论文库

Model selection and hyperparameter tuning pose a central challenge in off-policy evaluation (OPE) because estimation errors in fitted $Q$-functions and occupancy ratios cannot be directly evaluated using held-out transition data. Existing Bellman-residual-based validation methods often require auxiliary regressions, while many minimax-based methods require optimization over separate critic classes. We propose pairwise Bellman validation (PBV), a family of penalized Bellman tournament methods that reduces validation to computing held-out sample averages. PBV constructs critics from pairwise candidate differences, on the log scale for occupancy ratios, and selects the candidate with the smallest worst-case penalized score. This construction approximates oracle critics given by $Q$-function errors and log ratios of candidate to true occupancy ratios, whose Bellman and adjoint Bellman moments control estimation error. Critic penalties account for approximation error under candidate-class misspecification. We establish finite-sample oracle inequalities relative to the best available candidate, even under candidate-class misspecification. For $Q$-function selection, unweighted PBV avoids importance weighting but requires a coercivity condition on a quadratic form induced by the Bellman operator. Importance-weighted PBV (IW-PBV) uses marginalized importance weighting to exploit Bellman contraction under the target-policy occupancy measure, avoiding this condition. For occupancy-ratio selection, adjoint PBV (A-PBV) exploits KL contraction of the adjoint Bellman operator to obtain oracle guarantees under the offline data law. On Hopper and DeepOPE, unweighted PBV is competitive with LSTD-Tournament, often achieving higher accuracy on Hopper, while importance weighting can further improve performance under controlled overlap.

Autonomous SaR requires robots to explore unknown or partially known environments, avoid hazards, coordinate team activities, and guide discovered victims to safety. Improving individual capabilities of robots does not necessarily improve mission outcomes: safer motion can restrict the exploration behavior, faster discovery can increase the rescue demand, and effective guidance depends on victim responses. We introduce PA-SaR, a mathematical and architectural framework for coordinating heterogeneous local control, learning-based exploration, supervisory task allocation, and adaptive approach-lead victim guidance around mission-level objectives. PA- SaR distinguishes environmental ground-truth, robot-specific perceptions, and simulated victim behavior. This provides a common basis for integrating complementary mechanisms and analyzing their interactions across mission stages. We demonstrate the framework through single- and multi-robot simulation studies incorporating model-predictive and learning-based exploration, perception-based safety shielding, event-driven task allocation, and adaptive victim guidance. The results show trade-offs between collision avoidance and search completion, indicate that supervisory allocation improves search completion beyond information sharing alone, and demonstrate that increased exploration does not necessarily yield more rescues. Adaptive guidance increases rescues relative to continuous leading, without a detected improvement in mean or final victim trust or a reduction in mortality. These findings support evaluating local control, team coordination, and evacuation as related but distinct aspects of autonomous SaR within the considered simulation environments.

Vision-language models must reconcile visual evidence with memorized world knowledge when the two conflict. How they resolve this conflict shapes the reliability of multimodal systems, yet prior work characterizes it behaviorally without a component-level causal account. We combine activation patching across three granularities (residual stream, attention heads, and MLP sublayers) with model-component ablation studies and mechanistic analysis. Across three VLM families, we find that visual grounding emerges by default, whereas prior grounding depends on a small set of causally necessary attention heads (2.5-4.8%) concentrated in the second half of the network. These heads enable answers from stored world knowledge (e.g., "red" for a strawberry) despite conflicting visual input. Ablating them flips predictions from knowledge-grounded to visually grounded answers in 68-96% of cases under prior-knowledge prompts, but changes only 0.8-7.5% of visually grounded predictions, establishing an asymmetric causal structure. The identified heads decompose into routing heads, which modulate information flow, and writing heads, which directly project answer tokens into the residual stream. This structure is consistent across model families and scales, revealing a sparse causal circuit underlying perception-knowledge conflict in VLMs.