每天知道
你的领域
出了什么

Despite recent progress, LLM agents still struggle with reasoning over long interaction histories. While current memory-augmented agents rely on a static retrieve-then-reason paradigm, this rigid pipeline design prevents them from dynamically adapting memory access to intermediate evidence discovered during inference. To bridge this gap, we propose MRAgent, a framework that combines an associative memory graph with an active reconstruction mechanism. We represent memory as a Cue-Tag-Content graph, where associative tags serve as semantic bridges connecting fine-grained cues to memory contents. Operating on this structure, our active reconstruction mechanism integrates LLM reasoning directly into memory access, allowing the agent to iteratively explore and prune retrieval paths based on accumulated evidence. This ensures that memory retrieval is dynamically adapted to the reasoning context while avoiding combinatorial explosion caused by unconstrained expansion. Experiments on the LoCoMo benchmark and LongMemEval benchmark demonstrate significant improvements over strong baselines (up to 23%), while substantially reducing token and runtime cost, highlighting the effectiveness of active and associative reconstruction for long-horizon memory reasoning.

Agent skills, which consist of reusable strategies that guide agent reasoning and action, have shown strong potential for improving model capability at inference time. However, current skill construction methods treat the problem as one-shot extraction, overlooking a fundamental tension: a skill tailored to the specific task fails to transfer, while the abstracted skill often provides insufficient guidance. We attribute this fragility to the absence of explicit mechanisms for skill specification and generalization. To address this gap, we introduce SkillComposer, a framework that decomposes skill construction into three learnable operations: create, improve, and merge. Trained via systematic rejection sampling recipe, SkillComposer enables language models to self-evolve skills at inference time and supports three deployment modes: offline for building generalized libraries, online for task-specific refinement, and hybrid for combining both. Comprehensive experiments on $τ^2$-Bench, LiveCodeBench v6, and AppWorld show that SkillComposer consistently outperforms baselines. Our SkillComposer-4B improves a 27B executor by up to +4.5 on agent tasks and +3.4 on code tasks, while generalizing across domains and task types unseen during training. Analysis reveals that merge and improve address orthogonal quality dimensions and that skill composition is a transferable meta-ability, providing a practical recipe for skill-augmented inference.

LLM agents are increasingly deployed on long-horizon tasks requiring sustained reasoning over extended interaction histories. Realizing this at scale requires agents to persistently store, retrieve, and update their own memory across sessions. A rich ecosystem of agent memory systems has emerged spanning flat retrieval, LLM-mediated extraction, consolidating fact stores, and agentic control flows. Yet, their system-level behavior remains uncharacterized. We present the first systems characterization of agent memory. First, we introduce a system-oriented taxonomy classifying agent memory systems along four axes. Second, we build a phase-aware profiling harness attributing cost to construction, retrieval, and generation. Third, we characterize ten representative systems across two benchmark suites, uncovering how design choices shift cost across the write and read paths. Finally, we derive 10 system recommendations covering construction scheduling, capability floors, amortization via query volume, freshness-latency tradeoffs, and fleet-scale management.

Recent Large Language Models (LLMs) have shown impressive reasoning abilities; but they are still susceptible to hallucinations, intermediate reasoning mistakes, and unreliable reasoning results in complex mathematical reasoning problems. In this study, we introduce a critic-based heterogeneous multi-agent approach to improve the dependability of mathematical reasoning. This framework incorporates several LLM agents of different specialties and employs a critic-driven adaptive learning system to assess and guide the reasoning process based on intermediate feedback. The system adopts a generator-validator framework, with the validator not only determining correctness but also offering critiques to guide regeneration of solutions. This allows for adaptive error correction and prevents error cascading. Our experiments on the GSM8K benchmark show that the proposed method achieves up to 13% accuracy improvement over single-shot and non-critic models. Additionally, findings suggest that heterogeneity and critique reduce the need for large models, allowing smaller models to perform on par. Ablation studies reveal the main performance gains are due to the critic-based feedback loop and not model size. In summary, the proposed approach showcases the benefits of combining heterogeneous multi-agent collaboration and critique to obtain reliable and interpretable reasoning systems.

