This paper introduces Forecast-GS, a framework that predicts task-completed 3D states to improve language-guided robotic manipulation. Experiments on three real-world tasks show significant improvements over baselines, with automatic selection achieving up to 23/25 success rate, but highlighting room for improvement in automatic ranking.

Researchers propose ASIP-Planner, an adaptive UAV inspection framework for indoor environments with partial knowledge. It combines a segment-based global planner and a local view-angle adaptation module to handle unforeseen obstacles, achieving near-complete coverage with shorter trajectories in simulations and real-world tests.

ForceFlow is a force-aware reactive framework built on flow matching for contact-rich robot manipulation. It uses an asymmetric multimodal fusion architecture treating force as a global regulatory signal, combined with a joint prediction paradigm to achieve deep force-motion coupling. The framework hierarchically decomposes tasks into a vision-dominant approach stage and a touch-dominant interaction stage with a Vision-to-Force handover. Experiments on six real-world tasks show a 37% success rate improvement over ForceVLA, reduced cost, and strong zero-shot OOD generalization.

Researchers propose ECHO, a memory framework that maps VLA hidden states into a continuous hierarchical space using a hyperbolic autoencoder, building a semantic memory tree with background consolidation. It achieves a 12.8% absolute improvement on LIBERO-Long over the π0 baseline, enhancing long-horizon task performance and compositional generalization.

DenseTRF is a self-supervised framework that uses texture-centric attention to adapt surgical dense prediction models to new distributions without supervision, improving generalization across procedures.

ABRA is a new benchmark for radiology agents that must navigate an OHIF viewer and Orthanc DICOM server via 21 tools. It includes 655 tasks across 8 types and 3 difficulties. Testing 10 models reveals high execution but low outcome rates, with the bottleneck being perception, not tool orchestration.

To automate clinician-grade assessment reports for surgical procedures, researchers propose a Perception-Alignment-Reasoning framework featuring Hi-GaTA, a lightweight temporal adapter that efficiently compresses long video sequences into LLM-compatible visual prefix tokens. By pretraining Sur40k, a surgical-specific video encoder on 40,000 minutes of public videos, employing a temporal pyramid with text-conditioned dual cross-attention, and using cross-level gated fusion, the method achieves coherent report generation under limited supervision, outperforming strong MLLM baselines.

Political and social identities structure how people evaluate political information, yet computational tools often produce single scores that treat content as if it means the same to everyone. This paper develops the Perspectivist Visual Political Sentiment (PVPS) classifier, learning from ~82,000 evaluations by 5,575 U.S. adults to predict how identity-defined audiences evaluate the same image. Applied to studies of visual sentiment, PVPS reveals that perceived violence and emotional mechanisms behind protest image engagement change substantively when audience identity is accounted for.

LatentHDR proposes a framework that decouples scene generation from exposure modeling in latent space. A pretrained diffusion backbone produces a single coherent scene representation, while a lightweight conditional latent-to-latent head maps it to exposure-specific representations, enabling single-pass generation of a dense, structurally consistent exposure stack. This eliminates multi-pass diffusion, ensures cross-exposure alignment, and reduces computation by an order of magnitude. It supports both text- and image-conditioned HDR generation for perspective and panoramic scenes, achieving state-of-the-art dynamic range on SI-HDR benchmark.

This paper leverages the linear additivity property of VLM embedding spaces to decompose scene representations into foreground and background components. Using synthetic data, the proposed pre-training method achieves the first worst-group accuracy above 90% on Waterbirds under perfect spurious correlation (100% spurious correlation, no minority-group examples), with strong sim-to-real transfer and no need for real-world debiased data.

HiDream-O1-Image is a natively unified visual generative foundation model via pixel-space Diffusion Transformer, shifting from modular architectures to an end-to-end in-context visual generation engine. It maps raw image pixels, text tokens, and task-specific conditions into a single shared token space, eliminating the need for separate VAEs or disjoint text encoders. With only 8B parameters, it matches or surpasses much larger models like 27B Qwen-Image. Scaling to over 200B parameters, HiDream-O1-Image-Pro achieves state-of-the-art performance on multiple benchmarks.

Researchers release the first global dataset of agricultural field boundaries at 10m resolution for 2024 and 2025, comprising 3.17 billion polygons across 241 countries. Using a U-Net model on Sentinel-2 imagery, it achieves a mean pixel-level recall of 0.85 validated in 24 countries.

This paper proposes a principle-guided uncertainty supervision framework (PriUS) based on evidential learning, which enforces three perception-aligned principles—image contrast, corruption severity, and geometric complexity—during training to improve the interpretability and consistency of uncertainty estimates in medical image segmentation. Experiments on ACDC, ISIC, and WHS datasets show PriUS achieves more consistent uncertainty estimates while maintaining competitive segmentation performance compared to state-of-the-art methods.

ReAD is a reinforcement-guided capability distillation framework that accounts for capability interdependence. It infers task-essential capabilities, generates targeted supervision, and uses an uncertainty-aware contextual bandit to allocate distillation budget adaptively, improving downstream utility under fixed token budgets.

Researchers present Hebatron, an open-weight Hebrew LLM built on NVIDIA Nemotron-3 MoE architecture. It achieves 73.8% Hebrew reasoning accuracy with only 3B active parameters per forward pass, outperforming prior models and rivaling larger models like Gemma-3-27B, while providing 9x inference throughput and 65k token context.

