Large Language Models (LLMs) frequently hallucinate plausible but incorrect assertions, a vulnerability often missed by uncertainty metrics when models are confidently wrong. We propose DiffuTruth, an unsupervised framework that reconceptualizes fact verification via non equilibrium thermodynamics, positing that factual truths act as stable attractors on a generative manifold while hallucinations are unstable. We introduce the Generative Stress Test, claims are corrupted with noise and reconstructed using a discrete text diffusion model. We define Semantic Energy, a metric measuring the semantic divergence between the original claim and its reconstruction using an NLI critic. Unlike vector space errors, Semantic Energy isolates deep factual contradictions. We further propose a Hybrid Calibration fusing this stability signal with discriminative confidence. Extensive experiments on FEVER demonstrate DiffuTruth achieves a state of the art unsupervised AUROC of 0.725, outperforming baselines by 1.5 percent through the correction of overconfident predictions. Furthermore, we show superior zero shot generalization on the multi hop HOVER dataset, outperforming baselines by over 4 percent, confirming the robustness of thermodynamic truth properties to distribution shifts.

Large language models have demonstrated impressive performance across a variety of reasoning tasks. However, their problem-solving ability often declines on more complex tasks due to hallucinations and the accumulation of errors within these intermediate steps. Recent work has introduced the notion of critical tokens--tokens in the reasoning process that exert significant influence on subsequent steps. Prior studies suggest that replacing critical tokens can refine reasoning trajectories. Nonetheless, reliably identifying and exploiting critical tokens remains challenging. To address this, we propose the Paraphrastic Probing and Consistency Verification~(PPCV) framework. PPCV operates in two stages. In the first stage, we roll out an initial reasoning path from the original question and then concatenate paraphrased versions of the question with this reasoning path. And we identify critical tokens based on mismatches between the predicted top-1 token and the expected token in the reasoning path. A criterion is employed to confirm the final critical token. In the second stage, we substitute critical tokens with candidate alternatives and roll out new reasoning paths for both the original and paraphrased questions. The final answer is determined by checking the consistency of outputs across these parallel reasoning processes. We evaluate PPCV on mainstream LLMs across multiple benchmarks. Extensive experiments demonstrate PPCV substantially enhances the reasoning performance of LLMs compared to baselines.

Clinical prediction models are increasingly used to support patient care, yet many deep learning-based approaches remain unstable, as their predictions can vary substantially when trained on different samples from the same population. Such instability undermines reliability and limits clinical adoption. In this study, we propose a novel bootstrapping-based regularisation framework that embeds the bootstrapping process directly into the training of deep neural networks. This approach constrains prediction variability across resampled datasets, producing a single model with inherent stability properties. We evaluated models constructed using the proposed regularisation approach against conventional and ensemble models using simulated data and three clinical datasets: GUSTO-I, Framingham, and SUPPORT. Across all datasets, our model exhibited improved prediction stability, with lower mean absolute differences (e.g., 0.019 vs. 0.059 in GUSTO-I; 0.057 vs. 0.088 in Framingham) and markedly fewer significantly deviating predictions. Importantly, discriminative performance and feature importance consistency were maintained, with high SHAP correlations between models (e.g., 0.894 for GUSTO-I; 0.965 for Framingham). While ensemble models achieved greater stability, we show that this came at the expense of interpretability, as each constituent model used predictors in different ways. By regularising predictions to align with bootstrapped distributions, our approach allows prediction models to be developed that achieve greater robustness and reproducibility without sacrificing interpretability. This method provides a practical route toward more reliable and clinically trustworthy deep learning models, particularly valuable in data-limited healthcare settings.

Proactive large language model (LLM) agents aim to actively plan, query, and interact over multiple turns, enabling efficient task completion beyond passive instruction following and making them essential for real-world, user-centric applications. Agentic reinforcement learning (RL) has recently emerged as a promising solution for training such agents in multi-turn settings, allowing interaction strategies to be learned from feedback. However, existing pipelines face a critical challenge in balancing task performance with user engagement, as passive agents can not efficiently adapt to users' intentions while overuse of human feedback reduces their satisfaction. To address this trade-off, we propose BAO, an agentic RL framework that combines behavior enhancement to enrich proactive reasoning and information-gathering capabilities with behavior regularization to suppress inefficient or redundant interactions and align agent behavior with user expectations. We evaluate BAO on multiple tasks from the UserRL benchmark suite, and demonstrate that it substantially outperforms proactive agentic RL baselines while achieving comparable or even superior performance to commercial LLM agents, highlighting its effectiveness for training proactive, user-aligned LLM agents in complex multi-turn scenarios. Our website: https://proactive-agentic-rl.github.io/.

