每天知道
你的领域
出了什么

The emergence of Chain-of-Thought (CoT) reasoning has significantly enhanced the ability of large language models (LLMs) to tackle complex, multi-step tasks. However, when errors occur, current interaction approaches typically involve re-generating another response that may make mistakes again, or users laboriously flag the faulty step in follow-up turns that may get responses <You are right, I made a mistake here> followed by similar errors recurring. To address this issue, we propose an efficient human intervention mechanism for precisely correcting reasoning errors in LLMs, termed Deep Interaction. Our approach enables direct editing of the original response, allowing erroneous parts to be corrected while preserving accurate reasoning steps. We refine the edited CoT into a distilled prompt, which then steers the LLM along the corrected reasoning path. Experimental results show that our method achieves over a 25% improvement in correction success rate and reduces token usage by approximately 40% on STEM tasks reasoning compared to baseline approaches.

Although multimodal large language models (MLLMs) have achieved remarkable progress, understanding 3D spatial relationships from 2D images remains a critical challenge. Existing methods primarily rely on symbolic text tokens, which inherently lack the fidelity to represent continuous geometric information. While recent methods use latent representations to enhance reasoning, relying on a single latent type cannot adapt to the diversity of spatial tasks, leading to misalignment in complex geometric scenarios. To address these limitations, we propose GeoAnchor, an interleaved text-latent reasoning framework. GeoAnchor decomposes 3D spatial information into three complementary components: position latents for object grounding, direction latents for relational orientation, and geometry latents for scene structure. These components are recombined in a structured space to construct local evidence while capturing global context, enabling dynamic and interpretable reasoning. Furthermore, we introduce a collaborative training strategy that guides the model from local spatial perception to comprehensive 3D understanding. Extensive experiments on diverse and complex 3D reasoning tasks demonstrate that GeoAnchor outperforms the state of the art, validating its effectiveness and generalization capabilities.

Model-Based Systems Engineering (MBSE) relies on formal system models as primary technical artifacts for representing requirements, structure, and behavior across the system lifecycle. With the standardization of SysMLv2 as a textual language, interest is increasing in translating natural-language descriptions directly into executable models. For practical deployment, generated models must be accepted by industrial modeling environments, not merely satisfy grammar constraints. We present a conformance-checker-driven framework for reliable natural-language-to-SysMLv2 translation that enforces production-level acceptance as the termination condition. The system embeds a SysMLv2 conformance checker within a generate-check-repair loop. Each model is evaluated using the checker, and deterministic diagnostics are incorporated into revisions until zero conformance errors are achieved. Using the production checker as the oracle ensures the framework targets deployability rather than grammar plausibility. We evaluate the approach on the full SysMBench prompt set of 151 prompts across four large language model backends, yielding 604 prompt-model cases. Single-shot generation achieves 51.16% production-conformance acceptance, while our approach achieves 100.00% conformance. By elevating production conformance from a post-processing check to a control mechanism within generation, the framework converts probabilistic outputs into production-accepted SysMLv2 artifacts suitable for loading, visualization, and engineering use.

Although large language models (LLMs) have set benchmarks for zero-shot reasoning, their deployment remains cost-prohibitive and environmentally taxing. Small Language Models (SLMs) offer a sustainable alternative, but prone to errors, on tasks requiring complex, multi-hop logical grounding. We investigate a neuro-symbolic agentic framework to enhance the reasoning capabilities of SLMs, specifically Gemma 3 (1B, 4B) and Llama 3.2 (3B), using the CLUTRR kinship benchmark. Our approach transforms the SLM into a minimalist agent utilizing two specialized tool calls: extract_facts for symbolic triplet extraction and get_hint for expert reasoning via a Relational Graph Convolutional Network (RGCN). We evaluate these models across two configurations, both in an Oracle scenario with ground-truth triplets and a Realistic scenario relying on self-extracted knowledge. Our results reveal that while RGCN-derived hints provide a 1.5 - 2x performance gain over story-only baselines, the system is constrained by the extraction bottleneck and sequential deductive fragility, where early extraction errors compound over multi-hop chains. Furthermore, we identify a "distraction effect" in specific architectures where noisy, self-generated facts degrade performance despite the presence of expert hints. This work characterizes the challenges of symbolic grounding in low-resource agentic systems and provides a roadmap for iterative verification in neuro-symbolic agentic pipelines.

