We present XANE(3), a physics-based E(3)-equivariant graph neural network for predicting X-ray absorption near-edge structure (XANES) spectra directly from atomic structures. The model combines tensor-product message passing with spherical harmonic edge features, absorber-query attention pooling, custom equivariant layer normalization, adaptive gated residual connections, and a spectral readout based on a multi-scale Gaussian basis with an optional sigmoidal background term. To improve line-shape fidelity, training is performed with a composite objective that includes pointwise spectral reconstruction together with first- and second-derivative matching terms. We evaluate the model on a dataset of 5,941 FDMNES simulations of iron oxide surface facets and obtain a spectrum mean squared error of $1.0 \times 10^{-3}$ on the test set. The model accurately reproduces the main edge structure, relative peak intensities, pre-edge features, and post-edge oscillations. Ablation studies show that the derivative-aware objective, custom equivariant normalization, absorber-conditioned attention pooling, adaptive gated residual mixing, and global background term each improve performance. Interestingly, a capacity-matched scalar-only variant achieves comparable pointwise reconstruction error but reduced derivative-level fidelity, indicating that explicit tensorial channels are not strictly required for low intensity error on this dataset, although they remain beneficial for capturing finer spectral structure. These results establish XANE(3) as an accurate and efficient surrogate for XANES simulation and offer a promising route toward accelerated spectral prediction, ML-assisted spectroscopy, and data-driven materials discovery.

Historical approaches to Table Representation Learning (TRL) have largely adopted the sequential paradigms of Natural Language Processing (NLP). We argue that this linearization of tables discards their essential geometric and relational structure, creating representations that are brittle to layout permutations. This paper introduces the Platonic Representation Hypothesis (PRH) for tables, positing that a semantically robust latent space for table reasoning must be intrinsically Permutation Invariant (PI). To ground this hypothesis, we first conduct a retrospective analysis of table-reasoning tasks, highlighting the pervasive serialization bias that compromises structural integrity. We then propose a formal framework to diagnose this bias, introducing two principled metrics based on Centered Kernel Alignment (CKA): (i) PI, which measures embedding drift under complete structural derangement, and (ii) rho, a Spearman-based metric that tracks the convergence of latent structures toward a canonical form as structural information is incrementally restored. Our empirical analysis quantifies an expected flaw in modern Large Language Models (LLMs): even minor layout permutations induce significant, disproportionate semantic shifts in their table embeddings. This exposes a fundamental vulnerability in RAG systems, in which table retrieval becomes fragile to layout-dependent noise rather than to semantic content. In response, we present a novel, structure-aware TRL encoder architecture that explicitly enforces the cognitive principle of cell header alignment. This model demonstrates superior geometric stability and moves towards the PI ideal. Our work provides both a foundational critique of linearized table encoders and the theoretical scaffolding for semantically stable, permutation invariant retrieval, charting a new direction for table reasoning in information systems.

The transition from stateless model inference to stateful agentic execution is reshaping the systems assumptions underlying modern AI infrastructure. While large language models have made persistent, tool-using, and collaborative agents technically viable, existing runtime architectures remain constrained by materialization-heavy instantiation models that impose significant latency and memory overhead. This paper introduces Aethon, a reference-based replication primitive for near-constant-time instantiation of stateful AI agents. Rather than reconstructing agents as fully materialized objects, Aethon represents each instance as a compositional view over stable definitions, layered memory, and local contextual overlays. By shifting instantiation from duplication to reference, Aethon decouples creation cost from inherited structure. We present the conceptual framework, system architecture, and memory model underlying Aethon, including layered inheritance and copy-on-write semantics. We analyze its implications for complexity, scalability, multi-agent orchestration, and enterprise governance. We argue that reference-based instantiation is not merely an optimization, but a more appropriate systems abstraction for production-scale agentic software. Aethon points toward a new class of AI infrastructure in which agents become lightweight, composable execution identities that can be spawned, specialized, and governed at scale.

