arXiv:2603.07315v1 Announce Type: new Abstract: One common concern about advanced artificial intelligence is that it will prevent us from turning it off, as that would interfere with pursuing its goals. In this paper, we discuss an unorthodox proposal for addressing this concern: give the AI a (primary) goal of being turned off (see also papers by Martin et al., and by Goldstein and Robinson). We also discuss whether and under what conditions this would be a good idea.

arXiv:2603.07546v1 Announce Type: new Abstract: Aging bridge networks require proactive, verifiable, and interpretable maintenance strategies, yet reinforcement learning (RL) policies trained solely on reward signals provide no formal safety guarantees and remain opaque to infrastructure managers. We demonstrate COOL-MC as a tool for verifying and explaining RL policies for multi-bridge network maintenance, building on a single-bridge Markov decision process (MDP) from the literature and extending it to a parallel network of three heterogeneous bridges with a shared periodic budget constraint, encoded in the PRISM modeling language. We train an RL agent on this MDP and apply probabilistic model checking and explainability methods to the induced discrete-time Markov chain (DTMC) that arises from the interaction between the learned policy and the underlying MDP. Probabilistic model checking reveals that the trained policy has a safety-violation probability of 3.5\% over the planning horizon, being slightly above the theoretical minimum of 0\% and indicating the suboptimality of the learned policy, noting that these results are based on artificially constructed transition probabilities and deterioration rates rather than real-world data, so absolute performance figures should be interpreted with caution. The explainability analysis further reveals, for instance, a systematic bias in the trained policy toward the state of bridge 1 over the remaining bridges in the network. These results demonstrate COOL-MC's ability to provide formal, interpretable, and practical analysis of RL maintenance policies.

arXiv:2603.07598v1 Announce Type: new Abstract: Chain-of-thought (CoT) improves reasoning reliability but increases token cost, motivating post-training compression of explicit reasoning traces. However, the shortest sufficient reasoning is not universal: it depends on difficulty, model capacity, and training state, making fixed length targets brittle. In practice, naive RL-based compression can also undesirably shorten the user-facing answer, because a single completion-level learning signal leaks across the think/answer boundary. We propose Difficulty-Scaled Segment-Wise GRPO (DSS-GRPO), which decomposes returns into think and answer components, computes group-relative advantages per segment, and routes them with hard token masks so compression updates act only on think while answer alignment acts only on answer. DSS-GRPO uses prompt-wise within-group shaping and difficulty-aware scaling to encourage concise reasoning without collapsing answer behavior.

arXiv:2603.07868v1 Announce Type: new Abstract: Recent advances in Vision-Language Models (VLMs) have demonstrated impressive multimodal understanding in general domains. However, their applicability to decision-oriented domains such as hospitality remains largely unexplored. In this work, we investigate how well VLMs can perform visual question answering (VQA) about hotel and facility images that are central to consumer decision-making. While many existing VQA benchmarks focus on factual correctness, they rarely capture what information users actually find useful. To address this, we first introduce Informativeness as a formal framework to quantify how much hospitality-relevant information an image-question pair provides. Guided by this framework, we construct a new hospitality-specific VQA dataset that covers various facility types, where questions are specifically designed to reflect key user information needs. Using this benchmark, we conduct experiments with several state-of-the-art VLMs, revealing that VLMs are not intrinsically decision-aware-key visual signals remain underutilized, and reliable informativeness reasoning emerges only after modest domain-specific finetuning.

