Searcharxiv⌕ Search

arXiv subjects

Min Xu

Publications and source records attributed to Min Xu.

At least 19 recordsLinked to original sources

Know When to Zoom: How Small Is Too Small for Vision-Language Models?

Online vision-language models often miss small details in large images, because the service shrinks the image, or rejects it, before the model ever sees it. A car 20 pixels wide in an 8000 by 6000 drone photo reaches Claude about 6 pixels wide, and OpenAI's default rejects the photo outright. We measure how much detail survives for seven models from four vendors, using an eye exam of printed codes and 601 questions about small objects in real photos. The smallest text a model reads half the time grows from 6 pixels or less on a small image to a predicted 10 to 13 pixels on a 7680 by 4320 photo at default settings. Zooming wins much of the detail back. Splitting each photo into up to nine tiles raises accuracy by 11.1 to 14.5 percentage points for Claude and Gemini, and by avoiding rejections it lifts GPT-6.1 Sol from 50.7% to 83.7%. A two-step zoom, in which the model first names the region it needs, matches or beats tiles on Gemini and comes within 2.5 points of them on Claude, with about 40% of their image tokens. When a detail is lost, Claude and Gemini both say they cannot tell about half the time, but they fail differently in the rest: Gemini gives a confident wrong answer in 43% to 48% of these cases, Claude in only 5% to 9%. A check based on detail size, run before the call, misses most of the gain from tiles on Claude and Gemini, but Claude's own reply is a reliable trigger: tiling only after an unsure reply comes within 1.4 points of always tiling with 62% to 70% of the image tokens. We turn these findings into five rules for when an agent should zoom, and release the eye exam, every model answer and code at https://siddvoh.com/know-when-to-zoom.

cs.CV↗

Cross-Epoch Bias in SGD-based Leave-One-Out Influence: Analysis and Accumulative Correction

Training-data influence estimates support data selection and cleansing by predicting the effects of removing individual examples. The standard multi-epoch SGD-Influence Estimator (SGD-IE) aggregates contributions from a sample's repeated appearances along the recorded trajectory. When used to approximate finite deletion, it omits an interaction: later removals act on the parameter displacement accumulated from earlier removals. We introduce the Accumulative SGD-Influence Estimator (ACC-SGD-IE) to account for how earlier removals change the effect of removing the same sample again. It uses the sample's Hessian at subsequent appearances to correct the estimated parameter change. Our non-convex analysis shows that this correction removes SGD-IE's additional interaction contribution from the finite-horizon error bound: the omitted interaction produces a local error linear in the accumulated parameter change, while the remaining Taylor remainder is bounded quadratically under Lipschitz-Hessian assumptions. The bounds characterize how training propagates the remaining error. In non-convex experiments across clean and noisy tabular, text, and vision datasets, ACC-SGD-IE reduces validation-loss-change RMSE by 10.8% and improves Jaccard@10% by 15.2% on average relative to SGD-IE. These gains also translate into lower test misclassification rates after data cleansing on MNIST and CIFAR10 at the evaluated removal budgets. On noisy industrial datasets, it reduces aggregate parameter RMS error with a mean relative reduction of 53.4%. We further discuss computational tradeoffs, provide complementary theoretical analysis and experimental evaluation for strongly convex objectives, and explore potential extensions of the accumulative correction mechanism to other influence estimators.

cs.LG↗

DetouR: Plugging Association Routes into Search Agents for Multi-hop Question Answering

