arXiv:2602.19973v4 Announce Type: replace
Abstract: Shallow embeddings that use monads to represent effects are popular in proof-oriented languages because they are convenient for formal verification. Once shallowly embedded programs are verified, they are often extracted to mainstream languages like OCaml or C and linked into larger codebases. The extraction process is not fully verified because it often involves quotation -- turning the shallowly embedded program into a deeply embedded one -- and verifying quotation remains a major open challenge. Instead, some prior work obtains formal correctness guarantees using translation validation to certify individual extraction results. We build on this idea, but limit the use of translation validation to a first extraction step that we call relational quotation and that uses a metaprogram to construct a typing derivation for the given shallowly embedded program. This metaprogram is simple, since the typing derivation follows the structure of the original program. Once we validate that the typing derivation is valid for the original program, we pass it to a verified syntax-generation function that produces code guaranteed to be semantically related to the original program.
We apply this general idea to build SEIO*, a framework for extracting shallowly embedded F* programs with IO and refinement types to a deeply embedded simply typed lambda-calculus while providing formal secure compilation guarantees. Using two cross-language logical relations, we devise a machine-checked proof in F* that SEIO* guarantees Robust Relational Hyperproperty Preservation (RrHP), a very strong secure compilation criterion that implies full abstraction as well as preservation of trace properties and hyperproperties against arbitrary linked adversarial code. This goes beyond the state of the art in verified and certifying extraction, which so far has focused on correctness rather than security.
Science Journals
arXiv:2602.23659v2 Announce Type: replace
Abstract: In an age of financial system digitisation and the increasing adoption of digital currencies, Central Bank Digital Currencies (CBDCs) have emerged as a focal point for technological innovation. Privacy compliance has become a key factor in the successful design of CBDCs, extending beyond technical requirements to influence legal requirements, user trust, and security considerations. Implementing Privacy-Enhancing Technologies (PETs) in CBDCs requires an interdisciplinary approach, however, the lack of a common understanding of privacy and the essential technological characteristics restricts progress. This work investigates: (1) How privacy can be defined within the framework of CBDCs and what implications does this definition have for CBDCs design? and (2) Which PETs can be employed to enhance privacy in CBDC design? We propose a comprehensive definition for privacy that is mapped to the cryptographic landscape for feature implementation. The research is validated against case studies from 20 current CBDCs. The study shows that comprehensive privacy can be designed in the proposal stage, but that privacy does not reach the launched version of the CBDC. A failure analysis of abandoned privacy pilots identifies four root causes of this research-to-launch gap: regulatory visibility requirements, computational overhead, liability allocation, and institutional incentives.
arXiv:2602.24009v4 Announce Type: replace
Abstract: Jailbreak techniques for large language models (LLMs) evolve faster than benchmarks, making robustness estimates stale and difficult to compare across papers due to drift in datasets, harnesses, and judging protocols. We introduce JAILBREAK FOUNDRY (JBF), a system that addresses this gap via a multi-agent workflow to translate jailbreak papers into executable modules for immediate evaluation within a unified harness. JBF features three core components: (i) JBF-LIB for shared contracts and reusable utilities; (ii) JBF-FORGE for the multi-agent paper-to-module translation; and (iii) JBF-EVAL for standardizing evaluations. Across 30 reproduced attacks, JBF achieves high fidelity with a mean (reproduced-reported) attack success rate (ASR) deviation of +0.26 percentage points. By leveraging shared infrastructure, JBF reduces attack-specific implementation code by more than half relative to original repositories and achieves an 82.5% mean reused-code ratio. This system enables a standardized AdvBench evaluation of all 30 attacks across 10 victim models using a consistent GPT-4o judge. By automating both attack integration and standardized evaluation, JBF offers a scalable solution for creating living benchmarks that keep pace with the rapidly shifting security landscape.