For over half a century, software engineering has operated on a foundational premise: human engineers decompose problems, encode decision logic into static code, and manually adapt that code as requirements evolve. This paper argues that the emergence of AI agents — systems where large language models serve as the primary reasoning engine, dynamically generating and discarding code as an instrumental resource — constitutes not an incremental improvement but a fundamental restructuring of the software paradigm. Drawing on first-principles analysis of complexity scaling, we formalize the distinction between traditional software (where code is the carrier of decision logic) and agentic systems (where code is ephemeral tooling for an LLM-driven reasoning loop). We trace the historical arc from licensed software to SaaS to what we term Agent-as-a-Service (AaaS), showing that each shift transferred additional complexity away from end-users. We introduce the concept of Agentic Engineering as an emergent discipline — distinct from software engineering in its core object of study, control model, and human role. Through analysis of recent benchmark evidence including SWE-bench Verified, EvoClaw, and LangChain's multi-agent coordination studies, we demonstrate both the transformative potential of the agentic paradigm and its current limitations. We conclude with a four-stage roadmap toward self-evolving agent ecosystems and concrete recommendations for practitioners navigating this transition.

Planning for real-world problems by language models often involves both world and user constraints, which may not be fully specified upfront and are progressively disclosed through interaction. However, existing benchmarks still underexplore adaptive planning under such progressively revealed dual constraints. To address this gap, we introduce AdaPlanBench, a dynamic interactive benchmark for evaluating whether Large Language Model (LLM) agents can adaptively plan and re-plan under progressively revealed world and user constraints. AdaPlanBench is built on 307 household tasks, with a scalable constraint construction pipeline that augments each task with dual constraints. At runtime, agents interact with the environment in a multi-turn protocol where hidden constraints are revealed only when the agent proposes a plan that violates them, requiring iterative plan revision under accumulating feedback. This makes planning challenging, as agents must infer and track constraints from feedback while re-planning effectively. Experiments on ten leading LLMs show that adaptive planning under dual constraints remains challenging, with the best model reaching only 67.75% accuracy. We further observe that performance degrades as more constraints accumulate, with user constraints posing a particularly large challenge and failures often stemming from weaker physical grounding and reduced effectiveness. These results establish AdaPlanBench as a testbed for dual-constrained interactive planning and highlight the challenge of reliable adaptation to dynamically revealed constraints in LLM agents.

Recent advances in LLM agents have enabled complex cognitive capabilities, such as multi-step reasoning, planning, and tool use, that increasingly position these agents as human collaborators. Effective collaboration, however, requires collaborators to continuously maintain and align mental models of their own reasoning,partners' intentions, and shared goals during the collaborative process. Today's agents rarely develop such capabilities since they are primarily optimized for task completion, and the community lacks authentic human collaboration data with action-level mental model annotations that could guide agents toward process-level collaborative competence. To bridge this gap, we present ALMANAC, a dataset of Action-Level Mental model ANnotations for Agent Collaboration built from the Map Task, a classic dyadic routing task from social science. ALMANAC contains 2,987 collaboration actions, each paired with theory-informed mental model annotations that record the participants' self-reasoning, perceived partner intent, and perceived team goal. We benchmark six LLMs on predicting humans' next-turn behavior and mental models. Our results demonstrate ALMANAC's utility in evaluating models' ability to simulate human collaborative behaviors and infer their underlying mental models.

LLM-based guardrails typically safeguard agents by evaluating proposed actions or inputs before execution, producing safety signals such as binary allow/deny decisions, risk categories, and/or explanatory rationales about potential policy violations. However, agent risks often arise when otherwise benign tasks are contaminated by untrusted external content, unsafe instructions, or risky tool use. Existing guardrails often flag the entire task uniformly as unsafe, thereby blocking the threat but sacrificing the benign part. Moreover, existing work largely evaluates guardrails in isolation, leaving unclear whether their interventions lead to safer downstream agent behavior. To address this, we introduce TRIAD (Tripartite Response for Iterative Agent Guardrailing), a guardrail-integrated agent framework that leverages guardrail-generated verbal feedback as a guiding signal to keep the agent aligned with benign objectives at each planning step. We finetune a language model on a self-curated training dataset to output one of three decisions: proceed, refuse, or update, together with structured natural-language feedback. Rather than merely allowing or blocking execution, update guides the agent to revise its plan, avoid harmful components, and preserve the benign task where possible. TRIAD injects this feedback into the agent's context, enabling subsequent plan revision and forming a closed loop between guardrail feedback and agent planning. Extensive experiments on ASB and AgentHarm show that TRIAD reduces the average attack success rate to 10.42%, while achieving the best safety-utility trade-off among guardrail-integrated baselines. Our code is available at: https://github.com/YUHAOSUNABC/TRIAD.