We investigate how self-referential inputs alter the internal matrix dynamics of large language models. Measuring 106 scalar metrics across up to 7 analysis passes on four models from three architecture families -- Qwen3-VL-8B, Llama-3.2-11B, Llama-3.3-70B, and Gemma-2-9B -- over 300 prompts in a 14-level hierarchy at three temperatures ($T \in \{0.0, 0.3, 0.7\}$), we find that self-reference alone is not destabilizing: grounded self-referential statements and meta-cognitive prompts are markedly more stable than paradoxical self-reference on key collapse-related metrics, and on several such metrics can be as stable as factual controls. Instability concentrates in prompts inducing non-closing truth recursion (NCTR) -- truth-value computations with no finite-depth resolution. NCTR prompts produce anomalously elevated attention effective rank -- indicating attention reorganization with global dispersion rather than simple concentration collapse -- and key metrics reach Cohen's $d = 3.14$ (attention effective rank) to $3.52$ (variance kurtosis) vs. stable self-reference in the 70B model; 281/397 metric-model combinations differentiate NCTR from stable self-reference after FDR correction ($q < 0.05$), 198 with $|d| > 0.8$. Per-layer SVD confirms disruption at every sampled layer ($d > +1.0$ in all three models analyzed), ruling out aggregation artifacts. A classifier achieves AUC $0.81$-$0.90$; 30 minimal pairs yield 42/387 significant combinations; 43/106 metrics replicate across all four models. We connect these observations to three classical matrix-semigroup problems and propose, as a conjecture, that NCTR forces finite-depth transformers toward dynamical regimes where these problems concentrate. NCTR prompts also produce elevated contradictory output ($+34$-$56$ percentage points vs. controls), suggesting practical relevance for understanding self-referential failure modes.

Large Language Models (LLMs) have significantly advanced tool-augmented agents, enabling autonomous reasoning via API interactions. However, executing multi-step tasks within massive tool libraries remains challenging due to two critical bottlenecks: (1) the absence of rigorous, plan-level evaluation frameworks and (2) the computational demand of exploring vast decision spaces stemming from large toolsets and long-horizon planning. To bridge these gaps, we first introduce SLATE (Synthetic Large-scale API Toolkit for E-commerce), a large-scale context-aware benchmark designed for the automated assessment of tool-integrated agents. Unlike static metrics, SLATE accommodates diverse yet functionally valid execution trajectories, revealing that current agents struggle with self-correction and search efficiency. Motivated by these findings, we next propose Entropy-Guided Branching (EGB), an uncertainty-aware search algorithm that dynamically expands decision branches where predictive entropy is high. EGB optimizes the exploration-exploitation trade-off, significantly enhancing both task success rates and computational efficiency. Extensive experiments on SLATE demonstrate that our dual contribution provides a robust foundation for developing reliable and scalable LLM agents in tool-rich environments.

Large vision-language models (VLMs) often rely on familiar semantic priors, but existing evaluations do not cleanly separate perception failures from rule-mapping failures. We study this behavior as semantic fixation: preserving a default interpretation even when the prompt specifies an alternative, equally valid mapping. To isolate this effect, we introduce VLM-Fix, a controlled benchmark over four abstract strategy games that evaluates identical terminal board states under paired standard and inverse rule formulations. Across 14 open and closed VLMs, accuracy consistently favors standard rules, revealing a robust semantic-fixation gap. Prompt interventions support this mechanism: neutral alias prompts substantially narrow the inverse-rule gap, while semantically loaded aliases reopen it. Post-training is strongly rule-aligned: training on one rule improves same-rule transfer but hurts opposite-rule transfer, while joint-rule training improves broader transfer. To test external validity beyond synthetic games, we evaluate analogous defamiliarization interventions on VLMBias and observe the same qualitative pattern. Finally, late-layer activation steering partially recovers degraded performance, indicating that semantic-fixation errors are at least partly editable in late representations. Project page, code, and dataset available at https://maveryn.github.io/vlm-fix/.

Large vision-language models (LVLMs) achieve strong multimodal performance, but still suffer from hallucinations caused by unstable visual grounding and over-reliance on language priors. Existing training-free decoding methods typically apply calibration at every decoding step, introducing unnecessary computation and potentially disrupting stable predictions. We address this problem by identifying layer-wise hesitation, a simple signal of grounding instability reflected by fluctuations in token preference across intermediate layers. Based on this observation, we propose Hesitation-Triggered Differential Calibration (HTDC), a training-free decoding framework that preserves standard full-branch inference and activates calibration only at hesitation-prone steps. When triggered, HTDC contrasts the full branch with two lightweight probes, a visual-nullification probe and a semantic-nullification probe, to suppress hallucination-prone candidates while avoiding unnecessary intervention on stable steps. Experiments on representative hallucination benchmarks show that HTDC consistently reduces hallucinations while maintaining strong task accuracy, achieving a favorable trade-off between effectiveness and computational overhead.