Search agents answer multi-hop questions by interleaving reasoning with retrieval, deciding at each step what to search next. Existing agents follow semantic routes: they infer what is missing and phrase it as a query, reaching only what the model can name. Yet the decisive page is often one the question never names, frequently just one hyperlink away from evidence already retrieved. Following such association routes reaches more gold pages than semantic search under the same budget. Exploiting them raises two challenges: obtaining the association structure without costly offline graph construction, and keeping linked evidence in the reader's limited context, where relevance- and fusion-based selection discards it. We propose DetouR, a plug-in that reuses the links arriving with retrieved pages and adds no search or LLM call. Link Rerouting replaces near-repeat queries with routes to linked pages, chosen by a lightweight learned scorer; Chain-aware Context Selection runs personalized PageRank over these links so that evidence chains reach the reader intact. Across four agents, DetouR improves EM by 4.0 points on 2Wiki and 1.4 on HotpotQA on average. We also find that heavier LLM models struggle with this non-semantic evidence, adding little over these simple selectors or even falling behind. In summary, our work reveals association routes as a cheap complement to semantic routes that current search agents overlook: added as a plug-in, they improve every agent we test, in replay and in live runs, at no extra search or LLM cost.

cs.IR↗

MacroGS: Macro-Pixel Shared Gaussian Primitives for Arbitrary-Scale Light Field Super-Resolutio

Arbitrary-scale light field spatial super-resolution reconstructs high-resolution views at any requested scale from low-resolution multi-view observations. Gaussian primitives enable continuous reconstruction, but generating separate primitive sets for each view redundantly encodes shared scene structure across views and causes the primitive count grow linearly with the number of views. To avoid such redundancy, we propose MacroGS, a feed-forward decoder that jointly predicts a single shared Gaussian set per macro pixel rather than generating separate sets for each view. Disparity-dependent replacement and angle-conditioned appearance allow each set to be reused across all output views, keeping the predicted primitive count independent of the number of views at each requested scale. The decoder iteratively predicts per-primitive disparities, retrieves the corresponding multi-view features, and updates the primitive states. Gaussian splatting and residual refinement then reconstruct the output light field. Despite being learned without geometric supervision, the predicted disparities strongly correlate with underlying scene geometry. Experiments across five benchmarks demonstrate that MacroGS integrates effectively with diverse light field encoders, improving reconstruction quality at fixed scales while achieving leading performance across most evaluated arbitrary-scale settings. The decoder also achieves speedups of approximately $2$-$10\times$ over state-of-the-art implicit light field decoders and $10$-$18\times$ over the view-wise Gaussian baseline.

cs.CV↗

JEV-VAD: Efficient Semantic Endpointing via Task-First Conditioning and Decoder-Prefix Pruning

Semantic endpointing must distinguish a completed turn from a temporary pause, yet its binary output does not require the full generation pathway of an audio-language model. We propose JEV-VAD, an ASR-free framework that combines task-first conditioning, a scalar endpoint readout, and decoder-prefix pruning. The question and completion criteria are placed before audio tokens, allowing causal decoder attention to construct audio states conditioned on the endpointing objective. A dedicated scalar head maps the assistant-prefix state directly to a stop probability, removing transcript generation, answer decoding, and vocabulary-sized output projection. To reduce representation cost, decoder-prefix pruning physically retains only the initial decoder blocks. Matched head refitting then identifies which retained depths provide useful endpoint information under a common adaptation budget. The framework is built on Qwen3-ASR-0.6B and adapted through supervised fine-tuning followed by reinforcement learning for calibrated decisions. Together, these components address input conditioning, output complexity, and decoder depth within one endpointing framework. Experiments on EOT-bench show cross-language transfer from English endpoint supervision and useful discrimination in shallow prefixes, supporting a selectable quality-resource trade-off.

eess.AS↗

Mutual Colormap Dependence: How Recoloring Biomedical Images Changes Vision-Language Model Answers

