This research uncovers naturally occurring statistical signals in vision data that can be exploited like backdoor triggers without malicious insertion. By analyzing ImageNet, the authors identify patterns strongly linked to specific labels, use statistical controls to remove spurious correlations, and demonstrate that these signals directly and predictably alter model predictions. These statistical adversaries are more targeted than generic corruptions and transfer across architectures, suggesting vulnerabilities stem from dataset structure rather than model idiosyncrasies. The study recommends treating spurious structure as a latent attack surface.

Light-Omni is a multimodal agent framework for reflexive video understanding using dual contextual states, achieving 12.1x speedup and 2.6x GPU memory efficiency over M3-Agent, and serving as a memory system for MLLMs.

Ground3D-LMM is a unified model that integrates point cloud and RGB image inputs to enable 3D spatial conversations with explicit point grounding and metric measurements. It introduces the 3D Grounded Measurement task and a large-scale dataset with 2.5M QA pairs, setting a strong baseline for grounded, metric-aware 3D dialogue.

A gaze estimation method using only one camera and one light source is proposed, introducing a virtual light source and virtual glint. Performance is acceptable but degraded compared to two-light-source systems.

This paper presents a task-driven framework linking synthetic fog generation, image restoration, detection, and tracking. Results show fog degrades performance mainly through missed detections; fog-inclusive training offers the most consistent robustness gains.

This paper introduces CanvasAgent, a tool-augmented multimodal agent that learns to orchestrate heterogeneous visual tools through multi-turn interaction for complex image creation and editing. The authors also present CanvasCraft, a large-scale dataset with 140K executable trajectories and 10K RL task specifications. The agent is trained with supervised fine-tuning and then optimized with GRPO using a hybrid reward combining outcome- and process-level signals. Experiments demonstrate effectiveness in both final image quality and trajectory behavior.

NAVER LABS re-implements its IWSLT 2025 instruction-following pipeline for the IWSLT 2026 Shared Task (constrained condition, short audio track), adapting to mandated components: SeamlessM4T-v2-large as speech encoder and Qwen3-4B-Instruct as LLM backbone. The three-stage approach (projector alignment, text-only LoRA pre-training, multimodal merging) is preserved. Additionally, 100k synthetic instruction-following examples across ten speech-centric task types (10k per task) are constructed. The primary model achieves COMET 0.781 on EN-ZH speech translation and BERTScore-F1 0.346 on English SQA on the MCIF benchmark.

BaFCo is a benchmark dataset for Bangla form comprehension, comprising 200 multi-page complex Bangladeshi government forms from diverse sectors. It features 26 fine-grained and 5 coarse entity types. Evaluations of latest MLLMs show limitations in localizing granular form entities.

This study re-evaluates the traditional linear relationship between language model perplexity and automatic speech recognition word error rate. It finds that modern end-to-end ASR systems, with their built-in language modeling capacity, challenge this assumption. The paper examines whether external LMs still improve current systems, the linearity of the PPL-WER relation in log-log space, the effect of encoder context length, and how LLM perplexities fit with standard neural LMs. Additionally, internal language modeling in attention-based encoder-decoder systems is investigated, showing that ILM subtraction alters the observed relation, highlighting the importance of considering the decoder's internal LM when interpreting external LM quality.

ResonatorLM replaces attention with physics-inspired causal resonant functions, treating token sequences as a 1D latent field. On a 6M parameter model, it achieves 6.47x decode speedup at 32K tokens and 61.31% accuracy on WikiText vs 55.32% baseline. Accepted at ICANN 2026.

New research shows that prompt robustness in LLMs differs significantly between objective and subjective questions, with variations across models, datasets, and prompt modifications. The study warns against interpreting model responses to subjective questions as direct indicators of beliefs.

A new study decomposes the yes-no bias in LLMs using crossed symmetrization, finding that frontier models' internal moral stance is nearly format-invariant, while Claude models show significant order and lexical biases, GPT-5.5 and Gemini near zero. Bias shrinks with extended reasoning and follows surface wording, not the verdict.

A new study reveals that LLM conformity in benchmarks is largely due to repeated wrong answers rather than social influence. Introducing a no-source condition, researchers found that 66.5% of correct answers are changed to harmful ones even without a speaker, compared to 10.3% under plain re-ask. This suggests benchmarks should measure the speaker-free floor first.

A new method for structural sequence analysis based on Algorithmic Information Theory uses the Ladderpath approach to extract nested and hierarchical repetitions. Three distance measures combined with k-NN classifier outperform gzip and BERT on OOD and few-shot text classification.

This paper presents a workload-aware benchmark of KV-cache optimization techniques including KIVI, TurboQuant, SnapKV, and CaM, evaluated on Llama-3.1-8B-Instruct and Mistral-7B-Instruct-v0.3 models across multi-document QA, single-document QA, few-shot learning, and summarization tasks. Results show that compression ratio alone is a poor predictor of end-to-end performance. KIVI4 offers the most stable quality across models, SnapKV delivers the strongest long-context throughput, and CaM yields large gains on selected QA workloads but exhibits substantial workload sensitivity. The study motivates workload-aware selection of KV-cache mechanisms.

