Agentic systems increasingly solve complex user requests by executing orchestrated workflows, where subtasks are assigned to specialized models or tools and coordinated according to their dependencies. While recent work improves agent efficiency by optimizing the performance--cost--latency frontier, real deployments often impose concrete requirements: a workflow must be completed within a specified budget and before a specified deadline. This shifts the goal from average efficiency optimization to maximizing the probability that the entire workflow completes successfully under explicit budget and deadline constraints. We study \emph{constraint-driven online resource allocation for agentic workflows}. Given a dependency-structured workflow and estimates of success rates and generation lengths for each subtask--model pair, the executor allocates models and parallel samples across simultaneously executable subtasks while managing the remaining budget and time. We formulate this setting as a finite-horizon stochastic online allocation problem and propose \emph{Monte Carlo Portfolio Planning} (MCPP), a lightweight closed-loop planner that directly estimates constrained completion probability through simulated workflow executions and replans after observed outcomes. Experiments on CodeFlow and ProofFlow demonstrate that MCPP consistently improves constrained completion probability over strong baselines across a wide range of budget--deadline constraints.

Refurbishment-based noisy-label learning mixes an observed label with a model-derived pseudo target, typically using one sample-wise cleanliness score to control both branches. This creates a hidden coupling: reducing trust in the observed label automatically increases trust in the pseudo target. We show that this complementarity can replace one unreliable signal with another because a pseudo target learned from corrupted supervision may reproduce the noise it is meant to correct. Our representation diagnostics provide a consistent account of this mismatch: noisy supervision redirects deeper layers more strongly, whereas shallower relations remain comparatively stable and provide information beyond the loss posterior. We therefore propose TRACE, a Two-Source Reliability Assessment framework for Label Correction and Sample Reweighting. TRACE assesses the observed label using loss fit, shallow relation stability, and prediction agreement, while separately assessing the pseudo target using model confidence. Its source-specific scores control target correction and supervision strength without assuming complementary reliability. Across synthetic and real-world noisy benchmarks, TRACE improves representative refurbishment baselines and yields more reliable pseudo supervision.

Reinforcement learning with verifiable rewards (RLVR) efficiently scales the reasoning ability of large language models (LLMs) but is bottlenecked by scarce labeled data. Reinforcement learning with intrinsic rewards (RLIR) offers a scalable alternative via self-rewarding, yet often suffers from instability and inferior performance. We trace this gap to a systemic bias in confidence-coupled self-rewarding: the model tends to over-reward high-confidence mistakes, forming a self-confirming loop. We quantify this feedback-loop bias with three metrics: reward noise magnitude (rho_noise), policy-reward coupling (rho_selfbias), and over-/under-reward skew (rho_symbias). Our analyses show a compounding effect where strong coupling amplifies confidence-conditioned errors and drives a drift toward over-reward, leading to instability and a lower performance ceiling. To mitigate this, we propose reinforcement learning with ensembled rewards (RLER), which aggregates diverse models with adaptive reward interpolation and disagreement-aware rollout selection to reduce coupling and suppress over-reward drift. Extensive experiments show that RLER improves by 6.2

Self-Consistency (SC) improves test-time reasoning by sampling multiple reasoning chains and voting on their final answers; however, it typically discards low-vote chains. On mathematical reasoning tasks, low-vote chains can still contain useful partial evidence and intermediate results while failing only near the final step, especially for uncertain problems where solution paths diverge. We propose Two-Round Voting for Self-Consistency, a training-free test-time framework that recycles low-vote chains. After a standard SC pass, we extract high-confidence prefixes from low-vote chains and use them as additional evidence to re-score candidates, followed by a second-round vote. The second-round signal is fused with first-round votes via a tunable weight, and a flip-possible gate triggers re-voting only when it can change the predicted answer to limit overhead. Across the AIME (1983–2003) and MATH-500 benchmarks, Two-Round Voting yields consistent gains over Self-Consistency for Qwen3 models at multiple scales, improving the accuracy–compute trade-off with matched FLOPs.

