability of MLLMs but also impedes the thorough assessment of their high-order cognitive capabilities. To bridge this gap, we propose MetaphorVU-Bench, the first systematic and comprehensive benchmark dedicated to metaphorical video understanding. Through experiments, we find current MLLMs struggle with accurate metaphorical video understanding, lagging far behind human level, primarily due to defective cross-domain mapping. Motivated by this finding, we construct a metaphor knowledge graph as mapping augmentation and propose MetaphorBoost, an inference-time enhancement framework achieving consistent performance improvement. Our benchmark, analysis, and method provide useful insights and a foundation for future research on advancing MLLMs.

The generation of factually incorrect objects, commonly known as object hallucination, remains a persistent challenge in Large Vision-Language Models (LVLMs). Current approaches to address this issue - ranging from expensive data-driven fine-tuning and high-latency contrastive decoding to rigid attention head truncation - frequently compromise either computational efficiency or the continuity of the model's feature space. To overcome these limitations, we introduce a novel, training-free inference strategy that operates as a region-aware adaptive weighting mechanism to dynamically correct semantic drift without relying on abrupt heuristic truncations. By computing an outlier-resistant statistical midpoint across various attention heads, we establish a stable anchor for reliable visual representations. We then utilize the inter-head disagreement mapped across regions to dynamically determine intervention budgets, gently suppressing hallucination-inducing attention paths through a continuous penalty modulation. This recalibration process effectively rectifies visual-semantic misalignments while fully preserving generative fluency and language priors. Comprehensive evaluations on standard multimodal benchmarks, including CHAIR, POPE, and MME, reveal that our strategy substantially curtails both instance- and sentence-level hallucinations. The results demonstrate state-of-the-art performance against contemporary baselines, confirming our method's efficiency and algorithmic robustness. Our code will be public.

We present an approach to improve 3D vehicle labeling in self-driving applications through zero-shot inference of vehicle information, leveraging Vehicle Make and Model Recognition (VMMR) methods. The proposed approach utilizes a Vision Language Model (VLM) to both infer a vehicle's make, model, and generation from image crops, and output accurate 3D bounding box dimensions to seed manual labeling. We evaluate the impact of iterative prompt engineering and the choice of different VLMs on both vehicle bounding box inference and make/model/generation recognition. When compared to strong baselines, the proposed approach not only shows high accuracy, but also excels in mitigating specific failure modes where VLMs provide better dimensions than initial lidar-aided human annotated labels (e.g., in cases of significant vehicle occlusion). Experiments on both public and proprietary data strongly suggest that our conclusions are generalizable across different labelers and datasets. The results demonstrate that integrating VLMs into the labeling process can reduce manual labeling time while increasing label quality.

In recent years, multimodal large language models (MLLMs) have achieved remarkable progress, primarily attributed to effective paradigms for integrating visual and textual information. The dominant connector-based paradigm projects visual features into textual sequence, enabling unified multimodal alignment and reasoning within a generative architecture. However, our experiments reveal two key limitations: (1) Although visual information serves as the core evidential modality in MLLMs, it is treated on par with textual tokens, diminishing the unique contribution of the visual modality; (2) As generation length increases, particularly within a limited context window, the model's dependence on visual information progressively weakens, resulting in deteriorated vision-language alignment and reduced consistency between generated content and visual semantics. To address these challenges, we propose the Vision Inference Former (VIF), a lightweight architectural module that establishes a direct bridge between pure visual representations and the model's output space. Specifically, VIF continuously injects visual semantics throughout the decoding phase of the inference process, ensuring that the model remains firmly grounded in visual content during generation. We conduct experiments on 14 benchmark tasks covering general reasoning, OCR, table understanding, vision-centric evaluation, and hallucination. Experimental results show that VIF consistently improves model performance across diverse architectures while introducing minimal additional overhead. The code for this work is available at https://github.com/Dong-Xinpeng/VIF.

