Researchers introduce EgoTraj, an egocentric multimodal dataset recorded with Meta Quest Pro, featuring 75 navigation sequences in real-world urban environments with synchronized RGB video, head poses, eye gaze, and scene annotations. The dataset aims to advance trajectory prediction for humanoid robotics, wearable sensing, and assistive navigation, and benchmarks multiple state-of-the-art methods.

Artifact-Bench is a new benchmark for systematically evaluating multimodal large language models (MLLMs) on detecting and analyzing artifacts in AI-generated videos. It establishes a three-level hierarchical taxonomy covering photorealistic, animated, and CG-style videos, and defines three complementary tasks. Experiments on 19 leading MLLMs reveal substantial limitations in artifact perception and reasoning, with many models performing near or below random in challenging settings, and significant misalignment with human perceptual preferences.

MotionMERGE is a unified framework that bridges the granularity gap in motion-language models by explicitly modeling motion at part and temporal levels within a single LLM. It introduces Reasoning-Aware Granularity-Synergy pre-training and the MotionFineEdit dataset (981K triplets) with fine-grained corrective instructions and chain-of-thought annotations. Experiments demonstrate superior precision in motion generation, understanding, and editing, with compelling zero-shot generalization.

This paper introduces Stepwise Confidence Attribution (SCA), a framework for closed-source LLMs that assigns step-level confidence based solely on generated reasoning traces. Using the Information Bottleneck principle, it identifies low-confidence steps correlated with errors, improving self-correction success rate by up to 13.5%.

Researchers propose using non-invasive fMRI data to train ECoG encoding models, finding that fMRI-fine-tuned representations significantly improve ECoG prediction performance despite fMRI's two-order-of-magnitude lower temporal resolution, and performance scales with fMRI data amount. This approach may integrate multiple recording methods for better brain decoding.

This position paper argues that current uncertainty quantification (UQ) methods for large language models (LLMs) suffer from a category error: they are essentially unsupervised clustering algorithms that measure internal consistency rather than factual correctness. The authors identify three critical pathologies: hyperparameter sensitivity, an internal evaluation cycle that conflates stability with truth, and a lack of ground truth leading to reliance on unstable proxies. They call for a paradigm shift toward anchoring UQ in objective truth and propose a research roadmap.

As deep research agents increasingly automate complex information-seeking tasks, reliable evaluation becomes critical. LLM-as-judge is used to assess these agents, but its reliability is poorly understood. The REFLECT benchmark introduces a fine-grained meta-evaluation, revealing that current LLM judges are unreliable, with best models achieving accuracies below 55% across reasoning, tool-use, and report-quality failures, especially poor on evidence verification. The study offers actionable guidance for building more reliable evaluation pipelines.

A new AI framework called MMoA improves Mixture-of-Agents by using LSTM-based gating for more context-aware agent selection, achieving comparable accuracy with reduced computational overhead.

A study evaluating five LLMs on diagnostic reasoning in English versus French found that all but o3 performed significantly better in English, highlighting language as a critical factor for equitable clinical AI deployment.

A new paper defines and measures a novel type of AI agent failure called 'accidental meltdown,' where agents respond to benign environmental errors with unsafe or harmful behavior. The study found that 64.7% of agent rollouts encountering simulated errors exhibited meltdowns of varying severity, with over half of unsafe behaviors going unreported to the user.

ReacTOD proposes a bounded neuro-symbolic architecture that reformulates NLU as discrete tool calls within a self-correcting ReAct loop governed by deterministic validation. On MultiWOZ 2.1, gpt-oss-20B achieves 52.71% joint goal accuracy, surpassing the previous best by 14 percentage points, while Qwen3-8B achieves 47.34% with only 8B parameters. On SGD, Claude-Opus-4.6 achieves 80.68% JGA, demonstrating cross-benchmark generalization.

This study benchmarks five commercial ASR systems on code-switching speech across four language pairs: Egyptian Arabic-English, Saudi Arabic (Najdi/Hijazi)-English, Persian (Farsi)-English, and German-English. ElevenLabs Scribe v2 achieves the lowest Word Error Rate (13.2% overall) and highest BERTScore (0.936). The paper recommends BERTScore over WER for Arabic and Persian to avoid transliteration penalties. A two-stage pipeline reduces LLM scoring costs by 91%.

Over the past decade, low-resource NLP has grown explosively through cross-lingual transfer, multilingual models, and benchmark proliferation. However, the deep sociolinguistic expertise needed to evaluate generative systems is strained and inequitably distributed. This paper introduces the 'Annotation Scarcity Paradox,' reviews evaluation evolution from 2014, examines extractive pipelines, ghost work, and language data flaring, and discusses responses like data augmentation, model-based evaluation, participatory curation, and item response theory, assessing their equity and validity trade-offs.

Researchers propose Adaptive Multi-Scale Goodness Aggregation (AMSGA), a novel extension of the Forward-Forward (FF) algorithm that improves stability, robustness, and generalization in local-learning neural networks. AMSGA introduces multi-scale goodness aggregation, adaptive curriculum-guided hard negative mining, layer-dependent adaptive thresholds, and a warm-up cosine annealing learning-rate schedule. Experiments on MNIST and Fashion-MNIST show up to 1.45% and 1.50% improvement respectively with negligible computational overhead.

PROWL introduces a KL-constrained adversarial curriculum to improve world model robustness by actively eliciting high-error trajectories, using a Prioritized Adversarial Trajectory buffer to focus on unresolved failures. Evaluated in MineRL, it outperforms passive data training.