Vision-language models (VLMs) are being built to answer questions about biomedical images, yet they never see the measured intensities, only the colors that a colormap has assigned to them. For questions about image content, the colormap should not change the answer. We call the extent to which it does Mutual Colormap Dependence (MCD), the mutual information between the colormap and a model's answer, which is zero exactly when every colormap yields the same answer distribution. Because VLMs sample their answers, we compare recoloring with a control that uses the same number of calls to ask again about the unchanged gray image, so that ordinary resampling is not mistaken for an effect of the colormap. We test GPT-5.4, Gemini 3.1 Pro, Claude Sonnet 4.6, and Qwen3-VL-235B under six luminance-monotonic colormaps (gray, viridis, plasma, inferno, magma, and cividis) on 50 medical visual-question-answering items from 16 SLAKE images, 64 ChestX-ray14 chest radiographs, and 200 BBBC039 fluorescence-microscopy images for exact nucleus counting. Every prompt tells the model to ignore the colormap. Even so, with three answers per colormap on the chest radiographs, the answers of all four models depend on the colormap (all Holm-adjusted $p \leq 0.002$). For Claude, some colored map overturns the majority answer on 29 of the 64 radiographs, where resampling gray alone overturns about 4, and Qwen's accuracy falls from 0.54 under gray to between 0.32 and 0.41 under the colored maps, mostly because it abstains. With one answer per colormap, recoloring adds disagreement beyond repeated sampling in 5 of 12 model-panel cells under per-cell 95% intervals, and repeating the instruction to ignore color does not reliably reduce it. The colormap is therefore part of what a VLM sees, and it can change the answer.

eess.IV↗

Single-Radar Bilateral 3D Near-Field Imaging for Mobile Robots

Mobile robots operating in homes need reliable 3D perception of nearby furniture and surrounding structures, yet recovering this information on both sides of the robot's path with a single forward-mounted millimeter-wave radar remains challenging because of near-field wavefront curvature, viewpoint-dependent target visibility, and limited elevation diversity. This paper presents a motion-assisted framework for reconstructing structures on both sides of a robot's trajectory. Our near-field model relates scene reflectivity to stepped-frequency multiple-input multiple-output (MIMO) observations across frequencies and robot positions using spherical paths from transmitter (Tx) to scatterer to receiver (Rx). Bilateral Tx combining provides wideband directional weighting, while sliding-aperture back-projection coherently focuses the measurements using near-field propagation distances. Top-K power fusion combines local images with different visibility. The reconstruction provides lateral-forward support for multi-layer height extraction, followed by guarded adaptive height refinement using the original MIMO channels. Experiments in controlled, office/study, and corridor environments produce mean lateral-forward localization errors of 4.65, 5.63, and 5.98 cm, respectively, with peak-to-background ratios (PBRs) exceeding 13 dB. The framework recovers bilateral 3D scattering structure from one forward pass using a single radar, without relying on learned representations of the scene.

cs.RO↗

AgentSelect: Benchmark for Narrative Query-to-Agent Recommendation

LLM agents are rapidly becoming the practical interface for task automation, yet the ecosystem lacks a principled way to choose among an exploding space of deployable configurations. Existing LLM leaderboards and tool/agent benchmarks evaluate components in isolation and remain fragmented across tasks, metrics, and candidate pools, leaving a critical research gap: there is little query-conditioned supervision for learning to recommend end-to-end agent configurations that couple a backbone model with a toolkit. We address this gap with AgentSelect, a benchmark that reframes agent selection as narrative query-to-agent recommendation over capability profiles and systematically converts heterogeneous evaluation artifacts into unified, positive-only interaction data. AgentSelectcomprises 111,179 queries, 107,721 deployable agents, and 251,103 interaction records aggregated from 40+ sources, spanning LLM-only, toolkit-only, and compositional agents. Our analyses reveal a regime shift from dense head reuse to long-tail, near one-off supervision, where popularity-based CF/GNN methods become fragile and content-aware capability matching is essential. We further show that Part~III synthesized compositional interactions are learnable, induce capability-sensitive behavior under controlled counterfactual edits, and improve coverage over realistic compositions; models trained on AgentSelect also transfer to a public agent marketplace (MuleRun), yielding consistent gains on an unseen catalog. Overall, AgentSelect provides the first unified data and evaluation infrastructure for agent recommendation, which establishes a reproducible foundation to study and accelerate the emerging agent ecosystem.

