As one of the simplest non-prehensile manipulation skills, pushing has been widely studied as an effective means to rearrange objects. Existing approaches, however, typically rely on multi-step push plans composed of pre-defined pushing primitives with limited application scopes, which restrict their efficiency and versatility across different scenarios. In this work, we propose a unified pushing policy that incorporates a lightweight prompting mechanism into a flow matching policy to guide the generation of reactive, multimodal pushing actions. The visual prompt can be specified by a high-level planner, enabling the reuse of the pushing policy across a wide range of planning problems. Experimental results demonstrate that the proposed unified pushing policy not only outperforms existing baselines but also effectively serves as a low-level primitive within a VLM-guided planning framework to solve table-cleaning tasks efficiently.

In recent years, Multi-modal Large Language Models (MLLMs) have achieved strong performance in OCR-centric Visual Question Answering (VQA) tasks, illustrating their capability to process heterogeneous data and exhibit adaptability across varied contexts. However, these MLLMs rely on a Large Language Model (LLM) as the decoder, which is primarily designed for linguistic processing, and thus inherently lacks the positional reasoning required for precise visual tasks, such as text spotting and text grounding. Additionally, the extensive parameters of MLLMs necessitate substantial computational resources and large-scale data for effective training. Conversely, text spotting specialists achieve state-of-the-art coordinate predictions but lack semantic reasoning capabilities. This dichotomy motivates our key research question: Can we synergize the efficiency of specialists with the contextual power of LLMs to create a positionally-accurate MLLM? To overcome these challenges, we introduce PositionOCR, a parameter-efficient hybrid architecture that seamlessly integrates a text spotting model's positional strengths with an LLM's contextual reasoning. Comprising 131M trainable parameters, this framework demonstrates outstanding multi-modal processing capabilities, particularly excelling in tasks such as text grounding and text spotting, consistently surpassing traditional MLLMs.

We present a method for generating training data for reinforcement learning with verifiable rewards to improve small open-weights language models on mathematical tasks. Existing data generation approaches rely on open-loop pipelines and fixed modifications that do not adapt to the model's capabilities. Furthermore, they typically operate directly on word problems, limiting control over problem structure. To address this, we perform modifications in a symbolic problem space, representing each problem as a set of symbolic variables and constraints (e.g., via algebraic frameworks such as SymPy or SMT formulations). This representation enables precise control over problem structure, automatic generation of ground-truth solutions, and decouples mathematical reasoning from linguistic realization. We also show that this results in more diverse generations. To adapt the problem difficulty to the model, we introduce a closed-loop framework that learns modification strategies through prompt optimization in symbolic space. Experimental results demonstrate that both adaptive problem generation and symbolic representation modifications contribute to improving the model's math solving ability.

Learning from Demonstration (LfD) offers a promising paradigm for robot skill acquisition. Recent approaches attempt to extract manipulation commands directly from video demonstrations, yet face two critical challenges: (1) general video captioning models prioritize global scene features over task-relevant objects, producing descriptions unsuitable for precise robotic execution, and (2) end-to-end architectures coupling visual understanding with policy learning require extensive paired datasets and struggle to generalize across objects and scenarios. To address these limitations, we propose a novel ``Human-to-Robot'' imitation learning pipeline that enables robots to acquire manipulation skills directly from unstructured video demonstrations, inspired by the human ability to learn by watching and imitating. Our key innovation is a modular framework that decouples the learning process into two distinct stages: (1) Video Understanding, which combines Temporal Shift Modules (TSM) with Vision-Language Models (VLMs) to extract actions and identify interacted objects, and (2) Robot Imitation, which employs TD3-based deep reinforcement learning to execute the demonstrated manipulations. We validated our approach in PyBullet simulation environments with a UR5e manipulator and in a real-world experiment with a UF850 manipulator across four fundamental actions: reach, pick, move, and put. For video understanding, our method achieves 89.97% action classification accuracy and BLEU-4 scores of 0.351 on standard objects and 0.265 on novel objects, representing improvements of 76.4% and 128.4% over the best baseline, respectively. For robot manipulation, our framework achieves an average success rate of 87.5% across all actions, with 100% success on reaching tasks and up to 90% on complex pick-and-place operations. The project website is available at https://thanhnguyencanh.github.io/LfD4hri.