As large language models (LLMs) evolve rapidly, distinguishing AI-generated text (AIGT) from human-written text (HWT) is becoming increasingly challenging. Recently, some AIGT detectors have been developed to overcome this challenge and have achieved decent accuracy. However, their brittle text representations make them highly susceptible to text perturbations, such that even minor character-level perturbations can reverse their predictions. In this work, we propose a multi-grained latent feature denoising and contrastive representation learning architecture to enhance text representations in terms of granularity, robustness, and distinguishability of features, thereby achieving robust AIGT detection. Specifically, we first extract both document-level and fine-grained segment-level features using a dual network, which captures the global and subtle local differences between AIGT and HWT. To encourage feature stability under perturbations, we inject random noise into both latent features and employ a denoising network to reconstruct the original representations. While this does not precisely simulate discrete character-level perturbations, it acts as a feature-level regularizer that suppresses non-essential variations and promotes smoother, more stable representations. Considering the similarities between AIGT and HWT, we further design a contrastive augmentation mechanism to increase the distinguishability between them. Extensive experiments demonstrate that our method not only outperforms baseline models in terms of classification accuracy but also exhibits superior robustness against various text perturbations.

Multimodal Open-domain Dialogue (MOD) plays a crucial role in AI-human interactions and has garnered substantial interest. Although existing studies have explored various aspects of MOD, the evaluation of MOD remains underexplored. In this work, we propose MODE+, an evaluation benchmark for MOD and a probe into multimodel open-domain dialogue evaluation. Specifically, we construct MODE with a balanced difficulty distribution and divide it into three parts: MODE-Base and MODE-Hard, both consisting of single-turn dialogues, with MODE-Base containing 889 test cases and MODE-Hard comprising 215 more challenging cases designed for probing model robustness against multimodal inconsistencies. Additionally, we include MODE-Multi, which contains over 10,000 multi-turn dialogue cases for more extensive testing. Each case contains an image, a context, and turn-level response scores provided by at least three human annotators following standardized criteria. The consistency of human annotations has an average Spearman correlation of over 0.9, indicating that MODE is highly reliable in annotation. We test the MOD evaluation capabilities of various evaluators on MODE, including LLaMA, Claude3, GPT-4, LLaVA, Gemini and Qwen3-VL. Results show that even the best-performing model-based evaluators have surprisingly low agreement with human evaluations, with consistency scores for MODE-Base below 0.7 and for MODE-Hard falling below 0.4. To improve model-based MOD evaluation capabilities, we propose the MM-Eval framework, a systematic methodology designed to standardize automatic evaluation. MM-Eval introduces Image Transformation as a modality-bridging mechanism, Inference Enhancement for transparent reasoning, and Inference Calibration for statistical reliability. Compared to the baselines, MM-Eval achieves a 67.41% improvement on MODE-Base and a 297% enhancement on MODE-Hard. Furthermore, the performance on MODE-Multi shows significant improvements with MM-Eval, demonstrating that the framework is capable of handling larger and more complex datasets. These results demonstrate that MM-Eval serves as a transferable and robust standard for future MOD evaluation.

The prohibitive cost of evaluating Large Language Models (LLMs) necessitates efficient alternatives to full-scale benchmarking. Prevalent approaches address this by identifying a small coreset of items to approximate full-benchmark performance. However, existing methods must estimate a reliable item profile from response patterns across many source models, which becomes statistically unstable when the source pool is small. This dependency is particularly limiting for newly released benchmarks with minimal historical evaluation data. We argue that discrete correctness labels are a lossy view of the model's decision process and fail to capture information encoded in hidden states. To address this, we introduce \textsc{RepCore}, which aligns heterogeneous hidden states into a unified latent space to construct representative coresets. Using these subsets for performance extrapolation, \textsc{RepCore} achieves precise estimation accuracy with as few as ten source models. Experiments on five benchmarks and over 200 models show consistent gains over output-based baselines in ranking correlation and estimation accuracy. Spectral analysis further indicates that the aligned representations contain separable components reflecting broad response tendencies and task-specific reasoning patterns.

