This paper proposes R$^3$L, a framework addressing error accumulation in multi-hop relative spatial reasoning for 3D layout generation. It introduces invariant spatial decomposition, consistent spatial imagination through an imagine-and-revise loop, and supportive spatial optimization via global-to-local coordinate re-parameterization. Experiments demonstrate physically feasible and semantically consistent layouts. Code is open-sourced.

HumanNet is a one-million-hour human-centric video corpus covering both first- and third-person perspectives, fine-grained activities, human-object interactions, tool use, and long-horizon behaviors in diverse real-world environments. It provides interaction-centric annotations including captions, motion descriptions, and hand/body signals. A systematic data curation paradigm transforms unstructured internet video into a scalable substrate for representation learning, activity understanding, motion generation, and human-to-robot transfer. Initial validation shows that continued training with 1,000 hours of egocentric video from HumanNet outperforms 100 hours of real robot data from Magic Cobot on a Qwen VLM model.

A comprehensive survey on edge deep learning for computer vision and medical diagnostics, covering principles, hardware classification, lightweight design, model compression, applications, and future challenges.

Researchers propose a measure transport framework for visual text compression (VTC) to predict information loss. They decompose the cost into precision and coverage terms, enabling a label-free routing criterion and adaptive foveation. On 24 NLP datasets with Qwen3-4B, their method matches the per-dataset oracle on 70.8% of datasets, improving task scores by 3.3% while reducing token usage by 10.3%.

The paper introduces GoSkills, an inference-time group-structured retrieval method that transforms flat skill lists into compact, role-labeled execution contexts (Start, Support, Check, Avoid), improving skill-augmented agents' efficiency. Experiments on SkillsBench and ALFWorld show gains over baselines.

The paper introduces MultiSoc-4D, a benchmark dataset of over 58,000 Bengali social media comments annotated for category, sentiment, hate speech, and sarcasm. The authors use multiple LLMs (ChatGPT, Gemini, Claude, Grok) to annotate the data and discover a phenomenon called 'instruction-induced label collapse,' where LLMs overuse fallback labels, leading to high agreement but low detection of minority classes. For instance, they detected only 21% of hate speech and 25% of sarcasm instances compared to human annotators. The work highlights a 'label agreement illusion' validated by near-zero Fleiss' Kappa, and tests over 40 LLMs to show the bias propagates regardless of architecture.

A new study evaluates large language models' ability to adapt responses to the certainty of retrieved information, finding systematic limitations and proposing an interaction strategy that reduces obedience errors by 25% on average.

This paper presents a framework for developing Human-Centered Large Language Models (HCLLMs), integrating NLP, HCI, and responsible AI perspectives, arguing that human concerns must be addressed at every stage of the pipeline, from design to deployment, and concludes with a case study on the future of work.

Researchers propose MIST, a synthetic multi-turn voice-driven code generation dataset for IoT devices. Experiments reveal a significant gap between open- and closed-weight multimodal LLMs, with even frontier closed models having substantial room for improvement. The dataset and extensible generation framework are released to facilitate research on mixed-initiative voice assistants that reason about physical world constraints.

This paper presents TajPersLexon, a curated Tajik-Persian parallel lexical resource of 40,112 word and short-phrase pairs for cross-script lexical retrieval, transliteration, and alignment in low-resource settings. A comprehensive CPU-only benchmark compares three methodological families: lightweight hybrid pipeline, neural sequence-to-sequence models, and retrieval methods. Neural and retrieval baselines achieve 98-99% top-1 accuracy, while large multilingual sentence transformers fail on this exact lexical matching. The interpretable hybrid model offers a favorable accuracy-efficiency trade-off, achieving 96.4% accuracy in an OCR post-correction task. All experiments use fixed random seeds for full reproducibility. The dataset, code, and models will be publicly released.

IntentGrasp is a large-scale, multi-domain benchmark for evaluating LLMs' intent understanding, derived from 49 corpora across 12 domains. It includes a training set of 262,759 instances and two evaluation sets. Tests on 20 LLMs show performance far below human level (81.1%), but the proposed Intentional Fine-Tuning (IFT) yields significant gains and strong cross-domain generalization.

VITA-QinYu is the first end-to-end expressive spoken language model that supports natural conversation, role-playing, and singing. It uses a hybrid speech-text paradigm with multi-codebook audio tokens to enhance paralinguistic representation while maintaining modality separation. Trained on 15.8K hours of synthetic data, it outperforms peer models on role-playing and singing benchmarks and achieves state-of-the-art conversational accuracy and fluency. Code and models are open-sourced.

A study of 33 frontier LLMs using 1,500 MMLU items across six domains reveals that aggregate metacognitive scores mask within-model variation. Applied/Professional knowledge is easiest to monitor, while Formal Reasoning and Natural Science are hardest. Within-family profile clustering is significant for some families, and Gemma 4 shows substantial improvement over Gemma 3. The findings support domain-level screening before deployment.