arXiv:2602.24049v2 Announce Type: replace
Abstract: An automated algorithm to construct island divertors for stellarators is presented and is used to find divertors that meet heat load requirements determined by material limits. The algorithm uses just two initial conditions: two starting coordinates on the island separatrix chosen by the user. We leverage the simplicity of the algorithm to explore the divertor parameter space in a fixed magnetic equilibrium. Heat loads are approximated using the field line diffusion model implemented in the \texttt{FLARE} code. Divertor solutions that satisfy heat load requirements while maintaining a high power fraction captured are found using a parameter scan and a Bayesian optimization routine. The optimization finds divertors that perform the same as the parameter scan, but with a 95\% reduction in computational cost. The resulting divertors satisfy heat load requirements across varying cross-field heat diffusivities. Optimization over various islands in the equilibrium shows that low-elongation islands are the easiest to find divertors that satisfy heat flux requirements. This algorithm presents a simple parameterization for island divertors and facilitates further physics and optimization studies.
arXiv:2603.01875v3 Announce Type: replace
Abstract: Knowledge distillation (KD) is an essential technique to compress large language models (LLMs) into smaller ones. However, despite the distinct roles of the student model and the teacher model in KD, most existing frameworks still use a homogeneous training backend (e.g., FSDP and DeepSpeed) for both models, leading to suboptimal training efficiency. In this paper, we present a novel framework for LLM distillation, termed \textbf{KDFlow}, which features a decoupled architecture and employs SGLang for teacher inference. By bridging the training efficiency of FSDP2 and the inference efficiency of SGLang, KDFlow achieves full utilization of both advantages in a unified system. Moreover, instead of transferring full logits across different processes, our framework only transmits the teacher's hidden states using zero-copy data transfer and recomputes the logits on the student side, effectively balancing the communication cost and KD performance. Furthermore, our framework supports both off-policy and on-policy distillation and incorporates KD algorithms for cross-tokenizer KD through highly extensible and user-friendly APIs. Experiments show that KDFlow can achieve \textbf{1.44$\times$ to 6.36$\times$} speedup compared to current KD frameworks, enabling researchers to rapidly prototype and scale LLM distillation with minimal engineering overhead. Code is available at: https://github.com/songmzhang/KDFlow
arXiv:2603.02142v2 Announce Type: replace
Abstract: Scaling laws assume larger models trained on more data consistently outperform smaller ones -- an assumption that drives model selection in computer vision but remains untested in resource-constrained Earth observation (EO). We conduct a systematic efficiency analysis across three scaling dimensions: model size, dataset size, and input resolution, on rooftop photovoltaic (PV) detection in Madagascar, yielding 180 training runs across 60 configurations. Optimizing for model efficiency (mAP$_{50}$ per unit of model size), we find a consistent efficiency inversion: YOLO11N achieves the highest efficiency ($22\times$ higher than YOLO11X) with no accuracy penalty: it reaches the second highest absolute mAP$_{50}$ (0.459), outperforming all models except YOLO11S by a margin smaller than run-to-run variance, directly contradicting the scaling prior. Resolution is the dominant resource allocation lever: moving from 416 px to 1280 px at 10 % of the data matches the efficiency gain of collecting the full dataset at low resolution. These findings are robust to the deployment objective: small high-resolution configurations are Pareto-dominant across all 60 experimental setups in the joint accuracy-throughput space, leaving no tradeoff to resolve. In data-scarce EO, the scaling prior does not just fail: it inverts.
arXiv:2603.03537v2 Announce Type: replace
Abstract: Rapid acceleration and burst maneuvers in underwater robots depend less on maintaining precise resonance and more on force--velocity phase alignment during thrust generation. In this work, we investigate constrained-layer damping (CLD) as a passive mechanism for frequency-selective impedance shaping in soft robotic swimmers. Unlike conventional stiffness-tuning approaches, CLD selectively amplifies the dissipative component of bending impedance while preserving storage stiffness, passively shifting the impedance composition toward dissipative dominance as actuation frequency increases. We characterize this behavior through dry impedance measurements, demonstrate that CLD enhances thrust and alters force--motion phase relationships across Strouhal numbers in constrained propulsion tests, and validate that passive impedance shaping yields a nearly five-fold increase in peak acceleration and a three-fold increase in terminal velocity in unconstrained swimming trials. These results establish phase-oriented passive impedance modulation as a simple, control-free pathway for improving transient propulsion in soft robotic systems.
