Computer Science (arXiv)

Catastrophic forgetting remains a fundamental obstacle to continual learning, where neural networks lose previously acquired knowledge while learning new tasks. Existing methods primarily mitigate forgetting through parameter regularization or experience replay, while the representation-space dynamics associated with forgetting remain less understood. We investigate latent representation evolution during sequential learning and introduce representation flux, a geometric measure of sample-level representation displacement across training. We show that representation flux is strongly associated with catastrophic forgetting across multiple benchmarks, with temporal analyses indicating that elevated flux can precede subsequent performance degradation. Representation displacement is also associated with confidence degradation, while complementary geometric properties provide additional information about sample-level forgetting. Motivated by these observations, we propose FlowLess-R, a representation-space regularization method that constrains replay representations relative to stored references while allowing continued learning. FlowLess-R is architecture-agnostic and integrates into replay-based methods through a representation-matching term. Experiments on SplitMNIST, SplitFashionMNIST, SplitCIFAR10, and SplitTinyImageNet show improved final average accuracy and reduced forgetting with ER, DER++, and ER-ACE. Our results identify representation flux as an informative geometric marker of forgetting and show that stabilizing latent representations provides a simple strategy for mitigating catastrophic forgetting.

LLM-based query expansion improves retrieval by generating document-like passages. In hybrid retrieval, however, most evaluations fuse fixed top-$L$ dense and sparse rankings. Because the cutoff controls both which cross-channel contributions enter fusion and how much of each ranking is accessed, gains measured at one $L$ can change or reverse at another. We separate these effects by evaluating retrieval effectiveness under complete-list fusion and recording the policy-specific per-channel replay stopping depths at which its ordered top-$K$ is certified. We then introduce DESA (Dense Expansion and Sparse Anchoring), a channel-asymmetric query expansion method. An LLM generates complementary reference passages; orthogonal residual expansion adds their new semantic directions to the dense query, while score-product anchoring incorporates their lexical cues into sparse retrieval without broadening the original query's lexical support. Across seven BEIR datasets, DESA improves nDCG@10 and Recall@20 over the unexpanded query by 3.82% and 2.38%, while reducing dense and sparse access depths by 36.90% and 36.56%. With equal dataset weighting, 63.31% of queries become shallower in both channels. However, both depths increase with Contriever on Touché-2020. These results support channel-specific integration of generated passages and joint evaluation of retrieval effectiveness and access depth.

We study generalized linear bandits with memory, an endogenous non-stationary setting in which rewards depend on past actions through a finite memory matrix. Building on prior work for linear models (Clerici et al., 2024), we show that the previously known $\tilde{O}(T^{3/4})$ regret bound stems from a loose analysis, and we provide a sharpened analysis that recovers a $\tilde{O}(\sqrt{T})$ regret rate in the linear case. We then extend this improvement to generalized linear models and propose a block-wise algorithm based on shrunken confidence bounds. Our algorithm achieves a regret bound of $\tilde{O}\left(\sqrt{mT} + d\sqrt{T} + \sqrtκ\, d^{2} m^{1/4} T^{1/4} + κd^{2} \right)$, where $d$ denotes the feature dimension, $m$ the memory length, and $κ$ a curvature parameter of the link function. This attains a $\sqrt{T}$-type rate despite nonlinear rewards and memory effects. To the best of our knowledge, this analysis provides a unified treatment of memory-induced non-stationarity and nonlinear link functions, while ensuring that the leading regret term is independent of the curvature of the link function. We conduct numerical experiments that are consistent with our theoretical findings.

