TruEye is a novel model for fine-grained detection and localization of AI-generated or manipulated humans and scenes, distinguishing among five compositional categories of synthetic content. It runs over 100x faster than LLM-based competitors and outperforms state-of-the-art detectors on multiple datasets.

ReWorld is the first representation learning framework specifically designed for autonomous-driving world action models. It directly optimizes intermediate representations along three dimensions: future-predictive supervision on the generation module, cross-modal alignment and hard-negative supervision on the planning module. Experiments on nuScenes and NAVSIM show significant improvements in video generation quality and closed-loop planning performance, along with faster convergence.

Aloe-Vision introduces a family of open-source medical vision-language models trained on a large-scale quality-filtered dataset, achieving balanced performance and exposing vulnerabilities to adversarial inputs.

DMV-Bench is the first interactive benchmark for multimodal-agent visual memory, built on a home-furnishing e-commerce catalog of 1,000 product variants. Each product image carries a unique incidental cue; agents must recall cued products after long shopping sessions. The proposed DualMem architecture, maintaining parallel visual and verbal codes, outperforms baselines on Gemini 2.5 Flash and Qwen2.5-VL-7B.

SelectAnyTree is a promptable instance segmentation model that segments individual trees from 3D forest LiDAR point clouds with few clicks. It features a click-to-query prompt encoder and a Canopy Height Model-guided first prompt, with a state-space query decoder for efficient long-range context. Evaluated on seven diverse forest regions and a held-out dataset, it achieves 78.2 IoU from a single click, outperforming baselines by 24.8 points.

Researchers fine-tuned Gemini 2.5 Pro on 400 clinician-rated home videos using low-rank adaptation, achieving significant improvements in inter-rater reliability and ASD diagnosis accuracy, matching or exceeding clinician performance. The approach enables scalable behavioral feature extraction for autism assessment.

SemCityLoc is a novel aerial localization method that reframes pose estimation as structured surface registration between foundation-model visual priors and standardized LoD 3D city models. It eliminates reliance on precise GNSS or radiometric 3D reconstructions by aligning semantic surfaces and monocular depth with lightweight building models. The benchmark SemCityLockeD, combining centimeter-accurate UAV poses with LoD1-LoD3 models, is introduced. Experiments show up to 36% recall improvement and mean positional error reduction from 9.89m to 2.62m in urban canyons.

A new framework called Transformation-Aware Decoupling (TAD) improves viewpoint robustness in 3D scene graph generation by decoupling relation reasoning into viewpoint-sensitive and viewpoint-invariant components.

Ultra-fine entity typing (UFET) assigns highly specific types to entity mentions, but struggles with long-tail types due to reliance on sentence-level context. We present Narrative-UFET, which pairs each mention with an automatically generated narrative to provide cross-sentence context. Experiments show consistent improvements on long-tail types, especially when the entity's type changes across the narrative. Synthetic narratives outperform natural contexts, suggesting controlled discourse construction can surface implicit signals.

Ko-WideSearch is a Korean breadth-search benchmark built via an automated synthesize-and-verify pipeline, comprising 228 tables over 190 entities across 16 categories and three difficulty tiers. Evaluations on 20 web agents show consistent failure: agents recover sets but not rows (Item-F1 92.8 vs Row-F1 53.7), accuracy drops with difficulty, and open-ended free-text cells are the main bottleneck.

EntMTP is a training-free scheduler that dynamically switches tree-based attention topologies based on local generation entropy, enabling deep speculation in low-entropy regions and conservative speculation in high-entropy regions. It maximizes throughput without sacrificing quality, achieving 1.15x speedup over Hydra and peak 1.36x over Medusa on various benchmarks.

A new recurrent neural network architecture that pre-contextualizes tokens using a transformer block and correction network, achieving significant speedups over standard transformers while maintaining or improving performance.

Large language model (LLM) agents struggle to update facts in long-term interactions. Replacing full context with bounded memory drops accuracy from 92% to 77% even on frontier models. The gap scales with conversation length, not memory size. The authors introduce Supersede, a reinforcement learning environment that trains agents to prioritize current facts over superseded ones. Fine-tuning Qwen2.5-3B in this environment nearly doubles held-out accuracy (9.0% to 16.7%).

This study uses a developmental approach to investigate the statistical learning and mental representation of neural language models (NLM). A series of Generative Transformer models are trained on a synthetic grammar, and model states are saved at multiple stages. By analyzing changes in internal representations, the authors find that NLMs acquire the most abstract global statistical knowledge at the beginning of learning, and later acquire local statistical dependencies. This learning path contains many over-generalizations from the start, which are gradually constrained later. Based on this observation, a new framework is proposed to explain the statistical learning and language cognition of NLMs.

Team HSA_CORAL presents their submission to the FinCausal 2026 shared task, comparing three model families for cause-effect extraction from financial narratives via QA in English and Spanish. Supervised fine-tuning yielded the best results, with GPT-4.1 Mini achieving top scores on the English subtask and third on Spanish, highlighting the value of multilingual fine-tuning.