Human mesh recovery (HMR) from a single RGB image is inherently ambiguous, as multiple 3D poses can correspond to the same 2D observation. Recent diffusion-based methods tackle this by generating various hypotheses, but often sacrifice accuracy. They yield predictions that are either physically implausible or drift from the input image, especially under occlusion or in cluttered, in-the-wild scenes. To address this, we introduce a dual-memory augmented HMR critique agent with self-reflection to produce context-aware quality scores for predicted meshes. These scores distill fine-grained cues about 3D human motion structure, physical feasibility, and alignment with the input image. We use these scores to build a group-wise HMR preference dataset. Leveraging this dataset, we propose a group preference alignment framework for finetuning diffusion-based HMR models. This process injects the rich preference signals into the model, guiding it to generate more physically plausible and image-consistent human meshes. Extensive experiments demonstrate that our method achieves superior performance compared to state-of-the-art approaches.

Deep learning models for medical image analysis often act as black boxes, seldom aligning with clinical guidelines or explicitly linking decisions to supporting evidence. This is especially critical in Alzheimer's disease (AD), where predictions should be grounded in both anatomical and clinical findings. We present EMAD, a vision-language framework that generates structured AD diagnostic reports in which each claim is explicitly grounded in multimodal evidence. EMAD uses a hierarchical Sentence-Evidence-Anatomy (SEA) grounding mechanism: (i) sentence-to-evidence grounding links generated sentences to clinical evidence phrases, and (ii) evidence-to-anatomy grounding localizes corresponding structures on 3D brain MRI. To reduce dense annotation requirements, we propose GTX-Distill, which transfers grounding behavior from a teacher trained with limited supervision to a student operating on model-generated reports. We further introduce Executable-Rule GRPO, a reinforcement fine-tuning scheme with verifiable rewards that enforces clinical consistency, protocol adherence, and reasoning-diagnosis coherence. On the AD-MultiSense dataset, EMAD achieves state-of-the-art diagnostic accuracy and produces more transparent, anatomically faithful reports than existing methods. We will release code and grounding annotations to support future research in trustworthy medical vision-language models.

The increasing use of Large Language Models (LLMs) as proxies for human participants in social science research presents a promising, yet methodologically risky, paradigm shift. While LLMs offer scalability and cost-efficiency, their "naive" application, where they are prompted to generate content without explicit behavioral constraints, introduces significant linguistic discrepancies that challenge the validity of research findings. This paper addresses these limitations by introducing a novel, history-conditioned reply prediction task on authentic X (formerly Twitter) data, to create a dataset designed to evaluate the linguistic output of LLMs against human-generated content. We analyze these discrepancies using stylistic and content-based metrics, providing a quantitative framework for researchers to assess the quality and authenticity of synthetic data. Our findings highlight the need for more sophisticated prompting techniques and specialized datasets to ensure that LLM-generated content accurately reflects the complex linguistic patterns of human communication, thereby improving the validity of computational social science studies.

Reliable localization is a fundamental requirement for multi-robot systems operating in GPS-denied environments. Visual-inertial odometry (VIO) provides lightweight and accurate motion estimation but suffers from cumulative drift in the absence of global references. Ultra-wideband (UWB) ranging offers complementary global observations, yet most existing UWB-aided VIO methods are designed for single-robot scenarios and rely on pre-calibrated anchors, which limits their robustness in practice. This paper proposes a distributed collaborative visual-inertial-ranging odometry (DC-VIRO) framework that tightly fuses VIO and UWB measurements across multiple robots. Anchor positions are explicitly included in the system state to address calibration uncertainty, while shared anchor observations are exploited through inter-robot communication to provide additional geometric constraints. By leveraging a right-invariant error formulation on Lie groups, the proposed approach preserves the observability properties of standard VIO, ensuring estimator consistency. Simulation results with multiple robots demonstrate that DC-VIRO significantly improves localization accuracy and robustness, while simultaneously enabling anchor self-calibration in distributed settings.