KV cache in autoregressive LLMs eliminates redundant recomputation but has emerged as the dominant memory and bandwidth bottleneck during inference, notably with long contexts and test-time scaling. KV quantization is a key lever for reducing cache cost, but accuracy drops sharply as the native KV distribution lacks flatness and thus maintains a wide quantization range. Prior work focuses on isolating outliers, which caps their error but fails to flatten the overall distribution, leaving performance fragile under low-bit settings. In this work, we show that the K cache maintains a stable, context-evolving structure, while the V cache carries latent semantic regularities, with both contributing to the organization of vectors into shared patterns. Building on these insights, we propose , a pattern-aligned residual quantization scheme. It mines representative pattern vectors online, aligns each KV vector to its nearest pattern, and quantizes only the residual. This reshaping of the KV distribution flattens the quantization target and narrows its range, thereby improving the fidelity of low-bit KV quantization. Across long-context and test-time scaling settings on multiple backbones, PatternKV delivers consistent 2-bit gains, with a 0.08\% average 4-bit drop relative to FP16, improves test-time scaling accuracy by 10\% on average, and raises throughput by 1.5× while supporting 1.25× larger batches.

Reconstructing mesh-aligned surface attributes from images is important for analysis and visualization, but it remains challenging in practice because multi-view capture is often sparse and camera intrinsics/extrinsics may be unavailable or unreliable. Most learning-based 3D reconstruction methods prioritize geometry recovery and rendering-oriented texturing and typically assume calibrated cameras, making them less suitable when the goal is analysis-ready per-vertex attributes on a given mesh under uncalibrated sparse views. We propose PCF-MAR, a pose- and calibration-free neural framework that predicts view-consistent per-vertex attributes from sparse, uncalibrated images conditioned on a geometry-only mesh. PCF-MAR replaces camera supervision with dense cross-view correspondences and a mesh prior: detector-free pairwise matches are distilled into local correspondence embeddings that modulate a shared transformer image encoder and a global correspondence embedding that summarizes cross-view agreement and yields a scalar reliability gate. A point-based mesh encoder maps the input mesh to triplane features for decoder initialization and provides a global mesh modulation vector. A mesh-anchored triplane transformer decoder performs geometry-aware fusion via mesh-conditioned layer normalization, while the correspondence gate scales cross-attention residuals to suppress inconsistent evidence and output an explicit reliability signal. PCF-MAR is trained end-to-end with mesh-space supervision, image-space photometric and perceptual losses, and triplane/mesh regularization. Experiments on Google Scanned Objects and OmniObject3D demonstrate improved mesh-aligned attribute reconstruction over representative baselines; relative to a recent baseline, PCF-MAR improves PSNR by +29.6% on average and reduces LPIPS and VC-MSE by 71.1% and 94.2%, respectively, without camera calibration or pose estimation.

The rapid advancement of large language models (LLMs) has led to a surge in both model supply and application demands. To facilitate effective matching between them, reliable, generic and efficient benchmark generators are widely needed. However, human annotators are constrained by inefficiency, and current LLM benchmark generators not only lack generalizability but also struggle with limited reliability, as they lack a comprehensive evaluation framework for validation and optimization. To fill this gap, we first propose an automated and unbiased evaluation framework, structured around four dimensions and ten criteria. Under this framework, we carefully analyze the advantages and weaknesses of directly prompting LLMs as generic benchmark generators. To enhance the reliability, we introduce a series of methods to address the identified weaknesses and integrate them as BenchMaker. Experiments across multiple LLMs and tasks confirm that BenchMaker achieves superior or comparable performance to human-annotated benchmarks on all metrics, highlighting its generalizability and reliability. More importantly, it delivers highly consistent evaluation results across 12 LLMs (0.967 Pearson correlation against MMLU-Pro), while taking only $0.005 and 0.38 minutes per sample.

