Efficient Multimodal Inference

As Omni-modal large language models (Omni-LLMs) take in audio, video and text together, their KV cache memory cost grows. KV cache quantization is the de facto approach in text-only LLMs, but its application to Omni-LLMs remains unexplored. In this paper, we analyze how TurboQuant, a representative rotation-based KV cache quantization method, behaves on multimodal caches and identify two critical issues: temporal key drift and heterogeneous value geometry. To address these, we propose OmniKVQuant, a training-free framework that (i) sets the key quantization range over each short window of the input stream; and (ii) rotates values separately per modality. On Qwen2.5-Omni and Qwen3-Omni, OmniKVQuant enables 2-bit KV caches while substantially preserving performance across seven audio-visual benchmarks. We further provide a fused Triton decode kernel that unpacks the 2-bit cache during attention, so no dense FP16 cache is ever built. Code: https://github.com/kaistmm/OmniKVQuant

Universal multimodal embedding (UME) increasingly demands encoder's capacity for handling a broad range of tasks and modalities with increased complexity. Prior scaling methods either increase the representation size, retrieval effort, or scales the encoder into a heavy multimodal LLM. Recent works, such as Think-Then-Embed (TTE), explore scaling via reasoning tokens. However, embedding models are hard to scale up: increasing parameters directly tradeoffs for the large training batch size that contrastive learning needs, and retrieval has to be served under tight latency. Moreover, UME tasks are diverse in complexity, where scaling up embedders can bring significant redundant computation. In this work, we propose MOEMB, which instead scales UME along the expert axis through mixture-of-experts (MoE), growing encoder capacity while preserving single-vector, non-autoregressive encoding. Through a systematic study of the design space and training recipes for MoE-based UME, MoEMB sets a new state of the art on both MMEB-V2 and MRMR among models trained on public MMEB-family data: with only 3B active parameters, MoEMB surpasses TTE-based methods with >4x active parameters, using significantly less computes. To further improve the scalability and efficiency, we conduct the first comprehensive study of adaptive computation for MoE-based embedding, spanning diverse strategies across training-based and inference-only methods. Together, these results support expert scaling as an effective and efficient direction for UME, with adaptive computation further improving efficiency for MLLM-based embedding models towards large-scale retrieval and recommendation systems.

Recent multimodal Speech Emotion Recognition (SER) systems achieve high accuracy through interaction-heavy cross-modal transformers, but their computational cost limits deployment in latency-sensitive and resource-constrained surveillance systems. To address this challenge, we propose RAFM_SER++, a lightweight multimodal SER framework featuring an asymmetric Residual Attention Fusion Mechanism (RAFM). Rather than relying on computationally expensive bidirectional interactions, RAFM injects affective speech cues into semantic text representations through a one-directional residual attention pathway. Combined with a BYOL-inspired cross-modal alignment objective and attention-guided pooling, the proposed framework improves multimodal representation learning while maintaining low computational overhead. Experiments on the IEMOCAP and ESD benchmarks demonstrate that RAFM_SER++ consistently outperforms the HuBERT-Base baseline and achieves a superior accuracy-efficiency trade-off compared with the state-of-the-art MemoCMT. Specifically, RAFM_SER++ reduces trainable parameters by more than 60%, achieves faster inference (79.60 it/s), and attains BACC scores of 81.10% on IEMOCAP and 95.39% on ESD. These results indicate that lightweight asymmetric multimodal fusion is an effective alternative to interaction-heavy cross-modal transformers for real-time surveillance applications.

Recently, multimodal large-scale reasoning models have demonstrated remarkable capabilities in solving complex tasks through long Chains-of-Thought (M-CoT). However, excessively long reasoning trajectories incur substantial computational costs and significant KV-cache pressure. Existing CoT compression and alignment paradigms mainly rely on static rules or single-dimensional preferences, lacking fine-grained cross-modal constraints; as a result, they are prone to inducing visual laziness and hallucinatory reasoning. To address these issues, we propose Modality-Contrastive Preference Optimization (MCPO), a highly sample-efficient two-stage length-compression method that requires fewer than 900 training samples. In the compression stage, we introduce a step-level Normalized Cross-Modal Mutual Information (NCMI) pruning algorithm, which automatically identifies and removes visual-independent reasoning steps by comparing the reasoning discrepancies between with-image and no-image contexts. This significantly reduces redundancy and hallucinatory content in the reasoning chains. In the alignment stage, the model first undergoes supervised fine-tuning to achieve domain-adaptive initialization, followed by optimization using an asymmetric multimodal length-controlled preference loss. This objective adopts a highly nonlinear odds-ratio formulation that provides steep gradients in the with-image context to reinforce length constraints for preferred trajectories, while applying a scaled, flat-gradient linear difference in the no-image context to maintain modality consistency, thereby achieving stable cross-modal preference alignment. Extensive experiments on mainstream base models such as Qwen3-VL-Thinking show that our method can reduce CoT length by up to 69.5% and achieve up to 3.34x end-to-end inference speedup while preserving original accuracy.