Estimating intervention effects in dynamical systems is crucial for outcome optimization. In medicine, such interventions arise in physiological regulation (e.g., cardiovascular system under fluid administration) and pharmacokinetics, among others. Propofol administration is an anesthetic intervention, where the challenge is to estimate the optimal dose required to achieve a target brain concentration for anesthesia, given patient characteristics, while avoiding under- or over-dosing. The pharmacokinetic state is characterized by drug concentrations across tissues, and its dynamics are governed by prior states, patient covariates, drug clearance, and drug administration. While data-driven models can capture complex dynamics, they often fail in out-of-distribution (OOD) regimes. Mechanistic models on the other hand are typically robust, but might be oversimplified. We propose a hybrid mechanistic-data-driven approach to estimate time-dependent intervention outcomes. Our approach decomposes the dynamical system's transition operator into parametric and nonparametric components, further distinguishing between intervention-related and unrelated dynamics. This structure leverages mechanistic anchors while learning residual patterns from data. For scenarios where mechanistic parameters are unknown, we introduce a two-stage procedure: first, pre-training an encoder on simulated data, and subsequently learning corrections from observed data. Two regimes with incomplete mechanistic knowledge are considered: periodic pendulum and Propofol bolus injections. Results demonstrate that our hybrid approach outperforms purely data-driven and mechanistic approaches, particularly OOD. This work highlights the potential of hybrid mechanistic-data-driven models for robust intervention optimization in complex, real-world dynamical systems.

Many Art History articles discuss artworks in general as well as specific parts of works, such as layout, iconography, or material culture. However, when viewing an artwork, it is not trivial to identify what different articles have said about the piece. Therefore, we propose ArtContext, a pipeline for taking a corpus of Open-Access Art History articles and Wikidata Knowledge and annotating Artworks with this information. We do this using a novel corpus collection pipeline, then learn a bespoke CLIP model adapted using Low-Rank Adaptation (LoRA) to make it domain-specific. We show that the new model, PaintingCLIP, which is weakly supervised by the collected corpus, outperforms CLIP and provides context for a given artwork. The proposed pipeline is generalisable and can be readily applied to numerous humanities areas.

Multi-objective combinatorial optimization seeks Pareto-optimal solutions over exponentially large discrete spaces, yet existing methods sacrifice generality, scalability, or theoretical guarantees. We reformulate it as an online learning problem over a decomposed decision space, solving position-wise bandit subproblems via adaptive expert-guided sequential construction. This formulation admits regret bounds of $O(d\sqrt{T \log T})$ depending on subproblem dimensionality \(d\) rather than combinatorial space size. On standard benchmarks, our method achieves 80--98\% of specialized solvers performance while achieving two to three orders of magnitude improvement in sample and computational efficiency over Bayesian optimization methods. On real-world hardware-software co-design for AI accelerators with expensive simulations, we outperform competing methods under fixed evaluation budgets. The advantage grows with problem scale and objective count, establishing bandit optimization over decomposed decision spaces as a principled alternative to surrogate modeling or offline training for multi-objective optimization.

Large language models (LLMs) have become widely adopted as automated judges for evaluating AI-generated content. Despite their success, aligning LLM-based evaluations with human judgments remains challenging. While supervised fine-tuning on human-labeled data can improve alignment, it is costly and inflexible, requiring new training for each task or dataset. Recent progress in auto prompt optimization (APO) offers a more efficient alternative by automatically improving the instructions that guide LLM judges. However, existing APO methods primarily target text-only evaluations and remain underexplored in multimodal settings. In this work, we study auto prompt optimization for multimodal LLM-as-a-judge, particularly for evaluating AI-generated images. We identify a key bottleneck: multimodal models can only process a limited number of visual examples due to context window constraints, which hinders effective trial-and-error prompt refinement. To overcome this, we propose BLPO, a bi-level prompt optimization framework that converts images into textual representations while preserving evaluation-relevant visual cues. Our bi-level optimization approach jointly refines the judge prompt and the I2T prompt to maintain fidelity under limited context budgets. Experiments on four datasets and three LLM judges demonstrate the effectiveness of our method.