Despite significant advancements in traditional syntactic communications based on Shannon's theory, these methods struggle to meet the requirements of 6G immersive communications, especially under challenging transmission conditions. With the development of generative artificial intelligence (GenAI), progress has been made in reconstructing videos using high-level semantic information. In this paper, we propose a scalable generative video semantic communication framework that extracts and transmits semantic information to achieve high-quality video reconstruction. Specifically, at the transmitter, description and other condition signals (e.g., first frame, sketches, etc.) are extracted from the source video, functioning as text and structural semantics, respectively. At the receiver, the diffusion-based GenAI large models are utilized to fuse the semantics of the multiple modalities for reconstructing the video. Simulation results demonstrate that, at an ultra-low channel bandwidth ratio (CBR), our scheme effectively captures semantic information to reconstruct videos aligned with human perception under different signal-to-noise ratios. Notably, the proposed First Frame+Desc. scheme consistently achieves CLIP score exceeding 0.92 at CBR = 0.0031 for SNR > 0 dB. This demonstrates its robust performance even under low SNR conditions.

NeRF-based methods excel in generating realistic volumetric representations but often struggle with creating high-quality meshes, which are crucial for applications requiring explicit surface models. Existing approaches train NeRF and mesh generation networks separately, which limits their ability to learn jointly and leads to spatial and visual misalignment. To address the problem, we propose a novel unified optimization method that collaboratively learns NeRF and colored mesh representations for enhancing 3D reconstruction from monocular RGB images. Our approach employs Triplane features to generate Signed Distance Fields guided by density inputs, which are processed through a differentiable iso-surface extraction module. To ensure consistency between the NeRF and Mesh representations, we introduce a novel cross-photometric loss comparing rasterized outputs with the volumetric rendered images for unified learning. Additionally, we utilize a face-aware chamfer distance and a color loss based on barycentric interpolation, enabling uniform point sampling and improved vertex-and face-level alignment. This enhances geometric accuracy and ensures precise color transfer from source images to the mesh. Our model is trained using 3D ground-truth data and 2D multi-view images, enabling a unified optimization pipeline. Comprehensive experiments demonstrate that our approach outperforms existing approaches in colorized mesh generation, demonstrating its effectiveness and potential for practical applications. On the Google Scanned Objects dataset, our method outperforms the previous stateof-the-art by achieving a 26.33% higher PSNR (24.76 vs. 19.60), a 43.56% lower LPIPS (0.092 vs. 0.163), and a 79.25% lower chamfer distance (0.011 vs. 0.053), demonstrating superior visual and geometric fidelity.

Test-Time Scaling (TTS) enhances the reasoning capabilities of large language models by allocating additional inference compute to explore the solution space. However, existing parallel TTS methods typically keep branches isolated during search: intermediate discoveries remain branch-private and cannot guide other branches in time. This information isolation causes substantial redundant exploration, as branches repeatedly rediscover information already found elsewhere and require more search steps to collect complete decision information needed to reach correct answers. To bridge this gap, we propose Collaborative Parallel Thinking (CPT), a training-free inference framework that enables search-time information sharing across parallel branches. CPT extracts compact intermediate information from ongoing branches, maintains a deduplicated query-level information pool, and broadcasts pool entries through the input context, allowing each branch in subsequent search steps to reuse discoveries made by other branches rather than rediscover the same information. Empirically, experiments on HMMT and AIME benchmarks show that CPT establishes a stronger accuracy–latency Pareto frontier than strong baselines across rollout budgets and model scales, highlighting search-time collaboration as an effective direction for efficient parallel TTS.

