This paper introduces Orli (Ordered Regression of Lines), an end-to-end model that unifies text line detection and reading order prediction as a single image-to-sequence task. Trained on 196,691 pages across ten writing systems, Orli marginally exceeds state-of-the-art on cBAD line detection without dataset-specific training, achieves near-perfect coverage and ordering on multiple reading-order benchmarks zero-shot, and adapts to specialized out-of-domain layouts with limited fine-tuning. Code and weights are open-sourced.

Pinpoint is a new image geolocation method that combines internet photos and street-view imagery using a retrieve-and-rerank architecture. It trains a contrastive embedder on both Flickr and street-view data to learn a shared image-GPS space, then uses an attention-based reranker to rescore candidates. Without relying on multimodal LLMs, it achieves state-of-the-art results on multiple benchmarks.

A diffusion model fine-tuned for single-image reflection separation, jointly generating transmission and reflection layers with a cross-layer self-attention mechanism and disjoint sampling strategy, achieving state-of-the-art performance on real-world benchmarks.

A new benchmark, EVID-Bench, assesses search-based video misinformation detection. It includes 222 videos with 9 manipulation types in 3 categories. The best model achieves only 61.43% point-level and 43.24% video-level accuracy, with AI-generated manipulations being especially challenging.

This paper proposes a new approach to optimal transport flow matching by treating the prior as a design choice rather than a fixed input, avoiding the intractable computation of OT coupling in high dimensions. The authors identify low-frequency projection of natural images as a prior that admits an OT-optimal identity coupling with data, reducing the flow matching task to synthesizing high-frequency details. The method requires no modifications to the flow model and integrates naturally with latent-space models, classifier-free guidance, and one-step generation frameworks. Across benchmarks, trajectory curvature is reduced by more than 2x, improving generation quality in the few-step regime.

This paper proposes IN2R, a novel framework that shifts from discrete selection to synthesizing continuous soft prototypes to rectify noisy correspondence in cross-modal retrieval, leveraging the geometric stability of intra-modal data for graph reasoning, achieving state-of-the-art performance on multiple datasets.

The paper proposes SASA, a drift-resilient approach for weakly incremental learning in semantic segmentation. It uses learnable semantic anchors as rigid class-level references and elastic residual adaptation for stable yet flexible representation learning. A Spatial Label Arbitration mechanism filters noisy supervision and enforces the 'one object, one class' constraint. Experiments show state-of-the-art performance on standard benchmarks, especially in challenging multi-step incremental settings.

In embodied vision-language decision making tasks like robotic manipulation and navigation, VLMs and VLAs suffer from visual hallucinations due to inability to distinguish task-relevant objects from distractors. SceneDiver, a coarse-to-fine focus plan generation method, constructs a scene graph and progressively decomposes tasks to reduce hallucinations. A lightweight adapter distills this focus ability into VLAs for reactive control. Evaluations show reduced hallucinations and maintained efficiency.

MM-BizRAG is a novel multimodal retrieval-augmented generation approach for enterprise Q&A. It proactively extracts document structure via a structure-aware split and orientation-specific ingestion pipelines, enabling richer answers without finetuning. Experiments show up to 32% improvement over state-of-the-art baselines on enterprise and public benchmarks. The paper also introduces FastRAGEval, a cost-effective LLM judge metric. Accepted at ACL 2026 Industry Track.

This study investigates cross-prompt generalization in detecting AI-generated fake news using interpretable linguistic features such as lexical diversity, readability, and emotional intensity. A random forest classifier trained on one prompt's AI-generated articles and tested on another achieves AUC values between 0.988 and 1.000 across all six combinations, demonstrating robust detection under prompt variability.

This paper introduces ACAT, a web-based collaborative annotation tool that natively supports four ABSA workflows and provides an automated ETL pipeline to compute inter-annotator agreement metrics directly at export. Preliminary validation on 1,002 restaurant reviews shows a median annotation time of 31.58 seconds and IAA ranging from 0.78 to 0.86.

A large-scale empirical study assesses 284 interpretable linguistic features across 27 LLMs and ten text domains, finding that classifiers using only linguistic features can reliably distinguish AI-generated from human text, but many features are context-dependent while lexical richness remains robust across models and domains.

This study extends refusal steering methods to Mixture-of-Experts (MoE) large language models, demonstrating that steering performance is unaffected by the complex routing patterns of MoE. The authors propose two expert-aware refusal steering methods that leverage refusal-specific routing patterns and expert-specific directions to suppress normal refusal behavior. Results show that effective steering can be achieved based on a single expert's output, and that refusal signals differ from expert routing behavior, highlighting a significant role for attention.

A large-scale study finds that retrieval-augmented generation (RAG) yields only small and inconsistent improvements (1-2 points) over no-retrieval baselines in biomedical QA. The backbone model matters far more than retriever or corpus choice, and expert vs. layman sources perform similarly.

SALIMORY is a new framework that trains a single language model to manage cognitively-structured memory (user facts, preferences, working memory) for conversational agents, addressing the credit assignment bottleneck. Using hierarchical stage-wise process reward and reward-decomposed contrastive refinement, it supervises selective filtering, consolidation, and cue-driven recall end-to-end. It cuts memory-attributed failures by one-third, outperforms SOTA by over 10% in accuracy, and more than doubles the Good Personalization rate.

This article situates large language models (LLMs) within the longer history of computational approaches to concept analysis in the history, philosophy, and sociology of science (HPSS). It examines what LLMs add to existing methods, how they inherit longstanding problems, and reviews recent case studies. The first part reconstructs computational conceptual history before LLMs, covering early digital methods, distributional approaches, and lexical semantic change detection. The second part focuses on the LLM era, reviewing LLM-based work on lexical semantic change detection and relevant case studies, and revisiting methodological issues such as corpus construction, model choice, and operationalization.