arXiv:2603.03751v2 Announce Type: replace
Abstract: Cooperative object transport in unstructured environments remains challenging for assistive humanoids because strong, time-varying interaction forces can make tracking-centric whole-body control unreliable, especially in close-contact support tasks. This paper proposes a bio-inspired, interaction-oriented whole-body control (IO-WBC) that functions as an artificial cerebellum - an adaptive motor agent that translates upstream (skill-level) commands into stable, physically consistent whole-body behavior under contact. This work structurally separates upper-body interaction execution from lower-body support control, enabling the robot to maintain balance while shaping force exchange in a tightly coupled robot-object system. A trajectory-optimized reference generator (RG) provides a kinematic prior, while a reinforcement learning (RL) policy governs body responses under heavy-load interactions and disturbances. The policy is trained in simulation with randomized payload mass/inertia and external perturbations, and deployed via asymmetric teacher-student distillation so that the student relies only on proprioceptive histories at runtime. Extensive experiments demonstrate that IO-WBC maintains stable whole-body behavior and physical interaction even when precise velocity tracking becomes infeasible, enabling compliant object transport across a wide range of scenarios.
arXiv:2607.16143v1 Announce Type: new
Abstract: This paper presents a new implementation of the NeoSLAM algorithm. The proposed version is a complete rewrite of NeoSLAM into a modular architecture using modern frameworks that, together, enable real-time execution with minimal discarding of input data. This work also provides a comparative evaluation between NeoSLAM and RatSLAM across three datasets under varying environmental conditions. The experimental results highlight differences in mapping consistency and trajectory reconstruction, demonstrating the effectiveness and practical applicability of the proposed ROS2-based implementation. The results indicate that the new NeoSLAM outperforms the original in terms of processing throughput for real-time applications and achieves comparable performance to RatSLAM in terms of map reconstruction across the evaluated datasets.
arXiv:2607.16146v1 Announce Type: new
Abstract: Vision and touch are complementary modalities essential for robotic perception and manipulation. While vision provides global object context, touch offers precise local information at contact points. Integrating these modalities for contact localization, i.e., predicting the location of touch on an object's surface, poses significant challenges due to the need for accurate spatial alignment between tactile data and visual geometry. To address this challenge, we propose VTLoc, a novel visual-tactile framework that localizes contact points from tactile readings using a 3D point cloud as visual input. VTLoc introduces two key components: a geometric multi-modal alignment module, which reconstructs a pseudo-point cloud from fused visual-tactile features and aligns it with the visual point cloud to enforce spatial consistencies across modalities; and an iterative localizing updater, which iteratively refines the predicted contact location using fused visual-tactile features. Evaluated on a new benchmark of 100 real-world objects, VTLoc improves single-touch contact localization by reducing local-to-global correspondence ambiguity.
arXiv:2607.16139v1 Announce Type: cross
Abstract: The proliferation of Internet of Things (IoT) devices and wearable health monitors has created an urgent demand for ultra-low-power analog-to-digital converters (ADCs). Successive approximation register (SAR) ADCs are widely used in such applications, yet their energy efficiency remains constrained by the sequential bit-by-bit switching of the capacitive DAC (CDAC). The high-weight most significant bit (MSB) transitions dominate the total switching energy, and the rigid N -cycle conversion flow imposes a hard lower bound on latency per sample.This paper presents a Kalman filter-assisted data-predictive SAR ADC that replaces the first four comparator-driven decisions with a recursive state estimator. The Kalman filter predicts the 4 MSBs from the complete conversion history before each cycle begins, enabling simultaneous parallel switching of the MSB capacitors. This eliminates redundant CDAC transitions, shortens the quantization cycle by four clock periods, and reduces switching energy by approximately 50%. An optimized 4-bit MSB switching scheme further suppresses residual switching at the hardware level. The ADC, designed in a 180-nm CMOS process, supports configurable dual-mode operation, toggling between a conventional mode and the Kalman-driven predictive mode for robustness under erratic inputs. At 20 MS/s and a 1.8-V supply, the predictive mode reduces total power consumption by 50.3% (from 1.96 mW to 0.975 mW), with a measured SNR/SFDR of 57.88/74.51 dB at 504 kHz, confirming its suitability for energy-constrained wireless sensor networks.