Visual Foundation Models (VFMs) such as the Segment Anything Model (SAM) have significantly advanced broad use of image segmentation. However, SAM and its variants necessitate substantial manual effort for prompt generation and additional training for specific applications. Recent approaches address these limitations by integrating SAM into in-context (one/few shot) segmentation, enabling auto-prompting through semantic alignment between query and support images. Despite these efforts, they still generate sub-optimal prompts that degrade segmentation quality due to visual inconsistencies between support and query images. To tackle this limitation, we introduce PR-MaGIC (Prompt Refinement via Mask Decoder Gradient Flow for In-Context Segmentation), a training-free test-time framework that refines prompts via gradient flow derived from SAM's mask decoder. PR-MaGIC seamlessly integrates into in-context segmentation frameworks, being theoretically grounded yet practically stabilized through a simple top-1 selection strategy that ensures robust performance across samples. Extensive evaluations demonstrate that PR-MaGIC consistently improves segmentation quality across various benchmarks, effectively mitigating inadequate prompts without requiring additional training or architectural modifications.

Whole-slide image (WSI) classification in computational pathology is commonly formulated as slide-level Multiple Instance Learning (MIL) with a single global bag representation. However, slide-level MIL is fundamentally underconstrained: optimizing only global labels encourages models to aggregate features without learning anatomically meaningful localization. This creates a mismatch between the scale of supervision and the scale of clinical reasoning. Clinicians assess tumor burden, focal lesions, and architectural patterns within millimeter-scale regions, whereas standard MIL is trained only to predict whether "somewhere in the slide there is cancer." As a result, the model's inductive bias effectively erases anatomical structure. We propose Progressive-Context MIL (PC-MIL), a framework that treats the spatial extent of supervision as a first-class design dimension. Rather than altering magnification, patch size, or introducing pixel-level segmentation, we decouple feature resolution from supervision scale. Using fixed 20x features, we vary MIL bag extent in millimeter units and anchor supervision at a clinically motivated 2mm scale to preserve comparable tumor burden and avoid confounding scale with lesion density. PC-MIL progressively mixes slide- and region-level supervision in controlled proportions, enabling explicit train-context x test-context analysis. On 1,476 prostate WSIs from five public datasets for binary cancer detection, we show that anatomical context is an independent axis of generalization in MIL, orthogonal to feature resolution: modest regional supervision improves cross-context performance, and balanced multi-context training stabilizes accuracy across slide and regional evaluation without sacrificing global performance. These results demonstrate that supervision extent shapes MIL inductive bias and support anatomically grounded WSI generalization.

Large language models (LLMs) are increasingly used in daily applications, from content generation to code writing, where each interaction treats the model as stateless, generating responses independently without memory. Yet human writing is inherently longitudinal: authors' styles and cognitive states evolve across months and years. This raises a central question: can LLMs reproduce such temporal structure across extended time periods? We construct and publicly release a longitudinal dataset of 412 human authors and 6,086 documents spanning 2012--2024 across three domains (academic abstracts, blogs, news) and compare them to trajectories generated by three representative LLMs under standard and history-conditioned generation settings. Using drift and variance-based metrics over semantic, lexical, and cognitive-emotional representations, we find temporal flattening in LLM-generated text. LLMs produce greater lexical diversity but exhibit substantially reduced semantic and cognitive-emotional drift relative to humans. These differences are highly predictive: temporal variability patterns alone achieve 94% accuracy and 98% ROC-AUC in distinguishing human from LLM trajectories. Our results demonstrate that temporal flattening persists regardless of whether LLMs generate independently or with access to incremental history, revealing a fundamental property of current deployment paradigms. This gap has direct implications for applications requiring authentic temporal structure, such as synthetic training data and longitudinal text modeling.