This survey systematically reviews automated presentation coaching systems covering pronunciation, fluency, prosody, multimodal, and Q&A tools. It introduces a five-dimensional task taxonomy (segmental pronunciation, lexical stress, suprasegmental prosody, pacing, content faithfulness) and maps surveyed systems to reveal coverage gaps. Core methods include TTS-based exemplar generation and diagnostic assessment. Open challenges include scarcity of annotated corpora, accent-fair feedback, and low-latency real-time diagnostics.

This position paper argues that the term "machine unlearning" is overused in LLM research and should be reserved for dataset-defined deletion. Many tasks currently labeled as unlearning pursue different objectives and require separate terminology and evaluation methods. The authors call for stricter terminology aligned with explicit guarantees and reference models.

Researchers introduce an axiomatic evaluation framework for latent thought representations in LLMs, defining four axioms (Causality, Minimality, Separability, Stability) with quantitative metrics independent of downstream accuracy. Auditing 23 reasoning tasks, they find no candidate satisfies all axioms, representations cannot distinguish questions within the same task, and encode little beyond input embeddings. The failure persists across model families, indicating a structural gap.

This paper presents Privacy-Preserving Probabilistic Race/Ethnicity Estimation (PPRE) to enable fairness measurements with respect to race/ethnicity for U.S. LinkedIn members while preserving privacy. It combines the Bayesian Improved Surname Geocoding estimator with a sparse golden survey set of self-reported demographics, and applies secure two-party computation, differential privacy, and additive homomorphic encryption.

This study evaluates ten reduced-domain boundary treatments for predicting bottom-hole pressure (BHP) and CO2 plume migration in geological carbon storage. Conserving corner pore volume is critical; uniform treatments cause large errors, while corner corrections significantly improve accuracy. The gradual modifier with transmissibility correction performs best, achieving BHP NRMSE below 3.7% and plume IoU above 0.97 in both homogeneous and heterogeneous reservoirs.

A new study shows that activation patching, a key tool in mechanistic interpretability, conflates the natural indirect effect (NIE) with interaction effects (INT), which measure how a component's causal influence depends on other components. This can lead to misattribution of importance and instability in faithfulness scores. The authors demonstrate failures in GPT-2's IOI circuit, prove INT is inevitable but useful as a diagnostic, and show it scales with activation distance and decomposes into pairwise and higher-order interactions.

Researchers propose a geometry-conditioned Fourier neural operator (FNO) for the cubic nonlinear Schrödinger equation on two-dimensional flat tori with varying aspect ratios. The operator takes the real and imaginary parts of the solution along with the aspect ratio parameter ω² as input and learns the one-step solution operator. Experiments show it captures distinct dynamics on rational and irrational tori and reproduces Sobolev norm behavior. Ablation studies indicate that including ω² improves long-term predictive accuracy.

A new Transformer architecture called Prism Transformer replaces uniform head allocation with a progressive head schedule, increasing head count per layer to create a local-to-global representational hierarchy. It achieves consistent improvements on multiple zero-shot benchmarks without extra parameters or compute.

This paper studies regret minimization in Markovian bandits with non-observable states and constrained decision epochs. It introduces self-degrading Markovian bandits, proves that without prior knowledge, algorithms that switch arms rarely suffer super-logarithmic regret, and designs UCB-NOM which achieves nearly logarithmic regret, or O(log T) with prior knowledge.

PairSAE summarizes pairwise tensors via N-mode SVD into token-wise interaction roles, then uses a sparse autoencoder to learn shared token-level features that decode into both sequence and pair representations, enabling interpretability of protein co-folding models. Evaluated on Boltz-2 activations, features align with UniProt annotations and predict affinity.

Darts, a popular open-source time series library, introduces a unified interface for foundation models, enabling zero-shot forecasting, fine-tuning, and backtesting with minimal code changes.

This paper presents a new RANSAC scoring method that analytically marginalizes the inlier scale, eliminating the need for a user-supplied scale parameter. The method outperforms state-of-the-art on a benchmark of nearly 70,000 image pairs, staying robust under threshold miscalibration and achieving near-optimal accuracy with as few as two validation pairs.

OverFlowLight is a real-time framework that detects and prevents vehicle queue overflow at intersections using multi-modal sensing, and employs a hybrid control (rule-based + reinforcement learning) to dynamically generate clearing phases. Deployed at 43 intersections in three cities, it reduces overflow incidents by 60.4% and increases network throughput by 18.2%.

This paper compares agent-based and parameterized world models for LLM agents, proposing GILP, which combines a small parameterized backbone with LLM reasoning to reduce hallucinated state rates from 0.176 to 0.035 and increase success rates from 0.668 to 0.838 with only ~22% extra LLM calls.

This paper studies long-horizon rollout error in Graph World Models (GWMs). The authors formulate a unified fixed-edge and dynamic-edge GWM framework and develop graph-valued rollout bounds to separate topology-induced from model-induced amplification. They propose Error-Aware GWM combining spectral regularization, rollout consistency, and critical-node weighting. Experiments show rollout error and planning regret grow with horizon, dynamic-edge training is needed when structure evolves, and Error-Aware GWM prevents long-horizon divergence while preserving prediction accuracy.