arXiv:2607.15866v1 Announce Type: new
Abstract: Generative AI (GenAI) works via goal-directed computation, which differs fundamentally from human creative processes. This poses challenges for the intelligent support of creative experiences. We propose ``breakdowns'' as opportunities for the exchange of perspectives between human and machine. Breakdowns disrupt a flow and force us to consciously evaluate our ``being-in-the-world''. Between human and machine, breakdowns can function as openings for collaborative creative reflection. We are currently studying human-human creative interactions, to identify the markers of these inter-subjective openings, and to understand how they are used in a co-creative process. We present preliminary findings on breakdowns as a design principle for creativity support, prioritising human creative agency and meaningful reflection over automated content generation.
arXiv:2603.08337v2 Announce Type: replace
Abstract: Optimizing asset exchanges on blockchain-driven platforms poses a novel and challenging graph query optimization problem. In this model, assets represent vertices and exchanges form edges, recasting the graph query task as a routing problem over a large-scale, dynamic graph. However, the existing solutions fail to solve the problem efficiently due to the non-linear nature of the edge weights defined by a concave swap function. To address the challenge, we propose PRIME, a two-stage iterative graph algorithm designed for the Token Graph Routing Problem (TGRP). The first stage employs a pruned graph search to efficiently identify a set of high-potential routing paths. The second stage formulates the allocation task as a strongly convex optimization problem, which we solve using our novel Adaptive Sign Gradient Method (ASGM) with a linear convergence rate. Extensive experiments on real-world Ethereum data confirm PRIME's advantages over industry baselines. PRIME consistently outperforms the widely-used Uniswap routing algorithm, achieving up to 8.42 basis points (bps) better execution prices on large trades while reducing computation up to 96.7%. The practicality of PRIME is further validated by its deployment in hedge fund production environments, demonstrating its viability as a scalable graph query processing solution for high-frequency decentralized markets.
arXiv:2603.11593v2 Announce Type: replace
Abstract: Instruction-based image editing aims to modify specific content within existing images according to user-provided instructions while preserving non-target regions. Beyond traditional object- and style-centric manipulation, text-centric image editing focuses on modifying, translating, or rearranging textual elements embedded within images. However, existing leading models often struggle to execute complex text editing precisely, frequently producing blurry or hallucinated characters. We attribute these failures primarily to the lack of specialized training paradigms tailored for text-centric editing, as well as the absence of large-scale datasets and standardized benchmarks necessary for a closed-loop training and evaluation system. To address these limitations, we present WeEdit, a systematic solution encompassing a scalable data construction pipeline, two benchmarks, and a tailored two-stage training strategy. Specifically, we propose a novel HTML-based automatic editing pipeline, which generates 330K training pairs covering diverse editing operations and 15 languages, accompanied by standardized bilingual and multilingual benchmarks for comprehensive evaluation. On the algorithmic side, we employ glyph-guided supervised fine-tuning to inject explicit spatial and content priors, followed by a multi-objective reinforcement learning stage to align generation with instruction adherence, text clarity, and background preservation. Extensive experiments demonstrate that WeEdit outperforms previous open-source models by a clear margin across diverse editing operations.
arXiv:2603.12134v2 Announce Type: replace
Abstract: Magnetic relaxation drives plasma toward lower-energy equilibria under helicity constraints. In ideal magnetohydrodynamics (MHD), helicity is locally conserved, while resistive theories such as Taylor relaxation preserve only global helicity. This distinction has important implications for structure-preserving numerical methods. We compare three finite element formulations: an unconstrained scheme that does not conserve helicity, a mixed method based on finite element exterior calculus that preserves discrete local helicity on magnetically closed subdomains, and a Lagrange multiplier approach that enforces only global helicity conservation. Numerical experiments with magnetic knots and braids show that helicity-based constraints provide effective topological barriers when the relevant helicity-type invariant is nonzero, but do not fully characterize braided field-line topology when it vanishes. These results clarify both the strengths and the possible limitations of helicity-based structure-preserving finite element methods for magnetic relaxation.