arXiv:2603.07896v1 Announce Type: new Abstract: We introduce SMGI, a structural theory of general artificial intelligence, and recast the foundational problem of learning from the optimization of hypotheses within fixed environments to the controlled evolution of the learning interface itself. We formalize the Structural Model of General Intelligence (SMGI) via a typed meta-model $\theta = (r,\mathcal H,\Pi,\mathcal L,\mathcal E,\mathcal M)$ that treats representational maps, hypothesis spaces, structural priors, multi-regime evaluators, and memory operators as explicitly typed, dynamic components. By enforcing a strict mathematical separation between this structural ontology ($\theta$) and its induced behavioral semantics ($T_\theta$), we define general artificial intelligence as a class of admissible coupled dynamics $(\theta, T_\theta)$ satisfying four obligations: structural closure under typed transformations, dynamical stability under certified evolution, bounded statistical capacity, and evaluative invariance across regime shifts. We prove a structural generalization bound that links sequential PAC-Bayes analysis and Lyapunov stability, providing sufficient conditions for capacity control and bounded drift under admissible task transformations. Furthermore, we establish a strict structural inclusion theorem demonstrating that classical empirical risk minimization, reinforcement learning, program-prior models (Solomonoff-style), and modern frontier agentic pipelines operate as structurally restricted instances of SMGI.

arXiv:2603.07916v1 Announce Type: new Abstract: In recent advances, to enable a fully data-driven learning paradigm on relational databases (RDB), relational deep learning (RDL) is proposed to structure the RDB as a heterogeneous entity graph and adopt the graph neural network (GNN) as the predictive model. However, existing RDL methods neglect the imbalance problem of relational data in RDBs and risk under-representing the minority entities, leading to an unusable model in practice. In this work, we investigate, for the first time, class imbalance problem in RDB entity classification and design the relation-centric minority synthetic over-sampling GNN (Rel-MOSS), in order to fill a critical void in the current literature. Specifically, to mitigate the issue of minority-related information being submerged by majority counterparts, we design the relation-wise gating controller to modulate neighborhood messages from each individual relation type. Based on the relational-gated representations, we further propose the relation-guided minority synthesizer for over-sampling, which integrates the entity relational signatures to maintain relational consistency. Extensive experiments on 12 entity classification datasets provide compelling evidence for the superiority of Rel-MOSS, yielding an average improvement of up to 2.46% and 4.00% in terms of Balanced Accuracy and G-Mean, compared with SOTA RDL methods and classic methods for handling class imbalance.

arXiv:2603.08035v1 Announce Type: new Abstract: Reward modeling is essential for aligning Large Language Models(LLMs) with human preferences, yet conventional reward models suffer from poor interpretability and heavy reliance on costly expert annotations. While recent rubric-based approaches enhance evaluation transparency, they lack systematic quality control, yielding noisy and redundant criteria, failing to mitigate persistent biases (e.g., verbosity, position) in LLM evaluators, and creating a scalability-reliability trade-off. To address these limitations, we propose CDRRM (Contrast-Driven Rubric Reward Model), a framework built on a novel Contrast-then-Synthesis paradigm for high-quality rubric generation and guided preference judgment. CDRRM first conducts multi-dimensional contrastive profiling on preference pairs to identify causal discriminative factors, then synthesizes these insights into compact, context-aware rubrics to guide preference judg- ments. Extensive experiments on three authoritative benchmarks (RewardBench, RMBench, RMB) demonstrate that CDRRM achieves state-of-the-art performance across diverse domains and effectively mitigates aforementioned evaluation biases. Notably, our approach delivers exceptional data efficiency: training the rubric generator on only 3k high-quality samples empowers a frozen pre-trained judge model to outperform fully fine-tuned baselines. This work offers a scalable, interpretable, and data-efficient path for reward modeling.

arXiv:2603.08234v1 Announce Type: new Abstract: With the rapid advancement of large language models (LLMs), the safety of LLMs has become a critical concern. Despite significant efforts in safety alignment, current LLMs remain vulnerable to jailbreaking attacks. However, the root causes of such vulnerabilities are still poorly understood, necessitating a rigorous investigation into jailbreak mechanisms across both academic and industrial communities. In this work, we focus on a continuation-triggered jailbreak phenomenon, whereby simply relocating a continuation-triggered instruction suffix can substantially increase jailbreak success rates. To uncover the intrinsic mechanisms of this phenomenon, we conduct a comprehensive mechanistic interpretability analysis at the level of attention heads. Through causal interventions and activation scaling, we show that this jailbreak behavior primarily arises from an inherent competition between the model's intrinsic continuation drive and the safety defenses acquired through alignment training. Furthermore, we perform a detailed behavioral analysis of the identified safety-critical attention heads, revealing notable differences in the functions and behaviors of safety heads across different model architectures. These findings provide a novel mechanistic perspective for understanding and interpreting jailbreak behaviors in LLMs, offering both theoretical insights and practical implications for improving model safety.