On online advertising platforms, newly introduced promotional ads face the cold-start problem, as they lack sufficient user feedback for model training. In this work, we propose LLM-HYPER, a novel framework that treats large language models (LLMs) as hypernetworks to directly generate the parameters of the click-through rate (CTR) estimator in a training-free manner. LLM-HYPER uses few-shot Chain-of-Thought prompting over multimodal ad content (text and images) to infer feature-wise model weights for a linear CTR predictor. By retrieving semantically similar past campaigns via CLIP embeddings and formatting them into prompt-based demonstrations, the LLM learns to reason about customer intent, feature influence, and content relevance. To ensure numerical stability and serviceability, we introduce normalization and calibration techniques that align the generated weights with production-ready CTR distributions. Extensive offline experiments show that LLM-HYPER significantly outperforms cold-start baselines in NDCG$@10$ by 55.9\%. Our real-world online A/B test on one of the top e-commerce platforms in the U.S. demonstrates the strong performance of LLM-HYPER, which drastically reduces the cold-start period and achieves competitive performance. LLM-HYPER has been successfully deployed in production.

Behavior Trees (BTs) provide designers an intuitive graphical interface to construct long-horizon plans for autonomous systems. To ensure their correctness and safety, rigorous formal models and verification techniques are essential. Temporal BTs (TBTs) offer a promising approach by leveraging existing temporal logic formalisms to specify and verify the executions of BTs. However, this analysis is currently limited to offline post hoc analysis and trace repair. In this paper, we reformulate TBTs using a ternary-valued Signal Temporal Logic (STL) amenable for control synthesis. Ternary logic introduces a third truth value \textit{Unknown}, formally capturing cases where a trajectory has neither fully satisfied or dissatisfied a specification. We propose mixed-integer linear encodings for partial trajectory STL and TBTs over ternary logic allowing for correct-by-construction control strategies for linear dynamical systems via mixed-integer optimization. We demonstrate the utility of our framework by solving optimal control problems.

Designing robust reinforcement learning (RL) agents in the presence of imperfect reward signals remains a core challenge. In practice, agents are often trained with proxy rewards that only approximate the true objective, leaving them vulnerable to reward hacking, where high proxy returns arise from unintended or exploitative behaviors. Recent work formalizes this issue using r-correlation between proxy and true rewards, but existing methods like occupancy-regularized policy optimization (ORPO) optimize against a fixed proxy and do not provide strong guarantees against broader classes of correlated proxies. In this work, we formulate reward hacking as a robust policy optimization problem over the space of all r-correlated proxy rewards. We derive a tractable max-min formulation, where the agent maximizes performance under the worst-case proxy consistent with the correlation constraint. We further show that when the reward is a linear function of known features, our approach can be adapted to incorporate this prior knowledge, yielding both improved policies and interpretable worst-case rewards. Experiments across several environments show that our algorithms consistently outperform ORPO in worst-case returns, and offer improved robustness and stability across different levels of proxy-true reward correlation. These results show that our approach provides both robustness and transparency in settings where reward design is inherently uncertain. The code is available at https://github.com/ZixuanLiu4869/reward_hacking.

Spatial transcriptomics (ST) measures mRNA expression while preserving spatial organization, but multi-slice analysis faces two coupled difficulties: large non-rigid deformations across slices and inter-slice batch effects when alignment and integration are treated independently. We present INST-Align, an unsupervised pairwise framework that couples a coordinate-based deformation network with a shared Canonical Expression Field, an implicit neural representation mapping spatial coordinates to expression embeddings, for joint alignment and reconstruction. A two-phase training strategy first establishes a stable canonical embedding space and then jointly optimizes deformation and spatial-feature matching, enabling mutually constrained alignment and representation learning. Cross-slice parameter sharing of the canonical field regularizes ambiguous correspondences and absorbs batch variation. Across nine datasets, INST-Align achieves state-of-the-art mean OT Accuracy (0.702), NN Accuracy (0.719), and Chamfer distance, with Chamfer reductions of up to 94.9\% on large-deformation sections relative to the strongest baseline. The framework also yields biologically meaningful spatial embeddings and coherent 3D tissue reconstruction. The code will be released after review phase.

Memory is fundamental to social interaction, enabling humans to recall meaningful past experiences and adapt their behavior accordingly based on the context. However, most current social robots and embodied agents rely on non-selective, text-based memory, limiting their ability to support personalized, context-aware interactions. Drawing inspiration from cognitive neuroscience, we propose a context-selective, multimodal memory architecture for social robots that captures and retrieves both textual and visual episodic traces, prioritizing moments characterized by high emotional salience or scene novelty. By associating these memories with individual users, our system enables socially personalized recall and more natural, grounded dialogue. We evaluate the selective storage mechanism using a curated dataset of social scenarios, achieving a Spearman correlation of 0.506, surpassing human consistency ($ρ=0.415$) and outperforming existing image memorability models. In multimodal retrieval experiments, our fusion approach improves Recall@1 by up to 13\% over unimodal text or image retrieval. Runtime evaluations confirm that the system maintains real-time performance. Qualitative analyses further demonstrate that the proposed framework produces richer and more socially relevant responses than baseline models. This work advances memory design for social robots by bridging human-inspired selectivity and multimodal retrieval to enhance long-term, personalized human-robot interaction.

