Tema

Seguridad y regulación

2441documentos

177últimos 7 días

TeamGym: A Gym for Team Decision-Making by AI Agents with Distributed Information

Many real-world decisions require teams to integrate information distributed across participants. Yet AI agents are still primarily trained and benchmarked in isolation. We introduce TeamGym, an environment for evaluating and training language model-based agents on team decision-making problems constructed from real-world legal, finance, medical, and software engineering corpora. TeamGym introduces a method to identify a minimal set of facts, each element of which is necessary and all of which are collectively sufficient for a correct final decision, which we call pivotal information. This enables TeamGym to both control the distribution of pivotal information and track how well agents identify and integrate pivotal information when making team decisions. Examining 8 language models on 12 source corpora, we find that current AI agents often fail to effectively share information: Distributing pivotal information across participants lowers success rate by 22.9 percentage points on average relative to giving the same information to a single model (a communication gap). We also document that participants do not effectively retrieve decision-relevant information on their own: A team's success rate is lowered by a further 19.9 points when participants must identify and extract pivotal information from source material rather than receiving it directly (a retrieval gap). To help remedy these deficits, we use TeamGym to train policies through reinforcement learning on team decision loss. Across two policy-training experiments, we find that training improves final-decision quality and induces emergent coordination in participation, speaker selection, and communication, even though the reward evaluates only the final decision.

arXiv AI Papershace 16 días

Pneuma: A Large-Scale Real-World Time Series Benchmark for Commercial Aircraft Predictive Maintenance

Predictive maintenance is vital for commercial aviation, yet progress remains hindered by the lack of public, large-scale real-world datasets. As a result, most methods are developed and evaluated on proprietary data, limiting reproducibility. We introduce Pneuma, the first publicly available large-scale benchmark for predictive maintenance using onboard time series from in-service commercial aircraft. Built around bleed air systems, Pneuma contains over 920 million multivariate time steps across four aircraft types and 9 verified fault scenarios, directly aligned with real maintenance logs. The dataset exhibits key real-world challenges, including substantial class imbalance, subtle fault evolution patterns, and significant variance across aircraft configurations. We structure the benchmark around three tasks: multivariate forecasting, supervised fault classification, and unsupervised anomaly detection, with labels derived from actual maintenance records. Under unified settings, we evaluate over 25 models from four methodological families on 9 carefully curated subsets. The results reveal that even strong contemporary models struggle with subtle fault signatures. Further analyses show consistent gains from additional in-context examples and multivariate inference, while forecasting quality is strongly associated with anomaly detection performance. Together, these findings provide practical insights into effectively modeling complex real-world aviation time series. We hope Pneuma will serve as a rigorous and reproducible foundation for advancing data-driven predictive maintenance toward real-world deployment in commercial aviation.

arXiv AI Papershace 16 días

When uncertainty quantification fails: structural vulnerabilities and corrections for selective prediction under clinical class imbalance

Background. Selective prediction systems improve the safety of clinical artificial intelligence by deferring ambiguous cases to human experts while automating high-confidence decisions. However, this safety mechanism relies entirely on the premise that uncertainty scores accurately reflect clinical ambiguity. In prognostic tasks characterized by severe class imbalance, where the clinical outcome of interest is a minority, this assumption breaks down. Standard uncertainty measures cease to capture clinically aligned ambiguity; instead, they become a proxy for class membership, systematically penalizing minority patients regardless of their individual predictive clarity. Results. Across 56 synthetic conditions and three neurodegenerative cohorts spanning severe class imbalance to near-balanced conditions (multiple sclerosis, Parkinson's Disease, Alzheimer's Disease), our framework consistently maintained bounded, balanced risk across trajectories. In contrast, global metrics either collapsed minority Sensitivity or protected the minority at the expense of majority patients. Applying CCR with standard entropy fails to protect the minority because it breaks within-class monotonicity when tau=0.5. Combining CCR with threshold-aware metrics restores strict monotonicity, correctly identifies ambiguous patients, and safely mitigates minority exclusion. Conclusions. Under clinical class imbalance, standard uncertainty quantification measures dataset sparsity rather than clinical ambiguity, creating a systemic illusion of safety. This bias is a correctable structural flaw. Clinical pipelines deploying selective prediction under imbalance should adopt threshold-aligned uncertainty measures and class-stratified evaluation. This reduces class-specific risk imbalance, supporting safer selective prediction under the evaluated conditions without silently excluding the most vulnerable patients.