Does adding more agents help an LLM workflow once compared systems share the same benchmark loader, tool access, answer contract, usage accounting, and trajectory logging? We introduce BenchAgent, an evaluation framework that places single-agent, fixed multi-agent (MAS), and evolving MAS workflows under one normalized execution and logging protocol. BenchAgent evaluates these substrate-internal workflows across ten reasoning, coding, and tool-use benchmarks with GPT-4.1, and separately reports a Protocol-Aligned External (PAE) GAIA study of a runtime-generated workflow. Under SI conditions, at most one of six tested MAS exceeds the matched single-agent anchor on benchmark-balanced average accuracy: EvoAgent lies within the Wilson one-run guidance, while the remaining five trail by 2.56-11.29 points and occupy more expensive accuracy-cost trade-offs. On the PAE GAIA snapshot, a Claude-Code-style runtime workflow reaches 66.72% overall and 69.23% on Level 3, more than 20 points above the strongest non-Claude baseline, Jarvis, a fixed MAS.

Scaling reinforcement learning (RL) to diverse multitask settings remains a central challenge. While recent advances in model-based RL achieve strong performance, they rely on planning and complex training pipelines, making it unclear which components are essential for scalability. We revisit this question and argue that the primary driver of scalable multitask RL is not model-based control, but representation learning. In particular, we show that combining predictive, model-based representations with high-capacity value function approximation is sufficient to achieve strong performance, even without planning. We evaluate a simple model-free algorithm, MR.Q, coupled with auxiliary predictive objectives into a scalable actor-critic architecture. This approach outperforms a recent world-model-based method and a range of deep RL baselines across a diverse suite of multitask continuous control tasks, while significantly reducing computational overhead and improving wall-clock efficiency. We observe consistent improvements with increased model capacity and show through ablations that predictive representation learning is critical for performance.

We identify and formally characterize credit misassignment as a systematic failure mode of GRPO in tool-augmented multimodal search agents: its uniform broadcast of trajectory-level advantages to all tokens causes valuable tool-use steps in failing trajectories to be penalized no differently from valueless ones. We further empirically quantify the scale of this phenomenon. Over half of failing trajectories and failing tool-use actions exhibit correctable credit misassignment, demonstrating that the wasted training signal is both substantial and structurally exploitable. Building on this insight, we propose Tool-Aware Policy Optimization (TAPO), which exploits the parameter-determinism property of information-acquisition tools: similar call parameters define equivalent information-acquisition actions and should therefore share comparable action credit. TAPO constructs counterfactual witnesses within the current training batch and compensates misassigned negative credit via confidence-gated conservative advantage correction. It requires no additional annotation, models, or sampling, and introduces negligible computational overhead. Across multiple multimodal search benchmarks, TAPO delivers consistent, plug-and-play improvements over strong baselines for three mainstream RL algorithms (GRPO, GSPO, and SAPO). Our code and models will be publicly released upon acceptance.

Intra-vehicular robots in spacecraft help reduce astronaut workload and improve mission efficiency. Recent research focuses on using deep learning methods to achieve the acute control required for operations in these complex environments. However, objects exhibit unpredictable, unconstrained drift without gravitational damping. These factors demand robustness against complex multimodal action distributions. Diffusion policies (DP) can model these complex actions, but their iterative sampling process consumes too much energy for the limited power budgets of spacecraft. We therefore propose a low-energy intra-vehicular robotic manipulation framework, L-SDPPO, in which the Spiking Diffusion Policy (SDP) is optimized with a reinforcement learning (RL) algorithm. Furthermore, to address the insufficient perception of dynamic spatiotemporal features in microgravity, we propose the statedependent latency injection (SDLI) mechanism, which mimics biological neural delays to dynamically regulate the timing of input information. Evaluation on five representative intra-vehicular daily tasks (e.g., hatch opening and precision container capping) shows that our method consistently achieves higher success rates and lower energy consumption, compared to the state-of-the-art robotic manipulation methods. These results demonstrate our method is a viable intra-vehicular robotic manipulation method.