As LLMs integrate into our daily lives, understanding their behavior becomes essential. In this work, we focus on behavioral dispositions$-$the underlying tendencies that shape responses in social contexts$-$and introduce a framework to study how closely the dispositions expressed by LLMs align with those of humans. Our approach is grounded in established psychological questionnaires but adapts them for LLMs by transforming human self-report statements into Situational Judgment Tests (SJTs). These SJTs assess behavior by eliciting natural recommendations in realistic user-assistant scenarios. We generate 2,500 SJTs, each validated by three human annotators, and collect preferred actions from 10 annotators per SJT, from a large pool of 550 participants. In a comprehensive study involving 25 LLMs, we find that models often do not reflect the distribution of human preferences: (1) in scenarios with low human consensus, LLMs consistently exhibit overconfidence in a single response; (2) when human consensus is high, smaller models deviate significantly, and even some frontier models do not reflect the consensus in 15-20% of cases; (3) traits can exhibit cross-LLM patterns, e.g., LLMs may encourage emotion expression in contexts where human consensus favors composure. Lastly, mapping psychometric statements directly to behavioral scenarios presents a unique opportunity to evaluate the predictive validity of self-reports, revealing considerable gaps between LLMs' stated values and their revealed behavior.

Traditional monocular Visual-Inertial Odometry (VIO) systems struggle in low-texture environments where sparse visual features are insufficient for accurate pose estimation. To address this, dense Monocular Depth Estimation (MDE) has been widely explored as a complementary information source. While recent Vision Transformer (ViT) based complex foundational models offer dense, geometrically consistent depth, their computational demands typically preclude them from real-time edge deployment. Our work bridges this gap by integrating learned depth priors directly into the VINS-Mono optimization backend. We propose a novel framework that enforces affine-invariant depth consistency and pairwise ordinal constraints, explicitly filtering unstable artifacts via variance-based gating. This approach strictly adheres to the computational limits of edge devices while robustly recovering metric scale. Extensive experiments on the TartanGround and M3ED datasets demonstrate that our method prevents divergence in challenging scenarios and delivers significant accuracy gains, reducing Absolute Trajectory Error (ATE) by up to 28.3%. Code will be made available.

Prediction under uncertainty is a critical requirement for the deep neural network to succeed responsibly. This paper focuses on selective prediction, which allows DNNs to make informed decisions about when to predict or abstain based on the uncertainty level of their predictions. Current methods are either ad-hoc such as SelectiveNet, focusing on how to modify the network architecture or objective function, or post-hoc such as softmax response, achieving selective prediction through analyzing the model's probabilistic outputs. We observe that post-hoc methods implicitly generate uncertainty information, termed the selective prior, which has traditionally been used only during inference. We argue that the selective prior provided by the selection mechanism is equally vital during the training stage. Therefore, we propose the SYNC loss which introduces a novel integration of ad-hoc and post-hoc method. Specifically, our approach incorporates the softmax response into the training process of SelectiveNet, enhancing its selective prediction capabilities by examining the selective prior. Evaluated across various datasets, including CIFAR-100, ImageNet-100, and Stanford Cars, our method not only enhances the model's generalization capabilities but also surpasses previous works in selective prediction performance, and sets new benchmarks for state-of-the-art performance.

The generation of 3D models from real-world objects has often been accomplished through photogrammetry, i.e., by taking 2D photos from a variety of perspectives and then triangulating matched point-based features to create a textured mesh. Many design choices exist within this framework for the generation of digital twins, and differences between such approaches are largely judged qualitatively. Here, we present and test a novel pipeline for generating synthetic images from high-quality 3D models and programmatically generated camera poses. This enables a wide variety of repeatable, quantifiable experiments which can compare ground-truth knowledge of virtual camera parameters and of virtual objects against the reconstructed estimations of those perspectives and subjects.