The Identifiable Victim Effect (IVE) $-$ the tendency to allocate greater resources to a specific, narratively described victim than to a statistically characterized group facing equivalent hardship $-$ is one of the most robust findings in moral psychology and behavioural economics. As large language models (LLMs) assume consequential roles in humanitarian triage, automated grant evaluation, and content moderation, a critical question arises: do these systems inherit the affective irrationalities present in human moral reasoning? We present the first systematic, large-scale empirical investigation of the IVE in LLMs, comprising N=51,955 validated API trials across 16 frontier models spanning nine organizational lineages (Google, Anthropic, OpenAI, Meta, DeepSeek, xAI, Alibaba, IBM, and Moonshot). Using a suite of ten experiments $-$ porting and extending canonical paradigms from Small et al. (2007) and Kogut and Ritov (2005) $-$ we find that the IVE is prevalent but strongly modulated by alignment training. Instruction-tuned models exhibit extreme IVE (Cohen's d up to 1.56), while reasoning-specialized models invert the effect (down to d=-0.85). The pooled effect (d=0.223, p=2e-6) is approximately twice the single-victim human meta-analytic baseline (d$\approx$0.10) reported by Lee and Feeley (2016) $-$ and likely exceeds the overall human pooled effect by a larger margin, given that the group-victim human effect is near zero. Standard Chain-of-Thought (CoT) prompting $-$ contrary to its role as a deliberative corrective $-$ nearly triples the IVE effect size (from d=0.15 to d=0.41), while only utilitarian CoT reliably eliminates it. We further document psychophysical numbing, perfect quantity neglect, and marginal in-group/out-group cultural bias, with implications for AI deployment in humanitarian and ethical decision-making contexts.

The tumor microenvironment (TME) plays a central role in cancer progression, treatment response, and patient outcomes, yet large-scale, consistent, and quantitative TME characterization from routine hematoxylin and eosin (H&E)-stained histopathology remains scarce. We introduce OpenTME, an open-access dataset of pre-computed TME profiles derived from 3,634 H&E-stained whole-slide images across five cancer types (bladder, breast, colorectal, liver, and lung cancer) from The Cancer Genome Atlas (TCGA). All outputs were generated using Atlas H&E-TME, an AI-powered application built on the Atlas family of pathology foundation models, which performs tissue quality control, tissue segmentation, cell detection and classification, and spatial neighborhood analysis, yielding over 4,500 quantitative readouts per slide at cell-level resolution. OpenTME is available for non-commercial academic research on Hugging Face. We will continue to expand OpenTME over time and anticipate it will serve as a resource for biomarker discovery, spatial biology research, and the development of computational methods for TME analysis.

Visual localization is the task of estimating the camera pose of an image relative to a scene representation. In practice, visual localization systems are often cloud-based. Naturally, this raises privacy concerns in terms of revealing private details through the images sent to the server or through the representations stored on the server. Privacy-preserving localization aims to avoid such leakage of private details. However, the resulting localization approaches are significantly more complex, slower, and less accurate than their non-privacy-preserving counterparts. In this paper, we consider structureless localization methods in the context of privacy preservation. Structureless methods represent the scene through a set of reference images with known camera poses and intrinsics. In contrast to existing methods proposing representations that are as privacy-preserving as possible, we study a simple image obfuscation approach based on common image operations, e.g., replacing RGB images with (semantic) segmentations. We show that existing structureless pipelines do not need any special adjustments, as modern feature matchers can match obfuscated images out of the box. The results are easy-to-implement pipelines that can ensure both the privacy of the query images and the scene representations. Detailed experiments on multiple datasets show that the resulting methods achieve state-of-the-art pose accuracy for privacy-preserving approaches.