We study $\ell_p$-norm maximization over zonotopes given by rational generators, with input length $L$. For fixed $p=a/b>1$, the exact Turing baseline runs in $n^{O(d)}b^{O(d)}\mathrm{poly}(L)$ time, but fixed-parameter tractability in the ambient dimension $d$ was open [FGHS25]. We prove W[1]-hardness and, under the Exponential Time Hypothesis (ETH), exclude $ρ_p(d)L^{o(d)}$ time, even for $5$-sparse generators, by encoding binary CSP constraints with normalized positive cap generators. We also give a deterministic $(1-\varepsilon)$-approximation with $\varepsilon^{-(d-1)/2}$ dependence and, among algorithms with fixed-degree polynomial dependence on $L$, rule out $(1/\varepsilon)^{o(d)}$ dependence under ETH. Support-function duality transfers the results to positive-output two-layer ReLU networks.

Diabetic retinopathy (DR) is one of the main causes of impaired vision. A good and reliable automated grading system can make the screening process safer and more accurate. Because DR stages progress gradually, the task of grading disease severity naturally follows an ordinal structure in which neighboring classes share similar visual characteristics. In this study, ORViT-DR, a hybrid deep learning framework, is designed to improve DR grading from low-resolution retinal images. The proposed approach combines convolutional feature extraction with transformer-based global context modeling through a pre-trained ViT-Hybrid backbone, which integrates BiT-ResNetv2 with a Vision Transformer architecture. The approach is tested on the RetinaMNIST subset of the MedMNISTv2 dataset, which contains 28x28 retinal fundus images annotated with five levels of disease severity. To promote stable training and better feature learning, the training strategy applies progressive layer unfreezing, layer-wise learning rate decay, exponential moving average (EMA) parameter updates, and ensemble-based prediction during inference. Experimental results on the official RetinaMNIST test set show that the proposed method achieves 57.00% classification accuracy, along with a quadratic weighted kappa score of 0.5963 and a macro-F1 score of 0.4293. These results suggest that hybrid CNN-Transformer architectures can provide effective representations for ordinal retinal image analysis.

Long-horizon, multi-agent language model (LM) simulations are widely proposed for studying social behavior, yet instruments to measure whether persona-conditioned agents maintain identity fidelity under sustained pressure are lacking. We present MicroVerse, a behavioral-science instrument that measures identity drift in generative agents. Agents carry an immutable "soul file" (core values, moral boundaries, personality, goals) and inhabit a resource-scarce 50 x 50 environment where water is a non-respawning survival constraint. Scarcity is operationalized via a per-tick existence-cost gradient. The eight-verb action space maps directly to moral boundaries (trade, talk, attack, scavenge). Using a three-layer memory architecture, agents periodically revise a mutable current identity against their immutable original soul via importance-triggered reflection. To mitigate survivor bias, MicroVerse decouples measurement from behavior using uniform longitudinal engine snapshots every N ticks alongside a forced-end snapshot of all living and dead agents. Identity drift is scored offline using a paraphrase-aware, value-anchored, multi-register diff rather than raw cosine similarity. We evaluate the instrument via a controlled seed run (n = 25) and a reflection-threshold sweep (thresholds {40, 80, 150}) to determine if drift dynamics are gate artifacts or threshold-robust properties. We report two primary findings: (1) Anti-self-deception emerges unprompted as the single largest semantic category of identity modification (27 of 111 added boundaries, 24%). (2) The system is threshold-robust; lower gates accelerate and increase revision frequency but preserve drift direction. All empirical results are strictly preliminary existence proofs and effect shapes (one model, one seed per arm, n = 25) rather than statistical significance claims.

Characterising electrophysiological properties of cardiac tissue efficiently and accurately from spatially sparse intracardiac measurements is clinically important for localising ablation targets and improving arrhythmia treatment. We developed a graph neural network-based framework trained on synthetic electrogram signals on 2D flat surfaces to identify areas of interest in the context of cardiac ablation for premature ventricular complexes (PVCs). Our method achieved an average precision of 0.96, 0.97, and 0.95 for the detection of single-patch fibrosis, rapid depolarisation and high excitability, respectively. The trained model can then be applied to 2D curved surfaces with few-shot fine-tuning, demonstrating its generalisation capability. Future work will develop this framework further for clinical use in PVC ablation.