arXiv:2607.16157v1 Announce Type: new
Abstract: Shallow-water ocean forcing provides strong broadband passive Scholte-wave illumination along a seabed distributed acoustic sensing cable on the Texas Gulf Coast. We exploit this illumination through direct processing of the uncorrelated wavefield, long-duration frequency-wavenumber stacking, and a multi-aperture strategy that preserves high-frequency spatial localization while resolving low-frequency modes. Constant-wavenumber slices enable spatially consistent multimode tracking across successive array centers. The resulting dispersion spans approximately 0.3-4.5 Hz and supports 400 one-dimensional inversions along a 51-km cable segment. These profiles form a pseudo-2D shear-wave velocity model extending from the near seafloor to approximately 2 km depth. Broad shallow low-velocity intervals are consistent with softer incised-valley fill within stiffer Pleistocene deposits. The results demonstrate that existing seabed fiber infrastructure, when coupled with strong shallow-water passive illumination, can deliver broadband regional shear-wave imaging, while higher-frequency active sources remain necessary to resolve the uppermost several meters.
arXiv:2607.16161v1 Announce Type: new
Abstract: Modern GPU-accelerator platforms rely on multi-layer self-healing pipelines that span hardware, firmware, management software, and orchestration. When faults propagate across layer boundaries, they can bypass detection, corrupt diagnosis, or trigger conflicting remediations--yet conventional fault-injection campaigns test each layer in isolation. We present ADA-ST, an adaptive fault-injection methodology that uses a weighted fault-propagation graph to guide cross-layer scenario selection. We construct four-layer graphs for three successive platforms at a hyperscale operator: Alpha, Beta, and Gamma. Platform Alpha, a production system that accumulated 72,550 repair tickets over four years, provides the empirical foundation; 49% of those tickets involve cross-layer fault propagation. We show that existing static test campaigns cover only 20-25% of the modeled fault-propagation edges, leaving approximately three-quarters of the cross-layer attack surface unexercised. ADA-ST closes this gap through iterative, activity-guided scenario selection that maximizes marginal coverage gain per iteration, reaching full edge coverage within 10 iterations on Alpha, 12 on Beta, and 9 on Gamma. The Fault-Layer Abstraction Mapping (FLAM) transfers propagation knowledge across hardware generations with 100% fidelity from Alpha to Beta and 96% from Beta to Gamma. Physical spot-validation on the newest platform confirms all four tested propagation edges, revealing cross-layer vulnerabilities spanning telemetry blind spots, absence-based detection gaps, multi-signal correlation failures, and trust-without-verification propagation at the L2-to-L3 boundary.
arXiv:2512.10113v3 Announce Type: replace
Abstract: Dark personality traits have long been associated with antisocial and toxic online behaviors, yet their relationship with observable online activity remains unclear. We investigate the association between validated dark personality measures, self-reported experiences of online incivility, and linguistic and behavioral features extracted from real-world user activity. To this end, we developed a Web application that securely links responses to validated psychological questionnaires collected via Amazon Mechanical Turk with participants' Reddit activity. This yielded a dataset of nearly 57K comments (2.2M tokens) from 114 users, represented through a broad set of linguistic and behavioral features. Our analyses reveal a clear distinction between self-reported and observed behavior. Dark personality traits show consistent associations with self-reported engagement in uncivil interactions. However, no validated dark personality dimension significantly predicts text-derived toxicity or linguistic features. In contrast, self-reported experiences of engaging in or being targeted by toxic behavior are robustly reflected in users' language, exhibiting consistent associations with measures of negativity, moral framing, and emotional intensity. Taken together, these findings highlight a gap between stable personality traits and their manifestation in surface-level linguistic signals. While computational features effectively capture behavioral engagement in online incivility, they do not provide reliable proxies for underlying personality constructs within the present framework. Our results underscore the importance of grounding computational approaches in validated psychological measures and point to the need for richer, context-aware representations to better understand the relationship between personality and online behavior.
arXiv:2603.14160v2 Announce Type: replace
Abstract: In this paper, we propose a novel framework that allows therapists to teach robot-assisted rehabilitation exercises remotely via RGB-D video. Our system encodes demonstrations as 6-DoF body-centric trajectories using Cartesian Dynamic Movement Primitives (DMPs), ensuring accurate posture-independent spatial generalisation across diverse patient anatomies. Crucially, we execute these trajectories through a decoupled hybrid control architecture that constructs a spatially compliant virtual tunnel, paired with an effort-based temporal dilation mechanism. This architecture is applied to three distinct rehabilitation modalities: Passive, Active-Assisted, and Active-Resistive, by dynamically linking the exercise's execution phase to the patient's tangential force contribution. To guarantee safety, a Gaussian Mixture Regression (GMR) model is learned on-the-fly from the patient's own limb. This allows the detection of abnormal interaction forces and, if necessary, reverses the trajectory to prevent injury. Experimental validation demonstrates the system's precision, achieving an average trajectory reproduction error of 3.7cm and a range of motion (ROM) error of 5.5 degrees. Furthermore, dynamic interaction trials confirm that the controller successfully enforces effort-based progression while maintaining strict spatial path adherence against human disturbances.