Large language models can increasingly adapt educational tasks to learners characteristics. In the present study, we examine a multi-agent teacher-in-the-loop system for personalizing middle school math problems. The teacher enters a base problem and desired topic, the LLM generates the problem, and then four AI agents evaluate the problem using criteria that each specializes in (mathematical accuracy, authenticity, readability, and realism). Eight middle school mathematics teachers created 212 problems in ASSISTments using the system and assigned these problems to their students. We find that both teachers and students wanted to modify the fine-grained personalized elements of the real-world context of the problems, signaling issues with authenticity and fit. Although the agents detected many issues with realism as the problems were being written, there were few realism issues noted by teachers and students in the final versions. Issues with readability and mathematical hallucinations were also somewhat rare. Implications for multi-agent systems for personalization that support teacher control are given.

The analysis of DNA sequences has become critical in numerous fields, from evolutionary biology to understanding gene regulation and disease mechanisms. While deep neural networks can achieve remarkable predictive performance, they typically operate as black boxes. Contrasting these black boxes, axis-aligned decision trees offer a promising direction for interpretable DNA sequence analysis, yet they suffer from a fundamental limitation: considering individual raw features in isolation at each split limits their expressivity, which results in prohibitive tree depths that hinder both interpretability and generalization performance. We address this challenge by introducing DEFT, a novel framework that adaptively generates high-level sequence features during tree construction. DEFT leverages large language models to propose biologically-informed features tailored to the local sequence distributions at each node and to iteratively refine them with a reflection mechanism. Empirically, we demonstrate that DEFT discovers human-interpretable and highly predictive sequence features across a diverse range of genomic tasks.

Block-wise diffusion language models (DLMs) generate multiple tokens in any order, offering a promising alternative to the autoregressive decoding pipeline. However, they still remain bottlenecked by memory-bound attention in long-context scenarios. Naive sparse attention fails on DLMs due to a KV Inflation problem, where different queries select different prefix positions, making the union of accessed KV pages large. To address this, we observe that between consecutive denoising steps, only a small fraction of active tokens exhibit significant hidden-state changes, while the majority of stable tokens remain nearly constant. Based on this insight, we propose LOSA (Locality-aware Sparse Attention), which reuses cached prefix-attention results for stable tokens and applies sparse attention only to active tokens. This substantially shrinks the number of KV indices that must be loaded, yielding both higher speedup and higher accuracy. Across multiple block-wise DLMs and benchmarks, LOSA preserves near-dense accuracy while significantly improving efficiency, achieving up to +9 points in average accuracy at aggressive sparsity levels while maintaining 1.54x lower attention density. It also achieves up to 4.14x attention speedup on RTX A6000 GPUs, demonstrating the effectiveness of the proposed method.

The use of Large Language Models (LLMs) for reliable, enterprise-grade analytics such as text categorization is often hindered by the stochastic nature of attention mechanisms and sensitivity to noise that compromise their analytical precision and reproducibility. To address these technical frictions, this paper introduces the Weighted Syntactic and Semantic Context Assessment Summary (wSSAS), a deterministic framework designed to enforce data integrity on large-scale, chaotic datasets. We propose a two-phased validation framework that first organizes raw text into a hierarchical classification structure containing Themes, Stories, and Clusters. It then leverages a Signal-to-Noise Ratio (SNR) to prioritize high-value semantic features, ensuring the model's attention remains focused on the most representative data points. By incorporating this scoring mechanism into a Summary-of-Summaries (SoS) architecture, the framework effectively isolates essential information and mitigates background noise during data aggregation. Experimental results using Gemini 2.0 Flash Lite across diverse datasets - including Google Business reviews, Amazon Product reviews, and Goodreads Book reviews - demonstrate that wSSAS significantly improves clustering integrity and categorization accuracy. Our findings indicate that wSSAS reduces categorization entropy and provides a reproducible pathway for improving LLM based summaries based on a high-precision, deterministic process for large-scale text categorization.

PDF files are primarily intended for human reading rather than automated processing. In addition, the heterogeneous content of PDFs, such as text, tables, and images, poses significant challenges for parsing and information extraction. To address these difficulties, both practitioners and researchers are increasingly developing new methods, including the promising Retrieval-Augmented Generation (RAG) systems to automated PDF processing. However, there is no comprehensive study investigating how different components and design choices affect the performance of a RAG system for understanding PDFs. In this paper, we propose such a study (1) by focusing on Question Answering, a specific language understanding task, and (2) by leveraging two benchmarks from the financial domain, including TableQuest, our newly generated, publicly available benchmark. We systematically examine multiple PDF parsers and chunking strategies (with varied overlap), along with their potential synergies in preserving document structure and ensuring answer correctness. Overall, our results offer practical guidelines for building robust RAG pipelines for PDF understanding.