Vision-language models (VLMs) have recently emerged as a promising paradigm for video anomaly detection (VAD) due to their strong visual reasoning ability and natural language-based explainability. In this paper, we aim to address a key limitation of such pipelines, which perform segment-level inference independently owing to token constraints and reason without structured temporal context, allowing VLMs to interpret anomalies as deviations from evolving video dynamics rather than producing fragmented predictions and explanations. To specify, we propose a context-aware framework named LATERN, which reformulates VAD as a temporal evidence aggregation process. LATERN consists of two complementary modules: Context-Aware Anomaly Scoring (CEA) and Recursive Evidence Aggregation (REA). CEA introduces a novel image-grounded memory mechanism, which selectively chooses historical content via frame diversity and visual-textual alignment as expanded context to help generate reliable anomaly scores. Building upon these scores, REA performs recursive temporal aggregation to identify coherent anomaly intervals and produce event-level decisions and explanations grounded in visual-textual evidence. Extensive experiments on challenging benchmarks, including UCF-Crime and XD-Violence, show that LATERN enhances detection accuracy and explanation consistency for frozen VLMs during test time, while generating temporally coherent and semantically grounded event-level explanations.

Test-time scaling (TTS) methods have proven highly effective for LLMs, yet their application to vision-language models (VLMs) remains relatively underexplored. Existing VLM TTS methods largely require open-weight model access or expensive repeated sampling, and are evaluated primarily on multimodal mathematical and scientific reasoning benchmarks rather than general visual understanding tasks. In this paper, we propose Test-Time Hinting, a method that improves VLM performance via a single VLM call and requiring only black-box API access, which makes it broadly applicable to frontier closed-weight models. Our method is motivated by the observation that VLM errors tend to cluster around recurring failure patterns. We therefore train a lightweight hint generator model to predict, for a given test input, which "hint" should be prepended to the prompt, providing targeted contextual or procedural guidance that steers the VLM away from its characteristic failure modes. We show that Test-Time Hinting improves the accuracy of multiple closed-weight VLMs on natural-image VQA benchmarks and that these gains generalize to unseen benchmarks and VLMs without retraining the hint generator.

Edge deployment of Vision-Language Models (VLMs) faces a tradeoff between latency and accuracy: cloud execution provides high-quality predictions but incurs communication delay and energy cost, while edge-only execution is faster but less accurate due to limited model capacity. This trade-off is further complicated by heterogeneity in image quality and reasoning complexity, making static placement suboptimal. We present INAR-VL, a lightweight edge-cloud routing system for multimodal inference in a two-tier deployment. INAR-VL maintains complementary VLMs across edge and cloud and uses lightweight image and text complexity signals to guide routing and model selection, executing simple queries locally while offloading complex ones when beneficial. Evaluation on visual question answering shows that INAR-VL executes 36% of requests on the edge, reduces latency by 24%, lowers energy by 26%, and preserves 97% of cloud-level accuracy.

During MLLM decoding, attention often abnormally concentrates on irrelevant image tokens. While existing research dismisses this as invalid noise and forcibly redirects attention to compel focusing on key image information, we argue these tokens are critical carriers of visual and narrative logic, and such coercive corrections exacerbate visual-language imbalance. Adopting a "decoding-as-game" perspective, we reveal that hallucinations stem from an equilibrium imbalance between linguistic priors and visual information. We propose Adversarial Counter-Commonsense Equilibrium (ACE), a training-free framework that perturbs visual context via counter-commonsense patches. Leveraging the fact that authentic visual features remain stable under perturbation while hallucinations fluctuate, ACE implements a dynamic game decoding strategy. This approach precisely suppresses perturbation-sensitive priors while compensating for stable visual signals to restore balance. Extensive experiments demonstrate that ACE, as a plug-and-play strategy, enhances model trustworthiness with negligible inference overhead.