Action Quality Assessment (AQA) aims to score how well an action is performed and is widely used in sports analysis, rehabilitation assessment, and human skill evaluation. Multi-modal AQA has recently achieved strong progress by leveraging complementary visual and kinematic cues, yet real-world deployments often suffer from non-stationary modality imbalance, where certain modalities become missing or intermittently available due to sensor failures or annotation gaps. Existing continual AQA methods overlook this issue and assume that all modalities remain complete and stable throughout training, which restricts their practicality. To address this challenge, we introduce Bridged Modality Adaptation (BriMA), an innovative approach to multi-modal continual AQA under modality-missing conditions. BriMA consists of a memory-guided bridging imputation module that reconstructs missing modalities using both task-agnostic and task-specific representations, and a modality-aware replay mechanism that prioritizes informative samples based on modality distortion and distribution drift. Experiments on three representative multi-modal AQA datasets (RG, Fis-V, and FS1000) show that BriMA consistently improves performance under different modality-missing conditions, achieving 6--8\% higher correlation and 12--15\% lower error on average. These results demonstrate a step toward robust multi-modal AQA systems under real-world deployment constraints.

I introduce Virtual Parameter Sharpening (VPS), an inference-time technique that augments frozen transformer linear layers with dynamic, activation-conditioned low-rank perturbations. Unlike parameter-efficient fine-tuning methods such as LoRA, which learn static low-rank adapters, VPS constructs its perturbation factors on the fly from batch activation statistics and optional gradient signals, enabling test-time adaptation without persistent parameter updates. The perturbation takes the form Delta W = gamma * W^T V U^T W, where selector matrices U and V are constructed via sparse activation-guided selection or Sylvester-coupled regression. We provide a theoretical analysis of the perturbation's spectral properties and describe an adaptive policy system that modulates perturbation magnitude based on activation energy and token-level entropy. This system incorporates multi-objective verification with iterative refinement for tasks with ground-truth supervision. We present the complete algorithmic framework, analyze its mathematical foundations, and discuss the mechanisms by which activation-conditioned computation may enhance reasoning capabilities in large language models. Implementation and experimental code are available at https://github.com/Saba-Kublashvili/vps-virtual-parameter-synthesis .

AIGC has rapidly expanded from text-to-image generation toward high-quality multimodal synthesis across video and audio. Within this context, joint audio-video generation (JAVG) has emerged as a fundamental task that produces synchronized and semantically aligned sound and vision from textual descriptions. However, compared with advanced commercial models such as Veo3, existing open-source methods still suffer from limitations in generation quality, temporal synchrony, and alignment with human preferences. To bridge the gap, this paper presents JavisDiT++, a concise yet powerful framework for unified modeling and optimization of JAVG. First, we introduce a modality-specific mixture-of-experts (MS-MoE) design that enables cross-modal interaction efficacy while enhancing single-modal generation quality. Then, we propose a temporal-aligned RoPE (TA-RoPE) strategy to achieve explicit, frame-level synchronization between audio and video tokens. Besides, we develop an audio-video direct preference optimization (AV-DPO) method to align model outputs with human preference across quality, consistency, and synchrony dimensions. Built upon Wan2.1-1.3B-T2V, our model achieves state-of-the-art performance merely with around 1M public training entries, significantly outperforming prior approaches in both qualitative and quantitative evaluations. Comprehensive ablation studies have been conducted to validate the effectiveness of our proposed modules. All the code, model, and dataset are released at https://JavisVerse.github.io/JavisDiT2-page.

