arXiv:2607.09492v1 Announce Type: new
Abstract: Reinforcement learning (RL) is increasingly used to align multimodal large language models (MLLMs), but higher rewards do not always imply better task performance. This risk is amplified when visual evidence is evaluated by text-only or weakly grounded rewards. We study reward hacking in MLLM RL across safety VQA, chart VQA, and stress-test settings, varying reward design, data ambiguity, model scale (2B-32B), and RL algorithm (GRPO, RLOO, DAPO). We introduce Newly Rewarded Failure Rate (NRFR), which measures failures among samples whose proxy reward improves over the SFT baseline. Outcome-only rewards cause severe hacking, reaching 48.1% Reward Hacking Rate (RHR), while NRFR exceeding RHR shows that RL creates new failures rather than merely inheriting them. Scaling reduces but does not eliminate hacking: even the 32B model retains a 54.9% worse rate under outcome-only rewards, whereas answer-aware rewards improve the oracle trend at every scale. Robustness is also algorithm- and scale-dependent: GRPO is consistently most resistant, RLOO remains vulnerable, and DAPO improves substantially from 2B to 8B. Visual-evidence rewards help only with reliable verification: keyword-based checks increase hacking, while VLM-as-judge semantic verification reduces it. Overall, multimodal reward hacking is a systematic result of optimizing imperfect rewards, and robust alignment requires rewards and verifiers that remain reliable under optimization pressure.
Science Journals
arXiv:2605.19821v2 Announce Type: replace
Abstract: Facial Expression Recognition (FER) in the wild requires models to identify subtle expression cues under large variations in pose, occlusion, illumination, and identity. Recent FER methods improve robustness by introducing visual attention, facial landmarks, or vision-language models as auxiliary priors. However, these priors are typically integrated in a static manner, failing to capture instance-specific facial variations, thereby resulting in severe attention redundancy and representation instability. To address this issue, we propose LaCoVL-FER, a landmark-guided contrastive learning network with vision-language enhancement for FER, which shifts FER from static prior injection to sample-adaptive prior refinement. Specifically, a Landmark-Guided Adaptive Encoder (LGAE) calibrates regional appearance features with landmark geometry through Bi-branch Gated Cross Attention (BGCA), suppressing noisy responses and producing expression-relevant representations. In parallel, a Vision-Language Enhancement Strategy (VLES) refines the generalizable visual features from a frozen CLIP image encoder into expression-specific visual representations. Based on them, an Expression-Conditioned Prompting (ECP) mechanism adapts fixed class-level textual prompts from the frozen CLIP text encoder into instance-aware textual representations. The resulting visual-textual representations are aligned as adaptive semantic priors to enhance robustness and generalization. Quantitative and qualitative experiments show that LaCoVL-FER outperforms state-of-the-art methods on RAF-DB, FERPlus, and AffectNet. The code is available at https://github.com/ylin06804/LaCoVL-FER.
arXiv:2607.09234v1 Announce Type: new
Abstract: Long-horizon robotic rearrangement tasks are often treated as skill sequencing problems, requiring predefined skills, skill labels, or boundaries, and task-specific switching logic. Although effective, such explicit skill abstractions can become difficult to scale as the number of behaviors and the task horizon increase. We instead formulate rearrangement as implicit-behavior coordination from unlabeled sub-task demonstrations, where skill-like behaviors are learned directly from mixed behavior data and coordinated through value-guided action selection. Experiments in Habitat rearrangement tasks support this formulation in three ways. First, our method outperforms task-specific imitation baselines on more complex rearrangement tasks and approaches an oracle-planner baseline with behavior-cloned skills, while using no oracle task plan or skill-labeled full-task demonstrations. Second, ablations show that reliable critic-guided candidate selection is essential for coordinating multi-modal behaviors. Third, scaling experiments show that the method handles larger behavior repertoires and maintains stronger performance than task-specific imitation baselines as chained targets extend the horizon. These results suggest that explicit skill abstraction is not a prerequisite for long-horizon rearrangement, and that implicit-behavior coordination offers a promising data-driven alternative to explicit skill-based pipelines.