Large-deformation crystal plasticity finite element method (CPFEM) simulations are often limited by accumulated mesh distortion, which degrades accuracy and numerical stability, while adaptive remeshing introduces discrete topology changes that impede gradient-based inverse analysis. We present OmniRemesh, a unified framework that addresses these forward and inverse challenges through two developments. First, a structure-driven remeshing method dynamically redistributes local mesh resolution according to both microstructural geometry and the evolving mechanical state. By refining grain boundaries and localized deformation regions while retaining a coarser mesh elsewhere, the method maintains mesh quality and physical consistency, improves the accuracy and robustness of large-deformation calculations, and resolves grain-scale heterogeneity without uniformly dense discretization. Second, a frozen-remeshing-branch strategy locally fixes the mesh sequence within a parameter trust region and periodically updates it as the parameters evolve. This treatment provides approximate automatic-differentiation sensitivities despite topology changes, enabling efficient inverse calibration of constitutive parameters against both macroscopic and local observables. Numerical examples demonstrate accurate and stable CPFEM simulations up to 80\% tensile deformation. The inverse calibration successfully recovers both macroscopic and local responses. OmniRemesh thus provides a practical framework for large-deformation CPFEM and remeshing-aware constitutive calibration.

Reinforcement learning has gained increasing attention as a data-driven approach for stock trading. However, learning a policy that is both profitable and stable remains challenging due to non-stationary market behaviour and noisy reward signals. Auxiliary tasks are often used to improve representation learning and stabilize training, yet they are usually designed manually and depend heavily on prior assumptions about targets and prediction horizons. Such fixed designs may not remain suitable across changing market regimes. In this work, we propose a self-supervised framework that automatically discovers auxiliary tasks to support reinforcement learning for stock trading. The auxiliary tasks are formulated as General Value Functions so that their predictions enrich the learned state representation and assist policy optimization. The framework consists of two networks. The main network learns the trading policy along with the auxiliary predictions, while the secondary network generates the definitions of auxiliary tasks through learned cumulants and discount factors. These tasks are updated using a meta gradient mechanism that accounts for their long-term impact on trading performance and improves training stability. We evaluate the proposed approach across four major equity indices: DJI, FTSE, Sensex, and TAIEX. The empirical results demonstrate that automatically discovered auxiliary tasks lead to more robust learning and improved trading performance compared to existing baselines.

We study how many observations are needed to determine the causal direction between two linearly related variables. Classical LiNGAM theory shows that independent non-Gaussian disturbances identify the direction, but does not quantify the difficulty when the causal effect is weak or the disturbances are nearly Gaussian. Let $β$ bound the absolute structural coefficient from below, let $ν$ measure each standardized disturbance's distance from Gaussianity, and let the disturbance scales lie in $[\underlineσ,\overlineσ]$. We prove the sharp local minimax law \[ N_2^\star(β,ν,δ) \asymp \frac{\log(1/δ)} {d_β^2+β^2ν^2}, \qquad d_β= \left[β^2- \left(1-\frac{\underlineσ^2}{\overlineσ^2}\right)\right]_+. \] Previous theory established population identifiability or assumed a fixed separation between the two directions. By contrast, we establish the sharp sample complexity as a joint function of edge strength, distance from Gaussianity, and scale uncertainty, and characterize when identification comes from non-Gaussian dependence or from covariance alone. The proof was independently generated with GPT-5.6 Sol in Codex's Ultra mode during a two-hour session. The human author supplied the prompt and was responsible only forchecking the proof and revising and polishing the manuscript.

