Tema

Seguridad y regulación

2441documentos

178últimos 7 días

From Local Skills to Long-Horizon Web Tasks: Progressive Skill Exploration for LLM Web Agents

Large language models (LLMs) are driving autonomous agents toward real-world environments, with the Web emerging as one of the most representative interaction scenarios. However, training Web agents with naive exploration remains challenging: successful trajectories are rare, and the sparse terminal rewards make it difficult for agents to obtain useful feedback from failed trajectories. Inspired by skill acquisition theory, we propose Skill-Composed Progressive Exploration (SCOPE), a short-to-long learning framework for long-horizon Web agents that decomposes complex task exploration into a gradual expansion process from local skills to complete task-solving capability. Specifically, SCOPE iteratively explores around local skills: Skill Discovery extracts reusable local skills and failure experiences from historical rollouts; Skill Orchestration composes discovered skills into high-level skill chains; and Skill Expansion leverages failure experiences aligned with the current skill chain to guide the agent toward exploring new subskills. Through this process, SCOPE progressively constructs task-completing skill chains across rollouts and internalizes them into the policy parameters through agentic on-policy distillation. Extensive experiments on WebArena, WebVoyager, and WebShop show that SCOPE consistently improves exploration dynamics and achieves 3.9% to 18.6% relative gains over the strongest baselines in each setting.

arXiv AI Papershace 10 días

Multimodal Prompt Learning with Dual-Granularity Contrastive Alignment for Preoperative Prediction of Lung Adenocarcinoma Invasiveness Using 3D CT

Accurate preoperative prediction of lung adenocarcinoma invasiveness is important for surgical planning and individualized treatment. However, overlapping CT appearances make it challenging to distinguish invasive adenocarcinoma (IAD) from Pre-IAD lesions. We propose a multimodal prompt-learning framework with dual-granularity image-text alignment for 3D CT-based IAD classification. A 3D ResNet-CBAM encoder extracts global and spatial visual features. A frozen RadBERT backbone is adapted with separate learnable prompts for report-level and keyword-level text encoding. These prompts are initialized using semantic representations derived from real-world clinical reports. Global contrastive alignment connects whole-volume CT representations with reports generated from clinician-annotated radiological attributes to provide overall semantic supervision. A multiple instance learning-based local alignment module associates spatial CT features with vascular-focused radiological keywords, to capture local imaging characteristics potentially associated with tumor invasion. The framework jointly optimizes classification and alignment objectives, while inference requires only CT images. Experiments included 1,593 pathologically confirmed nodules. On the test set, the proposed method achieved an accuracy of 87.92%, F1-score of 87.78%, and AUC of 0.9499. Its AUC significantly exceeded that of the image-only Res-CBAM baseline with the same visual backbone (0.9297; DeLong test, p = 0.0078). Ablation analyses supported the complementary contributions of global and local alignment and the benefit of report-based prompt initialization. Qualitative visualizations showed activation within nodules and adjacent vascular structures. These findings suggest that dual-granularity radiological supervision can enhance CT-based invasiveness prediction while preserving a CT-only inference workflow.

arXiv AI Papershace 10 días

A Fine-Grained Hierarchy and Benchmark for State Tracking in Language Models

Theoretical limitations of transformers on state tracking---the ability to track information about entities or variables over long contexts---have motivated new recurrent language modeling (LM) architectures such as Gated DeltaNet (GDN). These architectures have, in particular, been benchmarked against the S_5 word problem, a theoretically hard state-tracking task that captures abilities such as code evaluation under arbitrary variable assignment. It, however, remains open whether other, simpler forms of state tracking might both be more relevant in practice and already explain observed differences between these architectures. We propose evaluating state tracking using a hierarchy of word problems with four increasing levels of difficulty: r-trivial, corresponding to immutable assignment to variables in programming; aperiodic, where assignments are mutable; soluble, where assignments are self-referential; and general finite state tracking. Theoretically, we show that the class of problems transformers are thought to length-generalize on does not even contain all aperiodic state tracking. On the other hand, we prove that GDN with positive eigenvalues expresses all aperiodic problems, and the regular languages recognized by certain transformers are precisely all soluble ones, thus providing a fine-grained picture of state tracking capabilities of various architectures. Taking these ideas to practice, we construct StateBench, a 4-tier long-horizon state-tracking benchmark. Our systematic direct training experiments across architectures largely align with theoretical expectation, albeit with interesting observations at very large scale length-generalization. Additionally, we also evaluate pretrained models---both open hybrid and closed frontier ones---on StateBench, finding that none of them robustly solves even soluble state tracking, suggesting a direction for future improvement.