arXiv:2603.08267v1 Announce Type: new Abstract: Bias in financial language models constitutes a major obstacle to their adoption in real-world applications. Detecting such bias is challenging, as it requires identifying inputs whose predictions change when varying properties unrelated to the decision, such as demographic attributes. Existing approaches typically rely on exhaustive mutation and pairwise prediction analysis over large corpora, which is effective but computationally expensive-particularly for large language models and can become impractical in continuous retraining and releasing processes. Aiming at reducing this cost, we conduct a large-scale study of bias in five financial language models, examining similarities in their bias tendencies across protected attributes and exploring cross-model-guided bias detection to identify bias-revealing inputs earlier. Our study uses approximately 17k real financial news sentences, mutated to construct over 125k original-mutant pairs. Results show that all models exhibit bias under both atomic (0.58\%-6.05\%) and intersectional (0.75\%-5.97\%) settings. Moreover, we observe consistent patterns in bias-revealing inputs across models, enabling substantial reuse and cost reduction in bias detection. For example, up to 73\% of FinMA's biased behaviours can be uncovered using only 20\% of the input pairs when guided by properties derived from DistilRoBERTa outputs.

arXiv:2603.08425v1 Announce Type: new Abstract: This paper presents IronEngine, a general AI assistant platform organized around a unified orchestration core that connects a desktop user interface, REST and WebSocket APIs, Python clients, local and cloud model backends, persistent memory, task scheduling, reusable skills, 24-category tool execution, MCP-compatible extensibility, and hardware-facing integration. IronEngine introduces a three-phase pipeline -- Discussion (Planner--Reviewer collaboration), Model Switch (VRAM-aware transition), and Execution (tool-augmented action loop) -- that separates planning quality from execution capability. The system features a hierarchical memory architecture with multi-level consolidation, a vectorized skill repository backed by ChromaDB, an adaptive model management layer supporting 92 model profiles with VRAM-aware context budgeting, and an intelligent tool routing system with 130+ alias normalization and automatic error correction. We present experimental results on file operation benchmarks achieving 100\% task completion with a mean total time of 1541 seconds across four heterogeneous tasks, and provide detailed comparisons with representative AI assistant systems including ChatGPT, Claude Desktop, Cursor, Windsurf, and open-source agent frameworks. Without disclosing proprietary prompts or core algorithms, this paper analyzes the platform's architectural decomposition, subsystem design, experimental performance, safety boundaries, and comparative engineering advantages. The resulting study positions IronEngine as a system-oriented foundation for general-purpose personal assistants, automation frameworks, and future human-centered agent platforms.

arXiv:2603.08455v1 Announce Type: new Abstract: When an RL agent's observations are gradually corrupted, at what drift rate does it "wake up" -- and what determines this boundary? We study world model-based self-monitoring under continuous observation drift across four MuJoCo environments, three detector families (z-score, variance, percentile), and three model capacities. We find that (1) a sharp detection threshold $\varepsilon^*$ exists universally: below it, drift is absorbed as normal variation; above it, detection occurs rapidly. The threshold's existence and sigmoid shape are invariant across all detector families and model capacities, though its position depends on the interaction between detector sensitivity, noise floor structure, and environment dynamics. (2) Sinusoidal drift is completely undetectable by all detector families -- including variance and percentile detectors with no temporal smoothing -- establishing this as a world model property rather than a detector artifact. (3) Within each environment, $\varepsilon^*$ follows a power law in detector parameters ($R^2 = 0.89$-$0.97$), but cross-environment prediction fails ($R^2 = 0.45$), revealing that the missing variable is environment-specific dynamics structure $\partial \mathrm{PE}/\partial\varepsilon$. (4) In fragile environments, agents collapse before any detector can fire ("collapse before awareness"), creating a fundamentally unmonitorable failure mode. Our results reframe $\varepsilon^*$ from an emergent world model property to a three-way interaction between noise floor, detector, and environment dynamics, providing a more defensible and empirically grounded account of self-monitoring boundaries in RL agents.