Tool-augmented large language model agents excel at long-horizon tasks, yet they are typically post-trained on fixed toolsets. When tasks demand new tools, these agents struggle to incorporate them effectively, and retraining from scratch is often impractical. We identify the core obstacle in such toolset expansion problem as behavioral inertia: the tendency of agents to fall back on familiar tools and established reasoning patterns despite having access to new ones. We demonstrate that injecting counterfactual anchor contexts at critical decision points can break this inertia, recovering failed trajectories by eliciting suppressed agent capabilities. To scale this insight, we propose ToolAnchor, a framework that uses teacher models to hypothesize these counterfactual contexts, verifies them via student rollouts, and internalizes the successful interventions through agentic post-training. Extensive evaluations across general AI assistant (GAIA), textual search (BrowseComp), and visual search (VDR-Bench) tasks demonstrate that ToolAnchor consistently exhibits competitive performance under expanded toolsets. Our work bridges the gap between static post-training and dynamic adaptation, charting a new path for scalable agentic reinforcement learning.

An agent harness is the external control layer that turns a base LLM into an executable agent by managing context, tools, orchestration, memory, decoding, and output handling. While harness design strongly affects agent behavior, most automatic improvement methods optimize narrower artifacts such as prompts, pipelines, or workflows, and deployed agents usually reuse a single global harness for all cases. We introduce MemoHarness, an adaptive harness optimization framework that learns from its own executions. MemoHarness decomposes the harness into six editable control dimensions, stores per-case diagnoses and distilled global patterns in a dual-layer experience bank, and adapts the learned harness to each test case using retrieved experience without test-time labels, feedback, or additional search. In our evaluation across shell-agent, code-generation, and analytical-reasoning benchmarks, MemoHarness improves over the fixed harnesses we compare against and shows selective transfer to unseen suites and base models. Its additional context can also remain cost-competitive when much of the retrieved experience is cacheable. These results provide evidence that execution experience is a practical substrate for building agent harnesses that are more adaptive than a single static configuration, while leaving broader claims about statistical robustness and component attribution to future work.

The Segment Anything Model (SAM) has demonstrated strong generalizability across a variety of segmentation tasks. However, SAM often struggles in situations where the target to be segmented is ambiguous. This poses a problem in medical imaging, where accurate delineation of targets such as tumors is vital, but even expert radiologists can disagree on the appropriate boundary for a target. Addressing this, we propose SARFA (Segment Anything with Radiomic Feature Alignment), a novel framework for improved medical image segmentation. Via probabilistic prompting, SARFA generates a diverse set of plausible masks for each input image and optimizes them with a radiomics-driven training objective based on Fréchet Radiomic Distance (FRD) and Direct Preference Optimization (DPO). By minimizing the FRD between masked predicted and ground truth regions within each image, SARFA encourages segmentation outputs whose anatomical and textural characteristics align with clinically meaningful ground truth representations, without relying solely on pixel-level overlap. Evaluated on computed tomography (CT) and magnetic resonance imaging (MRI) benchmarks, SARFA outperforms existing ambiguous segmentation methods, demonstrating the effectiveness of radiomic feature alignment and DPO-style candidate mask ranking as a training objective. Our code is available at https://github.com/tbwa233/SARFA.

Unequal availability of human preference data across languages poses a significant challenge for aligning large language models in multilingual settings. To address the lack of sufficient data in low-resource language alignment, we propose a meta-learning framework for Reinforcement Learning from Human Feedback and Direct Preference Optimization. By leveraging preference data from other languages, our framework learns a transferable initialization that enables effective adaptation to a target language with minimal data. We provide theoretical guarantees for both the meta-reward modeling and meta-policy optimization settings, and empirically demonstrate the effectiveness of our approach on multilingual benchmarks. In an extremely low-resource setting with only 100 target-language preference samples, our approach achieves up to $28%$ win-rate improvements over baseline methods, and consistently outperforms baselines across multiple target languages and model scales. Our approaches retain these advantages across different combinations of meta-training languages and varying linguistic distances from the target languages.