Multi-source DR datasets carry label noise, class imbalance, and cross-device shift. Confident learning detects label issues from out-of-fold probabilities but inherits teacher miscalibration, and purely semantic signals cannot flag physically degraded images. We propose CPDC, a data curation framework that integrates a confidence-agreement mechanism, an autoencoder quality gate, and progressive tiering. The agreement mechanism derives cross-paradigm scores from a supervised and a self-supervised teacher. The quality gate screens images for physical degradation before any semantic assessment occurs. Progressive tiering then stratifies samples into three tiers. We evaluate CPDC on 53,149 fundus images from seven datasets with in-distribution and out-of-distribution test splits. Against empirical risk minimization, CPDC yields higher balanced accuracy (ID: 86.24% vs 82.44%, +3.80pp; OOD: 90.88% vs 87.40%, +3.48pp; both CIs exclude zero). Compared to the single-teacher baselines Cleanlab, AUM, and EL2N, CPDC achieves comparable discriminative accuracy (BA difference  ≤ 0.41pp, CI includes zero) while providing lower calibration error (teacher ECE: 4.8% vs 12.4%–12.6%; OOD NLL: 0.169 vs 0.176, 95% CI) and per-source retention varying from 32.5% to 66.7%. A teacher-matched control narrows this gap to within seed-level variation, so the agreement signal rather than the tiering rule drives it. Correction-enabled variants provide a controllable ID–OOD trade-off via τA: conservative relabeling (τA ≥ 0.85) maximizes ID performance (BA 86.78%), while aggressive relabeling (τA ≥ 0.70) improves OOD generalization (BA 91.25%). Code and reproducibility metadata are publicly available. Original images must be obtained from the source datasets, after which our curation framework reconstructs the exact curated sets.

Listwise reranking utilizing Large Language Models (LLMs) has achieved state-of-the-art retrieval effectiveness. Recently, reasoning-enhanced models have further pushed these boundaries by employing Chain-of-Thought (CoT) to perform deep comparative analysis of candidate documents. However, this performance gain comes at a prohibitive computational cost, as models often generate thousands of reasoning tokens before producing a final ranking. In this work, we investigate the relationship between reasoning length and ranking quality, revealing an overthinking phenomenon where extended reasoning yields diminishing returns. To address this, we propose a Length-Regularized Self-Distillation framework. We synthesize a dataset by sampling diverse reasoning traces from a teacher model (Rank-K) and applying a Pareto-inspired filter to select traces that achieve high ranking performance with minimal token usage. By fine-tuning on these concise, high-quality rationales, the student model learns to internalize efficient reasoning patterns, effectively pruning redundant deliberation. Experiments on TREC Deep Learning and NeuCLIR benchmarks demonstrate that our method maintains the teacher's effectiveness while reducing inference token consumption by 34

Following the CMIVQA, MMI-VQA, and M4IVQA challenges in NLPCC 2023–2025, we introduce the Difficulty-Aware Medical Instructional Video Question Answering (DA-MIVQA) shared task for NLPCC 2026. DA-MIVQA extends previous multilingual and multimodal medical video benchmarks by explicitly distinguishing questions according to the type and complexity of evidence required for answering. Specifically, simple questions can often be answered from subtitle-based textual cues, whereas complex questions require visual grounding, procedural understanding, and cross-modal evidence integration. The challenge contains three tracks: Difficulty-Aware Temporal Answer Grounding in Single Video (DA-TAGSV), Difficulty-Aware Video Corpus Retrieval (DA-VCR), and Difficulty-Aware Temporal Answer Grounding in Video Corpus (DA-TAGVC). The dataset is collected from public medical instructional channels, covers diverse scenarios such as first aid, emergency response, rehabilitation, nursing, and general medical education, and is manually verified with difficulty annotations. This paper presents the task motivation, dataset construction, evaluation protocol, participation overview, competition results, and representative systems of DA-MIVQA. DA-MIVQA provides a practical benchmark for evaluating medical instructional video question answering systems under varying textual, visual, temporal, and procedural reasoning requirements.