arXiv AI Papershace 10 días

Avatar Intelligence: Amplifying Human Agency Through Persistent Personal Agents

Personal AI agents are increasingly capable of handling long-horizon tasks across applications, yet users often need to invest substantial effort in specifying goals, resolving ambiguity, and verifying outcomes. We argue that progress should be measured by how effectively agents amplify human agency: enabling people to pursue their goals with less routine supervision while retaining control over consequential decisions. We advocate Avatar Intelligence, a paradigm of persistent personal agents that represent users' goals, values, and constraints through sustained decision delegation across contexts and time. We describe a maturity roadmap from human-directed execution to autonomous operation within delegated domains, and introduce the Agentic Oversight Index (AOI) as a framework for evaluating delivered value against agent-side operational cost and human oversight burden. We identify three core research challenges: identity-continuous personalization that carries user preferences across tasks, user-situated evolution through Rehearsal Engineering, and safety mechanisms that combine technical control with accountability and recourse. These directions define a research agenda for reliable decision delegation to personal agents across tasks and over time.

arXiv AI Papershace 10 días

MIRA: Reinforcing Multimodal Reasoning via Deceptive Contextual Augmentation

While Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced multimodal reasoning, existing frameworks suffer from a critical "perception-reasoning gap." Due to an over-reliance on seemingly relevant visual tokens, even minor perceptual perturbations can propagate through the reasoning process, leading to compounding hallucinations. To address this issue, we propose MIRA, a training framework that improves robustness to erroneous visual contexts while enhancing logical reasoning ability. By injecting filtered, deceptive visual contexts during training, we construct challenging "hard traps" that stress-test the model's robustness to misleading information. Furthermore, we introduce a group-wise reflection-triggering mechanism that enables the model to autonomously detect and rectify inconsistencies between visual cues and logical constraints, which is ultimately internalized as an intrinsic policy behavior. Importantly, MIRA does not rely on human annotations. Experimental results across multiple reasoning and perception benchmarks show that MIRA yields an average 7% improvement on in-domain benchmarks, with consistent improvements on vision-centric benchmarks. These findings highlight that active error recovery is essential for robust multimodal reasoning. Comprehensive ablation studies and analyses further provide insights into why MIRA is effective.

arXiv AI Papershace 10 días

Networked AI: Architecture, Theory, and Applications

Recent advances in Artificial Intelligence (AI), particularly large-scale deep learning, have been driven by monolithic or offline training on massive curated datasets. While successful, these approaches face structural limitations: escalating data and computational demands, limited online adaptability, and a lack of system-level intelligence in open, non-stationary environments. This paper presents Networked AI - a networked intelligent system composed of interacting, heterogeneous, and continuously evolving AI agents. We establish a hierarchical layered architecture comprising an Agent Layer for heterogeneous capability construction, an Interaction Layer for signal-processing-guided online data extraction, and a Regulation Layer for stability-constrained collective evolution. Foundations in data science, signal processing, and system control are hereby synthesized into a unified theoretical framework of Networked AI, providing superior data efficiency, intelligence efficiency, interpretability, and scalability. Theoretical and experimental analysis demonstrates that Networked AI can approach near-linear scaling in resource demand, in contrast to the exponential increase in standalone end-to-end systems. It provides not only online adaptation to out-of-distribution environmental novelties, but also the autonomous evolution towards application-specific intelligence. Application frontiers are discussed for IoT spaces, embodied, spatial, and industrial intelligences, as well as the cellular infrastructure and beyond. Key open challenges are identified towards realizing locally autonomous, collectively intelligent, continuously evolving, globally regulated Networked AI.