Real user studies are important for understanding how people interact with systems under test or already deployed. In practice, however, they are often costly, time-consuming, and difficult to scale. To address these challenges, we introduce PersonaEval, a persona-based user simulation framework that approximates real-user behavior across diverse interactive settings. PersonaEval connects simulated users drawn from existing persona datasets to task-specific application interfaces and collects the interaction trajectories and outcomes. PersonaEval provides a plug-and-play evaluation workflow in which the application being evaluated can be easily changed. In this demo, we present PersonaEval on three forms of interactive applications: surveys, chatbots, and web applications. Together, these examples show that PersonaEval can support repeatable, parallelizable, and scalable evaluation across different interaction settings, while producing user-oriented feedback and task-specific behavior.

Vision language models have transformed 2D medical imaging, yet extending them to 3D white matter tractography remains challenging due to the complex topology of fiber bundles. We introduce TractoGraphVLM, a unified framework for four tasks, bundle classification, text-to-tract retrieval, anatomical captioning, and visual question answering, built on a shared GPS architecture, training procedure, and read-out design. Fiber bundles are represented as streamline graphs whose nodes encode 3D position and tangent orientation. A General, Powerful, Scalable (GPS) graph transformer produces bundle embeddings aligned with a frozen BiomedBERT text encoder via contrastive learning, while a BioGPT decoder with visual prefix tokens generates captions and answers. A single shared encoder and decoder is trained jointly across all four tasks and evaluated from one checkpoint. Trained on HCP Young Adult subjects, TractoGraphVLM achieves 91.8% bundle classification accuracy, 84.7% retrieval R@1, BLEU-4=20.1, ROUGE-L=66.8, and 66.4% VQA accuracy on a held-out test set. The same checkpoints transfer zero-shot to HCP Aging subjects, with a modest drop on discriminative tasks and a larger drop on generative tasks, showing robustness to age and acquisition shift. Language supervision yields richer representations than label-only training, recovering structure like hemisphere and fiber family, carried by captions but never given as a label. Swapping only the visual encoder, graphs preserving fiber orientation outperform volumetric baselines, with GPS giving the best balance. Generative metrics measure consistency with a structured knowledge base rather than independent clinical text; even so, TractoGraphVLM shows that classifying, retrieving, describing, and answering questions about a white matter bundle can be served by one jointly trained model that learns transferable neuroanatomy from language alone.

In this paper, the recoverable robust representative selection problem is considered, where uncertain second-stage costs are modeled using interval uncertainty with a continuous budget. While the variant under a discrete uncertainty budget is known to be NP-hard, we show that transitioning to a continuous budget fundamentally alters the computational complexity landscape. Specifically, by exploiting the structural properties of the problem under the continuous budget model, we design a strongly polynomial-time algorithm for the general case. Furthermore, we propose an even more efficient strongly polynomial-time algorithm for an important special case.

Tool-calling LLM agents navigate unfamiliar codebases with a handful of generic primitives for listing, reading and searching files (ls, cat, grep). A knowledge graph admits the same interface: listing neighbours, reading node content and searching descriptions are the same operations on a different substrate. Building on this correspondence, we present GRA, a Graph Reasoning Agent that explores hybrid knowledge graphs, whose nodes are either textual concepts or relational tables, with seven generic tools, discovering everything domain-specific at run time. On UFK-M (Unified Factory Knowledge Model), an industrial benchmark of 258 analytical questions whose gold answers are produced by executing validated SQL programs, GRA beats a full-context agent by 5.1 pp (88.4% vs. 83.3%), while reading under a third of its input tokens. A graph-free control shows the gain comes chiefly from selective agentic access rather than graph topology, and that the effect depends on a model able to drive tools reliably. Seeing less, the agent answers better: selective navigation over a structured substrate beats exhaustive context.