Deepfake detectors that achieve near-perfect scores on academic benchmarks collapse on real-world content: recent in-the-wild evaluations report AUC drops of 45-50% for state-of-the-art open-source models. We argue this gap is structural: static detectors are trained once against a moving generative frontier. We present BitMind Forensics (BMF), trained through Bittensor SN34, an open adversarial competition that continually refreshes the training distribution. We evaluate one dated export comprising image, general-video, and human-video checkpoints across nineteen public datasets: the canonical face-swap suites (FaceForensics++, Celeb-DF v1/v2/++, DFDC, DFD, UADFV, DF40) and recent in-the-wild and AI-generated-media benchmarks (Sumsub, Deepfake-Eval-2024, WildRF, Community Forensics, AIGCDetectBench, GenImage, AI-GenBench, AIGIBench, RAID, GenVidBench, GenVideo-100K). BMF reaches 0.936 AUC on Sumsub's original images and 0.872 pooled AUC over its full four-condition manipulation battery (1.4M images), staying robust under perturbation (0.855 JPEG, 0.799 downscaled), while GPEN enhancement improves detection (0.996). On Deepfake-Eval-2024, it matches the best commercial detector on images (0.915 vs 0.90) and exceeds it on video (0.822 vs 0.79), far above the best open-source detectors (0.56 and 0.63). It reaches 0.991 AUC on a 21-generator AI-image panel and 0.918 on GenVidBench, and exceeds the FF++-trained frontier on DFDC (0.947 vs 0.843) and Celeb-DF v2 (0.9985 vs 0.956), both contamination-audited, with statistical parity on Celeb-DF++. In a temporal study, successive dated exports improve on held-out media from generators absent from the static baseline's training (image 0.842 to 0.902; video 0.864 to 0.936). Our evaluation harness is public, and at publication the production API serves the exact evaluated snapshot for independent verification.

We address the problem of safely training an agent policy and deploying a good and safe policy, in settings where the environment dynamics are unknown and no suitable reward function is available. In the context of safety-critical environments, we consider traditional reinforcement learning impractical and resort to the resource of human input. We introduce DROPJ, a human-centred method for both safe training and deployment. We first learn a world model (a learned simulator) from a dataset of prior real-world trajectories. A human then plays the game in this learned simulator to extract several informative simulated trajectories. From these, we sample pairs of simulated trajectory segments and elicit from a human their preference over these segments, as well as a reason (justification) for their choice. We then train a reward model from these justified preferences and use it, together with the world model, to directly deploy the agent using model predictive control. Running real-user experiments, we find that generating informative simulated trajectories from a user significantly reduces the computational cost during training compared to other strategies, and can also improve the performance during deployment. In the context of training within a learned simulator, we show that the use of preferences rather than other types of feedback substantially improves the performance during deployment. We further demonstrate that safety justifications accompanying preferences can significantly enhance safety or prioritise user-prescribed aspects of safety associated with them during deployment.

We introduce Boogu-Image-0.1, an open-source unified multimodal understanding and generation model family, comprising Base, Turbo, Edit, and Edit-Turbo variants. It delivers competitive performance in high-quality text-to-image generation, fast inference, instruction-based editing, and bilingual (Chinese-English) text rendering. Closed-source multimodal systems like Nano-Banana-Pro and GPT-Image-2 achieve strong performance through system-level integration rather than a single model, yet their internal practices remain largely undisclosed. In this work, we demonstrate that targeted improvements in model understanding, data quality, and training pipelines, coupled with agentic inference-time scaling, can substantially enhance generation and editing performance even under highly constrained compute budgets. Comprehensive evaluations show that Boogu-Image-0.1 consistently matches or surpasses other open-source models across standard benchmarks, and achieves results approaching leading closed-source systems. Notably, this is accomplished with only 208.62 million unique images. The base model's theoretical training cost is only approximately \$400K. We share practical discussions that we believe are valuable to the broader research community, and release weights, code, and recipes under Apache 2.0 to advance the open ecosystem for unified multimodal understanding and generation. Our code is available here: https://github.com/Boogu-Project/Boogu-Image.

Recent diffusion-based video generation models have enabled high-quality personalized video customization through both tuning-based pipelines, which fine-tune a video diffusion model, and reference-based pipelines such as image-to-video generation. However, these capabilities raise serious concerns about personal privacy, identity ownership and intellectual property protection. Existing anti-customization works focus on protecting images, while protection for videos against both reference- and tuning-based customization remains largely underexplored. Protecting videos in this setting raises three challenges: (i) Image-level perturbations, optimized frame by frame, cannot survive temporal compression by 3D video VAE. (ii) A video-level perturbation optimized on a single video is vulnerable to temporal editing and fails to protect unseen videos. (iii) Temporally inconsistent perturbations are not robust to temporal attacks. To address these challenges, we propose Temporally Consistent Universal Adversarial Perturbations (TC-UAP), the first protection method against both reference- and tuning-based video customization. TC-UAP optimizes an identity-level multi-frame UAP over sliding windows from multiple videos, accounting for local temporal dependencies induced by temporal compression in video VAE and enabling a single perturbation to protect unseen videos of varying lengths. Moreover, we introduce intrinsic temporal modeling and an extrinsic surrogate temporal-attack loss, which make the perturbation temporally consistent and robust to unseen temporal attacks. Empirically, quantitative and qualitative results show that TC-UAP achieves the strongest identity protection compared with existing methods under both reference- and tuning-based video customization, and remains robust under multiple unseen temporal attacks.