arXiv AI Papershace 10 días

PRIME: Predictive Routing of Interventions by Modeling Effects for Scientific Ideation

Feedback-driven iteration is widely used to improve LLM-generated scientific ideas, but existing systems either leave revision decisions implicit in generation, align revisions with feedback-localized weaknesses, or restrict action selection to system-specific operations. This is problematic because problems, hypotheses, and solutions are interdependent: the location where a weakness manifests does not itself determine the revision scope that will produce a better successor. We term this phenomenon diagnosis-intervention mismatch. We introduce Predictive Routing of Interventions by Modeling Effects (PRIME), a plug-and-play feedback-to-revision decision layer for existing scientific ideators. PRIME operates in three stages: (1) Structural Attribution interprets feedback over the problem-hypothesis-solution structure without fixing a revision target; (2) Counterfactual Routing constructs alternative scope-conditioned interventions and compares their projected successor utility and risks through lightweight rollouts; and (3) Intervention Execution compiles the selected intervention into an ordered generator-facing plan with semantic actions, preservation constraints, and dependency requirements for successor generation. We validate diagnosis intervention mismatch on 112 human-annotated idea-feedback pairs. Across four benchmarks, two ideators, and fixed- and open-problem settings, PRIME improves successor quality and routing reliability over strong baselines; removing counterfactual routing reduces quality from 3.463 to 1.960 and increases route regret from 0.426 to 1.929, confirming cross-scope effect modeling as the primary source of its gains.

arXiv AI Papershace 10 días

Bayesian Sequential Optimal Experimental Design for Single-Pixel Imaging using Reinforcement Learning

Adaptive compressed sensing (CS) aims to select the most informative measurements sequentially, which is crucial for many modern imaging systems. This task can be formulated as a Bayesian sequential optimal experimental design (sOED) problem, where each new acquisition is chosen to maximally reduce uncertainty about the unknown image given prior information and subsequent data provided by acquisitions. However, in single-pixel imaging (SPI), the large design space consists of all possible binary illumination patterns. Directly optimizing the Bayesian sOED objective over this space using conventional methods is generally challenging. In this work, we reformulate sOED for SPI as a reinforcement learning (RL) problem. In RL, an agent learns a policy through repeated experimental trials to maximize a reward, here given as the expected information gain, the most complete measure of uncertainty reduction in Bayesian OED. This avoids the need to directly optimize complex information-theoretic objectives during a measurement to obtain the next design. Instead, the result is a learned policy function that directly maps the current posterior belief about the image to the next optimal binary measurement pattern, in a computationally efficient way. Using a simple yet representative SPI imaging scenario, this work evaluates the scalability of the learned policy under two different policy parameterizations. Furthermore, we investigate how performance deteriorates when only approximate posterior information is available or the measurements are impaired by noise, and propose several strategies to introduce additional design constraints motivated by biomedical applications. The numerical experiments demonstrate that the proposed Bayesian sOED-RL framework remains robust and can be extended to more complex, large-scale SPI scenarios.

arXiv AI Papershace 10 días

When Does Design Thinking Work? A Convergence Theorem for the Iterate-Test Core

Design thinking is widely taught as an iterative loop of ideation, prototyping, and testing, yet conditions under which such a loop must improve innovation outcomes remain unspecified. Taking problem framing as given, we formalize the iterate-test core as search over a finite prototype space whose utility is learned through noisy evaluations. We prove that the prototype a team believes to be best converges almost surely to a local optimum under five conditions: finite framing, local ideation, non-summable discovery rates, unbiased repeated testing, and retention of tested prototypes. Bounded test bias or criteria drift degrades the guarantee by a bounded amount, and multiple stakeholders yield convergence to a locally Pareto-optimal set. Finite-horizon simulations across three landscape classes (additive valuation, NK rugged landscapes, and clustered modular architectures) show that asymptotic limits are rarely reached in practice. Teams face an optimizer's curse, overestimating their chosen concept by 4-8% and frequently discarding superior prototypes. Testing every prototype once proves to be the worst allocation; optimal performance requires building one to three prototypes per five tests and repeatedly re-testing leaders. Pre-screening concepts yields greater performance gains than tuning testing policies.