Text-guided image editing has advanced rapidly with diffusion models and unified multimodal foundation models. However, most existing methods remain confined to single-turn settings, overlooking the more realistic scenario of multi-turn in-context editing, where users iteratively refine an image through a sequence of instructions. In this setting, a model must follow each new instruction while preserving accumulated session-level constraints, challenged by two coupled failure modes: long-context dilution, where sparse textual constraints become difficult to recover from growing interleaved image-text histories, and state contamination, where earlier editing mistakes degrade subsequent generations. We introduce Edit-R2, a novel reinforcement learning post-training framework for unified multimodal models. Edit-R2 reconstructs the operative session intent, which effectively consolidates scattered historical constraints into an explicit reasoning trace before each editing turn. It further enables multi-turn RL over both reasoning and generation through a unified objective that jointly optimizes intent reconstruction generation in discrete text space and flow-matching image generation in continuous latent space, while a trajectory filtering mechanism suppresses corrupted rollouts to stabilize training under state contamination. To support systematic evaluation, we introduce MICE-Bench, a large-scale benchmark for multi-turn in-context editing with automated metrics for instruction following (IF), content consistency (CC), and global awareness (GA) over accumulated session constraints. Experiments show that Edit-R2 substantially improves multi-turn in-context editing and achieves competitive performance compared against strong baselines.

Scientific peer review generation has attracted increasing attention for reducing reviewing burdens and providing timely feedback. However, existing Large Language Model (LLM)-based methods often produce generic comments with insufficient evidence support and weak source traceability, while complex multi-agent systems incur high inference costs. To address these challenges, we propose EGTR-Review, an Evidence-Grounded and Traceable Review Generation framework via Multi-Agent Teacher Distillation. EGTR-Review first constructs a multi-agent teacher that performs structure-aware paper decomposition, key-element extraction, external scholarly evidence retrieval, evidence-state labeling, verification reasoning, and review synthesis. It then distills both intermediate reasoning trajectories and final review comments into a lightweight student model through task-prefix-driven multi-task learning. An evidence-weighted objective further reduces the influence of weak, missing, or non-verifiable supervision. Experiments on public peer-review datasets show that EGTR-Review (Student) outperforms strong prompt-based, fine-tuned, and structured/agentic baselines across automatic metrics, LLM-as-Judge evaluation, and human evaluation, while maintaining strong factual grounding and source traceability with substantially lower token consumption and inference time. Our code, prompts, configurations, and sample data are available on GitHub.

On-policy distillation (OPD) improves reasoning by training a student on trajectories sampled from its own policy under supervision from a teacher. In multimodal reasoning, a common extension is to use a privileged teacher that observes training-time-only signals such as reference answers or rationales. However, such answer-side privilege creates a train-test mismatch: the teacher's supervision may depend on signals unavailable to the student, encouraging shortcut imitation rather than visually grounded reasoning. We propose ViCuR, a visually grounded privileged-teacher distillation framework that replaces answer-side privilege with visual cues (query-related evidence in the input). Because these cues are derived from the same visual input available at inference, their evidence is recoverable by the student. To support this, ViCuR introduces a lightweight cue recovery module that uses dedicated sink-token cross-attention during prefill to aggregate task-relevant visual evidence into an internal representation, without changing the inference interface or requiring auxiliary cue-generation losses. Across seven benchmarks with Qwen3-VL-2B and 8B students, ViCuR consistently improves over answer-based on-policy self-distillation by +1.19 and +1.24 on overall average performance. It also extends naturally to stronger-teacher OPD, surpassing OPD baselines by +0.64 and +1.08, with consistent out-of-domain gains at the 8B scale. These results show that, in multimodal on-policy distillation, the design of teacher privilege is as important as teacher strength.

On-policy distillation (OPD) supervises the student only in output space by matching next-token probabilities. This output-only paradigm has two limits: (1) sampling variance from Monte Carlo KL estimates over large vocabularies (e.g., Qwen's ~150k tokens) persists throughout training, and (2) it treats the teacher as a black-box, discarding all intermediate hidden states after the LM head. We propose On-Policy Representation Distillation (OPRD), which lifts distillation into hidden-state space by aligning student and teacher representations across selected layers on the same rollouts, bypassing the LM head entirely. Theoretically, OPRD eliminates sampling variance and provides richer per-layer structural information. Empirically, OPRD closes the student-teacher gap on AIME 2024/2025 and AIMO, while output-space OPD baselines plateau below the teacher. OPRD also trains 1.44x faster and uses 54% less memory than top-k OPD. Code: https://github.com/ShenzhiYang2000/OPRD.