Streaming video understanding is a critical capability for real-world applications, including embodied intelligence, autonomous driving, industrial monitoring, surveillance and early warning, and wearable assistants. However, processing continuous video streams with multimodal large language models (MLLMs) is computationally expensive. Existing efforts have explored reducing streaming overhead through visual token pruning, token merging, quantization, on-demand frame retrieval, and context offloading. However, most existing methods overlook the dimension of model depth. Repeatedly executing full-depth MLLM prefill over incoming frames is prohibitively expensive, incurring substantial computational overhead and causing the KV cache to grow at a rate directly proportional to the prefill depth. To address these challenges, we propose ShallowStream, a novel framework that leverages the shallow layers of an MLLM to simultaneously perform frame encoding and retrieval index building. During stream processing, ShallowStream maintains an always-on lightweight index using the KV cache of shallow layers. During query-time answering, we leverage the attention scores generated by the shallow layers to score context frames and employ a diversity-aware selection strategy to retrieve precise and comprehensive evidence. ShallowStream achieves performance on par with the strongest existing streaming methods, while reducing per-frame prefill latency and 10-second end-to-end latency by up to 52.1x and 11.9x, respectively. Our code is available at https://github.com/CURRENTF/ShallowStream.

Multimodal memory offers a scalable interface for long-video question answering, but existing methods often retrieve captions, frames, transcripts, summaries, or graph facts as isolated fragments. Although searchable, such fragments are not generation-ready: language models must reconstruct cross-modal and temporal alignments at inference time, when context is limited and attribution is difficult. We propose EM^2Mem, an event-centric multimodal memory framework that binds heterogeneous evidence to event anchors during memory construction. Each event-indexed memory cell aligns multimodal records, temporal context, graph-linked relations, semantic facts, and provenance, enabling compact evidence readout over grounded multimodal events rather than modality-specific fragments. Across three long-video QA benchmarks, EM^2Mem improves average accuracy over the strongest memory baseline by 2.0, 2.4, and 3.7 points, improves strict event-level Top-5 evidence recall by 7.0 points, and reduces per-query latency by 4.67 times and total inference tokens by 63.66% (The code will be integrated into https://github.com/zjunlp/LightMem).

Multimodal large language models (MLLMs) struggle with fine-grained Visual Search, the task of locating small or rare objects in high-resolution images. Existing remedies fall into two families: (1) Training-free methods based on attention or confidence scores are accurate but slow, since they require multiple MLLM queries per example. (2) Reinforcement Learning (RL) trained tool-use models are faster at inference but opaque, since their tool calls remain uncontrollable and hard to interpret. To overcome this, we propose \emph{VisLens} (Visual Focus via Logit Lens), a Visual Search method built on the logit lens, which decodes the semantics held in a hidden state by projecting it through the LLM head. VisLens further uses a lightweight tuned-lens that maps early hidden states into the final hidden state space, so visual tokens can be read out from early layers. These tokens are matched to target words in the query to generate a crop of the relevant region, which is fed back in alongside the original image to produce the final answer. The whole process, from decoding to the final answer, completes in a single forward pass without repeated queries. VisLens matches or exceeds prior baselines while delivering a substantial latency advantage, running $8.5$--$9.9\times$ faster than Thyme and up to $22.2\times$ faster than training-free multi-pass search methods.

Real-world knowledge resides in multimodal documents, necessitating retrieval-augmented generation (RAG) for accurate question answering. However, existing multimodal RAG models are primarily designed for single-image or closed-document settings and exhibit limited accuracy in realistic multi-image scenarios. Moreover, processing numerous retrieved images incurs substantial computational overhead from irrelevant visual tokens. To address these challenges, we introduce DocLongRAG, a large-scale dataset of 343K question--answer pairs, each associated with an average of 37.4 retrieved images to reflect authentic RAG workflows. Building on this dataset, we propose Doc-REFRAG, a question-guided framework that compresses visual tokens into coarse chunks and selectively expands question-relevant ones via a lightweight RL-based selector. Experiments on six benchmarks show that Doc-REFRAG outperforms eleven strong baselines, achieving state-of-the-art accuracy with significantly lower inference latency. Our resources are available at https://github.com/Collab-Gen/Doc-REFRAG.