arXiv:2603.16348v2 Announce Type: replace
Abstract: AI coding assistants are fundamentally reshaping software development by shifting developers' effort from writing code toward specifying intent through natural language prompts. In emerging chat-based development practices such as vibe coding, prompts mediate the transformation of human intent into executable software. While Requirements Engineering (RE) emphasizes capturing, validating, and evolving requirements, current prompting practices remain informal and ad hoc.
In this vision paper, we argue that prompts represent lightweight, evolving requirements artifacts that combine expressions of user needs with varying degrees of solution guidance. We use an existing conceptual model that decomposes prompts into three interrelated dimensions: Functionality and Quality (capturing intended system requirements), General Solutions (capturing architectural strategies and technology choices), and Specific Solutions (capturing implementation-level constraints and directives).
Building on this conceptualization, we formulate four research hypotheses concerning (i) the evolution of prompts over time, (ii) the influence of user characteristics on prompt evolution, (iii) the relationship between prompt content and requirements validation and verification activities, and (iv) the impact of prompt characteristics on requirements and resulting software quality. We envision an empirical research agenda combining real-world AI-assisted development data, corpus analysis, and controlled experimentation to investigate these hypotheses and derive evidence-based practices for requirements-aware prompt engineering. By reframing prompts through the lens of RE, we position prompting not merely as an interaction mechanism with AI systems, but as a central software engineering concern requiring systematic study.
arXiv:2603.19235v3 Announce Type: replace
Abstract: While Multimodal Large Language Models demonstrate impressive semantic capabilities, they often suffer from spatial blindness, struggling with fine-grained geometric reasoning and physical dynamics. Existing solutions typically rely on explicit 3D modalities or complex geometric scaffolding, which are limited by data scarcity and generalization challenges. In this work, we propose a paradigm shift by leveraging the implicit spatial prior within large-scale video generation models. We posit that to synthesize temporally coherent videos, these models inherently learn robust 3D structural priors and physical laws. We introduce VEGA-3D (Video Extracted Generative Awareness), a plug-and-play framework that repurposes a pre-trained video diffusion model as a Latent World Simulator. By extracting spatiotemporal features from intermediate noise levels and integrating them with semantic representations via a token-level adaptive gated fusion mechanism, we enrich MLLMs with dense geometric cues without explicit 3D supervision. Extensive experiments across 3D scene understanding, spatial reasoning, and embodied manipulation benchmarks demonstrate that our method outperforms state-of-the-art baselines, validating that generative priors provide a scalable foundation for physical-world understanding. Code is publicly available at https://github.com/H-EmbodVis/VEGA-3D.
arXiv:2603.19464v3 Announce Type: replace
Abstract: Robotic path planning problems are often NP-hard, and practical solutions typically rely on approximation algorithms with provable performance guarantees for general cases. While designing such algorithms is challenging, formally proving their approximation optimality is even more demanding, which requires domain-specific geometric insights and multi-step mathematical reasoning over complex operational constraints. Recent Large Language Models (LLMs) have demonstrated strong performance on mathematical reasoning benchmarks, yet their ability to assist with research-level optimality proofs in robotic path planning remains under-explored. In this work, we introduce the first benchmark for evaluating LLMs on approximation-ratio proofs of robotic path planning algorithms. The benchmark consists of 34 research-grade proof tasks spanning diverse planning problem types and complexity levels, each requiring structured reasoning over algorithm descriptions, problem constraints, and theoretical guarantees. Our evaluation of state-of-the-art proprietary and open-source LLMs reveals that even the strongest models struggle to produce fully valid proofs without external domain knowledge. However, providing LLMs with task-specific in-context lemmas substantially improves reasoning quality, a factor that is more effective than generic chain-of-thought prompting or supplying the ground-truth approximation ratio as posterior knowledge. We further provide fine-grained error analysis to characterize common logical failures and hallucinations, and demonstrate how each error type can be mitigated through targeted context augmentation.