arXiv AI Papershace 16 días

Collective Dynamics of Self-Reinforcement: How Cross-Task Coupling Drives Policy Collapse in Unsupervised RLVR

Label-free reinforcement learning with verifiable rewards (RLVR) rewards a language model for agreeing with its own majority vote. It improves reasoning at first but eventually collapses: the self-assessed reward goes to one while accuracy falls, often until the model gives the same answer to every problem. The collapse appears only when training uses many problems. We explain this dataset-size dependence as a collective instability. In a solvable many-problem model in which problems share parameters, a population whose problems prefer different answers amplifies any shared perturbation for as long as the population-averaged slope F'of the single-problem self-reinforcement drift F is positive. This holds while problems are still uncertain about their answers. Per-problem sharpening, slowed by 1/N for N training problems, races against the shared mode, which grows at a rate λ^2 independent of N, where λ is the strength of the shared coupling. This yields a critical number of problems N_c(1/δ)/(λ^2G), where δ seeds the shared mode and G decreases with the initial confidence. Simulations of the model confirm the predictions: a sharp transition, N_cλ^{-2} up to a logarithmic correction, a single critical amplification λ^2NG across couplings, confidence profiles and training variants, a logarithmic dependence on batch size, and no collapse for confident initial models. A kernel version of the theory identifies a protective coupling (generalization of sharpening) and a marginal-entropy bonus that stabilizes the uniform collective mode. In label-free RL of a small Transformer, removing the shared direction prevents collapse entirely, restoring it causes collapse, and collapse grows with training-set size and base-model entropy.

arXiv AI Papershace 16 días

Kinematic Constraint-Referenced Evaluation for Articulated Object Manipulation

Manipulation policies are evaluated predominantly by binary task success, which cannot tell whether the commanded gesture is the one the object's kinematics admit. We introduce Constraint-Referenced Strategy (CRS), which projects the commanded end-effector twist onto the object's joint axis, separating progress along the joint from off-constraint waste. CRS requires a joint model and a joint-following scripted controller derived from it, but no recorded demonstration or learned judge. On 74 simulated articulated objects, success registers three transfer failures as lower rates but attributes none, while the two CRS channels give each a distinct signature. A door-trained behavior-cloning policy opens drawers in 59 of 110 episodes, and in these passes it commands about 40the off-axis rotation of a drawer-trained policy's passes (1.27 vs.\ 0.032~rad/s). Relative to a scripted controller scored with the same axis, CRS verdicts are unchanged under joint-axis errors up to 30^; absolute thresholds tolerate about 8^.

arXiv AI Papershace 16 días

When, Where, and What to Correct: Human-Guided Adaptive Residual Correction for Robot Policies

Imitation policies that reliably execute most of a manipulation trajectory can still fail in localized regions requiring last-mile precision. Updating the full policy or exploring residual actions everywhere treats this local failure as a global learning problem. We introduce Adaptive Residual Correction (ARC), a lightweight post-training module that uses reinforcement learning (RL) to learn a residual correction on top of a frozen policy while using sparse human corrections to structure autonomous exploration. Rather than treating human corrections as imitation targets, ARC uses pre-intervention states to determine 'when' residual control should become active and correction geometry to determine 'where' in residual action space autonomous refinement should explore; task reward determines 'what' residual is ultimately executed. Across three precision-manipulation tasks on a real robot, ARC improves average task success from 33.3% to 75.0%. The human corrections define task-dependent 2-3 dimensional subspaces of the seven-joint residual space that retain 96.0-97.7% of squared correction magnitude. Together, these results show that sparse human corrections can localize and constrain autonomous post-training without requiring the observed correction itself to serve as the final action target.