Unified multimodal models jointly support understanding and generation, but incur substantial redundant computation across tokens, layers, and generation timesteps. Through token-importance probing, we identify an asymmetric core-expansion structure: understanding exhibits a stable importance component, while generation largely shares this component but requires progress-dependent corrections. We therefore propose CE-Router, which uses a task-shared core scorer and progress-conditioned generation expansions, optimized through generation decomposition and cross-task core alignment. At inference, CE-Router compacts token computation and supplies a learned routing signal to Unified Computation Scheduling, which coordinates layer skipping, FFN pruning, diffusion-head cache reuse, and denoising-step early exit. Experiments on two representative UMM architectures demonstrate consistent quality--efficiency improvements across both tasks, retaining 98.03% of dense understanding performance with a 1.93$\times$ end-to-end inference speedup.

A multimodal system may begin inference while holding only some of its inputs and may acquire the rest at a cost. With adaptive acquisition, the policy determines which inputs are ultimately observed, so we state risk control conditional on that terminal input pattern. Conditional calibration typically assumes that the grouping map is fixed independently of the calibration sample, a condition that policy-induced grouping does not satisfy. We characterize when pattern-conditional risk control remains valid and give two finite-sample constructions: threshold-free routing with calibration applied at the terminal pattern, and simultaneous validation of complete policy-pattern pairs, which lets calibration data select the deployed policy. A counterexample shows that validity proved for a calibration-independent grouping map need not transfer once the policy makes the terminal group calibration-dependent. We call the resulting method RouteCert. On a clinical electrocardiogram task with a staged, cost-ordered lead protocol, the deployed policy answers 71.2% of held-out patients, with an observed 7.4% disagreement with the cardiologist's diagnosis at 48.8% of the prespecified ordinal cost of acquiring each stage, and all three acquisition stages are validated separately. On masked multimodal benchmarks, validating pointwise at each terminal pattern holds observed worst-pattern selective risk, measured against the full-information reference decision rather than the true label, at 0.034, where a pooled design reaches 0.145 against a 0.10 cap, at a comparable answered fraction (0.350 vs 0.342); under the budget-matched simultaneous comparison, the answered fraction falls to 0.305.

Medical AI has demonstrated specialist-level diagnostic accuracy, yet these capabilities remain largely inaccessible in resource-constrained rural settings where bandwidth is scarce, compute is limited, and clinical decision-making requires integrating heterogeneous modalities. We introduce a cloud--edge collaborative architecture that addresses these constraints: lightweight, domain-specific models on the edge transform raw medical data into compact structured outputs, while a cloud LLM synthesizes these outputs into clinical summaries. An LLM-based orchestrator dynamically selects diagnostic tools based on patient context, promoting comprehensive modality coverage without processing irrelevant inputs. We evaluate on 20 multimodal clinical cases spanning cardiac, obstetric, trauma, and screening scenarios under three simulated network profiles (500,kbps--5,Mbps). The hybrid system achieves 98--99% diagnostic tool recall with 92--96% precision, matches or exceeds cloud-only baselines on clinical accuracy, and maintains bandwidth-invariant latency (25--35,s) at 4--15x lower token cost. These results highlight the role of architectural design in enabling efficient multimodal integration and improving factual grounding compared to cloud-only approaches under deployment constraints.

Omnimodal language models (OLMs) enable unified audio-visual understanding, but processing long joint token sequences makes inference computationally prohibitive. While recent token compression methods attempt to alleviate this burden, compressing modalities in isolation often destroys the temporal cross-modal anchors necessary for coherent reasoning. We introduce Omni2LoRA, a two-stage framework for efficient parametric memory compression via coherence-preserving context distillation that bypasses the token bottleneck entirely. First, a Perceiver hypernetwork processes intermediate representations from a frozen OLM to encode the multimodal context into a full-rank Low-Rank Adaptation (LoRA) adapter in a single forward pass. To prevent the resulting parameter footprint from scaling linearly with recording length, we optimize a discrete rank allocation policy via Group Relative Policy Optimization (GRPO) that uses a modality-ablated counterfactual reward to explicitly penalize the loss of audio-visual coherence, forcing the model to allocate its fixed sub-linear rank budget to synergistic cross-modal anchors rather than isolated visual features. Across three omnimodal backbones, Omni2LoRA operating at a 30% rank budget outperforms direct full-context inference and strong token-compression baselines (OmniZip, OMAC, O-MARC) on four audio-visual question answering benchmarks, improving average accuracy by 8-12% over the strongest baseline and remaining stable under compression ratios as tight as 75%, where token-pruning methods degrade sharply. By converting multimodal memory into a fixed-budget, reusable parameter state, our method drives answer-time multimodal-token load to zero, cutting per-query Time to First Token (TTFT) by up to 12x relative to full-context inference and amortizing to under 0.5s after a handful of queries.