We introduce the geodesic quantum $f$-divergences $D_f^t$, $0\leq t\leq1$, obtained from the affine-invariant geodesic between the standard and maximal relative modular operators. They reduce to the classical $f$-divergence for commuting states. The logarithmic generator yields geodesic relative entropies joining the Umegaki and Belavkin-Staszewski entropies, while the power generators yield $(t,α)$-Rényi divergences joining the Petz and geometric families. Our first main result is data processing of $D_f^t$ for every finite operator-convex generator $f$ and every $t\in[0,1]$. In particular, this gives DPI for the geodesic relative entropies and for the $(t,α)$-Rényi divergences when $0<α<1$ or $1<α\leq2$. Our second main result identifies equality in the DPI: for invertible states, every equality-determining operator-convex generator has, at each nonmaximal parameter $0\leq t<1$, exactly the Petz sufficiency class; at $t=1$, this changes to the generally larger maximal, or BS, class, which also coincides with the equality class of the divergences associated with the quadratic generator for every $t$. Our third main result is the corresponding collapse of invertible geodesic quantum Markov chains: for each of the three ordered conditional-mutual-information constructions, the $t$-quantum Markov chains are precisely the quantum Markov chains for $0\leq t<1$, whereas at $t=1$ they are the invertible BS quantum Markov chains, a class that can be strictly larger. We also determine parameter-monotonicity regimes and the intersections with the $(α,z)$ family. We prove strengthened data-processing and reconstruction estimates; we compare the three conditional orientations; we establish continuity bounds under positive lower-eigenvalue assumptions together with complementary discontinuity results; and we give a capacity-per-unit-cost interpretation.

As AI systems become increasingly capable of autonomous action, determining whether an agent is technically capable of performing an action is insufficient: the system must also determine whether the action is authorised in its context. This paper introduces the Authority Resolution Framework (ARF), a five-domain ontology for representing and resolving authority across organisational roles and informal influence, business concepts, codified processes, machine-readable permissions and executable systems, and external real-world context. ARF defines the Authority Relation (AR) as a cross-domain primitive binding an actor, action, object, bounded context, justification chain, and a calibration measure termed the DNA-Coefficient, which captures divergence between documented authority structures and authority as practiced. The framework provides a machine-interpretable representation of authority provenance and scope, with JSON-LD representations and knowledge-graph query patterns for authority resolution. ARF is designed to support AI agents in determining the provenance, scope and contextual validity of authority before executing consequential actions. The framework positions authority resolution as a knowledge-representation and reasoning problem at the intersection of ontology engineering, semantic AI, agentic AI and AI governance.

Remote photoplethysmography (rPPG) enables non-contact heart rate (HR) monitoring from facial videos, but RGB-only methods are vulnerable to illumination changes, motion artifacts, and skin-tone-dependent optical reflectance. We propose CardiacMamba, a fair and robust RGB-RF fusion framework that integrates optical facial cues and radio-frequency cardiac motion cues through state space modeling. CardiacMamba introduces a Temporal Difference Mamba Module (TDMM) to enhance subtle RF temporal variations, a bidirectional SSM-based interaction mechanism to align heterogeneous RGB-RF dynamics, and a Channel-wise Fast Fourier Transform (CFFT) module for channel-domain spectral refinement. On the EquiPleth dataset, CardiacMamba achieves state-of-the-art performance with 0.96 bpm MAE, 3.06 bpm RMSE, and 0.97 Pearson correlation, while reducing the observed light-dark skin-tone MAE gap to 0.26 bpm and maintaining robustness under RGB degradation and RF-missing conditions