Large language models (LLMs) often hallucinate in long-form generation. Existing approaches mainly improve factuality through post-hoc revision or reinforcement learning (RL) with correctness-based rewards, but they do not teach the model to estimate which parts of its generation are reliable. As a result, models may still state incorrect claims confidently in their responses. Recent advances in reasoning have significantly improved LLM performance, and have been leveraged to estimate confidence by incorporating calibration into RL objectives. However, existing approaches remain limited to a single scalar confidence for the entire response, which is insufficient for long-form generation where uncertainty varies across individual claims. To mitigate this problem, we propose CURE, a framework that improves long-form factuality by teaching LLMs to reason about uncertainty at the claim level. We first introduce a Claim-Aware Reasoning Protocol, which structures outputs into atomic claims paired with explicit confidence estimates. We then develop a multi-stage training pipeline that aligns model confidence with claims' correctness and then optimizes on factuality. The resulting calibrated confidence further enables selective prediction, allowing the model to abstain from uncertain claims at inference time. Experiments on four long-form factuality benchmarks show that CURE consistently improves factual accuracy over competitive supervised and RL baselines, while maintaining factual recall. In particular, it improves claim-level accuracy by up to 39.9% on Biography generation. These gains are accompanied by improved calibration, as reflected by a 16.0% increase in AUROC on FactBench.

Deep learning models may converge to suboptimal solutions despite strong validation accuracy, masking an optimization failure we term Trajectory Deviation. This is because as training proceeds, models can abandon high generalization states for specific data sub-populations, thus discarding previously learned latent features without triggering classical overfitting signals. To address this problem we introduce VISTA, an online self-distillation framework that enforces consistency along the optimization trajectory. Using a validation-informed Marginal Coverage score, VISTA identifies expert anchors, which are earlier model states that retain specialized competence over distinct data regions. A coverage-weighted ensemble of these anchors is integrated online during training, regularizing the loss landscape and preserving mastered knowledge. When evaluated across multiple benchmarks, VISTA demonstrates improved robustness and generalization over standard training and prior self-distillation methods, while a lightweight implementation reduces storage overhead by 90% without performance loss.

Visual token pruning is a widely used strategy for efficient inference in multimodal large language models (MLLMs), but existing work mainly evaluates it with task accuracy. In this paper, we study how visual token pruning affects model calibration, that is, whether predicted confidence matches actual correctness. Using LLaVA-1.5-7B on POPE and ScienceQA-IMG, we evaluate Expected Calibration Error (ECE), Brier score, and AURC under several pruning strategies, including SCOPE with different saliency weights, saliency-only pruning, FastV, and random pruning, across multiple token budgets. Our results show that pruning does not simply trade reliability for efficiency. On POPE, a pure-coverage setting in SCOPE achieves substantially lower ECE than the full unpruned model while maintaining similar accuracy. An internal alpha-sweep further shows a consistent trend: reducing the saliency weight improves calibration at all tested token budgets, while accuracy changes only slightly. In contrast, saliency-based pruning leads to worse calibration, and real FastV causes severe performance degradation in our setting. On ScienceQA-IMG, pruning also reduces ECE, with accuracy remaining stable or slightly improving. We additionally study the gap power exponent in coverage-based selection and find that its default setting is not always optimal. Overall, our results suggest that visual token pruning should be evaluated not only by accuracy, but also by confidence quality, especially for multimodal systems that need reliable decisions.