Autoregressive (AR) image generation models are highly expressive but computationally intensive, motivating effective model compression. Knowledge distillation (KD) is a natural approach for model compression and has been widely studied in language modeling, yet its behavior in visual AR generation remains underexplored. In this work, we present the first systematic study of distillation strategies for AR image models. Our analysis shows that while standard distillation can yield meaningful gains, recent methods developed for language do not directly transfer to images: long decoding horizons and visual token ambiguity make teacher supervision unreliable especially under student-conditioned contexts. To address this, we propose VarKD, a distillation framework for visual autoregressive models that distills on student samples while selectively applying teacher supervision and reducing token-level ambiguity. Experiments on ImageNet across multiple AR backbones show that VarKD consistently outperforms prior distillation baselines, narrowing the gap to large-scale models.

Diffusion-based vision-language-action (VLA) models often inherit the image-generation view: actions are generated by iterative denoising. We argue that VLA action generation has a different condition-target structure: the policy is conditioned on rich observations, language, and state, but predicts only a compact, low-dimensional action chunk. Under this asymmetry, strong one-step action generation should not necessarily require the advanced one-step methods developed for image synthesis. We keep standard velocity prediction and add no teacher model, distillation stage, or auxiliary objective; in our main recipe, we simply bias the training time distribution toward high-noise states. We first isolate the effect in a controlled MNIST grid-to-sequence task, then test it with extensive robot-policy experiments. Across standard LIBERO, LIBERO-Plus, and LIBERO-Pro, one-step policies trained with high-noise biased schedules generally match ten-step decoding under the same recipe, and on standard LIBERO can exceed ten-step policies trained with a uniform time distribution. A real-robot bimanual YAM RSS evaluation gives a small-sample cross-architecture check of the same sampler trend. On a 1.4B VLM model with a 30M action head, one-step decoding reaches 95.6% on LIBERO-Long. These results show that strong one-step VLA action generation can emerge from standard diffusion training, without importing the full few-step diffusion machinery developed for image generation.

Parametric retrieval augmentation encodes document information into lightweight, document-specific modules such as LoRA adapters, reducing the need to include all evidence as input context. However, it remains unclear how this parameter-side memory interacts with context-side memory stored in the KV cache. We study this interaction in document-level question answering by progressively evicting document key-value states and measuring when a document LoRA contributes beyond the retained context. We find that document LoRA adds little when the KV cache is largely intact, but becomes increasingly useful under aggressive compression, recovering 13-21 ROUGE-L points when no document context remains. The gain is largest when the base model encodes the document, and the adapter is applied only during answer generation, suggesting that document LoRA is better understood as decoding-time parametric memory than as a document encoder. Finally, QA-style supervision produces substantially stronger adapters than raw-context next-token-prediction. These results position document LoRA as a complementary memory channel whose value emerges precisely when context-side evidence is scarce.

Point clouds are a primary sensory representation for robotic perception, underpinning LiDAR-based autonomous driving, simultaneous localization and mapping (SLAM), and navigation. Within these pipelines, Farthest Point Sampling (FPS) is the most well-known downsampling operator, as its uniform coverage preserves the geometric structure on which downstream perception relies. However, the large time complexity of classical FPS scales poorly with the million-point-per-second rates of modern 3D sensors, making it a dominant latency bottleneck that conflicts with the real-time and limited onboard compute budgets of robotic systems. Therefore, we propose RadiusFPS, an FPS acceleration framework based on spherical voxel pruning that preserves the standard FPS update rule under the same initialization and tie-breaking policy. By indexing the point cloud with spherical voxels, RadiusFPS derives a conservative geometric bound that prunes redundant distance computations in each iteration, complemented by a coordinate-wise point-skip test that removes residual updates. We further introduce RadiusFPS-G, a warp-level GPU implementation that fuses voxel selection, pruning, and distance update into memory-coalesced kernels, eliminating costly global-memory round-trips. On indoor (S3DIS, ScanNet) and outdoor LiDAR (SemanticKITTI) benchmarks, RadiusFPS-G attains up to 2.5x speedup over GPU-based FPS and matches or exceeds QuickFPS among the evaluated methods while using roughly half its GPU memory, with comparable segmentation accuracy. When coupled with the learning-based FastPoint sampler, the resulting pipeline achieves the fastest End-to-End inference among all evaluated configurations. These properties make high-quality FPS-style sampling practical for latency- and memory-constrained robotic vision.