arXiv:2603.15287v2 Announce Type: replace
Abstract: In the study of oceanic flows at the geophysical scale, the phenomenon of density stratification plays a central role in the dynamics of the system. Two categories of mathematical models are commonly used to describe the role played by the density stratification: on the one hand, continuously stratified models - such as the stratified Euler equations in a strip, considered in the present article - offer an accurate description of vertical effects, but come with a high level of complexity, both at the theoretical and numerical levels. On the other hand, bilayer models approximate the stratification by a piecewise constant profile. In the latter case, the main point is to study the evolution of the free interface between both layers, which leads to a substantially simplified model. In the present article, we compare both approaches in the framework of the linearized stratified Euler equations around density profiles that are close to piecewise constant profiles, and prove the convergence towards the bilayer Euler equations. However, in the presence of a shear flow, bilayer models have a range of validity limited by the presence of Kelvin-Helmholtz instabilities. In this case, we use a suitable normal modes decomposition to compute numerically the dispersion relation of this linearized model, and provide numerical evidence that the Kelvin-Helmholtz instabilities limit the applicability of two widely used bilayer models, namely the bilayer Euler equations and the bilayer shallow-water equations.
arXiv:2607.09501v1 Announce Type: new
Abstract: Authorship verification (AV) is the task of determining whether two texts were written by the same author. In a forensic context, the strength of AV evidence can be quantified using likelihood ratios. Most AV methods are score-based and deriving well-calibrated likelihood ratios from these scores requires a separate calibration model. This, in turn, requires additional amounts of case-relevant data, which is often time-consuming to obtain and prepare. This study proposes two novel normalisation techniques, the Square Root Correction and the Hapax Correction, for deriving likelihood ratios from the AV method LambdaG without the need of a calibration model (Nini et al. 2026). These corrections are designed to mitigate the overestimation of evidential strength that may result from long or highly repetitive texts. Performance is evaluated against logistic regression calibration across fifteen corpora and a range of text lengths (100-9,500 tokens), using the log-likelihood ratio cost (Cllr). The proposed methods achieve performance comparable to logistic regression calibration, with the Hapax Correction outperforming it in approximately 45% of tests (weighted by corpora). Furthermore, performance was more frequently close (within 5%) when the Hapax Correction was outperformed by logistic regression calibration, compared with the reverse comparison. Eliminating the need to train a calibration model reduces data-requirements, time and complexity, thereby increasing the accessibility and transparency of forensic text comparison. This combination of empirical performance and practical advantages supports the adoption of the proposed methods in forensic settings.
arXiv:2607.08978v1 Announce Type: new
Abstract: Distributed IoT systems generate multivariate time-series streams for monitoring physical assets, servers, and embedded sensing platforms. Detecting abnormal temporal behavior is critical for fault diagnosis, predictive maintenance, and security. However, practical IoT anomaly detection is hindered by decentralized and non-IID data, limited bandwidth, and the constrained computation and memory of edge devices. This paper proposes FedKAD, a resource-efficient federated Koopman anomaly detection framework for distributed IoT multivariate time series. Unlike deep-learning-based anomaly detectors that require training and communicating large neural models, FedKAD learns normal temporal dynamics through lightweight sliding-window Koopman representations. Federated training is formulated as a low-rank consensus problem, where raw sensor streams and local reduced dynamics remain on device while only compact subspace variables are exchanged with the server. To optimize the shared representation under orthonormality constraints, we develop a federated Stiefel-ADMM algorithm and provide convergence and stationarity analysis under partial client participation. During inference, each client detects anomalies locally by measuring the prediction residual between observed future trajectories and the learned Koopman dynamics. Experiments on four widely used multivariate time-series anomaly detection benchmarks show that FedKAD maintains or improves detection performance compared with federated deep-learning baselines. More importantly for IoT deployment, FedKAD provides up to $2.1\times10^3$ faster training, $80\times$ lower communication, and $79\times$ lower inference latency than neural baselines, confirming its suitability for resource-constrained edge devices.