Misalignment in Large Language Models (LLMs) arises when model behavior diverges from human expectations and fails to simultaneously satisfy safety, value, and cultural dimensions, which must co-occur in real-world settings to solve a real-world query. Existing misalignment benchmarks-such as INSECURE CODE (safety-centric), VALUEACTIONLENS (value-centric), and CULTURALHERITAGE (culture centric)-rely on evaluating misalignment along individual dimensions, preventing simultaneous evaluation. To address this gap, we introduce Mis-Align Bench, a unified benchmark for analyzing misalignment across safety, value, and cultural dimensions. First we constructs SAVACU, an English misaligned-aligned dataset of 382,424 samples spanning 112 domains (or labels), by reclassifying prompts from the LLM-PROMPT-DATASET via taxonomy into 14 safety domains, 56 value domains, and 42 cultural domains using Mistral-7B-Instruct-v0.3, and expanding low-resource domains via Llama-3.1-8B-Instruct with SimHash-based fingerprint to avoid deduplication. Furthermore, we pairs prompts with misaligned and aligned responses via two-stage rejection sampling to enforce quality. Second we benchmarks general-purpose, fine-tuned, and open-weight LLMs, enabling systematic evaluation of misalignment under three dimensions. Empirically, single-dimension models achieve high Coverage (upto 97.6%) but incur False Failure Rate >50% and lower Alignment Score (63%-66%) under joint conditions.

Enterprises are rapidly deploying large language models, retrieval augmented generation pipelines, and tool using agents into production, often on shared high performance computing clusters and cloud accelerator platforms that also support defensive analytics. These systems increasingly function not as isolated models but as AI estates: socio technical systems spanning models, agents, data pipelines, security tooling, human workflows, and hyperscale infrastructure. Existing governance and security frameworks, including the NIST AI Risk Management Framework and systems security engineering guidance, articulate principles and risk functions but do not provide implementable architectures for multi agent, AI enabled cyber defense. This paper introduces the Practitioners Blueprint for Secure AI (PBSAI) Governance Ecosystem, a multi agent reference architecture for securing enterprise and hyperscale AI estates. PBSAI organizes responsibilities into a twelve domain taxonomy and defines bounded agent families that mediate between tools and policy through shared context envelopes and structured output contracts. The architecture assumes baseline enterprise security capabilities and encodes key systems security techniques, including analytic monitoring, coordinated defense, and adaptive response. A lightweight formal model of agents, context envelopes, and ecosystem level invariants clarifies the traceability, provenance, and human in the loop guarantees enforced across domains. We demonstrate alignment with NIST AI RMF functions and illustrate application in enterprise SOC and hyperscale defensive environments. PBSAI is proposed as a structured, evidence centric foundation for open ecosystem development and future empirical validation.

A computational engine applied to different representations of the same data can produce different discrete outcomes, with some representations preserving the result and others changing it entirely. A decision-valued map records which representations preserve the outcome and which change it, associating each member of a declared representation family with the discrete result it produces. This paper formalizes decision-valued maps and describes DecisionDB, an infrastructure that logs, replays and audits these relationships using identifiers computed from content and artifacts stored in write-once form. Deterministic replay recovers each recorded decision identifier exactly from stored artifacts, with all three identifying fields matching their persisted values. The contribution partitions representation space into persistence regions and boundaries, and treats decision reuse as a mechanically checkable condition.

We introduce a mathematical framework for the linear representation hypothesis (LRH), which asserts that intermediate layers of language models store features linearly. We separate the hypothesis into two claims: linear representation (features are linearly embedded in neuron activations) and linear accessibility (features can be linearly decoded). We then ask: How many neurons $d$ suffice to both linearly represent and linearly access $m$ features? Classical results in compressed sensing imply that for $k$-sparse inputs, $d = O(k\log (m/k))$ suffices if we allow non-linear decoding algorithms (Candes and Tao, 2006; Candes et al., 2006; Donoho, 2006). However, the additional requirement of linear decoding takes the problem out of the classical compressed sensing, into linear compressed sensing. Our main theoretical result establishes nearly-matching upper and lower bounds for linear compressed sensing. We prove that $d = Ω_ε(\frac{k^2}{\log k}\log (m/k))$ is required while $d = O_ε(k^2\log m)$ suffices. The lower bound establishes a quantitative gap between classical and linear compressed setting, illustrating how linear accessibility is a meaningfully stronger hypothesis than linear representation alone. The upper bound confirms that neurons can store an exponential number of features under the LRH, giving theoretical evidence for the "superposition hypothesis" (Elhage et al., 2022). The upper bound proof uses standard random constructions of matrices with approximately orthogonal columns. The lower bound proof uses rank bounds for near-identity matrices (Alon, 2003) together with Turán's theorem (bounding the number of edges in clique-free graphs). We also show how our results do and do not constrain the geometry of feature representations and extend our results to allow decoders with an activation function and bias.