Omni-modal LLMs jointly process audio, video, and text, but long multimodal sequences incur substantial prefill and KV-cache costs. Existing omni-modal compression methods primarily focus on pre-LLM token reduction, leaving modality-specific compression across the LLM boundary underexplored. We propose A-PACK, a two-stage framework that defers audio pruning until query-conditioned multimodal interactions emerge. Our analysis shows that audio exhibits higher task-relevant information density and representational diversity per token than video. We further find that local audio-visual dynamics provide a more effective cue for visual selection than token-wise matching. We therefore preserve audio and compress video with local dynamics before the LLM, then progressively prune low-relevance audio and visual tokens and their KV-cache entries inside the LLM. Across four benchmarks on Qwen2.5-Omni-7B/3B, A-PACK achieves the strongest average performance among the evaluated prior methods while reducing prefill FLOPs by up to 78% and improving decoding throughput by up to 2.21x.

Existing streaming multimodal models process observations incrementally but still follow a turn-based prefill-then-decode pattern, making them non-duplex: new observations cannot naturally enter an active generation stream. Proactive alternatives use micro-turn polling or external response gates, which fragment continuous interaction, decouple response timing from language generation, and complicate KV-cache-friendly serving. We introduce Aero Realtime, a 4B streaming multimodal model with a duplex architecture for realtime generation. Aero Realtime aligns video, audio, and textual output on a shared temporal grid, where each approximately 80-ms audio slot predicts either a lexical token or a silence token. This allows input and output to advance together, enabling one autoregressive objective to learn both when to respond and what to generate. During inference, Aero Realtime appends only the newest multimodal slot, carries forward the previous output state, and reuses the KV cache for efficient incremental execution. We further provide a complete training and serving recipe, including realtime QA construction, slot-aligned supervision, hardware-aware distributed training, and resumable inference. On four NVIDIA A6000 workstation GPUs, Aero Realtime maintains 84-ms median and 173-ms P95 processing lag over 20 minutes of a continuously streamed video, remaining within 200~ms of the source timeline. These results demonstrate the feasibility of fully aligned input-output modeling for duplex, proactive, and hardware-aligned multimodal interaction.

Multimodal temporal data are inherently irregular and uneven in information density, yet most models rely on uniform discretization, leading to inefficient representations. We propose \textbf{EvtGraph}, a unified framework that aligns computation with temporal salience under explicit budget constraints. EvtGraph reparameterizes sequences into event-level tokens via event-adaptive compression (EAMC), selects a compact subset with a node budget (NBC), and performs temporally constrained sparse graph reasoning (T2SG). This transforms dense sequences into structured computation over salient events, reducing complexity while preserving critical transitions. We show that this design provides a practical mechanism for allocating representational capacity under a fixed budget, yielding a consistent performance--efficiency trade-off, where a small budget is often sufficient in practice. Experiments on multimodal clinical (MIMIC-IV + CXR) and cross-domain benchmarks demonstrate that EvtGraph outperforms both Transformer-based and recurrent baselines while significantly improving efficiency. These results suggest that budget-constrained event-centric representation provides a general paradigm for learning from high-redundancy temporal data.

Omni-modal large language models (Omni-LLMs) have achieved remarkable performance on audio-visual understanding tasks, but processing long and highly redundant visual and audio token sequences incurs substantial computational overhead, demanding aggressive token compression for efficient deployment. Existing methods often degrade at low token budgets: pre-LLM compression may discard structurally important and globally distributed evidence, whereas inner-LLM compression often underexploits query-conditioned audio-visual collaboration. To address these limitations, we propose OmniPack, a training-free framework that coordinates structural compression before the LLM with task-relevant semantic refinement within the LLM. Before the LLM, OmniPack removes structural redundancy through modality-specific importance, global coverage, and similarity-aware merging. After sufficient multimodal interaction, it further consolidates diverse, task-relevant representations through textual guidance and audio-visual collaboration. Extensive experiments on five benchmarks with three Omni-LLM backbones demonstrate that OmniPack consistently achieves the best performance-efficiency trade-off across diverse retention ratios, outperforming all existing methods. Notably, on Qwen2.5-Omni-7B, OmniPack preserves 98.0% of the original performance while reducing FLOPs to 16.7%, and still retains 92.9% of the original performance with only 6.8% of the original FLOPs.