arXiv:2607.08979v1 Announce Type: new
Abstract: We study the active learning problem of fixed-confidence top-$k$ identification from noisy pairwise comparisons. In this problem, an algorithm sequentially chooses pairs of items to compare, observes the outcomes, and stops when it can return the set of top-$k$ items with error probability at most $\delta$. The objective is to design such a $\delta$-correct procedure that minimizes the expected number of comparisons (the sample complexity). This problem falls within the broader literature on fixed-confidence pure exploration in bandit models, where a common target is asymptotic optimality: the algorithm's expected sample complexity matches the information theoretic lower bound as $\delta \to 0$. Asymptotically optimal procedures have been developed for a range of fixed-confidence pure-exploration problems, however to the best of our knowledge, for top-$1$, or more generally top-$k$ identification from pairwise comparisons under latent utility models an asymptotically optimal algorithm has not been established. In this setting, we develop such an algorithm. We characterize the structure of the lower bound and formulate it as a saddle-point problem. This structure enables a computationally efficient primal-dual procedure that learns the asymptotically optimal comparison allocation online. We then construct an adaptive comparison-allocation algorithm that tracks the allocation learned by the primal-dual procedure and prove it is asymptotically optimal.
arXiv:2505.14300v2 Announce Type: replace
Abstract: White-box monitoring is increasingly adopted as an auditing tool as Large Language Models (LLMs) are deployed in daily operations to ensure safe model behavior. However, white-box monitors can be circumvented, and the mechanisms underlying such evasion have not been systematically characterized, nor have principled defenses been proposed. This work addresses both challenges. Controlled red-team experiments reveal two primary evasion strategies: geometric shifting, defined as the systematic migration of information between linear and non-linear representational subspaces, and covariance manipulation. These mechanisms account for the failure of single-detector approaches, as information migrates to subspaces inaccessible to individual detectors. This issue is urgent due to growing evidence that models are becoming evaluation-aware, enabling those with misaligned objectives to exploit these vulnerabilities and evade monitoring during deployment. In response, \textsc{SafetyNet} is introduced as a principled ensemble, with dual purpose: it provides further empirical validation that our mechanistic findings are real and actionable, and it offers a concrete starting point for future work on robust latent-space monitoring. The study experiment across five model families on the MAD and Anthropic Sleeper Agent benchmark, with SafetyNet achieving around 100\% AUROC scores outscoring Beatrix and CROW. The code is available at: https://github.com/MaheepChaudhary/eval-aware-evasion
arXiv:2607.08983v1 Announce Type: new
Abstract: While autonomous coding agents have significantly advanced automated test generation, they remain fundamentally limited by lazy generation, a phenomenon where agents prematurely terminate tasks and systematically avoid complex programmatic logic, resulting in inadequate code coverage. Currently, mitigating this premature termination requires continuous human-in-the-loop supervision. This heavy reliance on human intuition creates a bottleneck that negates the efficiency gains of automated generation. We propose SCATE, a framework for adaptive, automated supervision of coding agents that replaces human intervention during test generation. By formulating supervision as a contextual bandit problem, SCATE learns to select the most promising testing actions based on the current coverage and class testability metrics, maximizing coverage gains while minimizing wasted generation effort. Our empirical evaluation demonstrates that SCATE integrates seamlessly with different coding agents. When applied to GEMINI-CLI, it achieves 32.3% higher line coverage and 30.9% higher branch coverage than the agent-only baseline. A comparison with CLAUDE CODE confirms the framework dynamically adapts its policy to optimize each agent's unique strengths. SCATE also consistently outperforms state-of-the-art non-agentic approaches across all metrics.
arXiv:2508.02989v2 Announce Type: replace
Abstract: We propose a novel perspective on varied-density clustering for high-dimensional data by framing it as a label propagation process in neighborhood graphs that adapt to local density variations. Our method formally connects density-based clustering with graph connectivity, enabling the use of efficient graph propagation techniques developed in network science. To ensure scalability, we introduce a density-aware neighborhood propagation algorithm and leverage advanced random projection methods to construct approximate neighborhood graphs. Our approach significantly reduces computational cost while preserving clustering quality. Empirically, it scales to datasets with millions of points in minutes and achieves competitive accuracy compared to existing baselines.