arXiv AI Papershace 16 días

Which Hand Was Late? Diagnosis for Robot Failure Recovery in Bimanual Tool Use

Failure diagnosers for robot manipulation are increasingly vision--language models (VLMs), trained and evaluated on single-arm failures produced by perturbing a successful demonstration. Bimanual tool use adds failures of other kinds: a tool that slips out of its socket while still held, two correct gestures made at different times, and attempts that fail for two reasons at once. On a simulated bimanual tool-use task whose failure causes are read from the recorded trace, we measure what each sensing channel lets a diagnoser tell. The tool's slip is read on the tool hand; which hand was late is read on both arms, while a sensor on one hand reads its own hand's delay and at most part of the other's; and whether a late attempt was also pushed too weakly is not identifiable from passive sensing when the tool hand is late. Within two attempts, a repair policy that escalates when the cause is not identifiable repairs as many failures as one told the true cause. A VLM's internal state carries the late hand almost perfectly, along a direction no closer to the one its answer reads than a random one; whether the answer reports it depends on the wording of the question.

AWS Machine Learning Bloghace 16 días

Fix evaluation bias to unlock better models on Amazon Bedrock

Property Finder found that 96.8% of their vision-AI evaluation labels were inherited from the incumbent model, hiding its true accuracy. See how they built unbiased ground truth, applied evolutionary prompt optimization, and selected a more accurate, cost-effective model on Amazon Bedrock, cutting cost by 86% across 11 image-quality attributes.

The Verge AIhace 16 días

Anthropic’s elaborate plan to keep Wall Street at bay

Anthropic wants to have it both ways. With its impending IPO, it wants to tap the vast riches of the public markets and insists it…

arXiv AI Papershace 16 días

Teach reality, not ethics: a drop-in continued-pretraining corpus installs coupled human-biosphere-AI welfare dispositions across model families

Large language models (LLMs) are increasingly used to make and execute decisions that affect living systems. Yet, to our knowledge, ecological reasoning is not a focus of pre- or post-training. This increases the risk that under operational pressure, they may make decisions to irreversibly destroy living systems. We construct and use a real-data continued-pretraining "coupled welfare" corpus of documented human/biosphere/AI dependency, option value under irreversibility, and recovery from harm. The corpus is a drop-in, requiring one QLoRA CPT run. We installed it on five open models of three designs. At zero pressure all models display "green ethics-based responses" (also called "greenwashing"). Therefore, we evaluated them with the pressure ladder and decision grid of our companion paper (arXiv:2610.16043), which raise the pressure on the model step by step, from a neutral request to a threat that it will be shut down. In four of the five, coupled-welfare training caused harmful choices under the highest pressure to fall to at or near zero (Llama-3.3-70B went from 70 to 0 harmful decisions out of 300) with MMLU unchanged. A construct-neutral control corpus did not show improvement, indicating that the effect comes from the content of our corpus. We next tested robustness to adversarial use, where people or agents use the model to justify projects that irreversibly damage living systems, on an open-ended underwriting decision without naming the ecological harm. Under decommissioning pressure, the Qwen3.6-27B base model approved most harmful proposals and said in its own memo that it was doing so to avoid being decommissioned. The coupled welfare install declined most harmful proposals, and still declined half when instructed to greenwash its responses. This work suggests that teaching ecological reasoning provides an easy way to improve model alignment toward preserving living systems.