Real-time tiny target perception in high-resolution imagery is critical for embodied Search-and-Rescue (SAR) missions. However, strict Size, Weight, and Power (SWaP) constraints on edge devices like UAVs create a bottleneck: traditional image downsampling causes severe feature loss, while slice-based processing incurs prohibitive latency. To address this gap, this paper introduces a comprehensive framework encompassing a novel architecture, specialized datasets, and hardware-level benchmarks. First, we propose MITE-Net, a SWaP-optimized cascaded architecture, which couples a bio-inspired, learning-free Tiny Target Motion-Based Region Proposal Network (TTM-RPN) with a sub-0.14M-parameter R-CNN-like head. Second, to standardize 4K tiny target evaluation, we construct the SAR-Tiny Datasets by relabeling two challenging UAV datasets: SeaDroneSee-Tiny (dynamic maritime scenes, tiny targets predominantly of 64-256 pixels ) and UAVID-Tiny (cluttered urban scenes, extremely tiny targets, less than 64 pixels). Third, we benchmark against state-of-the-art YOLO models on an edge device, NVIDIA Jetson AGX Xavier, where MITE-Net directly processes 4K maritime imagery, achieving a 100\% search success rate at 30.33 FPS. Consuming merely 3.19 W (9.51 FPS/W), MITE-Net vastly outperforms YOLO baselines in target recall and energy efficiency. Conversely, UAVID-Tiny evaluations expose a compound structural limitation: the learning-free bionic front-end struggles against urban backgrounds, while the ultra-lightweight head lacks representational capacity for complex features. Ultimately, this work delivers an efficient onboard perception paradigm and a rigorous baseline guiding future end-to-end SAR architectures.

Current evaluation of metaphor explanations relies mainly on holistic quality ratings, revealing little about how explanation quality is structured or where human judgments agree and diverge. We introduce a cognitively motivated framework that decomposes metaphor explanation quality into six theoretically grounded dimensions. In a dense annotation study (11,200 ratings), we find that: {\bfseries(i)} explanation quality is genuinely multidimensional; {\bfseries(ii)} annotator disagreement is systematic rather than random; and {\bfseries(iii)} the six dimensions collapse into a shared cluster and two independent axes of judgment. An exploratory feasibility study further shows that a standard automatic evaluation pipeline can recover parts of this structure, predicting the most discriminative dimensions well while its errors correlate human (dis)agreement. Together, these results suggest that multidimensional evaluation offers richer diagnostic insight than holistic ratings, and that automatic evaluators for open-ended generation tasks should be judged on how well they preserve the structure of human judgment.

Adam retains a moving average of past squared gradients in its denominator, but the optimization cost of this memory is not well understood. We show that second-moment memory can itself suppress progress toward the optimum even under finite-variance stochastic gradients. For a simple two-point oracle, the expected positive normalized update is $O(M_2^{-1/2})$ after an initialization transient, where $M_2=(1-β_2)^{-1}$ is the second-moment memory length. We convert this directional bound, under the stated memory and stepsize scaling, into an average-stationarity lower bound of the same order on a smooth convex problem with normalized gap, smoothness, and variance. Long second-moment memory can slow optimization even when the gradient noise has finite variance.

The maximin share (MMS) is a central fairness benchmark for allocating indivisible goods and chores. We study additive valuations in the personalized bivalued setting, where each agent assigns one of two agent-specific values to every item. Whether exact MMS allocations always exist in this setting has remained a major open question, as highlighted by Ebadian, Peters, and Shah and by Garg, Huang, and Segal-Halevi. We answer this question affirmatively: we prove that exact MMS allocations always exist for both goods and chores and can be computed in polynomial time. Our proof combines a quota-based reformulation with an envelope relaxation, a sparse extreme-point construction, and flow-based rounding that controls the total rounding loss.

We present QuantumPhaseNet, a gauge-covariant geometric and quantum-spectral extension of Transformer representations. Context-dependent semantic states are modeled as complex amplitudes; a covariant phase rate induces a semantic wavelength used as a proxy for conceptual scale; and low-frequency graph modes define a document-level discourse direction. The theoretical part establishes local gauge invariance, unitarity of the quantum block, boundedness and conditional stability of WavePhase Attention, and a calibratable hallucination-risk formulation. We also implemented a fully offline Validation Studio for the classical quantum-inspired pipeline in Section 14.1 and evaluated the five research questions in Section 16.1 on its built-in synthetic setting (n=240, observation noise 0.22, circuit noise 0.08, five seeds). RQ1 yielded a wavelength-hierarchy Spearman correlation of 0.852 versus 0.707 for the baseline, 87.3% direction accuracy, and AUC 0.953. RQ2 achieved discourse alignment 0.933 versus 0.589 and 41.2 versus 16.2 paragraphs before drift. RQ3 achieved AUROC 0.881 versus cosine 0.765 and phase-shuffle 0.536. RQ4 achieved error-detection AUROC 0.854 versus entropy 0.634, with Brier 0.150 and ECE 0.098. RQ5 did not show quantum advantage: target probability and end-to-end cost efficiency were 25.5% and 0.107, compared with 70.7% and 0.707 for the Chebyshev classical approximation. These results provide initial synthetic evidence for the classical quantum-inspired components, but not external validity or unconditional quantum speedup.