Diffusion-based text-to-audio generative models such as AudioLDM achieve high perceptual quality and strong semantic consistency; however, their practical deployment is hindered by the substantial computational cost of the U-Net denoising backbone. In this work, we apply model pruning to improve the computational efficiency of AudioLDM, a U-Net-based text-conditioned audio latent diffusion model. We analyse parameter redundancy across U-Net convolutional blocks and evaluate a filter-pruning strategy. Pruning is guided by norm-based criteria and followed by lightweight finetuning to recover performance losses. Experimental results demonstrate that up to 83% of the parameters and 39% of the multiply-accumulate operations of U-Net have been reduced while maintaining, and in some cases improving, generation quality compared to the baseline unpruned network. We find that pruning affects AudioLDM's ability to generate certain sound events including safety-critical sounds such as gunshots, sirens, and explosions, as well as mechanical sounds such as drills and sewing machines, and other sounds such as sprays and tick-tocks, which are mostly recovered by lightweight finetuning of the pruned model.

We present AffectFlow-DINO, a multi-task learning system for the 11th ABAW challenge that extends a standard deterministic architecture with a conditional rectified-flow head to model the inherent ambiguity of in-the-wild facial behavior. Instead of predicting a single affect estimate, the model learns a conditional generative distribution, enabling uncertainty-aware one-to-many predictions through Monte Carlo sampling. The system jointly estimates continuous valence-arousal, classifies eight facial expressions, and detects twelve Action Units from static face images. Built on a frozen DINOv3 ViT-S/16 backbone, extensive ablation studies show that rectified-flow decoding consistently improves deterministic prediction, particularly for valence-arousal estimation (CCC-V $+0.058$). We further show that post-hoc threshold calibration effectively recovers performance on severely imbalanced rare classes (e.g., Fear: $3.8% \rightarrow 33.1%$) without retraining. Combined with backbone fine-tuning and flow retuning, the final model achieves $\mathbf{P_{MTL}=1.177}$, substantially outperforming the official challenge baseline of $P_{MTL}=0.45$.

Human cognition does not separate understanding and generation. A teacher at a whiteboard speaks and draws $together$, each modality reshapes the other. In this paper, we bring this coupled loop to artificial systems. Masked Diffusion Models (MDMs) are ideally suited to this task, yet existing samplers either decode text and image interleavedly or independently update them in parallel branches that share only previous-step history, but not the other modality's latest decisions $within$ the same step; combined with MDMs' inability to remask, cross-modal contradictions are neither detected nor repaired. We introduce $Self-Correcting Coupled Markov Jump Processes (SC-CMJP)$, a framework in which one modality's transition rates are functionals of the other modality's confidence score, as weighted by cross-modal attention. Furthermore, a remasking jump retracts commitments the moment cross-modal evidence turns against them. In conjunction with SC-CMJP, we introduce $CO_2Jump$ (Self-$\underline{\text{CO}}$rrecting $\underline{\text{CO}}$upled $\underline{\text{Jump}}$), a novel training-free single-pass sampler for joint multimodal geneneration. For training and evaluation purposes, we have created and will release three large-scale joint multimodal generation corpora: $\text{JEdit-1M}$, $\text{JMaze-200K}$, $\text{JNono-200K}$, with matching in- and out-of-distribution benchmarks. $CO_2Jump$ achieves best joint performance for image understanding and editing as well as visual reasoning (maze and nonogram solving). The performance of the sampler scales monotonically with the number of denoising steps, evidence that the benefits of cross-modal coupling $compound$ across the trajectory. Project page: https://coupled-jump.github.io