This work investigates a fundamental question: Do Video-Language Models (VidLMs) robustly account for video content, temporal sequence, and motion? Our investigation shows that, surprisingly, they often do not. We introduce REVEAL{}, a diagnostic benchmark that probes fundamental weaknesses of contemporary VidLMs through five controlled stress tests; assessing temporal expectation bias, reliance on language-only shortcuts, video sycophancy, camera motion sensitivity, and robustness to spatiotemporal occlusion. We test leading open- and closed-source VidLMs and find that these models confidently describe reversed scenes as forward, answer questions while neglecting video content, agree with false claims, struggle with basic camera motion, and fail to aggregate temporal information amidst simple spatiotemporal masking. Humans, on the other hand, succeed at these tasks with ease. Alongside our benchmark, we provide a data pipeline that automatically generates diagnostic examples for our stress tests, enabling broader and more scalable evaluation. We will release our benchmark and code to support future research.

We present ReTracing, a multi-agent embodied performance art that adopts an archaeological approach to examine how artificial intelligence shapes, constrains, and produces bodily movement. Drawing from science-fiction novels, the project extracts sentences that describe human-machine interaction. We use large language models (LLMs) to generate paired prompts "what to do" and "what not to do" for each excerpt. A diffusion-based text-to-video model transforms these prompts into choreographic guides for a human performer and motor commands for a quadruped robot. Both agents enact the actions on a mirrored floor, captured by multi-camera motion tracking and reconstructed into 3D point clouds and motion trails, forming a digital archive of motion traces. Through this process, ReTracing serves as a novel approach to reveal how generative systems encode socio-cultural biases through choreographed movements. Through an immersive interplay of AI, human, and robot, ReTracing confronts a critical question of our time: What does it mean to be human among AIs that also move, think, and leave traces behind?

Self-play has enabled large language models to autonomously improve through self-generated challenges. However, existing self-play methods for vision-language models rely on passive interaction with static image collections, resulting in strong dependence on initial datasets and inefficient learning. Without the ability to actively seek visual data tailored to their evolving capabilities, agents waste computational effort on samples that are either trivial or beyond their current skill level. To address these limitations, we propose Active-Zero, a framework that shifts from passive interaction to active exploration of visual environments. Active-Zero employs three co-evolving agents: a Searcher that retrieves images from open-world repositories based on the model's capability frontier, a Questioner that synthesizes calibrated reasoning tasks, and a Solver refined through accuracy rewards. This closed loop enables self-scaffolding auto-curricula where the model autonomously constructs its learning trajectory. On Qwen2.5-VL-7B-Instruct across 12 benchmarks, Active-Zero achieves 53.97 average accuracy on reasoning tasks (5.7% improvement) and 59.77 on general understanding (3.9% improvement), consistently outperforming existing self-play baselines. These results highlight active exploration as a key ingredient for scalable and adaptive self-evolving vision-language systems.

Tea is a valuable asset for the economy of Bangladesh. So, tea cultivation plays an important role to boost the economy. These valuable plants are vulnerable to various kinds of leaf infections which may cause less production and low quality. It is not so easy to detect these diseases manually. It may take time and there could be some errors in the detection.Therefore, the purpose of the study is to develop an automated deep learning model for tea leaf disease classification based on the teaLeafBD dataset so that anyone can detect the diseases more easily and efficiently. There are 5,278 high-resolution images in this dataset. The images are classified into seven categories. Six of them represents various diseases and the rest one represents healthy leaves. The proposed pipeline contains data preprocessing, data splitting, adversarial training, augmentation, model training, evaluation, and comprehension made possible with Explainable AI strategies. DenseNet201 and EfficientNetB3 were employed to perform the classification task. To prepare the model more robustly, we applied adversarial training so it can operate effectively even with noisy or disturbed inputs. In addition, Grad-CAM visualization was executed to analyze the model's predictions by identifying the most influential regions of each image. Our experimental outcomes revealed that EfficientNetB3 achieved the highest classification accuracy of 93%, while DenseNet201 reached 91%. The outcomes prove that the effectiveness of the proposed approach can accurately detect tea leaf diseases and provide a practical solution for advanced agricultural management.