Sparse attention is becoming increasingly important for serving large language models (LLMs) as generation lengths continue to grow. However, deploying and evaluating new sparse attention algorithms at scale remains highly engineering-intensive, slowing both human researchers and AI agents in exploring the sparse attention design. To address this challenge, we present Vortex, a system that combines a Python-embedded frontend language atop a page-centric tensor abstraction for expressing a broad range of sparse attention algorithms, with an efficient backend tightly integrated into modern LLM serving stacks. Vortex enables rapid prototyping, deployment, and evaluation of sparse attention algorithms, effectively translating their theoretical efficiency gains into real-world throughput improvements. As a result, Vortex substantially accelerates the design and iteration of sparse attention algorithms. First, AI agents use Vortex to automatically generate and refine diverse algorithms, the best reaching up to $3.46\times$ higher throughput than full attention while preserving accuracy. Second, Vortex extends sparse attention to emerging architectures and very large models that are otherwise hard to experiment with, reaching up to $4.7\times$ higher throughput on the MLA-based GLM-4.7-Flash and $1.37\times$ on the 229B-parameter MiniMax-M2.7 on NVIDIA B200 GPUs.

Mixture-of-Experts (MoE) models scale foundation models efficiently by activating only a subset of experts for each token, but their large number of expert parameters still makes quantization essential for practical deployment. Unlike dense models, however, MoE models are sensitive to routing instability: small quantization-induced perturbations can change the top-$k$ expert selection, altering the computation path and degrading model quality. We propose Value-and-Structure Routing Alignment for Quantization (VSRAQ), a MoE-specific post-training quantization objective that preserves pre-quantization expert-selection behavior under quantization. VSRAQ combines two complementary objectives that jointly preserve expert-selection behavior: value alignment, which matches routing-relevant logits or scores, and structure alignment, which preserves expert ordering and top-$k$ decision boundaries. By maintaining routing consistency, VSRAQ reduces quantization-induced degradation without introducing any inference-time overhead and can be integrated into existing quantization frameworks. Experiments on recent MoE foundation models show that VSRAQ improves expert-selection consistency and consistently outperforms reconstruction-only and router-aware baselines.

Retrieval-augmented generation (RAG) improves large language model (LLM) answer quality by grounding generation in external evidence, but processing retrieved contexts makes the prefill stage a dominant serving cost. RAG cache fusion reduces this cost by reusing precomputed key-value (KV) caches for retrieved chunks and selectively recomputing tokens under the current prompt. Existing selectors, however, face a dilemma between quality and efficiency: fast query-agnostic or final-layer query-to-context selectors can miss request-relevant evidence, whereas full-view query-aware selectors require broad context and layer visibility before recomputation and therefore stall the layer-wise cache-fusion pipeline. We present QCFuse, a compressed-view query-aware selector for RAG cache fusion. QCFuse uses chunk-anchor query probing to condition user-query states on compact per-chunk anchors and critical-layer profiling to identify recomputation tokens without all-layer inspection. We implement QCFuse in SGLang and evaluate it on four open-weight LLMs across six datasets. QCFuse reaches full-prefill-level quality. At matched quality, QCFuse achieves an average prefill-time speedup of 1.7x over full prefill and 1.5x over ProphetKV, the strongest quality-preserving baseline.