Existing training-free approaches for GUI grounding often rely on multiple inference runs, such as iterative cropping or candidate aggregation, to identify target elements. Despite this additional computation, each forward pass still independently interprets the instruction and parses the visual layout, without enabling progressive interaction among visual tokens. In this paper, we study what happens during GUI grounding in Vision-Language Models (VLMs) and identify a previously overlooked bottleneck. We show that grounding follows a two-stage paradigm: the prefill stage determines candidate UI elements, while the decoding stage subsequently refines the final coordinates. This asymmetry establishes prefill as the critical step, as errors in candidate selection cannot be effectively corrected during decoding. Based on this observation, we propose Re-Prefill, a training-free method that revisits inference by introducing an attention-guided second prefill stage to refine target selection. Specifically, visual tokens that consistently receive high attention from the query position, i.e., the final token, across layers are extracted as a preliminary target hypothesis and appended to the input, together with the instruction hidden states, enabling the model to deeply re-think its decision before coordinate generation. Experiments across four VLMs and five benchmarks, including ScreenSpot-Pro, ScreenSpot-V2, OSWorld-G, UI-Vision, and MMBench-GUI, demonstrate consistent improvements without additional training, with gains of up to 4.3% on ScreenSpot-Pro. Code will be available at https://github.com/linjiaping1/Re-Prefill.

Vision Language Models (VLMs) offer the exciting possibility of processing text as rendered images, bypassing the need for tokenizing the text into long token sequences. Since VLM image encoders map fixed-size images to a fixed number of visual tokens, varying rendering resolution provides a fine-grained compression knob. However, accuracy deteriorates quickly as compression increases: characters shrink below the vision encoder's effective resolution, making them indistinguishable. To address this, we propose LensVLM, an inference framework and post-training recipe that enables VLMs to scan compressed images, then selectively expand only the relevant images to their uncompressed form via learned tools. Building on Qwen3.5-9B-Base, LensVLM maintains accuracy comparable to the full-text upper bound at 4.3$\times$ effective compression and outperforms retrieval-based, text- and visual-compression baselines up to 10.1$\times$ effective compression across seven text QA benchmarks. LensVLM also generalizes to multimodal document and code understanding tasks, with the accuracy gain over baselines growing as compression increases. Our analysis validates this approach: training makes visual compression robust to rendering choices, and as compression grows the model increasingly relies on expanded content rather than unreliable visual reading. The analysis also yields practical tool-choice guidance: text expansion is preferable for rendered text, while high-resolution image expansion suits native documents whose layout cues carry task-relevant information.

Video text-based visual question answering (Video TextVQA) aims to answer questions by reasoning over visual textual content appearing in videos. Despite the strong multimodal video understanding capabilities of recent Video-LLMs, their performance on existing Video TextVQA benchmarks remains limited. To better understand this gap, we conduct an upper-bound analysis through frame-wise question answering, counting a sample as correct if any frame yields the right answer, which significantly outperforms direct video-based inference and reveals a substantial performance gap. The results suggest that the primary bottleneck lies in the localization of key question-relevant evidence, rather than in reasoning capacity itself. Building on this insight, we propose a question-guided agent framework that explicitly anchors the relevant keyframes before answering. The approach operates effectively in a training-free setting and consistently surpasses direct video inference. With additional supervised fine-tuning (SFT) and reinforcement learning (RL), it achieves an average improvement of +12.12 in accuracy and +11.15 in ANLS across benchmarks, establishing new state-of-the-art results. Our study underscores the critical role of explicit keyframe anchoring for advancing Video TextVQA. The code will be publicly released.

Vision-Language Models (VLMs) hallucinate objects that are not present, and a growing line of work tries to curb this by feeding the model its own generated caption as auxiliary evidence -- assuming that a caption, once available, is something to consume. We show this fails: naively appending a caption can lower accuracy rather than raise it, dropping Qwen2.5-VL-3B† on HallusionBench by nearly ten points. To understand why, we build GD-Probe, a diagnostic set that pairs a global and a detail question on the same image, so that any difference in caption effect is attributable to the question alone. Caption utility proves to be a per-query property: the same caption helps global questions and harms detail ones, through a single mechanism -- an embedded caption competes with the image for attention and pulls the model's evidence onto its own text -- whose sign is set by whether the caption covers the queried content. Crucially, this regime is readable from quantities the decoder already emits, with no attention access or grounding. We turn this into GEASS (Gated Evidence-Adaptive Selective Caption Trust), a training-free, logit-level module that decides per query how much of the caption to trust, gating it by the clean path's confidence, weighting it by the entropy reduction it induces, and raising the evidence bar when the two pathways disagree. Across four VLMs and two benchmarks (POPE and HallusionBench), GEASS improves over both vanilla inference and contrastive decoding under a single fixed setting, adding only two forward passes and no parameters.

