Proposes CPG-PAD, a framework that introduces model-level concept guidance into prompt learning for Presentation Attack Detection. It uses XAI to discover attack-relevant visual concepts and injects them into prompts, achieving state-of-the-art cross-domain performance on nine benchmarks.

AnchorSplat is a novel 3D-native refinement paradigm that operates directly on 3D structures, avoiding expensive optimization overhead of traditional pipelines. It enforces geometric consistency via a Point Anchor Mechanism and replaces iterative densification with single-pass multiplication, requiring no original multi-view images. Experiments show throughput up to 10^5 times faster than optimization methods with robust zero-shot generalization.

TurnNat is a likelihood-based framework for automatic evaluation of turn-taking naturalness in dyadic spoken dialogue. It uses a causal prediction model to compute negative log-likelihood of future voice activity, quantifying timing atypicality, and demonstrates effectiveness on a perturbation benchmark.

RuleChef is a framework that uses LLMs to generate executable rules for NLP tasks such as text classification, NER, and relation extraction. Rules are generated from task descriptions and labeled examples, then iteratively improved via additional examples and human feedback. LLMs are used only at learning time, resulting in a fast, deterministic, and inspectable rule system. Preliminary evaluation on classification and NER tasks shows promise, and the system is open-sourced under Apache 2.0.

The Office Comprehension Bench (OCB) is the first public benchmark to jointly evaluate LLM systems on Word, Excel, and PowerPoint comprehension over native file formats. It comprises two tracks: File Fidelity Q&A (structural/visual perception) and Domain Q&A (expert reasoning across 12 domains). Even the strongest frontier system achieves only 59.3% on Domain Q&A; increasing thinking depth within a tier yields no material improvement, while moving to a higher product tier offers modest gains. The dataset, evaluation tooling, judge prompt, and leaderboard are released.

Researchers propose RAGP, a novel prompt compression method that models text as a multiplex graph and uses Lévy walks for efficient pruning. It achieves a 49.3 average score on LongBench at 4x compression, outperforming existing LLM-based methods.

This paper reveals that count-based F1 can be artificially inflated by prompt framing without corresponding improvement in span localization—a gap termed F1 Inflation. It introduces ErrorBench, a controlled stress-test protocol. Experiments show anchored prompts cause up to 0.79 F1 Inflation. The findings recommend avoiding pre-populated error counts and reporting span-aware metrics alongside count-based ones.

A new study reveals that BPE tokenization fragments safety-critical words into subword units, creating exploitable gaps in LLM safety alignment. An optimization targeting this fragmentation flips refusal triggers on 80-100% of HarmBench prompts, with 48% producing genuinely harmful outputs. Defense attempts show DPO configurations fail to close ASR stably, while SFT leads to global collapse. The paper introduces Conv-Benign as a diagnostic tool.

Recent advances in speech synthesis have shifted to direct grapheme modeling, but grapheme-based models underperform in low-resource settings. SPARCLE is a speaker-aware grapheme representation model trained with contrastive learning to align graphemes with acoustic representations, reducing word error rates by half in extreme low-resource settings.

This paper proposes Kara, a sliding-window KV cache compression method that operates on recently generated context during decoding. It uses bidirectional attention to score and select informative KV pairs and a Token2Chunk module to flexibly preserve important semantic information. Experiments show Kara and the vLLM-based inference framework KvLLM significantly reduce KV cache memory and improve output throughput.

This paper proposes a provenance-based conceptual framework for detecting misalignment in LLM agents' tool invocations. The authors develop ProvenanceGuard, a multi-stage pipeline that analyzes three types of misalignment before tool execution. Evaluated on Agent-SafetyBench and WorkBench across 10 backbone LLMs, it reduces error rates from 42.9% to 1.8% and from 32.1% to 17.3%, respectively, while decreasing intervention burden on successful traces from 30.5% to 12.8%.

TokenScope is a new interactive tool that provides token-level explainability for LLMs during code generation, offering metrics, attention patterns, and counterfactual branching to explore model decisions.

This paper presents a proof-of-concept AI system that uses images and accident reports to assess railway crossing safety, achieving a macro F1 score of 0.757 for risk classification and an RMSE of 0.078 for safety scores.

A new attack called NightVision can estimate hidden dimension, depth, and parameter count of LLMs even when APIs are restricted to only single logit output, achieving low error rates.

A new monograph by Zhilin Zhao presents a unified, proof-oriented account of modern deep learning theory, bridging classical approximation, optimization, and generalization with contemporary topics like overparameterization, transformers, in-context learning, scaling laws, and emergence.

Researchers propose a method combining Sparse Random Projection and multinomial logistic regression for CNS tumor classification from DNA methylation data. It achieves 96% accuracy on a reference cohort and 86% (91-class) and 93% (family-level) on an independent clinical cohort, outperforming state-of-the-art by 4-5 percentage points.

IonSense-QKG enriches public lithium-ion battery datasets with quantum-relevant metadata, introducing a Quantum Readiness Score to help researchers select datasets suitable for hybrid quantum-classical machine learning. The framework provides query-based discovery and reproducible tools for data-centric quantum battery analytics.