Sign language is a primary communication channel for millions of Deaf and hard-of-hearing people, yet text-to-signer video generation remains costly because video diffusion models are expensive to train and evaluate. This paper presents Text2Sign, a text-conditioned diffusion model for short sign-language clips that runs on a single NVIDIA L4 GPU. It combines a frozen vision-language text encoder with a 3D encoder-decoder and factorized spatiotemporal attention to reduce the cost of full-video attention while preserving motion coherence. We compare convolution-only and transformer-style backbones, frozen pretrained and task-specific text encoders, and factorized versus full attention. On a signer-disjoint How2Sign split, the best short-run ablation reaches a validation loss of 0.0648, while a longer-run checkpoint reaches 0.00999. On a compact evaluation slice, the latter achieves an SSIM of $0.2403 \pm 0.0238$, a PSNR of $15.11 \pm 0.42$ dB, and temporal consistency of $1.0000 \pm 0.0000$ using 8-step DDIM sampling with a guidance scale of 5.0. It generates a 32-frame, $64 \times 64$ clip in 12.60 seconds, or 2.54 frames per second, with peak inference memory of 3.12 GB. A held-out denoising audit shows only weak prompt sensitivity: removing text increases late-timestep loss from 0.9875 to 0.9891, while shuffled prompts perform similarly to correct prompts. Frozen text conditioning therefore improves short-budget validation loss, but prompt-specific separation remains limited. The system is restricted to low-resolution, short clips and lacks expert linguistic evaluation, so it should be viewed as a single-GPU research baseline rather than a complete sign-language production system. Code is available at https://github.com/xiaruize0911/text2sign.

We present RAGthoven, our system for SemEval-2026 Task 1 (MWAHAHA), Subtask A (multilingual constrained humor generation in English, Spanish, and Chinese). RAGthoven decomposes creative text generation into a multi-stage large language model (LLM) pipeline (Planner, Best-of-N Writer, Reflector for self-critique, LLM-as-a-judge Judge) grounded in computational humor theory (Benign Violation Theory, Script-based Semantic Theory of Humor) and refined across ten experiments. In our final configuration, we augment the Planner with retrieval-augmented generation (RAG) from a curated joke corpus, seeding generation with diverse joke mechanisms. We also evaluate two agentic variants — ReAct-style sequential tool-calling (Exp09) and autonomous multi-branch orchestration (Exp10) — that expose the same four stages with a deterministic ConstraintAudit checker. Across four frontier models on a held-out 12-instance English sample, neither agentic variant produced outputs we judged superior to the non-agentic pipeline despite substantially higher tool-call budgets. RAGthoven shares Rank 1 with the Gemini 2.5 Flash baseline in all three languages, with overlapping organizer-reported confidence intervals. In Spanish, it leads the baseline by 42 raw Elo points (1182 vs. 1140), while in English (1045 vs. 1081) and Chinese (1045 vs. 1053) the baseline holds the higher raw rating within the same statistical tie. Together, these results suggest language-dependent diminishing returns from elaborate multi-stage prompt engineering and agentic scaffolding once a strong frontier model is in the loop.

While 3D Scene Graphs (3DSGs) provide crucial structured representations for embodied agents, conventional Ahead-of-Time, build-everything-then-filter pipelines conflict with the real-time, low-latency demands of edge platforms, inducing a perceptual saturation effect via severe observation redundancy. To resolve this, we present JITOMA (Just-In-Time On-demand Memory Activation), a closed-loop framework that unifies task reasoning, perception, and memory into a just-in-time growth process. Instead of exhaustively mapping the entire environment, JITOMA leverages a top-down task heatmap at the frontend to filter continuous observations, routing minimal streams to maintain a global foundation of low-cost, dormant anchors. Upon a cognitive query, the backend Large Language Model (LLM) parses the robotic intent to dynamically awaken task-relevant anchors, triggering resource-intensive operations — such as dense node captioning and functional inference — exclusively within the activated local subgraph. To evaluate these dynamic capabilities and study perceptual saturation trade-offs, we introduce JITOMA-Bench, a comprehensive suite for long-horizon multi-tasking and complex multi-step reasoning. Extensive experiments demonstrate that JITOMA substantially reduces active graph size and captioning latency, while maintaining stable processing time under long-horizon task switching.

The capability of a modern AI agent depends not only on its foundation model but also on its harness, which constructs prompts, manages state, invokes tools, and coordinates execution. As models, APIs, environments, and requirements evolve, the harness must be continually modified. Before such a change can be made, a developer or coding agent must identify all code locations that implement the target behavior. This is difficult because production harnesses are large, tightly coupled, and behaviorally distributed, while modification requests describe what the system should do and repositories are organized by files and modules. Code search, repository indexing, and long-context processing ease inspection, but still leave this behavior-to-code mapping to be recovered by hand. Behavior localization is therefore a central bottleneck in harness evolution. We introduce the Harness Handbook, a behavior-centric representation synthesized automatically from a harness codebase via static analysis and LLM-assisted structuring, linking each behavior to its corresponding source. We also introduce Behavior-Guided Progressive Disclosure (BGPD), which guides agents from high-level behaviors to relevant implementation details and verifies candidate locations against the current source. On diverse modification requests from two open-source harnesses, Handbook-Assisted planning improves behavior localization and edit-plan quality while using fewer planner tokens, with the largest gains on scattered sites, rarely executed paths, and cross-module interactions. Evolving complex agentic systems thus depends not only on generating edits, but also on determining where those edits should be made.