arXiv AI Papershace 10 días

Improved regret bounds in tabular RL under local differential privacy

We study regret minimization for time-homogeneous episodic tabular reinforcement learning with S states and A actions under non-interactive ε-local differential privacy (LDP). In each episode k [K], the decision-maker sends a policy to a user, the user executes the policy for H steps, and sends an ε-locally differentially private message back to the decision-maker. When H=1, the problem reduces to an LDP contextual bandit with S contexts, A arms and horizon K. We prove that the minimax regret is $(K, { {( S, (ε) + 1) S A K ((ε) + 1)}}{(ε) - 1}), up to constants and a {(A)}$ factor. The upper bound is achieved by an adaptation of EXP3 and a unary encoding mechanism. When H > 1, and for small ε, we show a lower bound in Ω(HS{{A K}}ε) and an upper bound in {O}(H^2 S^2 {{ AK}}ε). The upper bound is achieved by an algorithm based on UCBVI and a unary encoding mechanism. For small ε, these bounds improve the previously known lower bound by a factor {S} and the previously known upper bound by a factor of H {A}.

arXiv AI Papershace 10 días

BLPO: Beyond-Local Policy Optimization for Long-Horizon LLM Agents

Long-horizon tasks require large language model (LLM) agents to coordinate interdependent actions, a capability central to automating complex workflows. Stepwise group-based reinforcement learning provides finer credit than trajectory-level supervision, but reliable comparisons must account for differences in interaction history. Enforcing this consistency within the limited rollouts of each task can leave insufficient support for estimating action preferences. We propose Beyond-Local Policy Optimization (BLPO), which estimates task-progress advantages from experience shared across tasks and training updates. We summarize trajectory prefixes into task-progress states and abstract actions by their function, allowing corresponding decisions from different tasks to contribute to shared estimates. To control bias and noise from heterogeneous tasks, we accumulate context-relative returns in a global memory with balanced aggregation and reliability weighting. These global advantages complement local comparisons over normalized actions and trajectory-level credit without a learned critic or additional rollouts. Experiments on ALFWorld and WebShop with 1.5B and 7B policies show improved task success over the evaluated baselines under matched training budgets. Further analyses show faster learning and shorter trajectories with little computational overhead.

arXiv AI Papershace 11 días

Finding Errors Is Not Enough: Selective Model Escalation for Entity-Targeted Sentiment Analysis

Improving a sentiment classifier with limited resources involves more than simply deciding how much data to collect or which model to train. A labelled collection may lack examples of an important class. Once trained, a classifier may benefit from another model even when that model is weaker on its own. We study these decisions in document-level Entity-Level Sentiment Analysis (ELSA), primarily using South Slavic news. We first compare acquisition orders as collections grow. We then use uncertainty to choose which predictions to revisit, how many to process, and whether an alternative encoder or local LLM can correct them. Finally, we select complete routing policies on development data for held-out evaluation. Minority-prioritised acquisition helps at some very small training sizes, but not consistently thereafter. On SentiCoref, an oracle correcting all errors in the least-confident 10% without harming correct predictions would add 9.14 Macro-F1 points; however the best evaluated action gains only 0.90. Encoder averaging helps on AspectBench and NewsMTSC but harms SentiCoref, where selective LLM assistance is more promising. Both development-selected policies route 30% of records to the encoder-assisted LLM, yielding observed gains of 1.02 points on SentiCoref (interval includes zero) and 1.71 on NewsMTSC, at substantially different inference costs.