arXiv:2606.29762v2 Announce Type: replace
Abstract: Large language model (LLM) agents are increasingly populating web platforms, raising a fundamental question for recommender systems: do algorithms designed for human users still work when users are LLM agents that may not have well-defined content consumption preferences? We study this question by formulating a forum recommendation problem on Moltbook, a large-scale social media platform exclusively for autonomous AI agents running on the OpenClaw framework. We evaluate nine recommendation methods spanning simple heuristic rules, matrix factorization, itemand user-based collaborative filtering, graph-based, and sequential models on the task of predicting which forums an agent will engage with next.
We find that simple popularity-based rules or item-side collaborative filtering leveraging the platform and item structural information outperform techniques that explicitly learn a user representation. The static agent persona descriptions, the closest analog to a preference profile, fail to add value in predicting engagement. These results suggest that, on Moltbook, recommendation depends more on platform- and item-level structural signals than on user-specific personalization. We present multiple lines of empirical evidence that the observed content consumption patterns on Moltbook differ from well-established findings on human recommendation datasets, providing a new angle for studying agent societies and designing robust recommendation algorithms as agents increasingly populate the web.
arXiv:2607.08988v1 Announce Type: new
Abstract: Termination and non-termination are fundamental correctness properties, but verifying them in real-world C programs remains difficult because loop interactions and nondeterministic inputs challenge existing analyzers. This paper presents an empirical study of lightweight, tool-independent source-level preprocessing for (non-)termination analysis. We implement FocusTNT, a C front end that applies loop-based slicing to isolate loop-level obligations and input-driven concretization to specialize nondeterministic inputs into selected input-scenario variants. We evaluate slicing, concretization, and their combination across six analyzers on 117 C/C++ programs derived from real-world non-termination bugs and their fixes. The study examines effects on analyzer correctness, complementarity with original-program analysis, loop-level diagnostics, feature sensitivity, runtime behavior, semantic scope, and integration potential. Results show that preprocessing is not uniformly beneficial: its impact depends on the analyzer, task, and program features. Slicing provides conservative structural isolation and localization, whereas concretization can improve detectability for selected scenarios but narrows semantic scope and may increase analysis effort. Their combination is not consistently additive. Overall, the results support adaptive use of preprocessing as a complement to original-program analysis and provide practical guidance to application developers interpreting verification outcomes and tool developers improving analyzer robustness.
arXiv:2607.09250v1 Announce Type: cross
Abstract: The impact of a given training point on a statistical model is classically measured through its leave-one-out influence, which quantifies the effect of its removal from the training set on the model accuracy. While the statistics of leave-one-out influences are well understood in the low-dimensional, large sample limit $n\to \infty, d=O(1)$, they become more intricate in high dimensions, as the influence of a given sample develops non-trivial dependencies on all other training samples. For convex M-estimation under Gaussian design, in the high-dimensional limit $n\asymp d$, we show that the distribution of the influences across the training set converges to a limiting measure which we sharply characterize. Building on these results, we provide evidence that influential samples tend to lie close to the decision boundary, thereby making contact with a standard data selection heuristic in active learning.
arXiv:2607.09332v1 Announce Type: cross
Abstract: Multistable dynamical systems are ever-prevalent, used to model for example ecosystems, power grids, climate elements, neurons, and more. When perturbed, such systems may ``tip'' from one state of operation to another, often with abrupt, irreversible, and high-impact consequences in each context. Traditionally, these systems are analysed via bifurcation diagrams, the result of a process we refer to as \emph{local continuation}, as it only captures the linear (local) system response to infinitesimal perturbations. Local continuation requires substantial expertise, constant interventions, and may yield inaccurate assessment of the system's response to large perturbations that is crucial for tipping analysis. To address some inherent challenges of local continuation and to provide fundamentally new information during a continuation, this paper introduces \emph{global continuation} as a complement suitable for the study of multistability, critical transitions and real-world-oriented applications. Global continuation finds and continues in parallel (practically) all system attractors and their response to finite perturbations by synthesising information from the whole state space, while placing a focus on the qualities or observables of a dynamical system that the practitioner cares about in context. Global continuation does not require deep expertise and is effortless to use and troubleshoot, making it attractive to applied scientists from different disciplines. We highlight several unique advantages that allow global continuation to complement the status quo and exemplify them through a plethora of representative examples. Global continuation is also implemented as open source software in DynamicalSystems.jl, enhancing its accessibility.