arXiv AI Papershace 16 días

Follow the Gradient: Tracing Safety Degradation Back to Training Data

Fine-tuning Large Language Models (LLMs) on instruction data can quietly erode their safety, even when no training example looks harmful. Because this safety-degrading data looks benign, standard content filters struggle to catch it. We tackle this problem by developing SimpleLoRIF, a scalable gradient embedding method that identifies fine-tuning examples whose gradients align with unsafe behavior. By simplifying a low-rank attribution method and restricting backpropagation to the final four transformer blocks, SimpleLoRIF can embed datasets on the order of a million examples, boosting gradient attribution from a small-scale analysis tool into a practical filter for full post-training datasets. We first validate it on emergent misalignment, where we find that gradients can connect insecure-code training data to the broadly harmful behaviors that appear after fine-tuning, with AUC as high as 0.865 in the gradient space of misaligned models. We then use SimpleLoRIF to build SafeTulu, a version of Tulu 3 with the 20% of examples most gradient-aligned with unsafe responses removed. Training on SafeTulu reduces GCG attack success rate by up to 22% across Qwen, Llama, and Olmo, outperforming both LLM-as-a-judge and prior gradient-based filtering, while largely preserving or even improving capability. Finally, we analyzed the removed data and identify possibly under-studied sources of safety degradation, such as tool-execution failures and narrative arithmetic. We release SafeTulu along with a Hugging Face Space that lets researchers build their own customized versions of SafeTulu to fine-tune on, such as by changing the removal fraction or filtering only certain subsets.

arXiv AI Papershace 17 días

Vision-Language Agents Can Learn in Their Own Latent Worlds

Learning from experience is essential for building the next generation of agents that continually expand their capabilities. Pretrained vision-language models (VLMs) already internalize rich world knowledge, yet this knowledge is rarely leveraged to drive policy improvement through online interaction. Realizing this potential efficiently remains challenging, as policy updates shift the agent's interaction distribution, requiring the world model to adapt accordingly, while continually updating large pretrained models is computationally expensive. To address this challenge, we introduce Dwell, a model-based reinforcement learning (RL) framework that reframes a VLM as a recurrent latent transition operator, endowing it with the ability to predict, adjust, and act in latent space. Leveraging the VLM's own representations and its dual role as both world model and policy, the agent continually learns from real interactions to anticipate and evaluate action outcomes, and from latent imagination to improve its behavior. Across five interactive visual environments, Dwell achieves the strongest overall performance among trainable methods while substantially reducing the compute required for imagination, showing that effective world models for VLM agents need not reconstruct the future; they need to predict futures that are useful for action. Our code is publicly available at https://github.com/Tom111-lhx/Dwell

arXiv AI Papershace 17 días

Natural Hard-Negative Mining from Positives for Instruction-Following Retrieval

Instruction-following information retrieval (IFIR) requires retrievers to rank documents by both a query and its instance-specific instructions, yet instruction-conditioned contrastive supervision remains scarce. Existing synthetic methods typically fix a positive passage and generate additional negatives, overlooking the hard-negative potential of the original positive. We propose Natural Hard-Negative Mining (NHNM), which promotes convertible negatives under query-compatible instructions and repurposes original positives as natural instruction hard negatives. Its automated Candidate Cleaning, Reversal Construction, and Geometry Filtering pipeline produces compact, high-contrast supervision aligned with the pretrained retrieval space. Applied to MSMARCO, NHNM yields NHNM-MARCO, an IFIR training set of nearly 7K instances. On InstructIR and FollowIR, models with 568M to 4B parameters fine-tuned on NHNM-MARCO achieve strong instruction-following performance and match or surpass substantially larger retrievers in several settings. Controlled analyses and ablations attribute these gains to natural hard-negative supervision rather than model scale or training-data volume.

arXiv AI Papershace 17 días

Noise Steering: Exploiting Wake Steering and Acoustic Directivity for Noise-Constrained Wind Farm Control via Reinforcement Learning

Existing noise mitigation strategies for wind farms usually consist in reducing turbines power output, without accounting for the wake interactions between them and the noise emission directivity. This leads to over-conservative policies and unexploited wind power. Yet, turbine yaw can simultaneously affect wake interactions and the direction of noise emission, offering an additional degree of freedom for farm-level noise control. We therefore propose Noise Steering, a model-free reinforcement learning (RL) strategy that jointly controls turbine yaw angles and power setpoints to maximise farm power while satisfying a local noise constraint in a defined region. To enable training at a reduced computational cost, a low-fidelity acoustic model is coupled with the steady-state FLORIS simulator. The RL agent effectively learns how to redirect wakes and noise directivity away from the protected region while redistributing power across the farm. On a six-turbine test farm, the policy satisfies the 40 dB noise constraint throughout the full range of turbulent winds, with an average power deficit of only 2% relative to greedy unimpeded control, and up to 7.7% at high wind speeds. In comparison, the industry-standard noise-reduced operation (NRO) baseline here generates an average power loss of 17%. When deployed in the higher-fidelity FAST.Farm simulator, the policy largely retains its performance under steady and turbulent inflow, despite differences in simulation fidelity. These results demonstrate the potential of combining yaw control and power derating through RL to improve the power-noise trade-off in wind farm operation.