arXiv:2603.08575v1 Announce Type: new Abstract: The question of \emph{knowledge}, \emph{truth} and \emph{trust} is explored via a mathematical formulation of claims and sources. We define truth as the reproducibly perceived subset of knowledge, formalize sources as having both generative and discriminative roles, and develop a framework for reputation grounded in the \emph{conviction} -- the likelihood that a source's stance is vindicated by independent consensus. We argue that conviction, rather than correctness or faithfulness, is the principled basis for trust: it is regime-independent, rewards genuine contribution, and demands the transparent and self-sufficient perceptions that make external verification possible. We formalize reputation as the expected weighted signed conviction over a realm of claims, characterize its behavior across source-claim regimes, and identify continuous verification as both a theoretical necessity and a practical mechanism through which reputation accrues. The framework is applied to AI agents, which are identified as capable but error-prone sources for whom verifiable conviction and continuously accrued reputation constitute the only robust foundation for trust.

arXiv:2603.08706v1 Announce Type: new Abstract: Training large language models (LLMs) as autonomous agents often begins with imitation learning, but it only teaches agents what to do without understanding why: agents never contrast successful actions against suboptimal alternatives and thus lack awareness of action quality. Recent approaches attempt to address this by introducing self-reflection supervision derived from contrasts between expert and alternative actions. However, the training paradigm fundamentally remains imitation learning: the model imitates pre-constructed reflection text rather than learning to reason autonomously. We propose Agentic Critical Training (ACT), a reinforcement learning paradigm that trains agents to identify the better action among alternatives. By rewarding whether the model's judgment is correct, ACT drives the model to autonomously develop reasoning about action quality, producing genuine self-reflection rather than imitating it. Across three challenging agent benchmarks, ACT consistently improves agent performance when combined with different post-training methods. It achieves an average improvement of 5.07 points over imitation learning and 4.62 points over reinforcement learning. Compared to approaches that inject reflection capability through knowledge distillation, ACT also demonstrates clear advantages, yielding an average improvement of 2.42 points. Moreover, ACT enables strong out-of-distribution generalization on agentic benchmarks and improves performance on general reasoning benchmarks without any reasoning-specific training data, highlighting the value of our method. These results suggest that ACT is a promising path toward developing more reflective and capable LLM agents.

arXiv:2502.18775v1 Announce Type: cross Abstract: Glioma, the prevalent primary brain tumor, exhibits diverse aggressiveness levels and prognoses. Precise classification of glioma is paramount for treatment planning and predicting prognosis. This study aims to develop an algorithm to fuse the MRI images from T1, T2, T1ce, and fluid-attenuated inversion recovery (FLAIR) sequences to enhance the efficacy of glioma subclass classification as no tumor, necrotic core, peritumoral edema, and enhancing tumor. The MRI images from BraTS datasets were used in this work. The images were pre-processed using max-min normalization to ensure consistency in pixel intensity values across different images. The segmentation of the necrotic core, peritumoral edema, and enhancing tumor was performed on 2D and 3D images separately using UNET architecture. Further, the segmented regions from multimodal MRI images were fused using the weighted averaging technique. Integrating 2D and 3D segmented outputs enhances classification accuracy by capturing detailed features like tumor shape, boundaries, and intensity distribution in slices, while also providing a comprehensive view of spatial extent, shape, texture, and localization within the brain volume. The fused images were used as input to the pre-trained ResNet50 model for glioma subclass classification. The network is trained on 80% and validated on 20% of the data. The proposed method achieved a classification of accuracy of 99.25%, precision of 99.30%, recall of 99.10, F1 score of 99.19%, Intersection Over Union of 84.49%, and specificity of 99.76, which showed a significantly higher performance than existing techniques. These findings emphasize the significance of glioma segmentation and classification in aiding accurate diagnosis.