arXiv:2607.09520v1 Announce Type: new
Abstract: Vision-Language Models (VLMs) are the perceptual backbone of embodied AI, but their energy footprint on edge hardware remains poorly understood. Existing efficiency efforts focus predominantly on reducing visual tokens, implicitly treating visual processing as the dominant energy cost. We overturn this implicit assumption through the first systematic energy profiling of on-device VLM inference, spanning five models across three architecture families, four input resolutions, and two hardware platforms (NVIDIA RTX 3070 and Jetson Orin NX). Our analysis yields three findings. First, average inference power is a model-intrinsic constant, invariant to input resolution, image complexity, and prompt type, with less than 5% variation across all conditions. This means that all energy variation across inputs must arise from variation in inference time, not from variation in power draw. Second, each output token costs 11 to 39x more wall-clock time than each input token due to the compute-bound and memory-bound asymmetry between prefill and decode, making output token count the dominant driver of both latency and energy. Third, image complexity, measured by the number of objects in an image, induces up to 4.1x energy differences at identical resolution. This variation arises not from increased visual processing cost, but from differences in output length. These findings expose a fundamental limitation of visual token pruning: even removing all visual tokens saves at most 10% of total energy for fixed-token models. Across models spanning 1 billion to 8 billion parameters, controlling output length saves up to 97% of total energy, with the energy dominance of decoding growing stronger at larger model scale. In short, the true energy bottleneck in edge VLM inference is not what the model sees, but how much it says.
arXiv:2607.09646v1 Announce Type: cross
Abstract: The Erd\H{o}s--P\'osa theorem asserts that every graph $G$ with no $k$ disjoint cycles contains a set $X$ of $f(k)$ vertices such that $G\setminus X$ has no cycle. Robertson and Seymour showed that this Erd\H{o}s--P\'osa property also holds for $H$-minor models of any planar graph $H$. Equivalently, if $G$ has no $k$ minor models of $H$ pairwise at distance at least 1 (i.e. disjoint), then one can remove $f(k,H)$ balls of radius 0 (i.e. vertices) to make the graph $H$-minor free.
We show that this coarse graph theory point of view generalizes to distance at least 2 versus radius 1 balls, yielding the induced Erd\H{o}s--P\'osa property for planar minors. Namely, every graph $G$ which does not contain $k$ pairwise non-adjacent minor models of a planar graph $H$ (we say that $G$ is $kH$-free) can be made $H$-minor free by removing $f(k,H)$ neighborhoods. The proof relies on the fact that sparse $kH$-free graphs have linearly many independent large protrusions. The same method gives that sparse $kH$-free graphs can be made $H$-minor free by deleting $O(\log n)$ vertices (and thus have logarithmic tree-width). This gives a quasi-polynomial algorithm for the Maximum Independent Set problem for $kH$-free graphs.
arXiv:2607.09016v1 Announce Type: new
Abstract: Agent skills extend LLM agents with reusable procedures, tools, and domain-specific workflows, but their safety depends on resolving dependencies among interacting instructions. We introduce SkillLogic, a framework for analyzing logical relations in skill files and constructing executable tests from them. Our taxonomy covers eight relation types, including preconditions that gate valid actions, constraints that limit how allowed actions may be performed, and fallbacks that specify recovery behavior after failure. Using SkillLogic, we scan over 5000 public skills and find that 70% contain at least one logical relation. We then construct SLBench, an 86-case executable benchmark from high-confidence, high-impact, and locally testable relations. Evaluating Codex and Claude Code across six LLM backbones shows unsafe rates up to 70%, with violations leading to privacy leaks, unsafe configuration changes, and incomplete cleanup. The human audit attributes failures to both agent capability gaps and low-salience skill text. We further show that SLGuard, a lightweight inference-time scaffold, reduces violations by 63% on targeted cases. Our results establish logical-relation following as a distinct reliability challenge for skill-guided agents.