A study examined the impact of persona prompts on large language model agents playing an iterated Split or Steal game. Using four open models interacting with a Virtual Human, cooperation dominated, but model choice and persona type significantly influenced strategies.

This paper proposes SCISE, a framework that addresses structural isolation in graph clustering by combining community-aware sampling with constrained structural entropy. It introduces three components: SECC for optimizing structural information, CSampE for preserving global topology during batch training, and StructCL for learning high-order structural representations. Experiments on six benchmarks show state-of-the-art performance.

Researchers from Hong Kong Baptist University propose a novel distance metric and clustering algorithm for categorical data with both nominal and ordinal attributes. The method unifies the handling of both attribute types while preserving ordinal order information, and jointly learns distance weights and data partitions to avoid suboptimal solutions.

AdaStop introduces a cost-benefit framework for DNN testing, dynamically stopping labeling when the marginal fault discovery rate drops below a threshold, enabling detection of 65-84% of faults with only 9-31% of the labeling budget.

This paper proposes treating the execution harness of LLM agents as a learnable control layer, formalized as a finite-horizon Harness MDP, and trains a lightweight controller via offline advantage-weighted regression. Experiments show consistent improvement in verification behavior and selective gains in final task quality, surpassing baselines like behavior cloning.

A new controllability-observability framework compresses deep neural networks by analyzing hidden-state redundancy through dynamical system theory, achieving over 70% state compression on MNIST and CIFAR-10 with minimal accuracy loss.

This paper introduces exogenous dropout, a model-agnostic training method that randomly zeros entire exogenous channels, significantly improving robustness to noise, temporal misalignment, and missing data while maintaining clean accuracy. Experiments across multiple domains show it outperforms specifically designed robust architectures.

This paper explores the 'Granularity Paradox' in time-series forecasting, where finer temporal disaggregation improves in-sample diagnostics but degrades out-of-sample accuracy due to recursive error compounding over longer horizons. Benchmarking 10 models across six granularities on a 13-year procurement dataset reveals a non-monotonic threshold: recursive models (e.g., Holt-Winters) degrade severely at high frequencies, LSTM exhibits a U-shaped error curve, and linear regression remains stable. Standard pointwise metrics mask cumulative error; a consensus-dissensus diagnostic is introduced.

GAIA is a geometry-aware, infrastructure-anchored learning framework that addresses non-line-of-sight propagation, burst noise, and long-tail errors in UWB ranging by combining temporal range modeling, latent anchor-layout estimation, and deterministic distance projection. On a real-world outdoor UWB dataset, GAIA reduces range MSE by 18.4% and improves polygon IoU by 15.5% over PoseMLP, enabling accurate work-zone reconstruction.

Design-CP introduces two context-parallel inference strategies for RFdiffusion 3—1D row-sharding and 2D grid sharding with ring attention—to distribute quadratic activations across multiple GPUs, enabling the design of large protein nanoparticles on limited memory. The 2D sharding shows better wall-clock scaling for icosahedral assemblies and octahedral design is demonstrated on workstation-grade 16GB GPUs.

This paper introduces Statistically Meaningful Geometry (SMG), modeling over-parameterized learning systems as infinite-dimensional non-parametric Orlicz fiber bundles. It proves that under persistent out-of-distribution stimuli, continuous optimization fails, unmodeled variance accumulates as Active Acausal Tension, triggering a Gauge Symmetry Break (GSB) registered as a discrete step-jump in Structural G-Entropy. SMG provides a parameter-free, falsifiable dashboard to mathematically certify true intelligence and transform AI for Science into an engine of autonomous paradigm shifts.

This research investigates the feasibility of using large language models (LLMs) to generate synthetic consumer data for projective techniques. By comparing LLM and human responses on city tourism perceptions across multiple tasks, the study finds substantial overlap in broad topics but significant differences in style, linguistic structure, and diversity generation. Recommendations are provided for optimal LLM use and recognition of limitations.

ArtisanCAD is an industrial-level CAD agent that uses an executable CAD intermediate representation (CAD-IR) to distill expert knowledge, handle ambiguous or incomplete natural language prompts, and generate editable parametric B-Rep models. On the Text2CAD benchmark, CAD-IR improves generation from intermediate prompts by reducing mean Chamfer Distance from 14.83 to 9.88.

Akashic is a low-overhead memory system for LLM-based agent systems that uses MemAttention to organize context into bounded chunks and model semantic relationships, avoiding full history replay and improving accuracy, throughput, and sustainable request rate.

This research proposes moving memory storage inside the language agent's reasoning loop, reading and writing at every step to overcome network latency. Experiments show that in-process storage (~100μs) reduces redundant actions from 7.2/12 to 0.0/12 and improves recall from 0/5 to 3.6–4.8/5. The bottleneck shifts to embedding generation rather than storage.