In multimodal crack segmentation for industrial facilities, the key challenge is preventing missing modalities from degrading pixel-level performance while maintaining low computational cost. Existing methods struggle to address semantic degradation caused by missing modalities. We propose Compass, a lightweight network for robust crack segmentation under arbitrary missing modalities. Compass comprises Degradation Simulation Distillation (DSD), Needle Block, and Evidential Topology-Preserving Fusion (ETPF). DSD constructs a degradation simulation stream that mimics more severe missing conditions and performs reciprocal distillation with the original stream, decoupling complete perception from degradation adaptation. Within DSD, Feature-Aware Prototype Transmitter (FAPT) performs modality agnostic prototype-guided feature completion to maintain semantic integrity under incomplete modality conditions. As a lightweight backbone, Needle injects crack-direction cues into WKV modulation and combines connectivity-aware gating with anisotropic context probing for structure-aware modeling. ETPF fuses multimodal features via Dempster-Shafer evidential combination with uncertainty-gated decoding, preserving crack topology while suppressing unreliable features. Experiments on three datasets demonstrate state-of-the-art (SOTA) performance under diverse missing modality scenarios. Even with 90% depth modality missing on CrackDepth, Compass achieves F1 of 0.8216 and mIoU of 0.8434 with only 2.58M parameters. The code is available at https://github.com/Karl1109/Compass.

Reasoning in Multimodal Large Language Models (MLLMs) requires both fine-grained visual perception and rigorous logical deduction. Explicit text-based Chain-of-Thought (CoT) is computationally expensive and prone to visual hallucinations, while existing latent reasoning methods typically require costly training. Furthermore, directly adapting training-free LLM reasoning mechanisms to the multimodal setting yields unstable performance. We identify that this failure stems from their reliance on token-level entropy, which fundamentally conflates perceptual ambiguity (e.g., unclear visual details) with logical uncertainty (e.g., complex reasoning steps). To overcome this bottleneck, we present a novel training-free inference strategy for MLLMs that explicitly decouples perception and reasoning. We propose a novel metric, the vision-to-text attention ratio, to dynamically gauge the model's cognitive focus. Guided by this metric, our proposed framework, Attention-Guided Switching (AGS), adaptively triggers latent reasoning for perceptual tokens to preserve high-fidelity visual information in the continuous space, while enforcing explicit text generation for logical tokens to maintain structural anchoring. Extensive experiments demonstrate that our method achieves state-of-the-art performance, significantly improving both accuracy and inference efficiency by reducing autoregressive steps and latency. Code is released at https://github.com/swordAndSnow/MM26-AGS.

MLLM-based segmentation faces a core segmentation trilemma: high segmentation performance, preserved dialogue ability, and fast inference. Embedding-prediction methods may disrupt language modeling through pixel-level objectives, whereas next-token generation is inefficient for dense masks. We propose All-Mask Prediction, decoupling autoregressive dialogue from non-autoregressive mask prediction. Its binary instantiation, STAMP (Simultaneous Textual All-Mask Prediction), emits an in-vocabulary <SEG> trigger, fuses image-aligned mask tokens with corresponding patch features, and uses hybrid attention to classify all tokens as foreground or background in one pass. It thereby combines strong referring and reasoning segmentation with preserved multimodal ability and efficient inference. However, binary masks cannot retain multiple semantic or instance identities without repeated target-specific predictions. We therefore propose Structured All-Mask Prediction and develop STAMPlus. It generates a target list with explicit IDs and optional boxes, binds these IDs to a shared multi-class mask space, and jointly predicts all targets in one non-autoregressive pass. A single unified checkpoint retains STAMP's referring and reasoning capabilities while extending to open-vocabulary semantic, instance-aware, and remote-sensing small-target segmentation, where high-resolution mask-token scaling preserves finer spatial evidence. Across these settings, STAMPlus achieves state-of-the-art segmentation performance, preserves general multimodal instruction following, and reduces 12-category latency from 13.50s for repeated STAMP inference to 5.16s. Further analyses show that accurate target cues improve segmentation and learned spatial grounding benefits look-twice reasoning. Overall, STAMPlus resolves the trilemma beyond single-target prediction.