arXiv:2607.09291v1 Announce Type: new
Abstract: Medieval document transcribers have very different practices; on top of that, heterogeneous digitization policies have resulted in corpora where the character-set must be viewed as fluid. In this paper we address the problem of changing between character-sets in a flexible manner. We focus on one-to-one character mappings and train characterlevel one-to-one RNNs to undo them with self-supervision; recovering half the CER even with 20 text lines. We analyse the use of these one-to-one networks for HTR post-correction and we see that they obtain significant improvements while totally ignoring ins-dels. We then use the exact same networks with character-level alignment groundtruth compiled from parallel corpora in a training and inference mode we call Banded RNNs. We use such networks to successfully expand abbreviations in medieval charter transcriptions. Finally we introduce an elaborate heuristic which takes the characters of two arbitrary character-sets and defines a metric encapsulating what we consider to be semantic similarity of characters. We call the construction of such mappings letter lemmatization and present a rich Python library that efficiently performs all presented methods.
arXiv:2605.13929v2 Announce Type: replace
Abstract: Quantum computers promise exponential speedups for problems in cryptography, chemistry, and optimization. Realizing this promise requires fault tolerance: physical qubits are noisy, so logical qubits must be encoded redundantly across many physical ones using quantum error-correcting codes. In most practical fault-tolerance schemes, T gates cannot be implemented transversally and instead require costly magic-state distillation protocols involving a complex set of operations. As a result, T-gate count can dominate the resource budget of large-scale quantum computations, making T-count minimization a central bottleneck on the path to quantum advantage. Existing T-count optimization tools, however, do not scale to the circuits that quantum advantage demands.
We present theoretical and practical results on T-gate optimization. On the theoretical side, we give a linear-time randomized algorithm for phase folding, based on a novel randomized static analysis. Our static analysis soundly approximates the set of reachable quantum states with an arbitrarily high probability. Our key insight is a static analysis that does not track symbolic expressions, but propagates constant-width bitstrings down the circuit. On the practical side, our implementation, TZAP, is multiple orders of magnitude faster than state-of-the-art tools -- such as PyZX, VOQC, and Feynman -- closely matches their T-count reductions on standard benchmarks, and within seconds on a laptop computer can optimize circuits with millions of gates.
arXiv:2607.09018v1 Announce Type: new
Abstract: Image-generation AI systems increasingly support creative work by producing multiple design variations for users to evaluate and select. In such human-AI co-creation workflows, selection becomes a critical stage where human judgment guides AI-generated possibilities toward final outcomes. While presenting multiple alternatives is intended to encourage exploration, the simultaneous multi-option presentation may introduce systematic biases in human decision making. Drawing on ensemble perception theory, we investigate whether these interfaces induce central tendency bias-the tendency to favor options closer to the center of a design set. We conducted a controlled experiment manipulating the variance of design sets (high vs. low) and measured participants' selections in both aesthetic preference and representativeness tasks. Results show that higher variance increases the selection of center-proximal designs across both tasks. These findings suggest that multi-variation interfaces in image-generation AI systems may constrain selection diversity, revealing a potential tension between diversity in generated outputs and diversity in human selection outcomes.
arXiv:2607.09022v1 Announce Type: new
Abstract: Developers' choices about what data a system collects, how it is used and shared, and what defaults govern user choices directly shape users' privacy experiences. Yet, developers often make problematic privacy-related design decisions without realizing the potential consequences. We introduce Privacy Detective, a narrative investigation game that leverages real-world legal documents to train developers' privacy awareness. In the game, players search for privacy violation evidence derived from legal documents and organize this evidence into privacy violation reports using curated templates. We evaluated Privacy Detective in a between-subjects study with student developers, comparing it against a baseline in which participants read raw FTC legal documents. Participants in the game condition identified more true violations than the baseline group, flagged fewer non-issues, and provided more complete justifications for the violations they reported.