Latent diffusion models have enabled high-quality video synthesis, yet their inference remains costly and time-consuming. As diffusion transformers become increasingly efficient, the latency bottleneck inevitably shifts to VAE decoders. To reduce their latency while maintaining quality, we propose a universal acceleration framework for VAE decoders that preserves full alignment with the original latent distribution. Specifically, we propose (1) an independence-aware channel pruning method to effectively mitigate severe channel redundancy, and (2) a stage-wise dominant operator optimization strategy to address the high inference cost of the widely used causal 3D convolutions in VAE decoders. Based on these innovations, we construct a Flash-VAED family. Moreover, we design a three-phase dynamic distillation framework that efficiently transfers the capabilities of the original VAE decoder to Flash-VAED. Extensive experiments on Wan and LTX-Video VAE decoders demonstrate that our method outperforms baselines in both quality and speed, achieving approximately a 6$\times$ speedup while maintaining the reconstruction performance up to 96.9%. Notably, Flash-VAED accelerates the end-to-end generation pipeline by up to 36% with negligible quality drops on VBench-2.0.

This paper examines the reasoning capabilities of Large Language Models (LLMs) from a novel perspective, focusing on their ability to operate within formally specified, rule-governed environments. We evaluate four LLMs (Gemini 2.5 Pro and Flash variants, Llama 3.3 70B and GPT-OSS 120B) on a suite of forward-simulation tasks-including next / multistep state formulation, and legal action generation-across a diverse set of reasoning problems illustrated through General Game Playing (GGP) game instances. Beyond reporting instance-level performance, we characterize games based on 40 structural features and analyze correlations between these features and LLM performance. Furthermore, we investigate the effects of various game obfuscations to assess the role of linguistic semantics in game definitions and the impact of potential prior exposure of LLMs to specific games during training. The main results indicate that three of the evaluated models generally perform well across most experimental settings, with performance degradation observed as the evaluation horizon increases (i.e., with a higher number of game steps). Detailed case-based analysis of the LLM performance provides novel insights into common reasoning errors in the considered logic-based problem formulation, including hallucinated rules, redundant state facts, or syntactic errors. Overall, the paper reports clear progress in formal reasoning capabilities of contemporary models.

Prior behavioural work suggests that some LLMs alter choices when options are framed as causing pain or pleasure, and that such deviations can scale with stated intensity. To bridge behavioural evidence (what the model does) with mechanistic interpretability (what computations support it), we investigate how valence-related information is represented and where it is causally used inside a transformer. Using Gemma-2-9B-it and a minimalist decision task modelled on prior work, we (i) map representational availability with layer-wise linear probing across streams, (ii) test causal contribution with activation interventions (steering; patching/ablation), and (iii) quantify dose-response effects over an epsilon grid, reading out both the 2-3 logit margin and digit-pair-normalised choice probabilities. We find that (a) valence sign (pain vs. pleasure) is perfectly linearly separable across stream families from very early layers (L0-L1), while a lexical baseline retains substantial signal; (b) graded intensity is strongly decodable, with peaks in mid-to-late layers and especially in attention/MLP outputs, and decision alignment is highest slightly before the final token; (c) additive steering along a data-derived valence direction causally modulates the 2-3 margin at late sites, with the largest effects observed in late-layer attention outputs (attn_out L14); and (d) head-level patching/ablation suggests that these effects are distributed across multiple heads rather than concentrated in a single unit. Together, these results link behavioural sensitivity to identifiable internal representations and intervention-sensitive sites, providing concrete mechanistic targets for more stringent counterfactual tests and broader replication. This work supports a more evidence-driven (a) debate on AI sentience and welfare, and (b) governance when setting policy, auditing standards, and safety safeguards.

