Researchers propose TheProfessor, a multi-teacher extension of PromptKD for distilling vision-language models. Using an ensemble of a domain-finetuned teacher and a zero-shot teacher, confidence-weighted ensembling improves harmonic mean accuracy by 1.77 points on average, with significant gains on domain-shifted datasets like EuroSAT.

REALM provides the first unified red-teaming benchmark for physical-world vision-language models, integrating 12 attack methods, 3 defenses, and 13 models to enable fair comparison of vulnerabilities. Key findings include the effectiveness of text and typographic attacks and the limited protective role of model scale.

A new method called HeRA aligns attention heads individually in multimodal LLMs, improving performance and reducing hallucinations.

Vision-Language Models (VLMs) are brittle to negation, relying on shallow co-occurrence and susceptible to misleading cues. HANCLIP restructures the embedding space with hyperbolic geometry and an angular triplet objective to encode negations, trained on 20K quadruplets, improving negation benchmarks without degrading standard performance.

ABACUS is a unified vision-language model that handles object counting, crowd counting, referring-expression counting, and count-faithful image generation without any benchmark-specific training. Built on a 3B-parameter foundation model, it introduces three key innovations: density-aware adaptive zooming with objectness maps, a boundary-aware count policy via GRPO, and a cycle-consistent GRPO strategy. It achieves state-of-the-art results across seven benchmarks, outperforming task-specific specialists and larger generalist models.

A paradigm shift from spatial to spectral feature processing for small object detection, introducing a frequency-guided framework with three lightweight modules that achieves superior performance with 1/6 the parameters of YOLOv11.

Recent work finds that attention distributions used for vision-language consistency in VLMs suffer from decoding drift and structural token biases. To address this, researchers propose PV-TAM, which leverages prompt-side semantics and peak attention distributions to evaluate alignment, outperforming answer-side baselines across multiple datasets.

Modern video diffusion models achieve higher generation quality through scaling, but this also increases inference cost. The Sol Video Inference Engine is a training-free, agent-based acceleration framework that integrates five techniques—cache, sparse attention, token pruning, quantization, and kernel fusion—for instance-specific optimization. Tested on three models of varying sizes, it achieves over 2x end-to-end acceleration with near-lossless VBench quality.

This paper presents a geometry-informed computer vision pipeline that automates overtaking event detection from a single bicycle-mounted camera without multi-sensor configurations or explicit calibration. Validated on 315 real-world events, it achieves 97.8% recall with zero false positives. The system identifies overtaking intentions a mean of 2.44 seconds before vehicle passage, with 84.1% of events exceeding the 1.5-second human reaction time threshold. A preliminary calibration-free lateral distance estimation approach achieves mean absolute errors of 13-14 cm, sufficient for safety categorization.

Standard tokenizer evaluation metrics like fertility rate fail to capture morphological correctness for agglutinative languages. The QuechuaTok benchmark compares four tokenization strategies on Southern Quechua, using morphological boundary accuracy (MorphAcc) alongside traditional metrics. Results show that while BPE achieves the lowest fertility rate (1.636), its MorphAcc is only 6.67%, whereas the morphology-aware PRPE tokenizer reaches 83.33% MorphAcc, demonstrating that fertility rate alone is insufficient for agglutinative languages.

This study challenges the effectiveness of exact-match retrieval recall as a proxy for retriever quality. In tau-bench, retrieved policy clauses performed nearly as well as gold policies in downstream classification tasks, despite only 7% exact-match recall. The findings suggest that relying solely on recall may underestimate the practical utility of retrieved policies.

This study audits eight automatic scorers across three evaluation constructs, finding that no scorer transfers across all datasets. In the generated-answer attribution construct, metric rankings invert and an NLI scorer collapses on long-form tasks. A prompt-based LLM judge avoids collapse but is costly and non-deterministic. The research concludes that metric choice must be validated on the target dataset.

A new study evaluates six proprietary LLMs across 16 DSM-5 conditions using adversarial attacks, finding that safeguards only reliably hold for suicide and self-harm, with failure rates up to 100% for conditions like eating disorders, substance use disorder, and major depressive disorder. The authors call for clearly defined harm categories and targeted safeguards.

This paper proposes a training-free framework called IBA (Identify-Before-Answer) for Knowledge-Based Visual Question Answering (KB-VQA). It decouples entity identification from evidence ranking by prompting an MLLM to select high-confidence entities from candidate names, then using an off-the-shelf text re-ranker for evidence. Experiments on Encyclopedic-VQA and InfoSeek show consistent outperformance over fine-tuned multimodal reranking baselines with reduced training and inference complexity.

A new framework uses LLMs to quantify product desirability from qualitative feedback without explicit scores, achieving Pearson correlations up to 0.97 and classification accuracy up to 94% on PDT datasets. GPT-4o-mini matches larger models at 94% lower cost, and the system includes confidence ratings and explainable AI.

A new study shows that self-generated text recognition (SGTR) finetuning can effectively prevent and reverse emergent misalignment (EM) in large language models, outperforming benign finetuning methods. The research finds that EM results from destabilization of a model's aligned character rather than learning harmful content, and SGTR works by fortifying character consistency.