Agent memory is a systems problem for long-horizon agents. Practical deployments require retention of task state across extended conversations, recovery of user-specific facts and preferences across sessions, and accumulation of procedural knowledge from prior outcomes. These requirements extend beyond document retrieval: a memory layer must determine which interactions become durable state, how that state is scoped, how it is retrieved under latency constraints, and how it is revised or removed over time. This report studies Oracle Agent Memory as a database-native memory substrate built on Oracle Database. Three themes organize the discussion: memory as a lifecycle spanning ingestion, extraction, consolidation, retrieval, summarization, and revision or removal; a layered architecture that separates an active memory core from a passive memory-store interface with explicit scope control across users, agents, and threads; and evaluation methodology in which downstream task accuracy is complemented by memory-centric measures such as evidence retrieval, recall, latency, and estimated token use. The report summarizes LongMemEval results, reaching 93.8% accuracy, compares Oracle Agent Memory against flat-history baselines, using about 10.7x fewer tokens, and published or reported external baselines where available, and closes with implementation-oriented appendix material covering setup, thread lifecycle, and search semantics.

In online streaming video understanding, a video stream continues to arrive and queries may be issued at any time. Because streaming frames grow without bound, the system must continuously compress and retain information from the observed video prefix while future frames and future queries remain unknown. The core challenge is deciding what information to retain and how to organize the maintained history: as this history grows with the stream, memory cost increases and many redundant visual details are retained, whereas later queries often depend on specific entities, actions, and their temporal changes. To address this challenge, we introduce FOLIO, a training-free focused semantic memory system that records important parts of the stream in higher detail while keeping surrounding context compact. As the stream arrives, FOLIO updates memory at the segment level, guided by a dynamic focus state, combining a short-term visual buffer with a long-term semantic memory organized around observed entities and linked to a visual-evidence cache. At query time, lightweight hybrid retrieval combines direct matching over the structured memory with semantic query expansion. FOLIO achieves state-of-the-art performance, reaching 82.0/69.1 Perception/Backward accuracy on OVO-Bench with Qwen3-VL-8B and 74.5 overall accuracy on StreamingBench, while substantially reducing the cost of maintaining streaming memory by reserving detailed records for focused entities and storing surrounding context compactly.

We present a full-pipeline inference optimization for the MiMo-V2.5 model family, which combines Hybrid Sliding Window Attention (Hybrid SWA), sparse Mixture-of-Experts (MoE), and multimodal encoders. While Hybrid SWA can ideally reduce both attention compute and KVCache storage significantly compared to Full Attention, realizing these gains in production requires substantial engineering effort. We systematically optimize the KVCache system with layerwise prefetch, SWA-aware prefix cache trees, and specialized placement strategies, achieving strict $O(W)$ SWA storage and high cache hit rates. We further build GCache, a high-performance distributed cache infrastructure with RDMA-optimized networking, and develop a KVCache-affinity router to reduce computation while preserving load balancing. We also optimize for multimodal inputs, including GPU image preprocessing, parallel video decoding, and multimodal cache sharing. Together, these optimizations constitute the first large-scale LLM serving system in production that efficiently covers the Hybrid SWA + MoE + multimodal composite architecture.

On-device LLM inference faces a trilemma of response latency, limited hardware resources and user privacy. Full cloud inference delivers strong computing power but exposes user prompts and dialogue data, while standalone on-device inference is unfeasible for most consumer and embedded edge devices. This paper presents a privacy-centric edge-cloud collaborative LLM inference framework built on endpoint-authenticated KV cache. Local endpoints handle input preprocessing, embedding computation, adaptive feature optimization, KV cache authentication, speculative decoding and low-dimensional model head calculation, while the cloud conducts authenticated decoder inference, KV cache management, token verification and high-dimensional vocabulary projection. Endpoints fuse partial outputs, apply language-adaptive masking and sample target tokens. All transmitted data and truncated logits are quantized and AES-GCM encrypted for privacy, with core lightweight modules, draft parameters and cache access policies kept local to avoid leakage. The framework supports heterogeneous devices including CPU-only, GPU-equipped and embedded devices via optimized streaming, batching and quantized ONNX deployment. Evaluations demonstrate that the framework reduces per-token latency by up to 46.1% and downlink payloads by up to 67.4% over baseline split inference, retaining comparable performance to full cloud inference.