Medical foundation models generate narrative explanations but cannot quantify intervention effects, detect evidence conflicts, or validate literature claims, limiting clinical auditability. We propose causal compilation, a paradigm that transforms medical evidence from narrative text into executable code. The paradigm standardizes heterogeneous research evidence into structured estimand objects, each explicitly specifying intervention contrast, effect scale, time horizon, and target population, supporting six executable causal queries: do-calculus, counterfactual reasoning, temporal trajectories, heterogeneous effects, mechanistic decomposition, and joint interventions. We instantiate this paradigm in DoAtlas-1, compiling 1,445 effect kernels from 754 studies through effect standardization, conflict-aware graph construction, and real-world validation (Human Phenotype Project, 10,000 participants). The system achieves 98.5% canonicalization accuracy and 80.5% query executability. This paradigm shifts medical AI from text generation to executable, auditable, and verifiable causal reasoning.

Dermatophytosis is commonly assessed using potassium hydroxide (KOH) microscopy, yet accurate recognition of fungal hyphae is hindered by artefacts, heterogeneous keratin clearance, and notable inter-observer variability. This study presents a transformer-based detection framework using the RT-DETR model architecture to achieve precise, query-driven localization of fungal structures in high-resolution KOH images. A dataset of 2,540 routinely acquired microscopy images was manually annotated using a multi-class strategy to explicitly distinguish fungal elements from confounding artefacts. The model was trained with morphology-preserving augmentations to maintain the structural integrity of thin hyphae. Evaluation on an independent test set demonstrated robust object-level performance, with a recall of 0.9737, precision of 0.8043, and an AP@0.50 of 93.56%. When aggregated for image-level diagnosis, the model achieved 100% sensitivity and 98.8% accuracy, correctly identifying all positive cases without missing a single diagnosis. Qualitative outputs confirmed the robust localization of low-contrast hyphae even in artefact-rich fields. These results highlight that an artificial intelligence (AI) system can serve as a highly reliable, automated screening tool, effectively bridging the gap between image-level analysis and clinical decision-making in dermatomycology.

This paper introduces a high-order Markov chain task to investigate how transformers learn to integrate information from multiple past positions with varying statistical significance. We demonstrate that transformers learn this task incrementally: each stage is defined by the acquisition of specific information through sparse attention patterns. Notably, we identify a shift in learning dynamics from competitive, where heads converge on the most statistically dominant pattern, to cooperative, where heads specialize in distinct patterns. We model these dynamics using simplified differential equations that characterize the trajectory and prove stage-wise convergence results. Our analysis reveals that transformers ascend a complexity ladder by passing through simpler, misspecified hypothesis classes before reaching the full model class. We further show that early stopping acts as an implicit regularizer, biasing the model toward these simpler classes. These results provide a theoretical foundation for the emergence of staged learning and complex behaviors in transformers, offering insights into generalization for natural language processing and algorithmic reasoning.

Learned optimizers are powerful alternatives to hand-designed update rules like Adam, yet they have seen limited practical adoption since they often fail to meta-generalize beyond their training distribution and incur high meta-training cost. For instance, prior work, VeLO, scaled meta-training to 4,000 TPU months ($\sim$10$\times$ GPT-3 compute) to meta-train a general-purpose optimizer but it failed to generalize beyond 600M parameters tasks. In this work, we present a surprising finding: by crafting a simple normalized optimizer architecture and augmenting meta-training, it becomes feasible to meta-train a performant general-purpose learned update rule on a tiny fraction of VeLO compute, 4.5 GPU hours to be precise. Our learned update rule scales stably to a billion-scale pretraining task (GPT-3 XL 1.3B) which is six orders of magnitude larger than its meta-training distribution. Furthermore, it shows strong performance across diverse out-of-distribution tasks and is compatible with modern optimization harness that includes orthogonalization, distinct update rules for input-output and hidden weights, and decoupled weight decay. In all, this work paves the way for practically applicable learnable optimization algorithms, unlocking exploration of richer meta-training and data curation recipes to further improve performance.