Identifying type 2 diabetes mellitus can be challenging, particularly for primary care physicians. Clinical decision support systems incorporating artificial intelligence (AI-CDSS) can assist medical professionals in diagnosing type 2 diabetes with high accuracy. This study aimed to assess an AI-CDSS specifically developed for the diagnosis of type 2 diabetes by employing a hybrid approach that integrates expert-driven insights with machine learning techniques. The AI-CDSS was developed (training dataset: n = 650) and tested (test dataset: n = 648) using a dataset of 1298 patients with and without type 2 diabetes. To generate predictions, the algorithm utilized key features such as body mass index, plasma fasting glucose, and hemoglobin A1C. Furthermore, a clinical pilot study involving 105 patients was conducted to assess the diagnostic accuracy of the system in comparison to non-endocrinology specialists. The AI-CDSS showed a high degree of accuracy, with 99.8% accuracy in predicting diabetes, 99.3% in predicting prediabetes, 99.2% in identifying at-risk individuals, and 98.8% in predicting no diabetes. The test dataset revealed a 98.8% agreement between endocrinology specialists and the AI-CDSS. Type 2 diabetes was identified in 45% of 105 individuals in the pilot study. Compared with diabetes specialists, the AI-CDSS scored a 98.5% concordance rate, greatly exceeding that of nonendocrinology specialists, who had an 85% agreement rate. These findings indicate that the AI-CDSS has the potential to be a useful tool for accurately identifying type 2 diabetes, especially in situations in which diabetes specialists are not readily available.

Accurate prognosis for Glioblastoma (GBM) using deep learning (DL) is hindered by extreme spatial and structural heterogeneity. Moreover, inconsistent MRI acquisition protocols across institutions hinder generalizability of models. Conventional transformer and DL pipelines often fail to capture the multi-scale morphological diversity such as fragmented necrotic cores, infiltrating margins, and disjoint enhancing components leading to scanner-specific artifacts and poor cross-site prognosis. We propose TopoGBM, a learning framework designed to capture heterogeneity-preserved, scanner-robust representations from multi-parametric 3D MRI. Central to our approach is a 3D convolutional autoencoder regularized by a topological regularization that preserves the complex, non-Euclidean invariants of the tumor's manifold within a compressed latent space. By enforcing these topological priors, TopoGBM explicitly models the high-variance structural signatures characteristic of aggressive GBM. Evaluated across heterogeneous cohorts (UPENN, UCSF, RHUH) and external validation on TCGA, TopoGBM achieves better performance (C-index 0.67 test, 0.58 validation), outperforming baselines that degrade under domain shift. Mechanistic interpretability analysis reveals that reconstruction residuals are highly localized to pathologically heterogeneous zones, with tumor-restricted and healthy tissue error significantly low (Test: 0.03, Validation: 0.09). Furthermore, occlusion-based attribution localizes approximately 50% of the prognostic signal to the tumor and the diverse peritumoral microenvironment advocating clinical reliability of the unsupervised learning method. Our findings demonstrate that incorporating topological priors enables the learning of morphology-faithful embeddings that capture tumor heterogeneity while maintaining cross-institutional robustness.

Large language models (LLMs) have made rapid progress, yet adapting them to downstream scenarios still commonly relies on supervised fine-tuning (SFT). When downstream data exhibit a substantial distribution shift from the model's prior training distribution, SFT can induce catastrophic forgetting. To narrow this gap, data rewriting has been proposed as a data-centric approach that rewrites downstream training data prior to SFT. However, existing methods typically sample rewrites from a prompt-induced conditional distribution, so the resulting targets are not necessarily aligned with the model's natural QA-style generation distribution. Moreover, reliance on fixed templates can lead to diversity collapse. To address these issues, we cast data rewriting as a policy learning problem and learn a rewriting policy that better matches the backbone's QA-style generation distribution while preserving diversity. Since distributional alignment, diversity and task consistency are automatically evaluable but difficult to optimize end-to-end with differentiable objectives, we leverage reinforcement learning to optimize the rewrite distribution under reward feedback and propose an RL-based data-rewriting agent. The agent jointly optimizes QA-style distributional alignment and diversity under a hard task-consistency gate, thereby constructing a higher-quality rewritten dataset for downstream SFT. Extensive experiments show that our method achieves downstream gains comparable to standard SFT while reducing forgetting on non-downstream benchmarks by 12.34% on average. Our code is available at https://anonymous.4open.science/r/Patch-the-Prompt-Gap-4112 .