Vision-language models (VLMs) have demonstrated strong applicability in edge industrial applications, yet their deployment remains severely constrained by requirements for deterministic low latency and stable execution under resource limitations. Existing frameworks either rely on bloated general-purpose designs or force developers into opaque, hardware-specific closed-source ecosystems, leading to hardware lock-in limitation and poor cross-platform adaptability. Observing that modern AI agents can efficiently search and tune configurations to generate highly optimized low-level kernels for standard LLM operators, we propose EdgeFM, a lightweight, agent-driven VLM/LLM inference framework tailored for cross-platform industrial edge deployment. EdgeFM removes non-essential features to reduce single-request latency, and encapsulates agent-tuned kernel optimizations as a modular library of reusable skills. By allowing direct invocation of these skills rather than waiting for closed-source implementations, it effectively closes the performance gap long dominated by proprietary toolchains. The framework natively supports mainstream platforms including x86 and NVIDIA Orin SoCs, and represents the first end-to-end VLA deployment on the domestic Horizon Journey platform, enhancing cross-platform portability. In most cases, it yields clearly better inference performance than conventional vendor-specific toolchains, achieving up to 1.49 times speedup over TensorRT-Edge-LLM on the NVIDIA Orin platform. Experimental results show that EdgeFM delivers favorable end-to-end inference performance, providing an open-source, production-grade solution for diverse edge industrial scenarios.

Vision-Language Models (VLMs) exhibit strong performance in instruction following and open-ended vision-language reasoning, yet they frequently generate fluent outputs that are weakly grounded in visual evidence. Prior works have shown that instruction prompting further worsens this issue by amplifying language priors, especially when the visual signal is uncertain or ambiguous. To address this challenge, we propose a decoding framework that explicitly balances linguistic informativeness and visual faithfulness during generation. Our method, Instruction-Evidence Contrastive Dual-Stream Decoding (IECD$^2$), maintains two parallel probability distribution of tokens at each decoding step: an instruction-driven stream that promotes expressive and informative responses, and an evidence-driven stream that enforces strict grounding in the image. These two streams are adaptively fused using a symmetric KL-based contrastive gate, which suppresses tokens favored by language priors but unsupported by visual evidence, while preserving them when both distributions agree. We evaluate IECD$^2$ on multiple datasets spanning various generative vision-language reasoning tasks such as captioning and visual question answering on multiple datasets such as, POPE, MME, VQAv2, AMBER, and MSCOCO. IECD$^2$ demonstrates consistent improvements in task accuracy and reasoning performance with substantial reduction in hallucination compared to state-of-the-art decoding approaches.

Vision-Language Models (VLMs) are frequently undermined by object hallucination--generating content that contradicts visual reality--due to an over-reliance on linguistic priors. We introduce Positive-and-Negative Decoding (PND), a training-free inference framework that intervenes directly in the decoding process to enforce visual fidelity. PND is motivated by our key finding of a critical attention deficit in VLMs, where visual features are empirically under-weighted. Our framework corrects this via a dual-path contrast: The positive path amplifies salient visual evidence using multi-layer attention to encourage faithful descriptions, directly counteracting the attention deficit. Simultaneously, the negative path identifies and degrades the core object's features to create a strong counterfactual, which penalizes ungrounded, prior-dominant generation. By contrasting the model's outputs from these two perspectives at each step, PND steers generation towards text that is not just linguistically probable, but visually factual. Extensive experiments on benchmarks like POPE, MME, and CHAIR show that PND achieves state-of-the-art performance with up to 6.5% accuracy improvement, substantially reducing object hallucination while also enhancing descriptive detail--all without requiring any model retraining. The method generalizes effectively across diverse VLM architectures including LLaVA, InstructBLIP, InternVL, and Qwen-VL.