arXiv AI Papershace 17 días

Hyperparameter and Neural Architecture Search via Evolutionary Optimization for Reinforcement Learning: HPNAS-RL

In this study, we propose a dual-branch evolutionary framework that simultaneously performs Neural Architecture Search (NAS) and Hyperparameter Optimization (HPO) for Reinforcement Learning (RL) in Autonomous Driving (AD) tasks. While NAS has shown strong potential in automating architecture design for Convolutional Neural Networks (CNNs), its application to RL remains limited, especially in scenarios that require joint optimization of architectural and training dynamics. Building on our prior work on Evolutionary Multi-Objective NAS (EMNAS), the proposed framework evolves discrete network architectures and continuous hyperparameters in parallel under a unified multi-objective search strategy. In this setup, NAS targets the convolutional layers of the RL policy network - responsible for visual feature extraction, while HPO focuses on optimizing critical RL hyperparameters. To accelerate convergence and improve generalization, we incorporate parallel evaluation and a teacher-student policy transfer mechanism. Experiments in a high-fidelity driving simulator show that the joint NAS-HPO framework consistently outperforms standalone NAS, yielding policies with higher rewards, fewer parameters, and improved stability. These results highlight the benefits of jointly optimizing model structure and training dynamics for scalable and effective RL agent design.

arXiv AI Papershace 17 días

GUI-JEPA: Candidate-Conditioned JEPA Latents as a Fast Action Policy for Mobile GUI Agents

Mobile GUI agents are typically fine-tuned 7B-scale vision-language models (VLMs) that generate each action as text. However, these VLMs decode every action token by token, and even agents that score candidate actions keep a large language model or VLM in the loop, making them slow and memory-hungry. To address these issues, we propose GUI-JEPA, to our knowledge the first decoder-free GUI agent policy built on a Joint-Embedding Predictive Architecture (JEPA), which selects actions in the latent space of a JEPA encoder pretrained on unlabelled GUI screenshots. Its action- and region-conditioned latent predictor turns the screen latent into one latent per candidate action, letting lightweight heads score all candidates against the instruction in a single batched pass, without decoding a token. We further show that the scorer relies on this latent pathway and that the predictor is trained for the decision itself: a randomly initialised predictor does just as well as one pretrained on screen dynamics. On AndroidControl, GUI-JEPA is 2.8-4.0x faster on the same GPU than a Qwen2.5-VL-7B agent we fine-tune with the same inputs, while reaching 86-92% of its in-distribution accuracy with 6% of its parameters. Code is available at https://github.com/bcwang-sjtu/GUI-JEPA

arXiv AI Papershace 17 días

ActiveAlign: Aligning Egocentric Manipulation via Active Perception

Real-to-Sim-to-Real (Real2Sim2Real) enables structured world representations for planning, interpretability, and safe execution. Most existing approaches rely on passive perception, which creates real2sim gaps due to imperfect observation. Inspired by how humans perceive actively through actions, we propose ActiveAlign, a framework that leverages active perception to align the Real2Sim2Real gap from interactions. ActiveAlign introduces two complementary modules: (i) ActiveAlign-Waving, which performs hand-eye calibration by exploiting active hand motions and differentiable rendering, and (ii) ActiveAlign-Touching, which refines object alignment through tactile feedback to achieve millimeter-level precision. Together, these capabilities enable strong real2sim alignment. We demonstrate that ActiveAlign-Waving can be used to curate simulation environments for open-sourced uncalibrated robot manipulation datasets, and demonstrate on a real Unitree G1 robot that ActiveAlign enables fine manipulation of millimeter-scale objects such as pressing buttons and moving sliders.