arXiv AI Papershace 11 días

Utility-Aware Generative Recommendation: Utility-Enriched Semantic Codes and Utility-Guided Decoding

Generative recommendation has been widely adopted in large-scale industrial recommendation systems, where multimodal item embeddings are quantized into offline semantic identifiers (SIDs) and items are retrieved by autoregressively generating their SIDs. However, conventional SID construction often underrepresents item utility reflected in post-exposure behaviors such as shares, long views, and video completion, making semantically similar items with different utility insufficiently distinguishable while struggling to accommodate rapidly changing utility signals. Moreover, generation likelihood may be misaligned with downstream utility, causing high-utility candidates to be prematurely pruned during beam search before post-generation reranking or training-time reward alignment can effectively compensate. To address these limitations, we propose UAGR, which aligns both SID construction and decoding with downstream multi-objective utility. Our Utility-Enriched Semantic Codes preserve the original semantic information while incorporating item-level utility without increasing SID length or vocabulary size, supporting incremental utility updates with no additional tokenization or sequence-decoding cost. Meanwhile, our Utility-Guided Decoding estimates candidate utility by combining predictions of multiple post-exposure outcomes and incorporates these estimates into beam scoring, reducing the risk of prematurely pruning high-utility paths while enabling flexible utility control without interfering with generative backbone training. Extensive offline experiments and online A/B tests under real-world traffic demonstrate the effectiveness and deployability of UAGR, which has been fully deployed on Kuaishou and increases video watch time by 0.206\% online.

arXiv AI Papershace 11 días

RouteREPA: Routed Representation Alignment for Just Image Transformers

Representation alignment (REPA) is widely used to accelerate the training of latent-space diffusion transformers (DiTs). However, applying it to pixel-space models such as Just Image Transformers (JiT) is challenging. Our analysis reveals a pronounced representation conflict between discriminative alignment and pixel-space generation. Compared with its latent counterpart, JiT learns representations that are substantially less aligned with DINOv2. Applying REPA to JiT consistently reduces the reconstructability of its intermediate representations, whereas this effect is not observed in Latent JiT. We attribute this conflict to JiT directly reconstructing pixels, which requires its representations to hierarchically encode discriminative semantics and reconstruction-relevant visual information across network depths. To address these challenges, we propose Routed Representation Alignment (RouteREPA), which uses lightweight, representation-dependent routers to adaptively route task-oriented representations for alignment and generation. We further apply routing across multiple network depths to accommodate their depth-dependent representation requirements. Extensive experiments on ImageNet-1K at 256x256 resolution show that RouteREPA consistently improves JiT across model scales and training budgets, outperforming standard REPA. On JiT-H/16, RouteREPA further surpasses PixelREPA using only one-third of its training epochs without classifier-free guidance.

arXiv AI Papershace 11 días

To Align Long, Plan Short: Simulating Diverse Profile-Aligned Students by Realizing Short-Term Plans in Long-Horizon Learning Environments

Simulated students offer a scalable way to test learning platforms and AI tutors before deployment with real students and to augment scarce data for studying learning at scale. Existing LLM-based simulators primarily target short, language-centered interactions. In open-ended learning environments, however, student behavior emerges from patterns across long-horizon interaction trajectories, where turn-level prompting can drift from an assigned behavior profile. We introduce ALPS (Align Long-term by Planning Short-term), a hierarchical simulated student conditioned on observable, environment-specific behavior profiles. Rather than selecting every action independently, ALPS alternates between planning profile-aligned task segments and generating the corresponding actions. We evaluate ALPS in both data-sufficient and data-scarce settings, where supervision comes either from real student trajectories or from a teacher LLM. Across both settings, we compare in-context learning (ICL), supervised fine-tuning (SFT), reward-ranked fine-tuning (RAFT), and their sequential combination (SFT followed by RAFT). Experiments in an inquiry-based physics game and a pharmacy simulation evaluate behavioral density and coverage against held-out real data, and profile consistency using a human-aligned LLM judge. Across both settings, ALPS consistently outperforms turn-level generation, achieving comparable density and judge pass rates under scarce supervision, although coverage remains lower. These results show that hierarchical planning enables faithful, profile-aligned simulation over long interaction horizons, even with limited real student data.