Stochastic sampling strategies are widely adopted in large language models (LLMs) to balance output coherence and diversity. These heuristics are often inherited in Multimodal LLMs (MLLMs) without task-specific justification. However, we contend that stochastic decoding can be suboptimal for Visual Question Answering (VQA). VQA is a closed-ended task with head-heavy answer distributions where uncertainty is usually epistemic, arising from missing or ambiguous visual evidence rather than plausible continuations. In this work, we provide a theoretical formalization of the relationship between model calibration and predictive accuracy, and derive the sufficient conditions for greedy decoding optimality. Extensive experiments provide empirical evidence for the superiority of greedy decoding over stochastic sampling across multiple benchmarks. Furthermore, we propose Greedy Decoding for Reasoning Models, which outperforms both stochastic sampling and standard greedy decoding in multimodal reasoning scenarios. Overall, our results caution against naively inheriting LLMs decoding heuristics in MLLMs and demonstrate that greedy decoding can be an efficient yet strong default for VQA.

Vision-Language Models (VLMs) are frequently undermined by object hallucination, generating content that contradicts visual reality, due to an over-reliance on linguistic priors. We introduce Positive-and-Negative Decoding (PND), a training-free inference framework that intervenes directly in the decoding process to enforce visual fidelity. PND is motivated by our finding of an attention imbalance in VLMs, where visual features are under-weighted. Our framework introduces a dual-path contrast: a positive path that amplifies visual evidence and a negative path that constructs counterfactuals to penalize prior-dominant generation. By contrasting outputs from both paths during decoding, PND steers generation toward visually grounded results. Experiments on POPE, MME, and CHAIR demonstrate state-of-the-art performance without retraining.

Vision-Language Models (VLMs) have demonstrated strong capability in a wide range of tasks such as visual recognition, document parsing, and visual grounding. Nevertheless, recent work shows that while VLMs often manage to capture the correct image region corresponding to the question, they do not necessarily produce the correct answers. In this work, we demonstrate that this misalignment could be attributed to suboptimal information flow within VLMs, where text tokens distribute too much attention to irrelevant visual tokens, leading to incorrect answers. Based on the observation, we show that modulating the information flow during inference can improve the perception capability of VLMs. The idea is that text tokens should only be associated with important visual tokens during decoding, eliminating the interference of irrelevant regions. To achieve this, we propose a token dynamics-based method to determine the importance of visual tokens, where visual tokens that exhibit distinct activation patterns during different decoding stages are viewed as important. We apply our approach to representative open-source VLMs and evaluate on various datasets, including visual question answering, visual grounding and counting, optical character recognition, and object hallucination. The results show that our approach significantly improves the performance of baselines. Project page: https://cxliu0.github.io/AIF/.

Vision-Language Models (VLMs) exhibit strong visual reasoning capabilities, yet they still struggle with 3D understanding. In particular, VLMs often fail to infer a text-consistent goal 6D pose of a target object in a 3D scene. However, we find that with some inference-time techniques and iterative reasoning, VLMs can achieve dramatic performance gains. Concretely, given a 3D scene represented by an RGB-D image (or a compositional scene of 3D meshes) and a text instruction specifying a desired state change, we repeat the following loop: observe the current scene; evaluate whether it is faithful to the instruction; propose a pose update for the target object; apply the update; and render the updated scene. Through this closed-loop interaction, the VLM effectively acts as an agent. We further introduce three inference-time techniques that are essential to this closed-loop process: (i) multi-view reasoning with supporting view selection, (ii) object-centered coordinate system visualization, and (iii) single-axis rotation prediction. Without any additional fine-tuning or new modules, our approach surpasses prior methods at predicting the text-guided goal 6D pose of the target object. It works consistently across both closed-source and open-source VLMs. Moreover, when combining our 6D pose prediction with simple robot motion planning, it enables more successful robot manipulation than recent Vision-Language-Action models (VLAs). Finally, we conduct an ablation study to demonstrate the necessity of each proposed technique.