Retrieval-Augmented Generation remains the dominant pattern for giving LLMs persistent memory, but a visible cluster of personal wiki-style memory architectures emerged in April 2026 -- design proposals from Karpathy, MemPalace, and LLM Wiki v2 that compile knowledge into an interlinked artifact for long-term use by a single user. They sit alongside production memory systems that the major labs have shipped for over a year, and an active academic lineage including MemGPT, Generative Agents, Mem0, Zep, A-Mem, MemMachine, SleepGate, and Second Me. Within a 2026 landscape of emerging governance frameworks for agent context and memory -- including Context Cartography and MemOS -- this paper proposes a companion-specific governance profile: a set of normative obligations, a time-structured procedural rule, and testable conformance invariants for the specific failure mode of entrenchment under user-coupled drift in single-user knowledge wikis built on the LLM wiki pattern. The design principle is that personal LLM memory is a companion system: its job is to mirror the user on operational dimensions (working vocabulary, load-bearing structure, continuity of context) and compensate on epistemic failure modes (entrenchment, suppression of contradicting evidence, Kuhnian ossification). Five operations implement this split -- TRIAGE, DECAY, CONTEXTUALIZE, CONSOLIDATE, AUDIT -- supported by memory gravity and minority-hypothesis retention. The sharpest prediction: accumulated contradictory evidence should have a structural path to updating a centrality-protected dominant interpretation through multi-cycle buffer pressure accumulation, a failure mode no existing benchmark captures. The safety story at the single-agent level is partial, and the paper is explicit about what it does and does not solve.

Large Vision-Language Models (LVLMs) increasingly rely on retrieval to answer knowledge-intensive multimodal questions. Existing benchmarks overlook conflicts between visual and textual evidence and the importance of generating deflections (e.g., Sorry, I cannot answer...) when retrieved knowledge is incomplete. These benchmarks also suffer from rapid obsolescence, as growing LVLM training sets allow models to answer many questions without retrieval. We address these gaps with three contributions. First, we propose a dynamic data curation pipeline that preserves benchmark difficulty over time by filtering for genuinely retrieval-dependent samples. Second, we introduce VLM-DeflectionBench, a benchmark of 2,775 samples spanning diverse multimodal retrieval settings, designed to probe model behaviour under conflicting or insufficient evidence. Third, we define a fine-grained evaluation protocol with four scenarios that disentangle parametric memorization from retrieval robustness. Experiments across 20 state-of-the-art LVLMs indicate that models usually fail to deflect in the presence of noisy or misleading evidence. Our results highlight the need to evaluate not only what models know, but how they behave when they do not, and serve as a reusable and extensible benchmark for reliable KB-VQA evaluation. All resources will be publicly available upon publication.

Worm-inspired robots provide an effective locomotion strategy for constrained environments by combining cyclic body deformation with alternating anchoring. For compliant robots, however, the interaction between deformable anchoring structures and the environment makes predictive modeling and deployable gait optimization challenging. This paper presents an experimentally grounded modeling and optimization framework for a compliant worm robot capable of traversing corrugated pipes. First, a hybrid dynamic locomotion model is derived, in which the robot motion is represented by continuous dynamics within a corrugation groove and discrete switching of anchoring positions between adjacent grooves. A slack-aware actuation model is further introduced to map the commanded gait input to the realized body-length change, and an energy model is developed based on physics and calibrated with empirical power measurement. Based on these models, a multi-objective gait optimization problem is formulated to maximize average speed while minimizing average power. To reduce the fragility of nominal boundary-seeking solutions, a kinematic robustness margin is introduced into the anchoring-transition conditions, leading to a margin-based robust gait optimization framework. Experimental results show that the proposed framework captures the dominant locomotion and energy-consumption behavior of the robot over the tested conditions, and enables robust gait optimization for achieving speed-power trade-off.

The proliferation of sophisticated generative models has significantly advanced the realism of synthetic facial content, known as deepfakes, raising serious concerns about digital trust. Although modern deep learning-based detectors perform well, many rely on spatial-domain features that degrade under compression. This limitation has prompted a shift toward integrating frequency-domain representations with deep learning to improve robustness. Prior research has explored frequency transforms such as Discrete Cosine Transform (DCT), Fast Fourier Transform (FFT), and Wavelet Transform, among others. However, to the best of our knowledge, the Curvelet Transform, despite its superior directional and multiscale properties, remains entirely unexplored in the context of deepfake detection. In this work, we introduce a novel Curvelet-based detection approach that enhances feature quality through wedge-level attention and scale-aware spatial masking, both trained to selectively emphasize discriminative frequency components. The refined frequency cues are reconstructed and passed to a modified pretrained Xception network for classification. Evaluated on two compression qualities in the challenging FaceForensics++ dataset, our method achieves 98.48% accuracy and 99.96% AUC on FF++ low compression, while maintaining strong performance under high compression, demonstrating the efficacy and interpretability of Curvelet-informed forgery detection.