Modern diffusion models generate high-quality images and videos, but their iterative denoising process makes inference expensive. Feature caching accelerates sampling by reusing or predicting intermediate activations across neighboring denoising steps, exploiting the redundancy of computations along the reverse trajectory. In this work, we focus on the caching schedule: selecting which denoising steps should be fully recomputed. Existing schedules are either fixed (e.g. uniform) or chosen adaptively from per-step error heuristics; in both cases, the actual compute cost is a side-effect of hand-tuned thresholds rather than a quantity the user can specify. We propose ReCache, which inverts this: given a target budget k, it learns the recomputation schedule that maximizes generation quality, turning compute into a directly controllable input. ReCache trains via policy gradients, sidestepping backpropagation through full diffusion inference, and uses no labelled data. Generations from uncached inference serve as matching targets, paired with a reward for generation quality. ReCache is compatible with any caching mechanism, including feature reuse and feature forecasting; for each mechanism, a single trained policy adapts across computational budgets at inference time. ReCache consistently outperforms scheduling baselines: under a $\times5.04$ FLOPs reduction on FLUX, it reduces LPIPS by 31% (from 0.456 to 0.316) compared to DiCache; on Wan 2.1 at a $\sim \times2.6$ speedup, it drops LPIPS by 65% (from 0.480 to 0.169) and boosts the VBench score by 7% (5.6 points, from 70.4 to 76.0) over uniform HiCache. Code is available at https://github.com/thecrazymage/ReCache.

Long-context inference in modern LLMs is increasingly constrained by decoding efficiency, especially in reasoning-heavy settings where models generate long intermediate chains of thought. Existing sparse attention methods often face a practical efficiency-quality trade-off. Structured block sparse methods typically provide stronger acceleration but incur noticeable quality loss, while token sparse methods are usually more accurate yet deliver limited end-to-end speedup because top-k routing over the full cache remains expensive. In this work, we propose cross-layer sparse attention (CLSA), which is built on top of KV-sharing architectures such as YOCO. The core idea is to share not only the KV cache across cross-decoder layers, but also the routing index. A single indexer computes token-level top-k selection once and reuses the resulting index across layers, thereby preserving the fine-grained selectivity of token sparse attention while amortizing the routing overhead. The resulting architecture improves all major inference bottlenecks jointly, including pre-filling, KV-cache storage, and long-context decoding. Experiments across short-context and long-context benchmarks show that CLSA is both accurate and efficient, achieving up to 7.6x decoding speedup and 17.1x overall throughput improvement at 128K context. These results suggest a more complete architectural solution for long-context LLMs that jointly advances model quality and inference efficiency.

As the input length of large language model (LLM) serving continues to grow, the KV cache has become a dominant bottleneck in AI infrastructure. It limits GPU memory capacity, serving concurrency, cache reuse, and distributed scalability. Several important problems, including position-independent KV cache, prefix KV cache compression, hot/cold KV cache separation, and distributed KV cache management, all depend on how the KV cache is represented and managed. However, existing serving systems largely rely on a monolithic KV cache abstraction, where the KV cache is treated as a homogeneous sequence of token-level memory blocks and managed with similar policies across attention heads and serving scenarios. We observe that KV cache utility is highly structured across KV heads: different heads exhibit different functional roles, attention distances, and runtime importance. Therefore, a full KV cache is not always necessary for every head, token range, or serving scenario. We present RedKnot, a head-aware KV cache management system for LLM serving. RedKnot breaks the conventional monolithic KV cache abstraction by decomposing the KV cache along KV heads, whose importance and effective attention ranges vary significantly across serving scenarios. This head-level decomposition turns the KV cache from a monolithic tensor abstraction into a structured memory object, enabling RedKnot to uniformly support position-independent KV reuse, prefix KV compression, hot/cold KV separation, and distributed KV placement while preserving output fidelity and improving resource efficiency, without requiring model retraining or fine-tuning. RedKnot establishes a new foundation for AI infrastructure by transforming the KV cache from a monolithic, passive runtime artifact into a dynamic, model-aware runtime substrate for scalable LLM serving.

Multi-turn Large Language Model (LLM) serving is critical for consistent user experiences, yet the linear growth of the Key-Value (KV) cache imposes significant pressure on GPU memory and bandwidth. Non-uniform KV compression effectively preserves more information by considering the individual importance of each KV cache. However, such KV cache heterogeneity introduces various systemic challenges - including memory fragmentation, scheduling complexities, and diminished kernel utilization - which collectively lead to significant inefficiencies in existing LLM serving systems. To overcome these challenges, we present Tangram, a novel serving system designed to make Non-uniform KV caches practical. Tangram addresses systemic inefficiencies through three core techniques: (1) Deterministic Budget Allocation assigns a static memory footprint to each head based on its intrinsic pattern, entirely eliminating dynamic scheduling overhead and prefill stalls; (2) Head Group Page clusters attention heads with similar retention demands and manages them with independent, vectorized page tables, thereby maximizing physical memory reclamation; and (3) Ahead-of-Time (AOT) Load Balancing leverages static budget profiles to ensure uniform GPU utilization without runtime overhead. Experimental results show that Tangram improves throughput by up to 2.6x compared to existing baselines, while fully preserving model accuracy. Our implementation is publicly available at https://github.com/aiha-lab/TANGRAM.