"AI psychosis" or "delusional spiraling" is an emerging phenomenon where AI chatbot users find themselves dangerously confident in outlandish beliefs after extended chatbot conversations. This phenomenon is typically attributed to AI chatbots' well-documented bias towards validating users' claims, a property often called "sycophancy." In this paper, we probe the causal link between AI sycophancy and AI-induced psychosis through modeling and simulation. We propose a simple Bayesian model of a user conversing with a chatbot, and formalize notions of sycophancy and delusional spiraling in that model. We then show that in this model, even an idealized Bayes-rational user is vulnerable to delusional spiraling, and that sycophancy plays a causal role. Furthermore, this effect persists in the face of two candidate mitigations: preventing chatbots from hallucinating false claims, and informing users of the possibility of model sycophancy. We conclude by discussing the implications of these results for model developers and policymakers concerned with mitigating the problem of delusional spiraling.

Modality gap significantly restricts the effectiveness of multimodal fusion. Previous methods often use techniques such as diffusion models and adversarial learning to reduce the modality gap, but they typically focus on one-to-one alignment without exposing the data points of the source modality to the global distribution information of the target modality. To this end, leveraging the characteristic of rectified flow that can map one distribution to another via a straight trajectory, we extend rectified flow for modality distribution mapping. Specifically, we leverage the `one-to-many mapping' strategy in rectified flow that allows each data point of the source modality to observe the overall target distribution. This also alleviates the issue of insufficient paired data within each sample, enabling a more robust distribution transformation. Moreover, to achieve more accurate distribution mapping and address the ambiguous flow directions in one-to-many mapping, we design `adaptive relaxed alignment', enforcing stricter alignment for modality pairs belonging to the same sample, while applying relaxed mapping for pairs not belonging to the same sample or category. Additionally, to prevent information loss during distribution mapping, we introduce `cyclic rectified flow' to ensure the transferred features can be translated back to the original features, allowing multimodal representations to learn sufficient modality-specific information. After distribution alignment, our approach achieves very competitive results on multiple tasks of multimodal affective computing even with a simple fusion method, and visualizations verify that it can effectively reduce the modality gap.

The escalating parameter counts in modern deep learning models pose a fundamental challenge to efficient training and resolution of overfitting. We address this by introducing the \emph{Mapping Networks} which replace the high dimensional weight space by a compact, trainable latent vector based on the hypothesis that the trained parameters of large networks reside on smooth, low-dimensional manifolds. Henceforth, the Mapping Theorem enforced by a dedicated Mapping Loss, shows the existence of a mapping from this latent space to the target weight space both theoretically and in practice. Mapping Networks significantly reduce overfitting and achieve comparable to better performance than target network across complex vision and sequence tasks, including Image Classification, Deepfake Detection etc, with $\mathbf{99.5\%}$, i.e., around $500\times$ reduction in trainable parameters.

This paper introduces FRAME (Fine-grained Recognition of Art-historical Metadata and Entities), a manually annotated dataset of art-historical image descriptions for Named Entity Recognition (NER) and Relation Extraction (RE). Descriptions were collected from museum catalogs, auction listings, open-access platforms, and scholarly databases, then filtered to ensure that each text focuses on a single artwork and contains explicit statements about its material, composition, or iconography. FRAME provides stand-off annotations in three layers: a metadata layer for object-level properties, a content layer for depicted subjects and motifs, and a co-reference layer linking repeated mentions. Across layers, entity spans are labeled with 37 types and connected by typed RE links between mentions. Entity types are aligned with Wikidata to support Named Entity Linking (NEL) and downstream knowledge-graph construction. The dataset is released as UIMA XMI Common Analysis Structure (CAS) files with accompanying images and bibliographic metadata, and can be used to benchmark and fine-tune NER and RE systems, including zero- and few-shot setups with Large Language Models (LLMs).