Visual token pruning reduces the inference cost of multimodal large language models, but a fixed token ratio is poorly matched to text-rich inputs. In OCR-centric tasks, decisive evidence can be a small number, label, or field whose relevance is specified by the question; indiscriminate pruning can erase that evidence while retaining visually salient but irrelevant regions. We present ET-Prune, a training-free framework that casts pruning as evidence allocation. It derives question-conditioned evidence from a decoder-side partial query-key block, safeguards text-like spatial regions, and converts evidence uncertainty and density into a sample-specific token floor. Three progressive middle-layer events then move the sequence toward this budget, retaining more tokens for diffuse or text-dense evidence and pruning concentrated evidence more aggressively. At the observed point estimates from one deterministic pass per configuration, ET-Prune leads or ties among pruned methods in all six backbone-benchmark comparisons at roughly half tokens. On OCRBench-v2, it leads the strongest pruned baselines by 1.80 and 0.68 percentage points on Qwen3-VL-8B and InternVL3.5-8B, respectively, while retaining about half of the visual tokens; on MMBench v1.1, it reaches 0.8467 circular exact-matching accuracy versus 0.8437 for Vanilla at 54.45% average visual-token retention. These results show a favorable observed quality-cost trade-off for evidence-aware dynamic budgeting in text-rich multimodal inference.

Unified multimodal object tracking has achieved remarkable robustness by leveraging complementary sensor data (e.g., RGB, Thermal, Depth), yet the heavy computational burden of state-of-the-art models hinders their deployment on resource-constrained edge devices. In this work, we identify the prediction head as a critical but often overlooked efficiency bottleneck. By strategically streamlining the decoder architecture, we unlock the potential for real-time inference but simultaneously introduce a capacity gap between the lightweight student and the heavy teacher. To resolve this, we conduct a systematic analysis of 17 distillation strategies and introduce a Dual-Alignment Distillation framework. Our key insight is that effective compression requires decoupling knowledge transfer into two complementary streams: (1) Spatial Representation Alignment, which employs feature distillation to sharpen the student's spatial focus on foreground targets ("Where to track"); and (2) Semantic Distribution Alignment, which utilizes logit-based distillation to align decision boundaries and transfer discriminative dark knowledge ("What to track"). Extensive experiments across five benchmarks demonstrate that our approach significantly outperforms complex state-of-the-art methods. Notably, our distilled model achieves 91.5% MPR on RGBT234 and operates at 54 FPS on a single RTX 4090, representing a 5x speedup over the teacher model while maintaining superior accuracy.

Rendering source code as images offers a promising way to reduce the input costs of Multimodal Large Language Models (MLLMs). Adjusting image resolution can trade visual token cost against content fidelity. However, resolution scaling alone overlooks two sources of inefficiency: blank regions created by line breaks and indentation, and code regions irrelevant to the current instruction. Moreover, the best compression setting varies across inputs, tasks, and models, limiting fixed-ratio strategies. We propose CodeShrink, an adaptive visual compression framework with three components. Blank-Free Rendering replaces whitespace-dependent layouts with compact layouts and explicit structural markers, removing layout-induced tokens. Adaptive Compression Configuration uses a lightweight agent trained with reinforcement learning to predict a per-input setting that balances token efficiency and readability. Dominant Token Selection jointly analyzes the instruction and code image to prune task-irrelevant visual tokens during inference. We evaluate CodeShrink on code question answering, clone detection, and code completion. CodeShrink reduces visual token use by up to 71.2% while matching or exceeding uncompressed text-only inputs, and consistently outperforms text-based and visual compression baselines across all three tasks. These results show that combining layout compaction, adaptive configuration, and instruction-aware pruning can make multimodal code understanding more efficient. Our code is available at https://github.com/vinsontang1/CodeShrink.

Universal multimodal embedding (UME) maps heterogeneous multimodal inputs into a shared embedding space. Existing UME models either form embeddings through single forward encoding or add computation through explicit rationale tokens and latent autoregressive states. Although token expansion can improve complex matching, serial generation increases retrieval latency and makes the final embedding depend on generated intermediate states. This raises a different question: can useful computation be expanded along model depth while keeping the token workspace fixed? We analyze positive-negative similarity separation at every layer of independently trained UME models and observe a shared progression: early layers contextualize multimodal inputs, a contiguous middle-to-late stage forms retrieval-discriminative features, and the final layers map them into the embedding space. Based on this finding, we propose ReLoop-UME, which executes the early layers once, recurrently reuses a parameter-shared retrieval-forming block, and applies the final mapping layers after the last loop. Learnable Retrieval Registers provide persistent retrieval-specific states that accumulate and exchange evidence across loops, with the final register serving as the embedding readout. On MMEB-V2 and MRMR, ReLoop-UME consistently improves retrieval across different backbones while running 44.9x faster than UME-R1 and 1.5x faster than PLUME.