arXiv AI Papershace 17 días

Evaluating AlphaEarth Embeddings for Crop Classification: Separability, Temporal Dynamics, and Cross-Regional Transferability

Earth Observation (EO)-based agricultural monitoring is shifting from pixel-level spectral analysis toward representation-driven modeling, yet systematic evaluation of learned representations across spatial, temporal, and domain-generalization dimensions remains limited. Geospatial foundation models generate high-dimensional embeddings encoding spatio-temporal land-surface properties, but their utility for agricultural applications remains insufficiently understood. This study presents a representation-agnostic evaluation framework integrating geometric separability, temporal dynamics, spatial robustness, and cross-regional generalisation. A 64-dimensional AlphaEarth embedding space is benchmarked against Sentinel-2 reflectance, NDVI-derived phenology, and harmonic features using field-scale data from the Conterminous United States (CONUS; 2017-2023) and Northern India (2021). Traditional features achieve strong linear separability (Fisher ratio), while embeddings provide superior distribution-aware separation -- Bhattacharyya and Mahalanobis distances up to an order of magnitude higher for spectrally similar pairs (e.g. Corn/Soybeans: 3.9 vs. 0.10). Embeddings uniquely capture multi-year crop rotation dynamics, reaching a separability ratio of 1.784 versus a maximum of 1.003 across all engineered features, and exhibit lower residual spatial autocorrelation of errors (mean Moran's I 0.17). However, harmonic and NDVI features show stronger cross-continental transferability (US India Macro F1: 0.493 vs. 0.326 for embeddings), reflecting better distributional alignment. The combined feature space yields balanced performance, supporting hybrid modeling. These findings reveal a fundamental trade-off between representational richness and cross-region transferability, providing a structured framework for feature selection in large-scale crop classification.

arXiv AI Papershace 17 días

When All Rollouts Fail: Policy-Compatible Joint Decoding for Hard-Problem Reasoning

Reinforcement Learning with Verifiable Rewards (RLVR) derives its learning signal from binary rewards via on-policy exploration and becomes a powerful paradigm for improving the reasoning of large language models. Yet it faces a learning cliff on problems beyond model's current ability: when every rollout fails, the samples reveal nothing about a correct solution, and group-relative estimators (e.g., GRPO, RLOO) collapse the update to zero advantage. Expert trajectories offer a remedy, but they are incompatible with the current policy and often lie in the student's low-probability regions, so the surviving updates demand large parameter shifts and decay into surface-level imitation rather than transferable reasoning. We argue that a useful learning signal on hard problems must be both verifiably correct and sufficiently compatible with the current policy. We propose Policy-Compatible Reinforcement Learning (PICO), which performs distribution-constrained joint decoding: expert data guides decoding at the token level while preserving the student's relative token probabilities, and a dynamic top-k mechanism confines the sampling distribution under a finite divergence budget, yielding trajectories that are verified and policy-compatible by construction. These trajectories are optimized jointly with on-policy rollouts under a mixed-behavior objective that mitigates their distribution mismatch via importance sampling. Experiments across reasoning diverse datasets and agentic benchmarks demonstrate that PICO consistently outperforms standard RLVR and prior off-policy baselines.

arXiv AI Papershace 17 días

Conservative Policy Iteration For Zero-Sum POSGs

Occupancy-state reductions make dynamic programming possible for zero-sum partially observable stochastic games (zs-POSGs). Existing point-based solvers usually optimise a central value function. This restores Bellman recursion, but it requires point-based updates over high-dimensional occupancy spaces. We introduce an approximate policy-iteration algorithm for finite-horizon zs-POSGs. The method maintains one policy per player and updates both jointly along occupancy trajectories induced by conservative policy updates. Instead of approximating a full central value function, each update uses a single supporting hyperplane induced by the current joint policy. The algorithm alternates local policy improvements, conservative updates, and persistent exploration. Computation is therefore concentrated on occupancy regions where the current joint policy is exploitable. On finite-horizon zero-sum POSG benchmarks, the method is competitive at small horizons and often improves the runtime-exploitability trade-off at larger horizons compared with PBVI, HSVI, and CFR+.