Attention-based KV cache eviction (H2O and its descendants) compresses the memory-constrained state of a long-context model by ranking tokens on accumulated attention mass, treated here as signal energy, and keeping the heaviest. On schema-dense input streams such as nested JSON, this score acts as a non-stationary filter that disproportionately retains noise: a non-content sink role (delimiters or whitespace) carries an order of magnitude more energy than any content role, and structural KEY tokens are over-retained at roughly 1.8x the rate of the answer-carrying VALUE tokens, collapsing exact-match accuracy from 88% to 0% at a 5% budget as the signal-to-noise ratio of the retained state degrades. A counterfactual experiment establishes that suppressing KEY tokens is the best deployable filter. Our retraining-free, role-conditional allocation over SnapKV's windowed score, governed by a single tuned hyperparameter, closes 63-98% of the H2O gap at sub-20% budgets and, at higher budgets, modestly matches or exceeds full-cache accuracy — a small, seed-sensitive denoising effect (borderline significant at B=0.50; not distinguishable from zero at B=0.30 over four seeds). A 15 MB linear role probe supplies these labels at negligible inference cost, though matching parser-level downstream accuracy remains open.

Procedural material creation underpins applications in digital content creation, visual effects, and 3D asset design. Achieving high-quality results requires more than reproducing node graphs — it demands understanding the process by which experts construct materials. We formulate procedural material generation as retrieval-time process reasoning over expert demonstrations, elevating process to a first-class representation beyond graph-only synthesis. Concretely, we represent expert workflows as process traces: textual records of construction steps, parameters, and design intent. To instantiate this idea, we use a pretrained LLM-based ProcessSynthesizer to synthesize a process trace aligned with a user's intent and a pretrained LLM-based Compiler to ground the process trace into an executable Blender material graph. Because procedural expertise is most naturally conveyed through demonstrations, we leverage tutorial videos as a source of process knowledge and extract textual, LLM-compatible traces using automated video analysis tools. In an expert study with five Blender artists (avg. 7.5 years of experience), materials generated by reflecting expert demonstrations were found to produce workflows requiring fewer edits, and more closely match professional design strategies than methods operating solely on static artifacts. A user study with 150 participants further shows that our approach achieves superior generation and editing performance compared to prior procedural systems. All code, models, and data will be available at https://materialapprentice.github.io

Industry 5.0 emphasises human-centric industrial system design, placing additional demands on modelling tools. Multi-level modelling (MLM) can directly represent three or more abstraction levels, but this comes at the cost of more complex semantic constraints that model correctness depends on. Large Language Models (LLMs) have been increasingly studied in model-driven engineering, but this evidence rests entirely on two-level modelling tasks, and whether it generalises to MLM, whose semantics differ in kind, remains untested. This paper presents the first empirical study of this question. We have three commercial LLMs (GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro) generate multi-level models for the MULTI Warehouse Challenge in the SLICER language under six prompting strategies, yielding 90 generated models compared against a manually validated reference using fourteen metrics. Syntactic correctness is within reach, but semantic correctness is only partially achieved, with Instantiation/Specialisation Correctness ranging from 52% to 79%. Models reproduce content stated explicitly in the task text, but rarely complete structure and constraints the text implies without stating. Prompting strategies trade off precision against completeness, and self-checking functions mainly as a rule checker rather than reliably improving alignment with the reference design. Among the three LLMs, Claude shows the most balanced profile. These results clarify the boundaries of current LLM capability for MLM and inform the design of human-centred, AI-assisted modelling workflows for Industry 5.0.

Training large language models at the multi-billion to trillion parameter scale is confined to datacenters, where data-parallel (DP) and model-parallel (MP) techniques presume homogeneous accelerators, high-speed interconnects, and a single orchestrating entity. Frontier model development is thereby concentrated among the few groups able to assemble such clusters. Meanwhile, an enormous pool of compute remains unusable for training: consumer and professional GPUs that are heterogeneous, preemptible, individually owned, and connected only by the internet. We present Agora, a system that makes efficient use of this compute. Agora combines bandwidth-efficient pipeline-parallel model sharding over internet-grade links with multi-party, fault-tolerant collective operations. Each participant holds only one stage of the model, and no single party ever possesses the full weights. We term this setup Protocol Learning: it enables collectively trained, collectively owned models, opening a path to open-source frontier training with economic sustainability. This report presents the outcome of a research effort spanning communication-efficient parallelism, asynchronous optimization, and fault-tolerant systems design. It culminates in the first demonstration of its kind: Pluralis-8B, an open, permissionless pretraining run of an 8.6B-parameter model on 500B tokens of FineWeb-Edu. The model was trained over 40 days by 330 contributor nodes, predominantly consumer GPUs on internet connections, joining and leaving throughout. The run sustained ~170k tokens/s and 4.2 tokens per TFLOP of pooled compute, 63% of the efficiency of a centralized H100 baseline, and converged to within a small margin of a centralized reference run.