arXiv:2507.11202v1 Announce Type: cross Abstract: Multimodal Emotion Recognition (MER) often encounters incomplete multimodality in practical applications due to sensor failures or privacy protection requirements. While existing methods attempt to address various incomplete multimodal scenarios by balancing the training of each modality combination through additional gradients, these approaches face a critical limitation: training gradients from different modality combinations conflict with each other, ultimately degrading the performance of the final prediction model. In this paper, we propose a unimodal decoupled dynamic low-rank adaptation method based on modality combinations, named MCULoRA, which is a novel framework for the parameter-efficient training of incomplete multimodal learning models. MCULoRA consists of two key modules, modality combination aware low-rank adaptation (MCLA) and dynamic parameter fine-tuning (DPFT). The MCLA module effectively decouples the shared information from the distinct characteristics of individual modality combinations. The DPFT module adjusts the training ratio of modality combinations based on the separability of each modality's representation space, optimizing the learning efficiency across different modality combinations. Our extensive experimental evaluation in multiple benchmark datasets demonstrates that MCULoRA substantially outperforms previous incomplete multimodal learning approaches in downstream task accuracy.

arXiv:2603.06589v1 Announce Type: cross Abstract: Model calibration and debiasing are fundamental to the reliability and fairness of large scale recommendation systems. We introduce the Isotonic Layer, a novel, differentiable framework that integrates piecewise linear fitting directly into neural architectures. By partitioning the feature space into discrete segments and optimizing non negative slopes via a constrained dot product mechanism, we enforce a global monotonic inductive bias. This ensures model outputs remain logically consistent with critical features such as latent relevance, recency, or quality scores. We further generalize this architecture by parameterizing segment wise slopes as learnable embeddings. This enables the model to adaptively capture context specific distortions, such as position based CTR bias through specialized isotonic profiles. Our approach utilizes a dual task formulation that decouples the recommendation objective into latent relevance estimation and bias aware calibration. A major contribution of this work is the ability to perform highly granular, customized calibration for arbitrary combinations of context features, a level of control difficult to achieve with traditional non parametric methods. We also extend this to Multi Task Learning environments with dedicated embeddings for distinct objectives. Extensive empirical evaluations on real world datasets and production AB tests demonstrate that the Isotonic Layer effectively mitigates systematic bias and enhances calibration fidelity, significantly outperforming production baselines in both predictive accuracy and ranking consistency.

arXiv:2603.06600v1 Announce Type: cross Abstract: Vision Language Models (VLMs) are prone to errors, and identifying where these errors occur is critical for ensuring the reliability and safety of AI systems. In this paper, we propose an approach that automatically generates questions designed to deliberately induce incorrect responses from VLMs, thereby revealing their vulnerabilities. The core of this approach lies in fuzz testing and reinforcement finetuning: we transform a single input query into a large set of diverse variants through vision and language fuzzing. Based on the fuzzing outcomes, the question generator is further instructed by adversarial reinforcement fine-tuning to produce increasingly challenging queries that trigger model failures. With this approach, we can consistently drive down a target VLM's answer accuracy -- for example, the accuracy of Qwen2.5-VL-32B on our generated questions drops from 86.58\% to 65.53\% in four RL iterations. Moreover, a fuzzing policy trained against a single target VLM transfers to multiple other VLMs, producing challenging queries that degrade their performance as well.