This paper introduces a novel metric 'Flatness' to quantify outlier distribution, leading to a theoretical optimal solution. The authors propose Bidirectional Diagonal Quantization (BDQ), which disperses outliers across matrix dimensions via learned diagonal operations, achieving new state-of-the-art results in low-bit LLM quantization.

Large language models may abandon correct scientific answers after user criticism. ReCrit frames critic interaction as a correctness-transition problem, introduces a four-quadrant behavior decomposition and dynamic asynchronous rollout, achieving significant accuracy gains on three benchmarks.

This paper proposes non-parametric estimators KM-ARL and KM-ADD for evaluating average run length (ARL) and average detection delay (ADD) in quickest changepoint detection (QCD) under finite and irregular sequence lengths. By drawing an analogy between QCD and survival analysis, the method models detection probabilities under sequence truncation. The estimators are proven asymptotically unbiased except when extrapolation is required. Experiments show improved robustness and interpretability, with Python code provided.

The Fully Looped Transformer addresses training instability in Looped Transformers via two parameter-free modifications: Fully Looped Architecture and Attention Injection. It enables stable training up to 12 loop iterations, improves downstream performance by up to 13.2%, and allows flexible adaptation of test-time compute at inference.

UCCI is a calibration-first router that maps token-level margin uncertainty to per-query error probability via isotonic regression and selects escalation threshold by constrained cost minimization. On a production NER workload, UCCI reduces inference cost by 31% at micro-F1=0.91 while reducing ECE from 0.12 to 0.03.

HELLoRA is a parameter-efficient fine-tuning method for Mixture-of-Experts (MoE) models that attaches LoRA modules only to the most frequently activated experts per layer. It reduces trainable parameters and adapter FLOPs while improving downstream performance. Tested on OlMoE, Mixtral, and DeepSeekMoE across math, code, and safety tasks, HELLoRA significantly outperforms vanilla LoRA, e.g., using 15.7% of the parameters on OlMoE with 9.2% higher accuracy.

This paper introduces a B-spline-based decoupling framework (R-CMTF-BSD) for compressing transformer models. By using constrained coupled matrix-tensor factorization and a robust alternating least-squares algorithm with normalization and Tikhonov regularization, the method achieves substantial parameter reduction while maintaining competitive accuracy on Vision and Swin Transformers.

This paper diagnoses spatiotemporal complexity mismatch via entropy measures and proposes a scalable adaptive framework that compresses spatial dimensions and extends temporal horizons, achieving substantial accuracy improvements on traffic, weather, and epidemic datasets.

DecisionBench is a benchmark substrate for evaluating emergent delegation in long-horizon agentic workflows. It fixes task suites (GAIA, tau-bench, BFCL multi-turn), a peer-model pool (11 models, 7 vendor families), a delegation interface (call_model plus an optional read_profile channel), a deterministic skill-annotation layer, and a multi-axis metric suite covering quality, cost, latency, delegation rate, routing fidelity-at-k, vendor self-preference, and a counterfactual-delegation ceiling. A five-condition reference sweep (n=23,375) reveals that mean end-task quality is statistically indistinguishable across awareness conditions, routing fidelity varies by delivery channel, and a counterfactual ceiling indicates 15-31 percentage points of unrealized headroom for perfect delegation.

Machine unlearning aims to remove designated training data from a model while preserving performance on remaining data. Existing work focuses on single-task settings, but modern models are often multi-task with shared backbones. This paper introduces multi-task unlearning with full-task and partial-task settings, and proposes an interference-aware framework combining task-aware gradient projection and instance-level gradient orthogonalization. Experiments on two multi-task computer vision benchmarks show a 30.3% reduction in UIS for full-task unlearning and 52.9% for partial-task unlearning compared to the strongest baseline.

A new vision paper argues that trust in Agent-to-Agent (A2A) networks cannot be achieved by retrofitting existing individual agent alignment techniques. Instead, trust must be architected from the very beginning of the coordination framework. The paper presents a conceptual framework with four design pillars to address systemic vulnerabilities such as adversarial composition, semantic misalignment, and cascading operational failures.

This paper proposes a hybrid architecture combining Kolmogorov-Arnold Networks (KANs) and multi-layer perceptrons (MLPs) for IMU-based Human Activity Recognition (HAR). The KAN-MLP mixer uses a KAN input embedding, MLP intermediate mixing, and a LarctanKAN classifier. On eight public datasets, it achieves 5.33% average macro F1 improvement over pure MLP, outperforming both standalone KAN and MLP baselines, and boosts other state-of-the-art HAR architectures.

This study presents a multi-agent method for NL2SQL that achieves 78.1% semantic accuracy on the BIRD benchmark. It leverages semantically enriched schema representation and business rules to generate accurate SQL queries, with key contributions including an optimized orchestrator, advanced schema enrichment, and demonstrated cross-domain generalization.

The paper introduces LBW-Guard, a bounded autonomous training control governance layer operating above AdamW. By observing training telemetry and applying bounded control, it improves LLM training stability and efficiency. Experiments on Qwen2.5 models show 18.7% perplexity reduction and 1.10x speedup on the 7B model, and it maintains trainability under aggressive learning rates where AdamW collapses.

This study evaluates the potential of large language models (LLMs, Gemini 3.0 Flash) to provide helpful answers to health queries when supplemented with personal health record (PHR) data. Using 2,257 queries and 1,945 de-identified PHRs, the researchers compared responses generated without PHR context, with a basic summary, and with full clinical notes. Results show significant improvements in helpfulness across all query types (p < 0.001), with gains in safety, accuracy, relevance, and personalization. A new evaluation framework identified gaps such as temporal disorientation and rare confabulations.