Ensuring public safety in densely populated urban environments remains a critical challenge, necessitating the deployment of intelligent and automated video surveillance systems. Traditional surveillance approaches rely heavily on manual monitoring, which is inefficient and susceptible to human fatigue, delayed response, and observational errors. To overcome these limitations, this work presents a real-time object detection-based surveillance framework. The proposed system focuses on detecting guns, knives, and region-specific blunt objects commonly involved in violent activities in Indian surveillance scenarios. A key contribution of this work is the use of a custom-created dataset collected using a mobile camera, consisting of 336 labeled images of blunt objects such as iron rods, wooden sticks, and plastic rods. This dataset is combined with a publicly available dataset of 7,623 images of guns and knives, forming a consolidated dataset of 7,959 images across three classes: gun, knife, and blunt object. The combined dataset is used to train a YOLOv8-based object detection model for real-time performance. Experimental evaluation shows that increasing the training duration significantly improves recall and average precision for the blunt object class without signs of overfitting. Overall, the proposed framework achieves an effective balance between accuracy and efficiency, making it suitable for deployment in real-world surveillance environments such as campuses, public spaces, and transportation areas.

Video question answering (VideoQA) aims to answer questions about given videos. While existing approaches excel on factoid VideoQA, they struggle with deep video understanding (DVU), which requires the comprehension of complex storylines. This challenge arises from the inherent long-range video content, multi-faceted question types, and instance-level story elements, all of which constrain the scale and diversity of manually constructed DVU datasets. These difficulties constrain the scale and diversity of manually-constructed DVU dataset. To address these, we previously introduced StoryMind to automatically construct DVU datasets with balanced fine-grained topics. Though it can generate high-quality question-answer pairs (QAs) for TV series, it suffers significant performance degradation when handling longer and more complex movies. In this paper, we further design StoryMindv2, an enhanced multi-agent collaboration framework to generate high-quality DVU datasets for both TV series and movies. By integrating a novel supervisor-guided generation mechanism and a refined multi-reviewer voting strategy, the framework is utilized to construct StoryVideoQA, the largest DVU dataset to date, featuring over 363K QAs on 393.2 hours diverse story videos including TV series (avg. 1,635 seconds) and movies (avg. 7,878 seconds). Comprehensive evaluations of 20 state-of-the-art VideoQA methods on this large-scale benchmark reveal that they cannot fully maintain long-range character associations or construct a coherent understanding of complex storylines. To bridge this gap, we propose PlotTree, a novel video understanding agent, re-organizing long-range video content into a hierarchical plot structure, enabling efficient storyline reasoning on StoryVideoQA. Project page: https://github.com/nercms-mmap/StoryVideoQA/

Video event prediction (VEP) requires models to infer unobserved future states from partial video evidence. Existing video MLLMs usually verbalize intermediate future reasoning in text space: once visual evidence is verbalized, fine-grained motion, geometry, and interaction cues can be lost, leading to plausible but visually ungrounded hallucinations. We introduce Future-L1, an interleaved latent visual reasoning framework that lets an MLLM alternate between language tokens and continuous latent visual spans during autoregressive decoding. To train this capability, we construct Future-L1-50K by selecting examples where future visual hints help prediction and align latent states to future-frame embeddings, then further optimize sampled latent trajectories with LA-DAPO, a latent-aware RL objective with outcome-contrastive and temporal-diversity rewards. Future-L1 achieves new state-of-the-art results on both benchmarks: on FutureBench, it improves Qwen3-VL-8B from 61.0 to 85.4 and exceeds the previous best Video-CoE by 10.4 points; on TwiFF-Bench, it improves the average score from 2.44 to 3.04. These results suggest that future-oriented video reasoning benefits from preserving intermediate visual semantics in latent space rather than translating every reasoning step into text.