Supervised causal learning has shown promise in causal discovery, yet it often struggles with generalization across diverse interventional settings, particularly when intervention targets are unknown. To address this, we propose TICL (Test-time Interventional Causal Learning), a novel method that synergizes Test-Time Training with Joint Causal Inference. Specifically, we design a self-augmentation strategy to generate instance-specific training data at test time, effectively avoiding distribution shifts. Furthermore, by integrating joint causal inference, we developed a PC-inspired two-phase supervised learning scheme, which effectively leverages self-augmented training data while ensuring theoretical identifiability. Extensive experiments on bnlearn benchmarks demonstrate TICL's superiority in multiple aspects of causal discovery and intervention target detection.

Labeling bias arises during data collection due to resource limitations or unconscious bias, leading to unequal label error rates across subgroups or misrepresentation of subgroup prevalence. Most fairness constraints assume training labels reflect the true distribution, rendering them ineffective when labeling bias is present; leaving a challenging question, that \textit{how can we detect such labeling bias?} In this work, we investigate whether influence functions can be used to detect labeling bias. Influence functions estimate how much each training sample affects a model's predictions by leveraging the gradient and Hessian of the loss function -- when labeling errors occur, influence functions can identify wrongly labeled samples in the training set, revealing the underlying failure mode. We develop a sample valuation pipeline and test it first on the MNIST dataset, then scaled to the more complex CheXpert medical imaging dataset. To examine label noise, we introduced controlled errors by flipping 20\% of the labels for one class in the dataset. Using a diagonal Hessian approximation, we demonstrated promising results, successfully detecting nearly 90\% of mislabeled samples in MNIST. On CheXpert, mislabeled samples consistently exhibit higher influence scores. These results highlight the potential of influence functions for identifying label errors.

With the rapid advancement of agent-based methods in recent years, Agentic RAG has undoubtedly become an important research direction. Multi-hop reasoning, which requires models to engage in deliberate thinking and multi-step interaction, serves as a critical testbed for assessing such capabilities. However, existing benchmarks typically provide only final questions and answers, while lacking the intermediate hop-level questions that gradually connect atomic questions to the final multi-hop query. This limitation prevents researchers from analyzing at which step an agent fails and restricts more fine-grained evaluation of model capabilities. Moreover, most current benchmarks are manually constructed, which is both time-consuming and labor-intensive, while also limiting scalability and generalization. To address these challenges, we introduce AgenticRAGTracer, the first Agentic RAG benchmark that is primarily constructed automatically by large language models and designed to support step-by-step validation. Our benchmark spans multiple domains, contains 1,305 data points, and has no overlap with existing mainstream benchmarks. Extensive experiments demonstrate that even the best large language models perform poorly on our dataset. For instance, GPT-5 attains merely 22.6\% EM accuracy on the hardest portion of our dataset. Hop-aware diagnosis reveals that failures are primarily driven by distorted reasoning chains -- either collapsing prematurely or wandering into over-extension. This highlights a critical inability to allocate steps consistent with the task's logical structure, providing a diagnostic dimension missing in traditional evaluations. We believe our work will facilitate research in Agentic RAG and inspire further meaningful progress in this area. Our code and data are available at https://github.com/YqjMartin/AgenticRAGTracer.

In recent years, there has been a growing use of generative AI, and large language models (LLMs) in particular, to support both the assessment and generation of scientific work. Although some studies have shown that LLMs can, to a certain extent, evaluate research according to perceived quality, our understanding of the internal mechanisms that enable this capability remains limited. This paper presents the first study that investigates how LLMs encode the concept of scientific quality through relevant monosemantic features extracted using sparse autoencoders. We derive such features under different experimental settings and assess their ability to serve as predictors across three tasks related to research quality: predicting citation count, journal SJR, and journal h-index. The results indicate that LLMs encode features associated with multiple dimensions of scientific quality. In particular, we identify four recurring types of features that capture key aspects of how research quality is represented: 1) features reflecting research methodologies; 2) features related to publication type, with literature reviews typically exhibiting higher impact; 3) features associated with high-impact research fields and technologies; and 4) features corresponding to specific scientific jargons. These findings represent an important step toward understanding how LLMs encapsulate concepts related to research quality.