Continuous inference over concurrent video streams imposes substantial compute and memory demands on vision-language model (VLM) serving. Streaming inference uses sliding windows to maintain a bounded context of recent video, but processing each window independently repeats visual encoding and large language model (LLM) prefilling for similar and overlapping content. Existing optimizations provide limited coordination across these stages and often rely on model-specific training, profiling, or model-generated signals. We present CodecSight, a streaming VLM serving system that uses codec metadata as shared runtime guidance across visual encoding and LLM prefilling, without model-specific training or offline profiling. Codec-derived change signals guide patch pruning before visual encoding, reducing both visual computation and the number of downstream visual tokens. Codec-defined frame types guide selective key-value (KV) refresh across windows, while positional correction enables reuse of the remaining cached keys. Across three VLMs and four video workloads, our vLLM-based implementation supports up to $3.3\times$ as many concurrent streams and achieves up to a $5.3\times$ speedup in average time-to-first-token relative to the state-of-the-art baselines. It also reduces executed FLOPs by up to 93%, with a maximum task-quality decrease of 4.64 percentage points.

Large Vision-Language Models (LVLMs) exhibit strong performance on single-image tasks. However, their performance degrades significantly when handling multi-image inputs. While this degradation has been observed in prior work, its nature remains poorly understood. We empirically observe visual elements from different images become entangled in the model's representations and responses. We refer to this phenomenon as cross-image information leakage. To address this issue, we propose FOCUS, a training-free and architecture-agnostic method. FOCUS masks all but one image with random noise, guiding the model to focus on the single clean image. This process is applied across the target images to obtain logits under partially masked contexts. These logits are aggregated and then refined using a noise-only reference input, which suppresses the leakage and yields more accurate outputs. FOCUS consistently improves performance on diverse multi-image benchmarks. We further show that FOCUS generalizes to video understanding, extending its applicability beyond static multi-image inputs. This demonstrates that FOCUS offers a general solution for enhancing multi-image reasoning without additional training or architectural modifications.

Vision-language models (VLMs) have achieved impressive performance on multimodal inference tasks, but the cost remains a significant challenge due to the large number of vision tokens processed during the prefill stage. Existing token pruning methods often rely on utilizing the static attention patterns directly, failing to exploit the dynamic internal signals within VLMs. To address the issue, we propose AdaptInfer, a novel plug-and-play framework for vision token pruning. First, we introduce a dynamic text-guided pruning mechanism that construct soft priors over text-token importance on each pruning layer, allowing more informed scoring of vision tokens at each stage. Second, we observe a highly consistent distribution of cross-modal attention shifts by architecture, which inspires us to introduce a efficient data-driven schedule which determines the pruning locations automatically. Experimental results have verified the effectiveness and generalization of the proposed method. Under the same token budget, AdaptInfer surpasses existing approaches in accuracy. For instance, AdaptInfer maintains averagely 99.4% accuracy on Qwen2-VL while 70% of the prefilling vision token overhead is reduced. The source code is available on: https://github.com/weiczh02/AdaptInfer-base.

Despite the impressive capabilities of Large Vision-Language Models (LVLMs), they remain susceptible to hallucinations, where generated content is inconsistent with the input image. Existing training-free hallucination mitigation methods often suffer from unstable performance and high sensitivity to hyperparameter settings, which limits their practicality and broader adoption. In this paper, we propose Decoding with Inter-layer Consistency via Layer Aggregation (DCLA), a training-free decoding mechanism that requires no retraining, fine-tuning, or access to external knowledge bases. Specifically, DCLA constructs a dynamic semantic reference by aggregating representations from previous layers and uses it to correct semantically deviated layers, thereby enforcing inter-layer consistency. Experiments across seven LVLMs and multiple benchmarks demonstrate the generality of DCLA: it surpasses standard decoding by 28.58 MME points on LLaVA1.5-7B and 42.6 MME points on Qwen2.5-VL, while improving POPE accuracy by 2.74 percentage points in the strongest setting.