Teams measuring whether large language models (LLMs) recommend a brand face a reproducibility problem: ask the same question twice and the answer moves. Practice resamples each prompt a few times (commonly five) and averages, treating within-prompt resampling as the source of the noise. But a measured brand score moves for at least four separable reasons: within-prompt resampling, prompt paraphrase, model identity, and query language. We specify a crossed random-effects (generalizability-theory) decomposition that partitions the total variance of a response-level brand outcome into these four sources, and embed the components in a decision-study allocation that returns how many repeats, paraphrases, models, and languages to buy for a target reliability. We apply it to a fully crossed corpus of 12,933 LLM responses on 20 Central and Eastern European brands, 8 languages, and 3 models (GPT-5.2 and Gemini 3 Flash in parametric mode, Perplexity in grounded retrieval), with a stability subset of 1,435 cells resampled about five times. The outcome is per-response multilingual sentiment polarity. Query language is the largest systematic facet (26.5% of the variance of one response) against 1.5% for brand identity (ICC 0.0146), so a single AI answer carries almost no brand-discriminating signal. Once a cell term isolates pure resampling, resampling is 34.8% of variance and the brand-in-context interaction 29.6%; brand-by-language is 8.6% (a bilingual penalty) while brand-by-model and brand-by-prompt are near zero. Per unit of query budget, adding languages and models reduces relative-error variance far more than adding repeats: a repeat past the fifth reduces it by only 0.0003. Brand-ranking reliability stays low, near 0.01 for a single answer and about 0.36 at the full crossed design, so reliability is bought by spreading across languages and models, not by repeating one prompt.

When labeled data are scarce, off-the-shelf diffusion models can augment training sets for few-shot medical image classification, but not all generated samples are equally useful for the downstream task. Existing approaches largely improve synthetic data by increasing realism, diversity, or domain adaptation, while overlooking a more fundamental question: how should sample usefulness for classification be measured and optimized? We address this with Class-Contrastive Influence (C2I), a criterion that quantifies a sample's usefulness through its gradient-based influence on the classifier. We find that effective samples exhibit a strong C2I gap: their loss gradients align with validation gradients from the same class and oppose those from other classes. Our analysis further suggests that such high-C2I samples are hard, boundary-proximal examples that help refine the decision boundary and improve robustness. Building on this insight, we fine-tune diffusion models with reinforcement learning using a C2I-based reward to steer generation toward class-informative samples. Across several few-shot medical imaging benchmarks, C2I-guided generation improves downstream accuracy and robustness over diffusion-based augmentation baselines, showing that synthetic augmentation is most effective when guided by task usefulness rather than image quality alone.

Video thumbnails are a key factor for attracting user clicks on video platforms, and are increasingly supported by automation. However, existing thumbnail generation methods typically produce generic results shared across users, overlooking the diversity of individual preferences. We therefore introduce personalized video thumbnail generation, a novel task that aims to create thumbnails tailored to user-specific preferences. It is challenging in two aspects: (i) identifying visual anchors (i.e., key frames) from each video to guide the generation, which requires a balance between personalization and informativeness that existing highlight detection methods fail to achieve; and (ii) generating personalized thumbnails that are both visually coherent and faithful to the original video. As a response, we propose a two-stage framework that tightly couples preference-aware retrieval with controllable generation. In the first stage, a personalized highlight retriever captures fine-grained user-video interactions and incorporates video semantics through summarization, enabling the selection of diverse visual anchors aligned with both user preferences and video contexts. In the second stage, a VLM-guided diffusion pipeline transforms these anchors into thumbnails by extracting and injecting semantically grounded visual cues, improving personalization while preserving visual coherence and fidelity. Experiments on two public datasets show our method delivers state-of-the-art performance compared with both retrieval-based and generative baselines. A user study further demonstrates improved click preference, highlighting its effectiveness in enhancing user engagement. The code is available at https://github.com/hezy18/PVTG.