A new study investigates how discourse-role labels such as 'Reference:', 'Instruction:', and 'Example:' affect language models' reliance on context. Using a paired fixed-content probe on 500 MMLU-Pro items, the researchers found that labels can shift misleading adoption rates by 56-84 percentage points. Labels like Instruction and Reference increase adoption, while Example suppresses it. The study recommends that RAG benchmarks report and control wrapper labels.

POLARIS is a training recipe for small open-weight models that significantly improves long-form creative writing. By combining a frontier LLM judge with human-reference injection in a GRPO framework, the resulting 9B model matches the performance of much larger models and exhibits strong length generalization.

arXiv:2606.04050v1 Announce Type: new Abstract: Existing quantization methods are fundamentally limited by rigid, integer-based bit-widths (e.g., 2, 3-bit), resulting in a ``deployment gap" where Large Language Models cannot be optimally fitted to specific memory budgets. To bridge this gap, we introduce LiftQuant, a novel framework that enables continuous bit-width control for true Pareto-optimal deployment. The core innovation is a ``lift-then-project" mechanism which approximates low-dimensional weight vectors by projecting a simple 1-bit lattice from a higher-dimensional ``lifted" space. Crucially, the effective bit-width is determined simply by the ratio of the lifted dimension to the original dimension, which allows the bit-width to be tuned quasi-continuous as the dimension is a flexible structural parameter. This projection generates a structured yet non-uniform codebook, capturing the expressive power of Vector Quantization (VQ). While beneficial over VQ, LiftQuant's decoding path relies solely on linear transformations and 1-bit uniform quantizers, retaining hardware-friendly nature. This flexibility is transformative: LiftQuant enables a 70B LLM to be compressed to 2.4 bits to precisely fit a 24GB GPU, where its performance significantly surpasses state-of-the-art 2-bit models fitted on the same device. Our code and ckpt is available at https://github.com/Heliulu/LiftQuant.

This paper extends the Maximal Update Parametrization (μP) to Gated Delta Networks, an efficient linear architecture. By rigorously propagating coordinate-size estimates through the forward pass, gating mechanisms, and recurrent state dynamics, the authors derive scaling rules that enable stable learning-rate transfer across model widths under AdamW and SGD, whereas standard parametrization fails.

This paper defines Bayes-sufficient representations for a fixed supervised decision problem, introduces the Bayes quotient, and distinguishes sufficiency from minimality. Experiments validate the framework.

This paper presents a methodology for inverse design of critical experiments using deep neural network surrogate modeling and nonparametric gradient optimization to maximize neutronic similarity (c_k). The model combines a U-Net encoder-decoder with a novel multigroup attention pooling layer, enabling gradient-based optimization of material assignments in a geometry grid. Applied to a TN-LC transportation cask with HALEU fuel, the method achieves c_k scores of 0.97757, 0.81324, and 0.93276 for three configurations, demonstrating the potential of deep learning in accelerating nuclear technology validation.

A systematic study of projection sharing in the query-key-value (QKV) attention mechanism of transformers, showing that sharing key and value projections (Q-K=V) reduces KV cache by 50% with only 3.1% perplexity degradation. Combining with grouped-query attention (GQA) or multi-query attention (MQA) achieves 87.5% and 96.9% cache reduction, respectively, enabling practical on-device inference. Experiments span synthetic tasks, vision, and language modeling. Code is publicly available.

This paper presents a sharp pseudospectral analysis of transient amplification in coupled gradient descent algorithms, which are common in bilevel optimization, two-time-scale stochastic approximation, and adversarial training. The authors prove bounds on the Kreiss constant for block-triangular Jacobians and provide a finite-horizon iteration-complexity bound. Their theory reveals a non-asymptotic regime invisible to traditional spectral analysis, confirmed by experiments on linear-quadratic problems and neural network training.

This position paper argues that real-world RL systems must adopt continual learning over the train-then-fix paradigm, identifying four sources of post-deployment non-stationarity and highlighting successful continual RL examples.

The IEEE P3109 draft standard defines a parameterized family of binary floating-point formats and associated operations, tailored for machine learning. These formats allow efficient representation in few bits, with parameters including width, precision, signedness, and infinities. Operations are exception-free, using stochastic rounding, and signal exceptions via NaN. The standard also introduces kappa-approximation for approximate implementations and is mechanically verified.

A new study using XGBoost and eight routine clinical features from the ADNI dataset achieves high-precision three-class classification of normal cognition, mild cognitive impairment, and Alzheimer's disease, with a macro AUC of 0.982 and explainability via SHAP values.

This paper investigates the challenge of determining when to interrupt autonomous AI agents during long-horizon tasks. Using the HEART affective dynamics model, the authors evaluate four types of intervention triggers and report three major findings: a state saturation trap where frustration indicators quickly max out, a capability floor for LLM-based judges, and, most critically, very low inter-rater reliability among human annotators, questioning the validity of using single-annotator F1 scores as optimization targets.

This study examines how mathematicians use AI for formalizing mathematical proofs. Through surveys and a controlled user study, it finds that people desire AI assistance while retaining high-level control, using AI improves formalization accuracy, and users tend to employ multiple AI tools flexibly.

A new benchmark, Curation-Bench, tests whether generalist coding agents can autonomously curate training data. While out-of-the-box agents match published baselines within ten iterations, they tend to tune local variants rather than explore new policy families. Scaffolded agents that require citing and adapting prior methods autonomously discover a policy that outperforms baselines at one-tenth the data budget, suggesting that structured method adaptation is key for reliable data curation automation.