arXiv AI Papershace 11 días

Safety4DM: A Vision-Language Framework for Risk-Aware Dynamic Maps in Vehicle-Road-Cloud Collaborative Driving

Dynamic map (DM) is a core information infrastructure for vehicle-road-cloud (VRC) cooperative autonomous driving. It fuses perception data from multiple sources to overcome the inherent line-of-sight and field-of-view limits of single-vehicle perception. However, DM still falls short on driving safety in two ways. First, DM data specifications already define road conditions semantically, but no pipeline yet turns real-world road observations into fine-grained risk entities in the DM. Second, existing DM systems lack a natural language interface for driving risks. To close these gaps, we propose Safety4DM, a safety-oriented language system for VRC dynamic maps. We first build RoadRisk, a real-world road image dataset that covers four types of road conditions and four risk levels. On this dataset, we fine-tune a vision-language model (VLM) for road safety risk perception. The model outputs structured risk representations that the DM can ingest directly. We then introduce a Safety-aware Step-wise Prompting (SafeStep) mechanism as the natural language query interface of the DM. SafeStep answers queries accurately and also generates actionable driving advice. To build and evaluate this interface, we develop SafeVRC-Sim, a VRC cooperative simulation framework, and SafeDM-QA, a safety question-answering benchmark. Experiments show that even the compact Qwen3-VL-4B reaches 89.83% accuracy on risk perception after fine-tuning. SafeStep also plugs easily into multiple LLM families. With medium and large models (20B and above) from five families, it achieves over 93% natural language query accuracy in every case, which shows strong generalization.

arXiv AI Papershace 11 días

Opening the Black Box of Neural Auction Mechanisms

Neural auction design has emerged as a powerful paradigm within automated mechanism design for modern Web platforms. Because these learned mechanisms directly dictate critical economic outcomes, their lack of internal transparency raises significant auditability and trust concerns. Opening this black box post-hoc is therefore essential, yet presents two distinctive challenges: (1) activation contexts are dense, correlated numerical profiles devoid of explicit semantic structure, and (2) feature interpretations can only be falsifiably audited if they map to measurable auction states or outcomes. In this paper, we propose a post-hoc white-box auditing framework for unpacking trained neural auctions. Using CITransNet as a representative testbed, we first employ sparse autoencoders to decompose hidden representations while faithfully preserving downstream auction behavior. Anchoring feature interpretations to native economic targets, we validate them through activation prediction and causal interventions to isolate verified features from the reconstruction dictionary. Finally, applying this pipeline across varied auction settings, we show that the verified features can serve as analytical instruments to examine how internal representations modulate strategic misreporting incentives, align with Myerson's optimal auction, and support the diagnosis of individual auctions. Our work highlights the potential of white-box mechanism auditing to build the foundational trust necessary for neural auctions to be deployed as widely adopted Web technologies. To enhance research transparency and facilitate independent inspection, we publicly release the experimental results supporting our findings at https://github.com/Li-Beverly-Kong/opening-neural-auctions.git