While visual token pruning is essential for efficient Multimodal Large Language Models (MLLMs), existing training-free methods suffer from a critical limitation: they rely on static, instantaneous heuristics to perform irreversible filtering. This approach ignores the hierarchical nature of MLLMs, where token importance often evolves dynamically rather than remaining fixed across layers. Consequently, tokens essential for deep-layer reasoning are often prematurely discarded by shallow-layer estimates. To address this, we propose Trend-aware Pruning, a novel framework that elevates pruning from a local snapshot decision to a temporal trajectory modeling problem. Instead of relying on isolated scores, our method captures the momentum of attention flow. This enables a dynamic rectification mechanism that selectively reactivates "late-blooming" tokens, those initially undervalued but exhibiting rising semantic importance, thereby preventing the loss of critical visual cues. Extensive experiments demonstrate that our approach achieves a superior efficiency-performance trade-off across diverse multimodal tasks. Notably, it reduces visual tokens by over 77.8%, retaining only approximately 23 tokens in the final layer while maintaining competitive performance, offering a robust and reversible solution for high-efficiency multimodal inference.

Question answering (QA) over irregular clinical time series (ICTS) plays a pivotal role in a wide range of healthcare applications. Although recent multimodal time-series large language models (LLMs) have shown considerable promise in general-purpose time-series QA, they remain poorly equipped to model the sparsity, asynchrony, and irregular sampling patterns of clinical observations. To fill this gap, we propose ClinPRISM, a cost-effective multimodal LLM reasoning framework for question answering over ICTS data. First, we devise an irregularity-aware multi-scale encoder to capture sparse clinical evidence at diverse temporal scales. Then, we propose a temporal evidence distiller to integrate representations across these scales and compress them into a small number of LLM-compatible tokens. Moreover, we introduce a progressive alignment strategy that sequentially aligns the irregular trajectories with the LLM's textual embedding space. To facilitate training, we construct 30,000 clinical time series paired with multi-scale descriptions, together with 41,000 instruction-tuning instances spanning 11 tasks. Using a 4-billion-parameter LLM backbone, ClinPRISM achieves state-of-the-art performance on the held-out evaluation benchmark while using only 16 time-series tokens and achieving an average inference latency of 0.15 seconds per question.

Recent multimodal large language models (MLLMs), such as Qwen2.5-VL and InternVL3, generate large numbers of vision tokens for high-resolution inputs, leading to substantial computational cost. Existing vision token pruning methods either depend on cross-modal attention and cannot prune before the prefill stage, or rely on diversity estimation with high computational overhead. We observe that attention scores from both vision and text tokens peak at modality separator tokens, suggesting that these separators bridge the two modalities. Based on this observation, we propose SepPrune, an efficient, training-free, plug-and-play pruning method that uses the separator token as a unified query to rank and select informative vision tokens. SepPrune reuses the LLM's built-in projection parameters and requires no architectural changes. Experiments on Qwen2.5-VL-7B show that SepPrune achieves state-of-the-art performance, retaining 96.3% of the original accuracy while removing 80.2% of vision tokens.

Emerging Omni-modal Large Language Models (OmniLLMs) enable unified understanding of text, audio, and video, but their long audio-video token sequences introduce substantial memory and inference costs. Existing compression methods mainly focus on selecting important tokens under fixed budgets, leaving the preceding budget-allocation problem underexplored. We show that direct query-to-audio/video similarity is unreliable for inter-modal budget allocation, and that uniform intra-modal budgets can miss key evidence while retaining redundant content. To address these limitations, we propose OmniDelta, a training-free, skill-driven framework that couples intent-aware inter-modal allocation with content-aware intra-modal allocation. OmniDelta first constructs audio and video skill pools to shift the fixed retained-token budget according to query demand, then reallocates modality budgets over audio segments and video frames using local complexity and temporal redundancy. The resulting local budgets can be combined with existing pruning strategies, preserving the total retained-token ratio while changing where the budget is spent. Experiments on four audio-video benchmarks with two Qwen2.5-Omni models show that OmniDelta establishes a new accuracy-efficiency Pareto frontier across pruning ratios. At 25% token retention on Qwen2.5-Omni-7B, OmniDelta reduces GPU memory by 22.0% and achieves a 1.64x end-to-end speedup over full-token inference.