arXiv:2603.26556v2 Announce Type: replace
Abstract: Converting a pretrained Transformer into a more efficient hybrid model through distillation offers a promising approach to reducing inference costs. However, achieving high-quality generation in distilled models requires careful joint design of both the student architecture and the distillation process. Many prior distillation works evaluate downstream multiple-choice benchmarks by ranking candidate answers with log-likelihood rather than requiring autoregressive generation, which can obscure important differences in model quality. For example, on overlapping benchmarks, we show that a 7B distilled model that nearly matches its teacher to within 0.2 pp under log-likelihood scoring falls behind by 20.8 pp when it must generate answers autoregressively.
We investigate this phenomenon with GenDistill, a multi-stage pipeline we designed for distilling a pretrained Transformer into an efficient Hybrid Kimi Delta Attention (Hybrid-KDA) student. Using it as a controlled testbed on Qwen3-0.6B, we systematically ablate six design axes (training objective, loss masking, training duration, dataset selection, parameter freezing, and architecture choice) and evaluate every choice under both log-likelihood and generation-based protocols. We find that log-likelihood-based evaluation consistently underestimates the gap between teacher and student, and can in some cases reverse the ranking of design choices, so conclusions drawn from perplexity-only evaluation may be misleading. Among the factors we study, dataset selection, completion-only masking, and freezing attention layers during post-training have the largest impact on generation quality.
Our best distillation recipe, using a Hybrid-KDA model as the student, retains 86-90% of teacher accuracy on knowledge benchmarks while reducing KV cache memory by up to 75% and improving time-to-first-token by 2-4x at 128K-token contexts.
arXiv:2603.27556v2 Announce Type: replace
Abstract: Open-Vocabulary Object Detection (OVOD) has achieved remarkable success in generalizing to novel categories. However, this success often rests on the implicit assumption of domain stationarity. In this work, we revisit the OVOD paradigm and study a key vulnerability: the fragile coupling between visual manifolds and textual embeddings under distribution shifts. We first formulate Domain-Generalized Open-Vocabulary Object Detection (DG-OVOD) as an evaluation protocol for open-vocabulary recognition under visual shifts. Through empirical analysis, we observe that visual shifts can destabilize the latent cross-modal space, causing novel-category visual signals to drift away from their semantic anchors. Motivated by these observations, we propose Progressive Domain-invariant Cross-modal Alignment (PICA). PICA departs from uniform training by introducing a multi-level curriculum based on ambiguity and signal strength. It constructs a quality-adjusted curriculum over pseudo-word prototypes, refined by sample reliability and visual consistency, to encourage more stable cross-domain modality alignment. Our findings suggest that OVOD robustness under domain shifts is closely linked to the stability of the latent cross-modal alignment space. Our work provides a DG-OVOD evaluation protocol and a practical perspective on building more generalizable open-vocabulary systems beyond static laboratory conditions.
arXiv:2603.29139v3 Announce Type: replace
Abstract: Recent advances in large language models (LLMs) have enabled agentic systems to translate natural-language intent into executable scientific visualization (SciVis) tasks. Despite rapid progress, the community lacks a principled and reproducible benchmark for evaluating these emerging SciVis agents in realistic, multi-step analysis settings. We present SciVisAgentBench, a comprehensive and extensible benchmark for evaluating scientific data analysis and visualization agents. Our benchmark is grounded in a structured taxonomy spanning four dimensions: application domain, data type, complexity level, and visualization operation. It currently comprises 108 expert-crafted cases covering diverse SciVis scenarios. To enable reliable assessment, we introduce a multimodal outcome-centric evaluation pipeline that combines LLM-based judging with deterministic evaluators, including image-based metrics, code checkers, rule-based verifiers, and case-specific evaluators. We also conduct a validity study with 12 SciVis experts to examine the agreement between human and LLM judges. Using this framework, we evaluate representative SciVis agents and general-purpose coding agents to establish initial baselines and reveal capability gaps. SciVisAgentBench is designed as a living benchmark to support systematic comparison, diagnose failure modes, and drive progress in agentic SciVis. The benchmark is available at https://scivisagentbench.github.io/.