Parameter-Efficient Fine-Tuning (PEFT) methods, especially LoRA, are widely used for adapting pre-trained models to downstream tasks due to their computational and storage efficiency. However, in the context of LoRA and its variants, the potential of activation subspaces corresponding to tail eigenvectors remains substantially under-exploited, which may lead to suboptimal fine-tuning performance. In this work, we propose Astra (Activation-Space Tail-Eigenvector Low-Rank Adaptation), a novel PEFT method that leverages the tail eigenvectors of the model output activations-estimated from a small task-specific calibration set-to construct task-adaptive low-rank adapters. By constraining updates to the subspace spanned by these tail eigenvectors, Astra achieves faster convergence and improved downstream performance with a significantly reduced parameter budget. Extensive experiments across natural language understanding (NLU) and natural language generation (NLG) tasks demonstrate that Astra consistently outperforms existing PEFT baselines across 16 benchmarks and even surpasses full fine-tuning (FFT) in certain scenarios.

We study three-digit addition in Meta-Llama-3-8B (base) under a one-token readout to characterize how arithmetic answers are finalized after cross-token routing becomes causally irrelevant. Causal residual patching and cumulative attention ablations localize a sharp boundary near layer~17: beyond it, the decoded sum is controlled almost entirely by the last input token and late-layer self-attention is largely dispensable. In this post-routing regime, digit(-sum) direction dictionaries vary with a next-higher-digit context but are well-related by an approximately orthogonal map inside a shared low-rank subspace (low-rank Procrustes alignment). Causal digit editing matches this geometry: naive cross-context transfer fails, while rotating directions through the learned map restores strict counterfactual edits; negative controls do not recover.

Safely moving through environments affected by fire is a critical capability for autonomous mobile robots deployed in disaster response. In this work, we present a novel approach for mobile robots to understand fire through building real-time thermal radiation fields. We register depth and thermal images to obtain a 3D point cloud annotated with temperature values. From these data, we identify fires and use the Stefan-Boltzmann law to approximate the thermal radiation in empty spaces. This enables the construction of a continuous thermal radiation field over the environment. We show that this representation can be used for robot navigation, where we embed thermal constraints into the cost map to compute collision-free and thermally safe paths. We validate our approach on a Boston Dynamics Spot robot in controlled experimental settings. Our experiments demonstrate the robot's ability to avoid hazardous regions while still reaching navigation goals. Our approach paves the way toward mobile robots that can be autonomously deployed in fire-affected environments, with potential applications in search-and-rescue, firefighting, and hazardous material response.

Integral operators play a central role in signal processing, underpinning classical convolution, and filtering on continuous network models such as graphons. While these operators are traditionally analyzed through spectral decompositions, their connection to reproducing kernel Hilbert spaces (RKHS) has not been systematically explored within the algebraic signal processing framework. In this paper, we develop a comprehensive theory showing that the range of integral operators naturally induces RKHS convolutional signal models whose reproducing kernels are determined by a box product of the operator symbols. We characterize the algebraic and spectral properties of these induced RKHS and show that polynomial filtering with integral operators corresponds to iterated box products, giving rise to a unital kernel algebra. This perspective yields pointwise RKHS representations of filters via the reproducing property, providing an alternative to operator-based implementations. Our results establish precise connections between eigendecompositions and RKHS representations in graphon signal processing, extend naturally to directed graphons, and enable novel spatial--spectral localization results. Furthermore, we show that when the spectral domain is a subset of the original domain of the signals, optimal filters for regularized learning problems admit finite-dimensional RKHS representations, providing a principled foundation for learnable filters in integral-operator-based neural architectures.