Researchers introduce a discrete diffusion model for antibody sequence design that uses germline-absorbing states to reduce germline bias and enables classifier-guided conditional generation, achieving significant improvement in non-germline residue prediction.

This study examines the stability of counterfactual impact evaluation results for refugee matching in the United States using a range of off-policy evaluation methods. Results show consistent magnitude and statistical significance across scenarios, aligning with earlier findings.

Using 2020-2023 CRSS data, this study develops a hybrid framework identifying restraint non-use, vehicle age, speeding, and driver impairment as key risk factors for tree-involved crash severity, with critical interactions, supporting targeted interventions.

A new architecture called Toeplitz MLP Mixer (TMM) replaces attention with triangular-masked Toeplitz matrix multiplication, achieving sub-quadratic complexity (O(d n log n) time, O(d n) space) while retaining more input information and improving copying ability, information retrieval, and in-context learning accuracy over comparable architectures.

A hierarchical ensemble pipeline is introduced to address anomaly detection in multivariate telemetry data from the European Space Agency (ESA). The method integrates shapelet-based and statistical feature extraction, per-channel modeling, intra-channel stacking, and cross-channel aggregation. Using time-series cross-validation and two-level masking to prevent leakage, the pipeline achieved strong generalization on the ESA-ADB benchmark, effectively detecting subtle anomalies. It was awarded 2nd place in the Spacecraft Anomaly Challenge final round.

Flow matching generates data by integrating a learned velocity field, where NFE determines inference cost. This paper analyses the velocity Jacobian decomposition into strain rate and vorticity, proving strain causes exponential error amplification while vorticity contributes linearly, and proposes weighted Jacobian regularization that reduces error significantly at low NFE.

This paper introduces Positive-and-Negative Decoding (PND), a training-free inference framework that reduces object hallucination in vision-language models by amplifying visual evidence and penalizing prior-dominant generation, achieving state-of-the-art performance on POPE, MME, and CHAIR benchmarks.

This paper introduces SwiGAN, a Conditional GAN-based model that generates future spatio-temporal trajectories of the Soil Wetness Index (SWI) to assess drought risk in France, supporting adaptive risk management and insurance strategies up to 2050.

LKV introduces an end-to-end differentiable optimization approach for KV cache compression, learning task-optimized global budgets and token importance without heuristic proxies. It achieves state-of-the-art performance on LongBench and RULER, with near-lossless results at 15% cache retention on LongBench.

Paper proposes RateQuant, a method that uses rate-distortion theory to solve distortion model mismatch in mixed-precision KV cache quantization. On Qwen3-8B at 2.5 average bits, it reduces KIVI's perplexity from 49.3 to 14.9 (70% reduction), requiring only 1.6 seconds for calibration.

SCALAR framework studies AI-researcher interactions on quantum field theory and string theory problems. Multi-turn dialogue outperforms single-shot attempts; critique strategy effectiveness depends on Actor-Critic pairing.

Weblica proposes a framework for constructing reproducible and scalable web environments for training visual web agents using HTTP-level caching and LLM-based environment synthesis. Scaling RL to thousands of environments and tasks, the Weblica-8B model outperforms open-weight baselines on multiple benchmarks with fewer inference steps and is competitive with API models.

This paper introduces a method to determine when a language model's answer becomes stable before it is verbalized. Using a finite-answer projection, the authors find that the model's preference stabilizes 17-31 tokens before the answer is parseable in delayed-verdict tasks.

This survey proposes a novel evolutionary framework for LLM agent memory mechanisms, formalizing three stages: Storage (trajectory preservation), Reflection (trajectory refinement), and Experience (trajectory abstraction). It identifies three core drivers: long-range consistency, dynamic environment challenges, and continual learning. The paper specifically explores proactive exploration and cross-trajectory abstraction in the Experience stage, offering design principles and a roadmap for next-generation LLM agents.

This paper formalizes deployment-time learning (DTL) as a third stage in the LLM lifecycle and presents CASCADE, a framework that equips LLM agents with an evolving episodic memory and contextual bandit mechanism for experience reuse without modifying parameters. Across 16 tasks, CASCADE improves macro-averaged success rate by 20.9% over zero-shot prompting.

A new method detects coalition structures in multi-agent AI by analyzing mutual information of hidden states and applying spectral partitioning, validated in RL and LLM settings.

This paper proposes representing the evolving reasoning state as an epistemic state graph and defines an 'order-gap' to decide when to stop iterating. A necessary and sufficient condition for the linearized order-gap to be non-degenerate near the fixed point is provided. Applications include agent loops, tree-of-thought reasoning, theorem proving, and continual learning.