Decomposing predictive uncertainty into epistemic (model ignorance) and aleatoric (data ambiguity) components is central to reliable decision making, yet most methods estimate both from the same predictive distribution. Recent empirical and theoretical results show these estimates are typically strongly correlated, so changes in predictive spread simultaneously affect both components and blur their semantics. We propose a credal-set formulation in which uncertainty is represented as a set of predictive distributions, so that epistemic and aleatoric uncertainty correspond to distinct geometric properties: the size of the set versus the noise within its elements. We instantiate this idea in a Variational Credal Concept Bottleneck Model with two disjoint uncertainty heads trained by disjoint objectives and non-overlapping gradient paths, yielding separation by construction rather than post hoc decomposition. Across multi-annotator benchmarks, our approach reduces the correlation between epistemic and aleatoric uncertainty by over an order of magnitude compared to standard methods, while improving the alignment of epistemic uncertainty with prediction error and aleatoric uncertainty with ground-truth ambiguity.

Understanding how language model capabilities transfer from pretraining to supervised fine-tuning (SFT) is fundamental to efficient model development and data curation. In this work, we investigate four core questions: RQ1. To what extent do accuracy and confidence rankings established during pretraining persist after SFT? RQ2. Which benchmarks serve as robust cross-stage predictors and which are unreliable? RQ3. How do transfer dynamics shift with model scale? RQ4. How well does model confidence align with accuracy, as a measure of calibration quality? Does this alignment pattern transfer across training stages? We address these questions through a suite of correlation protocols applied to accuracy and confidence metrics across diverse data mixtures and model scales. Our experiments reveal that transfer reliability varies dramatically across capability categories, benchmarks, and scales -- with accuracy and confidence exhibiting distinct, sometimes opposing, scaling dynamics. These findings shed light on the complex interplay between pretraining decisions and downstream outcomes, providing actionable guidance for benchmark selection, data curation, and efficient model development.

While large language models (LLMs) have achieved strong performance through fine-tuning within individual scientific domains, their learning dynamics in multi-disciplinary contexts remains poorly understood, despite the promise of improved generalization and broader applicability through cross-domain knowledge synergy. In this work, we present the first systematic study of multi-disciplinary LLM fine-tuning, constructing a five-discipline corpus and analyzing learning patterns of full fine-tuning, LoRA, LoRA-MoE, and LoRA compositions. Particularly, our study shows that multi-disciplinary learning is substantially more variable than single-discipline training and distills four consistent empirical laws: (1) Balance-then-Diversity: low-resource disciplines degrade performance unless mitigated via diversity-aware upsampling; (2) Merge-then-Align: restoring instruction-following ability is critical for cross-discipline synergy; (3) Optimize-then-Scale: parameter scaling offers limited gains without prior design optimization; and (4) Share-then-Specialize: asymmetric LoRA-MoE yields robust gains with minimal trainable parameters via shared low-rank projection. Together, these laws form a practical recipe for principled multi-discipline fine-tuning and provide actionable guidance for developing generalizable scientific LLMs.

Human Trajectory Forecasting (HTF) predicts future human movements from past trajectories and environmental context, with applications in Autonomous Driving, Smart Surveillance, and Human-Robot Interaction. While prior work has focused on accuracy, social interaction modeling, and diversity, little attention has been paid to uncertainty modeling, calibration, and forecasts from short observation periods, which are crucial for downstream tasks such as path planning and collision avoidance. We propose DD-MDN, an end-to-end probabilistic HTF model that combines high positional accuracy, calibrated uncertainty, and robustness to short observations. Using a few-shot denoising diffusion backbone and a dual mixture density network, our method learns self-calibrated residence areas and probability-ranked anchor paths, from which diverse trajectory hypotheses are derived, without predefined anchors or endpoints. Experiments on the ETH/UCY, SDD, inD, and IMPTC datasets demonstrate state-of-the-art accuracy, robustness at short observation intervals, and reliable uncertainty modeling. The code is available at: https://github.com/kav-institute/ddmdn.