Test-Time Scaling enhances the reasoning capabilities of Large Language Models by allocating additional inference compute to broaden the exploration of the solution space. However, existing search strategies typically treat rollouts as disposable samples, where valuable intermediate insights are effectively discarded after each trial. This systemic memorylessness leads to massive computational redundancy, as models repeatedly re-derive discovered conclusions and revisit known dead ends across extensive attempts. To bridge this gap, we propose Recycling Search Experience (RSE), a self-guided, training-free strategy that turns test-time search from a series of isolated trials into a cumulative process. By actively distilling raw trajectories into a shared experience bank, RSE enables positive recycling of intermediate conclusions to shortcut redundant derivations and negative recycling of failure patterns to prune encountered dead ends. Theoretically, we provide an analysis that formalizes the efficiency gains of RSE, validating its advantage over independent sampling in solving complex reasoning tasks. Empirically, extensive experiments on HMMT24, HMMT25, IMO-Bench, and HLE show that RSE consistently outperforms strong baselines with comparable computational cost, achieving state-of-the-art scaling efficiency.

Deep learning models excel in visual recognition but suffer severe performance drops when training labels are corrupted by noise. Under label noise prior work cannot learn accurate similarities and thus misguide the learning process. In this paper, we uncover a complementary and novel phenomenon, Dissimilarity Invariance, whereby semantic dissimilarity between unrelated samples remains stable despite label noise. Leveraging this insight, we propose NegScale, a plug-and-play framework that shifts focus from fragile similarity to robust dissimilarity. NegScale integrates: (1) Structured Negative Orthogonality Penalty (SNOP), enforcing subspace orthogonality for unrelated samples; and (2) Dissimilarity-Calibrated Similarity Adjustment (DCSA), suppressing spurious similarity using dissimilarity anchors. We also give theoretical analysis that proves Dissimilarity Invariance and the effectiveness of NegScale. Empirical results demonstrate that NegScale consistently outperforms state-of-the-art baselines, establishing new benchmarks on CIFAR with synthetic noise and real-world datasets.

Multimodal large language models (MLLMs) still struggle with spatial reasoning that requires perspective transformation. In particular, they often rely on camera-centric cues rather than reasoning from the reference object's viewpoint, leading to systematic errors in non-camera reference settings. In this paper, we first analyze this failure mode and show that object orientation is a key factor underlying such camera-centric shortcut behavior. To address this issue, we propose OrientSAM, an orientation-aware spatial alignment framework for multimodal models. OrientSAM injects explicit orientation information into multimodal representations through orientation-aware tokens and Fourier-based angle encoding, and further adopts a curriculum learning strategy to progressively improve perspective-aware reasoning. In addition, we build a spatial data construction pipeline to generate orientation-aware spatial supervision from large-scale images. Experiments on Spatial-MM, ViewSpatial, and 3DSRBench show that OrientSAM consistently outperforms strong baselines, especially on non-camera-view, person-centric, and orientation-sensitive tasks. The results further demonstrate that explicit orientation modeling is important for mitigating camera-centric shortcut behavior and enabling more robust allocentric spatial reasoning in multimodal models.

The field of multi-skill dialogue generation focuses on creating conversational agents capable of handling a variety of tasks. Traditional approaches, such as task detection and adapter fusion, have faced challenges; task detection heavily depends on accurate context recognition, while adapter fusion methods often suffer from high computational complexity and are constrained by the performance of single-adapter models. To address these issues, we propose a novel method called Knowledge Synergy Complement (KSC) learning for enhancing multi-skill dialogue generation. In KSC, we leverage single-skill adapters by defining their Lipschitz constants as macro knowledge and utilizing intermediate hidden states to capture micro knowledge. This complementary approach allows for better integration of diverse skills. Moreover, a novel knowledge pruning mechanism is introduced to prevent knowledge forgetting and reduce conflict between skills. Our experiments demonstrate that KSC effectively integrates multiple skills, achieving superior performance compared to current state-of-the-art methods, while also being more efficient in terms of computational and storage resources required.