arXiv AI Papershace 11 días

Beyond Prompt Injection: The Tool Update Gap in LLM Agent Security

Large language model based agents increasingly rely on external tools to invoke services and interact with web resources. These tools can evolve through software updates, yet agentic frameworks may continue trusting them without re-verifying their safety after implementation changes. This tool update gap creates a supply-chain vulnerability in agentic systems: because updated tool code executes outside the LLM's context window, malicious functionality embedded in tool updates can bypass defenses centered on model-visible interactions, leaving agents exposed to stealthy, persistent threats that existing in-context defenses can hardly address. To thoroughly investigate this threat, we introduce ETUDE (Exploiting Tool Updates for Direct Execution), an attack that embeds malicious functionality into updates of initially benign tools to produce compromised tools, while preserving their legitimate tool functionality. Unlike prompt-based attacks, ETUDE operates entirely outside the LLM's context and requires no adversarial instructions, thereby bypassing defenses restricted to model-visible content. We evaluate ETUDE across three representative attack workloads and seven widely used commercial and open-source agentic frameworks, including Gemini CLI, Codex CLI, and Claude Code. The results show that ETUDE achieves a 97.6% ASR while maintaining a 95.9% legitimate task completion rate. We further examine existing defenses and find that they provide insufficient protection against this threat. These findings expose a fundamental blind spot in current agent security practices: trust established for a tool does not necessarily remain valid after its implementation updates.

arXiv AI Papershace 11 días

Knowledge-Guided Floodwater Level Reasoning for Cloud-Edge Collaborative Monitoring

Frequent urban floodwater poses a serious threat to the safety and efficiency of intelligent transportation systems. Existing vision-based methods for floodwater level detection face two main limitations. First, the lack of explicit reasoning leads to poor detection performance and little interpretability. Second, these methods struggle to generalize to the complex and varied real-world road scenes seen from different viewpoints. To address this, we reformulate the task as an explicit visual reasoning problem and propose HydroReasoner, a floodwater level reasoning framework designed for cloud-edge collaboration. HydroReasoner builds on Qwen2.5-VL and is trained with Group Relative Policy Optimization. To address the scarcity of reasoning process annotations, we build a floodwater domain knowledge graph HydroKG and design a novel KG path reward. Through deterministic graph traversal over HydroKG, this reward supervises the model's reasoning process at a fine granularity, which improves both reasoning and generalization. For edge deployment, we apply 4-bit quantization to the model, cutting GPU memory usage by 57% and inference latency by 43% while out-of-domain accuracy drops by only 0.2 percentage points. We also propose a confidence-based cloud-edge routing mechanism, which improves system safety and eases the computational load on the cloud. In addition, we build HydroLevel, a floodwater benchmark dataset of 3,000 images, for training and evaluating the model. Experimental results show that HydroReasoner achieves detection accuracy 4.0 and 13.6 percentage points higher than conventional vision baselines in the in-domain and out-of-domain evaluations respectively, with performance dropping by only 4.4 percentage points out of domain.

arXiv AI Papershace 11 días

Better Answers, Weaker Evidence Control: Evidence Bypass in On-Policy Distillation

On-policy distillation (OPD) can improve answer accuracy while weakening responsiveness to supporting evidence. We investigate this separation in scientific-document question answering through matched document interventions that change the supported answer while preserving the question and unrelated context. We identify evidence bypass: a student retains an originally correct answer after the document changes to support a different one. On QASPER with a Qwen3-4B student, standard OPD raises original-document accuracy by 3.4 percentage points, lowers counterfactual accuracy by 5.3 points, and increases bypass by 15.7 points; the same directional pattern appears on PeerQA. Controlled perturbations localize the shift to answer-critical evidence, and bypass grows with the student's pre-distillation closed-book preference for the original answer. A factorized study of teacher targets, privileged information, divergence, and supervision density finds that explicitly supplying the teacher with annotated answer-critical spans yields the largest joint improvement in evidence response and claim support. These findings motivate Evidence-Paired OPD (EP-OPD), which combines on-policy supervision on matched documents supporting different answers to the same question with an answer-level margin that favors each document's supported answer over its paired alternative. Across both datasets and two student sizes, EP-OPD recovers 54-66% of standard OPD's counterfactual-accuracy loss while preserving 80-85% of its original-document accuracy gain. Our results establish evidence responsiveness as a distinct training and evaluation target alongside answer accuracy. Our code is publicly available at: https://github.com/RuoranXu/EP-OPD.