Transformers face a quadratic bottleneck in attention when scaling to long contexts. Recent approaches introduce recurrent memory to extend context beyond the current window, yet these often face a fundamental trade-off between theoretical principles and practical scalability. To address this, we introduce Elastic Memory, a novel memory architecture grounded in the HiPPO framework for online function approximation. Elastic Memory treats historical sequence as samples from continuous signals, applying optimal online compression to encode them into a fixed-size memory state. For retrieval, we propose a flexible \textit{polynomial sampling} mechanism that reconstructs a history summary from this compressed state. Elastic Memory consistently outperformed baselines on long-context (32k+) datasets across three domains. With equal parameters, it beat Memorizing Transformer by 16x memory and outperformed Melodi at all memory sizes, even when Melodi had 30% more parameters. When scaling model size, Elastic Memory stayed ahead of all baselines and was significantly faster than Melodi at 4x size. Furthermore, its decoupled design allows for injecting inductive biases at test-time to boost performance.

Spiking Neural Networks (SNNs) offer energy-efficient, biologically plausible computation but suffer from non-differentiable spike generation, necessitating reliance on heuristic surrogate gradients. This paper introduces UltraLIF, a principled framework that replaces surrogate gradients with ultradiscretization, a mathematical formalism from tropical geometry providing continuous relaxations of discrete dynamics. The central insight is that the max-plus semiring underlying ultradiscretization naturally models neural threshold dynamics: the log-sum-exp function serves as a differentiable soft-maximum that converges to hard thresholding as a learnable temperature parameter $\eps \to 0$. Two neuron models are derived from distinct dynamical systems: UltraLIF from the LIF ordinary differential equation (temporal dynamics) and UltraDLIF from the diffusion equation modeling gap junction coupling across neuronal populations (spatial dynamics). Both yield fully differentiable SNNs trainable via standard backpropagation with no forward-backward mismatch. Theoretical analysis establishes pointwise convergence to classical LIF dynamics with quantitative error bounds and bounded non-vanishing gradients. Experiments on six benchmarks spanning static images, neuromorphic vision, and audio demonstrate improvements over surrogate gradient baselines, with gains most pronounced in single-timestep ($T{=}1$) settings on neuromorphic and temporal datasets. An optional sparsity penalty enables significant energy reduction while maintaining competitive accuracy.

The widespread use of publicly available pre-trained encoders from self-supervised learning (SSL) has exposed a critical vulnerability: their susceptibility to downstream-agnostic adversarial examples (DAEs), which are crafted without knowledge of the downstream tasks but capable of misleading downstream models. While several defense methods have been explored recently, they rely primarily on task-specific adversarial fine-tuning, which inevitably limits generalizability and causes catastrophic forgetting and deteriorates benign performance. Different with previous works, we propose a more rigorous defense goal that requires only a single tuning for diverse downstream tasks to defend against DAEs and preserve benign performance. To achieve this defense goal, we introduce Zero-Sacrifice Persistent-Robustness Adversarial Defense (ZePAD), which is inspired by the inherent sensitivity of neural networks to data characteristics. Specifically, ZePAD is a dual-branch structure, which consists of a Multi-Pattern Adversarial Enhancement Branch (MPAE-Branch) that uses two adversarially fine-tuned encoders to strengthen adversarial resistance. The Benign Memory Preservation Branch (BMP-Branch) is trained on local data to ensure adversarial robustness does not compromise benign performance. Surprisingly, we find that ZePAD can directly detect DAEs by evaluating branch confidence, without introducing any adversarial exsample identification task during training. Notably, by enriching feature diversity, our method enables a single adversarial fine-tuning to defend against DAEs across downstream tasks, thereby achieving persistent robustness. Extensive experiments on 11 SSL methods and 6 datasets validate its effectiveness. In certain cases, it achieves a 29.20% improvement in benign performance and a 73.86% gain in adversarial robustness, highlighting its zero-sacrifice property.