This paper introduces the Normalized Context Utilization (NCU) metric to quantify contextual information gain in RAG systems. Experiments show that small language models match or outperform large models in strict factual extraction, while a commercial API overrides external evidence in nearly half of adversarial conflicts.

Proposes ModTGCN, a modularity-aware GNN for text classification that jointly optimizes cross-entropy and a modularity-based auxiliary objective to promote class-coherent document communities while preserving discriminative representations. Achieves consistent gains on five benchmarks, with larger improvements on low-homophily datasets.

EXPO-SQL proposes an execution-based clause-level policy optimization method that assigns fine-grained rewards to each clause in a SQL query by analyzing execution results, including error messages and incremental clause-wise execution, addressing the issue of insufficient learning signals caused by coarse-grained query-level rewards in existing RL methods. Experiments show it significantly outperforms existing supervised fine-tuning, prompting, and RL methods on multiple Text-to-SQL benchmarks.

A new method called Degeneracy Distillery automatically detects and resolves degenerate parameter combinations using Fisher information estimation and flattening, without requiring real data observations. It reduces simulation budgets for neural posterior estimation by up to 10x.

A deep learning approach using a multivariate time series graph neural network (MTGNN) reconstructs GRACE-like terrestrial water storage anomalies back to 1940 by learning from ERA5 meteorological data. The model achieves a basin-mean correlation of 0.94 and reproduces El Niño/La Niña events, using fewer predictors than existing methods.

Molecular surfaces encode the geometric and physicochemical patterns that determine antibody-antigen recognition, central to epitope prediction. However, existing methods rely on sequences or backbone structures and struggle to capture discontinuous, surface-driven epitopes. This study presents SurfBind, a surface-centric learning framework for epitope prediction that operates directly on molecular surface representations. SurfBind integrates geometric and physicochemical cues through a Transformer-based architecture with patch-level surface modeling, binder-aware cross-attention, and a hierarchical coarse-to-fine prediction paradigm. Experiments on challenging epitope identification benchmarks, including SAbDab and DB5.5, demonstrate that SurfBind achieves state-of-the-art performance and strong generalization across unseen antibodies and conformational states, highlighting the value of interaction-aware surface modeling for understanding the crucial mechanisms of protein-protein interactions.

A new study conducts a uniform re-evaluation of causal direction methods on the Tuebingen dataset, introducing a parameter-free compression baseline that achieves 74.7% accuracy and ties with top methods, revealing mechanisms that inflate published figures.

This paper proposes a novel meta-learning strategy called MEDIC that considers implicit gradient matching for both inter-domain and inter-class task splits simultaneously. It addresses the imbalance issue in one-vs-all classifiers for open set domain generalization, achieving better decision boundaries and outperforming prior methods while maintaining competitive closed set generalization.

A gray-box workflow, PC-MCMC-CIGP, integrates spike-and-slab topology sampling, hard conservation and thermodynamic screening, and a Chemical-Informed Gaussian Process (CIGP) residual model for extracting interpretable governing equations from sparse noisy chemical time-series data. On the H2+Br2 benchmark, it distinguishes elementary radical pathways; on styrene epoxidation, it improves yield by 12.5% over the baseline. A 10-seed acquisition study reveals trade-offs: PC-EI reduces low-yield BO suggestions, while EI-style criteria yield the strongest final performance.

A new approach to physical neural networks places trainable nonlinear functions on connections rather than using scalar weights, achieving low-power continuous control tasks with far fewer nodes. The design, implemented on analogue arrays, shows task-dependent benefits and projects to 30 microwatts in CMOS.

This paper presents a comprehensive survey on federated causal discovery and inference, proposing multi-dimensional taxonomies and highlighting the integration of causal structure learning and effect estimation in a unified pipeline, addressing challenges like privacy and data heterogeneity.

This paper investigates whether six offline RL losses (SFT, RFT, DFT, RIFT, Offline GRPO, DPO) are mechanistically distinct in weight-space geometry when used for reasoning distillation. Using identical math rollouts from Qwen3-4B, they find SFT, RFT, and RIFT have nearly colinear deltas; DFT diverges; Offline GRPO adds orthogonal components; and DPO lies in a near-orthogonal subspace with highest accuracy but a mode-connectivity barrier.

This paper presents an automated large-scale search pipeline for heterogeneous 4-Expert Mixture-of-Experts (MoE4) architectures within the LEMUR neural network dataset ecosystem. Over 28 days on an NVIDIA RTX 4090, the pipeline generated 4,463 candidate models and evaluated 1,021. A critical coverage bias was discovered: due to alphabetical enumeration, the search space was anchored to the AirNet family. Within this scope, ShuffleNet and MobileNetV3 ensembles achieved highest average accuracy of 0.632, while FractalNet and MNASNet were identified as low-yield families.

This study constructs large-scale algorithm co-occurrence networks in NLP using deep learning on full-text papers. It analyzes network structure and centrality to assess collective influence over decades, finding that classic, high-performing, and cross-period algorithms dominate, and declining influence first loses core network position.