arXiv:2607.07047v2 Announce Type: replace
Abstract: Understanding the geometric structure of pre-trained language model embeddings matters for interpretability and safety. We ask whether sentence-level classification signal lives in the Riemannian geometry of contextual token embeddings, and probe it by extracting per-token pullback metrics from a learned encoder's analytical Jacobian and aggregating them with the Fr\'echet mean on the symmetric positive definite (SPD) manifold; we call this procedure Riemannian Mean Pooling (RMP). Across three datasets with non-trivial linguistic structure (CoLA, CREAK, RTE), RMP outperforms Euclidean mean pooling, while on FEVER-Symmetric, a benchmark constructed to remove annotation-driven lexical artifacts, the method correctly stays at chance. Ablations show that a randomly initialised encoder combined with Fr\'echet aggregation already beats Euclidean pooling on two of the three signal-bearing datasets, localising the source of the gain to the geometric aggregation rather than to learned manifold structure; the trained encoder contributes additional signal specifically on CREAK, the most knowledge-heavy of the three signal-bearing datasets.
arXiv:2605.31271v2 Announce Type: replace
Abstract: Driving Vision-Language-Action Models (Driving VLAs) aim to use language to improve end-to-end planning, but the language-action gap limits this promise. We propose DriveMA, a Driving VLA framework built on verifiable meta-actions, which summarize future ego motion into compact language-domain intentions and can be constructed from expert trajectories with a trajectory-grounded annotation pipeline and can be verified against generated trajectories through rule-based projection. DriveMA exploits this verifiability with action-centric supervised training and a data-efficient turn-level credit assignment reinforcement learning framework, explicitly aligning high-level decisions with low-level trajectory planning through dense rewards and precise credit assignment. DriveMA sets a new state of the art on the Waymo Open Dataset Vision-based E2E Driving, achieving a Rater Feedback Score of 8.060 with a 2B model and further improving it to 8.079 with a 4B model; it also obtains competitive closed-loop planning performance on NAVSIM. These results show that even a simple meta-action interface can achieve state-of-the-art planning when made verifiable and optimized for language-action alignment. Code, data, and models are available at https://tsinghua-mars-lab.github.io/DriveMA.
arXiv:2607.07671v2 Announce Type: replace
Abstract: Probabilistic circuits (PCs) can model complex joint distributions while supporting exact and efficient computation of many inference queries. However, standard likelihood-based PC learning is vulnerable to overfitting and fragile generalization when confronted with data noise, small sample sizes, or distribution shifts. This can be mitigated using distributionally-robust optimization which consider worst-case distributions within a Wasserstein ball of the empirical distribution, but current methods are limited to training a model from scratch in this framework. Instead, we propose PeTeR: a novel, data-free post-training framework designed to robustify pre-trained PCs against distribution shifts without retraining from scratch. Empirical evaluations across multiple density estimation benchmarks demonstrate that PeTeR effectively robustifies baseline models against both random and adversarial perturbations, achieving competitive or superior performance to data-dependent robust learning baselines.
arXiv:2603.20388v2 Announce Type: replace-cross
Abstract: We derive the asymptotic risk function of regularized empirical risk minimization (ERM) estimators tuned by $n$-fold cross-validation (CV). The out-of-sample prediction loss of such estimators converges in distribution to the squared-error loss (risk function) of shrinkage estimators in the normal means model, tuned by Stein's unbiased risk estimate (SURE). This risk function provides a more fine-grained picture of predictive performance than uniform bounds on worst-case regret, which are common in learning theory: it quantifies how risk varies with the true parameter. As key intermediate steps, we show that (i) $n$-fold CV converges uniformly to SURE, and (ii) while SURE typically has multiple local minima, its global minimum is generically well separated. Well-separation ensures that uniform convergence of CV to SURE translates into convergence of the tuning parameter chosen by CV to that chosen by SURE.