arXiv:2603.06603v1 Announce Type: cross Abstract: Data duplication during pretraining can degrade generalization and lead to memorization, motivating aggressive deduplication pipelines. However, at web scale, it is unclear what constitutes a ``duplicate'': beyond surface-form matches, semantically equivalent documents (e.g. translations) may induce redundant training signals once models become sufficiently capable. Practically, this means that semantic duplicates operate increasingly like exact duplicates during training. We present evidence that duplication is scale-dependent in two ways. First, as model capability increases, cross-entropy loss gradients for semantically equivalent documents become more aligned. Smaller models, by contrast, produce gradients that reflect surface similarity (e.g., shared tokens) rather than semantic similarity. Second, we embedded all 192 million FineWeb-Edu-Dedup documents using EmbeddingGemma-300m. For moderate corpus sizes, the cosine similarity between nearest-neighbors follows an isotropic power law baseline. However, as corpus size grows to hundreds of billions of tokens, the nearest-neighbor similarities deviate sharply, indicating accelerated semantic collisions. Finally, controlled pretraining on data sampled with replacement from pools of finite unique documents shows that limited uniqueness yields mild degradation for small models, but rapidly increasing loss penalties for larger models, breaking naive scaling extrapolation. We derive explicit scaling laws that allow practitioners to estimate deviation from expected scaling due to limited semantic uniqueness of the pretraining corpus. Our results identify and resolve an unstudied source of scale-dependence, allowing for more accurate prediction at scale.

arXiv:2603.06607v1 Announce Type: cross Abstract: Multi-agent deep reinforcement learning (DRL) has emerged as a promising approach for radio resource allocation (RRA) in cellular vehicle-to-everything (C-V2X) networks. However, the multifaceted challenges inherent to multi-agent reinforcement learning (MARL) - including non-stationarity, coordination difficulty, large action spaces, partial observability, and limited robustness and generalization - are often intertwined, making it difficult to understand their individual impact on performance in vehicular environments. Moreover, existing studies typically rely on different baseline MARL algorithms, and a systematic comparison of their capabilities in addressing specific challenges in C-V2X RRA remains lacking. In this paper, we bridge this gap by formulating C-V2X RRA as a sequence of multi-agent interference games with progressively increasing complexity, each designed to isolate a key MARL challenge. Based on these formulations, we construct a suite of learning tasks that enable controlled evaluation of performance degradation attributable to each challenge. We further develop large-scale, diverse training and testing datasets using SUMO-generated highway traces to capture a wide range of vehicular topologies and corresponding interference patterns. Through extensive benchmarking of representative MARL algorithms, we identify policy robustness and generalization across diverse vehicular topologies as the dominant challenge in C-V2X RRA. We further show that, on the most challenging task, the best-performing actor-critic method outperforms the value-based approach by 42%. By emphasizing the need for zero-shot policy transfer to both seen and unseen topologies at runtime, and by open-sourcing the code, datasets, and interference-game benchmark suite, this work provides a systematic and reproducible foundation for evaluating and advancing MARL algorithms in vehicular networks.

arXiv:2603.06612v1 Announce Type: cross Abstract: Pass@k and other methods of scaling inference compute can improve language model performance in domains with external verifiers, including mathematics and code, where incorrect candidates can be filtered reliably. This raises a natural question: can we similarly scale compute to elicit gains in truthfulness for domains without convenient verification? We show that across five benchmarks and models, surprisingly, it cannot. Even at 25x the inference cost of naive sampling, polling-style aggregation yields no consistent accuracy gains over single-sample baselines and often amplifies shared misconceptions. We find that under uncertainty, models are better at predicting what other models will say within model ensembles than at identifying what is true, revealing a separation between social prediction and truth verification. Across models and benchmarks, aggregation fails to provide a robust truth signal because language model errors are strongly correlated. The source of correlation goes beyond any individual benchmark: we show that even when conditioned on out of distribution random strings and asked to produce pseudo-random outputs, different models produce correlated outputs. Confidence-based weighting provides no benefit because self-reported confidence fails to reliably distinguish correct from incorrect answers. These results delineate a boundary for inference-time scaling: in verified domains, additional samples provide more candidates for a verifier to filter; in unverified domains, additional samples merely reinforce shared misconceptions.