Knowledge-intensive multimodal question answering (KI-MMQA) sits at the intersection of three expensive primitives: long visual token sequences, dense retrieval over large external corpora, and full cross-modal fusion. Existing systems pay all three costs uniformly per query, even though only a small fraction of visual content and retrieved knowledge is actually relevant to any given question. We introduce SKIP (Salient Knowledge-Injected Pathways), a unified inference architecture that routes computation along sparse pathways jointly conditioned on the question, the image, and a difficulty estimate. SKIP combines question-guided visual token pruning, region-conditional sparse retrieval, bipartite sparse cross-attention, and speculative knowledge verification with an adaptive budget controller that allocates compute proportional to predicted question difficulty. We derive an information-bottleneck bound showing that the optimal visual sparsity rate scales as $O(1/\sqrt{N})$ under realistic question-image mutual-information assumptions, with retained accuracy guarantees. Across five KI-MMQA benchmarks (OK-VQA, A-OKVQA, InfoSeek, Encyclopedic-VQA, and ViQuAE), SKIP matches or exceeds the accuracy of strong dense baselines while using $3.4$--$6.8\times$ fewer FLOPs and $2.7\times$ less end-to-end latency. Code available at: https://pmlrbd.github.io/skip/

Multimodal large language models (MLLMs) have enabled long-form video understanding at a scale that was not previously possible. However, the density of relevant content decreases sharply as video sequence length increases, and exposing the model to more irrelevant content measurably reduces its accuracy. In this paper, we address the problem of maximizing query-relevant information in a frame subset selected at inference time, without training. FORGE (Frame Orthogonality in Relevance Geometry) is a model-agnostic method that induces a query-conditioned geometry on a pretrained multimodal embedding space, unifying relevance and diversity into a single objective. In this space, frames that cover independent query-relevant directions are far apart, and selecting the subset of maximum information captures diverse query-relevant content within the budget. Experiments on Video-MME and LongVideoBench at budgets of 16, 32, and 64 frames show that FORGE improves the unified keyframe selection score by 11.0-15.3 points over the strongest training-free baseline and up to doubles keyframe recall (0.415 vs. 0.204 at K=64 on Video-MME). The gains extend to question answering, where accuracy improves in every evaluated setting across eight open-source MLLMs spanning 4B to 32B parameters, by up to 8.7 points over uniform sampling and 5.2 points over the strongest baseline. Our findings suggest that aligning the embedding space with the query's high-dimensional structure is a promising direction for inference-time video understanding.

Advancing multimodal retrieval-augmented generation (RAG) for complex document understanding presents a formidable dual dilemma of accuracy and efficiency, particularly in graph RAG. Processing structurally sparse yet visually dense layouts, such as extracting a tiny data marker from a financial chart, often incurs computationally prohibitive token overhead while still triggering catastrophic hallucination. However, multimodal Graph RAG pipelines rely on graph-construction stages that assume Vision-Language Models (VLMs) can resolve sparse semantics within high-density layouts. We challenge this assumption, revealing that forcing VLMs to localize visual evidence, interpret semantics, and extract relations triggers a "Visual Attention Sink," a mechanism driving catastrophic semantic loss, while full-page processing incurs massive computational overhead. Controlled interventions verify that this failure is boundary-driven rather than content-specific and that semantic anchoring mitigates it. To fundamentally correct this flawed paradigm, we introduce DeCoRAG, a multimodal Graph RAG pipeline that shifts knowledge processing from coupled visual-semantic reasoning to "Cognitive Decoupling." Rather than passively processing raw pixels, its graph-construction stage establishes a macroscopic Semantic Anchor to neutralize the attention sink. This anchor subsequently drives our Region-Aware Pruning and Cropping (RAP-Crop) mechanism, shifting the reasoning space from dense, noisy backgrounds to purified, intent-driven semantic clusters. The resulting graph supports hybrid retrieval and answer generation. Across complex document benchmarks, DeCoRAG improves the semantic pass rate by up to 12.5 percentage points over the strongest baseline and generalizes to DocVQA. RAP-Crop reduces offline graph-construction prompt tokens by 40.8% without sacrificing end-to-end accuracy.