This study systematically characterizes a multiprobe grid algorithm for ANN search, revealing a d-scaling crossover on GloVe embeddings where grid maintains constant dimensional scaling while other methods degrade, with near-linear query scaling and lower indexing cost, suggesting competitiveness in high-dimensional settings and informing efficient transformer design.

This paper introduces a domain knowledge-based graph convolution network for ECG recognition, incorporating PRQST landmarks as domain knowledge. A double-stream directed graph models intra- and inter-cycle spatial and temporal dependencies. On the First Chinese ECG Intelligent Competition dataset, the method achieves an overall F1 score of 88.1% and 76.3% for rare categories, outperforming state-of-the-art models.

This paper studies adversarial robustness in programming-by-example systems, introducing a new attack mode where an adversary observes the synthesizer and chooses examples that most damage the returned program. The authors formalize worst-case corruption for finite version spaces, implement search algorithms, and propose Version-Space Partition Aggregation (VPA) as a defense. Experiments show that low-margin tasks have an adversarial robustness dimension missed by random-noise evaluations, and VPA only helps when clean semantics maintain a partition vote margin, which often fails in realistic tasks.

This paper proposes I²RiMA, a network that constructs spatial covariance matrices at each frequency point, maps them to the SPD tangent space, and uses frequency cluster aggregation and intra-inter slice attention to improve cross-subject EEG stress detection accuracy, achieving 82.78% balanced accuracy on three datasets with only 1.60M parameters.

The research proposes a multilayer Q-matrix-embedded neural network for cognitive diagnosis (M-QCDNet), which integrates the structural interpretability of cognitive diagnostic models (CDMs) with deep learning neural networks. M-QCDNet uses the Q-matrix as a structural prior to organize item-skill relationships, ensuring latent mastery profiles remain interpretable and consistent with cognitive theory. A loss function with L2 penalty balances predictive performance and structural alignment. Corresponding interpretable alignment-based metrics quantify the degree to which predicted skill activations correspond to item-level skills. M-QCDNet offers practical benefits for classroom practice, enabling early detection of learning difficulties and supporting mastery-based interventions. By embedding diagnostic validity into model design, M-QCDNet bridges psychometric transparency and neural flexibility, advancing interpretable, fair, and actionable AI for cognitive diagnostics.

Procedural Memory Distillation (PMD) captures cross-episode signals from RL rollouts, organizes them into three-level memory, and distills into policy via a memory-conditioned self-teacher, outperforming SDPO on benchmarks.

The study audits MedAgentBench v1/v2, finds a 41.7% silent-finish ceiling, and constructs MAB-v3 (508 tasks, 8.9% ceiling). Training Qwen3-8B reveals two structural barriers: a capability ceiling and a format-knowledge barrier. Pure RL achieves 18.2% pass@1 vs. 34.1% for rule-based SFT, a 15.9 pp gap entirely attributable to these barriers. A decision/format-knowledge/lookup taxonomy predicts RL learnability.

This paper explores using Reinforcement Learning with Verifiable Rewards (RLVR) to train small language models for tool-use in enterprise SaaS workflows, addressing the limitations of next-token prediction. In five synthetic environments simulating Jira and Confluence APIs, RL-trained models improved average reward from 0.35–0.92 to 0.95–1.00 on four non-degenerate scenarios, with the largest gain on Confluence page creation (0.35 to 1.00). Limitations include the scalability of hand-crafting rewards and a saturating reward in one scenario.

Researchers adapt a mixture-of-experts diffusion language model, DiffusionGemma-26B, for medical visual question answering, benchmarking it against its autoregressive sibling. The diffusion model matches or exceeds AR performance, decodes 3.5-4.4x faster, and offers any-order infill for drafting radiology reports, a capability inherently absent in autoregressive models.

CreativityNeuro is a data-free method that enhances divergent thinking in LLMs via contrastive weight steering, achieving up to 14 human percentile points improvement on creativity tests and reducing mode collapse.

This paper proposes a difficulty-routed service-control architecture for autonomous customer-service agents performing operational tasks like refunds and cancellations. A lightweight router keeps routine sessions on a low-cost baseline path and routes operationally coupled sessions to an escalated workflow that uses conflict-aware communication and write-triggered reconsideration to focus safeguards before consequential backend writes. Evaluated on retail and airline tasks, the method improves reliability consistently on conflicting requests, and improvements are not due to indiscriminate interaction expansion.

Understanding large, complex codebases remains challenging. Agent4cs is a multi-agent framework that summarizes codebases bottom-up using three agents: summarization, keyword extraction, and quality assurance. Evaluated on 7 models, it improves semantic consistency by 8% and keyword coverage by up to 38%.

Wiola is a fully original small language model architecture built from first principles, unrelated to existing families like GPT, LLaMA, Mistral, or Falcon. It introduces five novel components: Spiral Rotary Positional Encoding (SRPE), Gated Cross-Layer Attention (GCLA), Adaptive Token Merging (ATM), Dual Stream Feed-Forward (DSFF), and WiolaRMSNorm. Released in four sizes (120M to 1.5B parameters), it is fully compatible with HuggingFace Transformers.