This paper presents the RETUYT-INCO participation in the BEA 2026 shared task on rubric-based short answer scoring for German. The team introduced Meta-prompting, where an LLM creates a custom prompt from training examples to grade new answers. They also used classic ML, fine-tuning open-source LLMs, and prompting techniques. Official results: Track 1: 6th/8 QWK 0.729; Track 3: 4th/9 QWK 0.674; Track 4: 4th/8 QWK 0.49.

ReVision trains multimodal language models to remove redundant visual patches between consecutive screenshots, reducing token usage by ~46% on average while improving success rate by 3%. This efficiency gain enables agents to process longer trajectories with fewer tokens, and reveals that performance saturation in visual history is due to inefficient token representation.

A new study reveals that instructions in language models primarily affect the output production stage, not the input processing stage. By probing task-specific information layer-wise across five binary judgment tasks, researchers found an asymmetry: information in sample tokens remains stable and weakly correlated with behavior, while information in output tokens varies substantially and strongly correlates with behavior. Attention-based interventions confirm that blocking instruction flow to all subsequent tokens reduces both behavior and information in output tokens, whereas blocking it only to sample tokens has minimal effect. The asymmetry generalizes across model families and tasks, and becomes sharper with model scale and instruction-tuning.

A systematic evaluation of social bias in a pretrained LLM trained with DP-SGD, comparing DP model against non-DP baselines across four paradigms. DP reduces bias in sentence scoring tasks but this improvement does not generalize. There is a discrepancy between logit-level and output-level bias, and decreasing memorization does not necessarily reduce unfairness.

The Bicameral Model couples two frozen language models via a trainable neural interface on hidden states, enabling continuous communication without text serialization. It achieves dramatic accuracy gains on arithmetic, logic puzzles, and mathematical reasoning tasks.

This paper decomposes an evolutionary mixture-of-LoRA system on a ~150M-parameter substrate into three factors: a router rewrite, a per-domain leave-one-out evaluation scope, and a lifecycle of death, inheritance, mutation, and slot reallocation. The router rewrite accounts for the entire improvement over a static baseline (+0.0426 nat log-PPL), while the lifecycle is a net drag (approx -0.028 nat). A synthetic sandbox reveals a substrate-conditional boundary: evolutionary search on the routing channel is only beneficial when adapters are pre-aligned to the task.

This paper presents ClinicalBench, a 400-question benchmark for evaluating assertion-aware retrieval in clinical QA, and EpiKG, a knowledge graph that enriches facts with assertion and temporality labels. Experiments show significant improvements over dense-RAG baselines, with physician adjudication revealing that 56% of auto-generated reference answers are defective, highlighting the necessity of human review.

A new paper introduces a validity-diversity framework showing that diversity collapse in LLMs stems from two forms of miscalibration during decoding: order calibration (valid tokens not reliably ranked above invalid) and shape calibration (probability mass over-concentrated on few valid tokens). Experiments across 14 models confirm that diversity issues are not just sampling heuristic flaws but fundamental calibration problems.

This paper proposes a plug-and-play framework that projects ESM-2 protein language model representations onto protein contact graphs and applies SoftBlobGIN, a lightweight Graph Isomorphism Network with differentiable Gumbel-softmax substructure pooling, to perform structure-aware message passing and learn coarse functional substructures for downstream prediction tasks. It achieves 92.8% accuracy and 0.898 macro-F1 on enzyme classification, and improves residue AUROC from 0.885 to 0.983 on binding-site detection. The framework requires no retraining of the language model, adds only ~1.1M parameters, and provides directly auditable structural explanations, making protein language model predictions more transparent.

Reinforcement learning alignment methods often suffer from reward hacking, reducing diversity. TMPO addresses this by matching trajectory-level reward distributions rather than maximizing scalar rewards. It introduces Softmax Trajectory Balance (Softmax-TB) and Dynamic Stochastic Tree Sampling, achieving a 9.1% improvement in generative diversity over state-of-the-art methods.

ξ-DPO proposes a novel direct preference optimization method that addresses the joint hyperparameter tuning issue in SimPO by introducing a ratio reward margin (ξ). The method redefines rewards to yield a bounded and interpretable margin that can be derived from the initial reward distribution, eliminating trial-and-error tuning.

Diffusion Language Models (dLLMs) promise high parallelism but are constrained by stringent confidence thresholds. Through token-level analysis, we find many tokens converge early but fail to meet confidence standards, indicating conservatism. We propose LEAP, a training-free, plug-and-play method that uses future context filtering and multi-sequence superposition to detect early-converging tokens, enabling reliable early decoding. LEAP reduces denoising steps by ~30% and, combined with dParallel on GSM8K, achieves 7.2 tokens per step while maintaining precision, breaking reliance on high-confidence priors.

This paper introduces the Hierarchical Multi-view HAAR (HMH) framework for heterophilous graph classification. HMH learns signed affinities via a heterophily-aware encoder, constructs a soft graph hierarchy, and applies sparse orthonormal Haar wavelet bases for spectral filtering at each level, effectively mitigating hub-dominated aggregation, oversmoothing, and oversquashing. Experiments show up to 3% improvement on node classification and 7% points on graph classification while maintaining near-linear scalability.

A new method called Vertex-Softmax provides tighter bounds for certified verification of transformer attention by exactly optimizing softmax over interval constraints, improving verification rates and lower bounds on multiple datasets while being computationally efficient.

Proposes RPSFT, which penalizes changes in projected top-k singular-vector blocks of pretrained weight matrices to limit unnecessary rotation while preserving task adaptation, improving in-domain/OOD trade-off.