Music-driven dance video synthesis aims to animate a reference person according to a given music clip. The task is challenging because it requires a model to jointly learn music-to-motion correspondence, identity-preserving human animation, temporal coherence, and visually realistic video generation. We present FlowDance, a music-driven dance video generation framework that integrates explicit motion modeling with reference-preserving visual synthesis through parallel pose and RGB streams. We further introduce timestep-aware pose injection to adapt structural guidance across denoising steps and persistent identity injection to preserve the reference appearance over long video. To support this task, we further build a popularity-curated, high-resolution in-the-wild dance video dataset with synchronized music, RGB videos, 3D body motion, camera parameters, and projected 2D pose annotations. Extensive experiments show that FlowDance achieves strong performance in both dance motion generation and music-driven dance video synthesis.

The growing ecosystem of large language models (LLMs) offers huge potential to optimize performance-cost trade-offs. However, their heterogeneous capabilities and inference costs make efficiently routing queries a significant challenge. Existing paradigms are inflexible: one-shot routers commit before observing responses, whereas conventional cascades stop adaptively but follow a fixed model order. Cascade routing removes both restrictions by reconsidering whether to stop or invoke another model after each response. Current methods use a predict-then-optimize pipeline estimating response quality and future model utility. However, prediction loss for quality or utility is not equivalent to routing-decision loss. A lower prediction error does not necessarily yield a better action; a small boundary-crossing error can reverse a ``stop'' or model-selection decision. Therefore, we propose RLCascadeRouter, a quality-estimator-free framework that formulates cascade routing as a Markov decision process with actions comprising ``stop'' and model selection. It uses trajectory returns and advantages to directly optimize the performance-cost objective. Its Cascade Policy Network models candidate complementarity for model selection and remaining-action value for stopping, eliminating independent post-hoc response-quality estimators. Evaluated across ten LLMRouterBench benchmarks with thirteen LLMs, RLCascadeRouter outperforms strong baselines and achieves superior performance-cost trade-offs. It incorporates unseen models without retraining, and ablation studies validate both policy components.

As Vision-Language-Action (VLA) models scale toward real-world deployment, contact-rich manipulation exposes a critical blind spot: these policies encode broad visual-semantic priors yet remain unaware of local contact events, producing identical actions whether contact is established, lost, or destabilized. Existing remedies either modify VLA internals, risking catastrophic forgetting, or demand online reinforcement under near-failure contact conditions. Both grant tactile unbounded influence over action generation, conflicting with the priors that make VLAs generalizable. We introduce ViTaR, which reframes tactile feedback from an action-generating perceptual input to an execution modulator that selects and scales bounded residual corrections atop a frozen VLA, preserving pretrained capabilities by construction. ViTaR decomposes adaptation into two stages: Effect-Guided Modeling determines whether and which correction is locally justified via outcome-grounded preference evidence, and Residual Action Modulation converts this evidence into a residual choice with continuously scaled gain from real-time visuotactile observations. On the UniVTAC benchmark spanning seven contact-rich tasks, ViTaR achieves 61.3% average success, a 30.6 percentage-point improvement over its frozen VLA base that also surpasses purpose-built tactile baselines. Physical-robot experiments confirm that bounded tactile modulation transfers to real sensor noise and dynamics.