cs.AI↗

X-Fly: Learning to Reach and Stop in Full-Chunk Aerial Vision-Language Navigation

Aerial vision-language navigation (VLN) agents built on large vision-language models are expensive to query. Executing a complete predicted action chunk before replanning amortizes this cost, reducing model inference from 195 to 42 ms per executed waypoint for our policy. It also moves the states at which the policy is queried away from expert demonstrations, which supervise neither recovery from these states nor termination at the learner's own arrival. On OpenFly test-seen, no-stop failures account for 47.78% of full-chunk behavior cloning (BC) outcomes. We present X-Fly, a staged training framework for full-chunk aerial VLN. Recovery DAgger labels learner-visited decision states with smooth chunks that reconnect to future expert segments and with STOP at the first in-region state. Flow-GRPO then jointly refines the chunked flow controller and a Bernoulli STOP policy from same-task closed-loop outcomes, supplemented by outcome-verified stopping supervision. With eight controls per decision, X-Fly reaches 74.18% SR on test-seen and 51.73% on test-unseen, compared with 16.45% and 18.43% for full-chunk BC and 32.04% and 29.46% for DreamFly.

cs.RO↗

Two-disjoint-cycle-cover edge bipancyclicity in Cartesian products

For two even integers $r_1\leq r_2$, a graph $G$ of even order is \emph{2-DCC edge $[r_1,r_2]$-bipancyclic} if for every ordered pair $(e,f)$ of independent edges and every even integer $\ell\in[r_1,r_2]$, there exist two vertex-disjoint cycles $C_1$ and $C_2$ with $V(C_1)\cup V(C_2)=V(G)$, $|C_1|=\ell$, $|C_2|=|V(G)|-\ell$, $e\in E(C_1)$, and $f\in E(C_2)$. We study this property for Cartesian products $G=H_1\square\cdots\square H_d$ with $d\geq2$, where each factor is $K_2$ or an even-order bipartite edge-Hamiltonian graph. We prove that $G$ is 2-DCC edge $[4,|V(G)|/2]$-bipancyclic whenever $d\geq4$, and also for $d\in\{2,3\}$ under explicit Hamiltonian-cycle conditions on $H_i$. Here $1\leq i\leq d$. As applications, we give sharp characterizations for undirected toroidal meshes, star-cube networks, and hyperstar networks.

math.CO↗

SpeakerTurn: Target-Speaker Semantic Turn Detection in Multi-Speaker Environments

Semantic completeness alone cannot determine when an assistant should respond in a multi-speaker environment: a completed utterance may belong to an interfering speaker, while the designated user is still speaking, acknowledging, or requesting time. We therefore study semantic turn inference conditioned on speaker identity. We propose SpeakerTurn, a framework that jointly predicts the target transcript and turn state from a speech mixture and an independent target-speaker reference. Built on CocktailASR-1, it links semantic decisions to the corresponding target content through joint supervision, distinguishing complete, incomplete, backchannel, wait, and target-absent states, with a separate interruption task. We also examine whether structured chain-of-thought (CoT) supervision improves decisions by representing speaker presence and semantic cues. On a synthetic multi-speaker test set, SpeakerTurn achieves five-state macro F1 of 0.811 and target-transcription character error rate (CER) of 4.20%, correctly resolving all ten reference-swapped interruption pairs. Joint-format transcription and native-interface recognition nevertheless behave differently, and CoT improves validation transcription without improving turn macro F1. These results show why target attribution, semantic turn discrimination, and recognition retention require separate assessment when adapting speech language models for speaker-conditioned interaction.

eess.AS↗

Localizing Concepts in Visual Autoregressive Models

Understanding the internal routing of visual concepts is crucial for the safe and controllable adaptation of generative models. While concept localization has been widely studied in diffusion models, the emerging paradigm of next-scale visual autoregressive models remains largely unexplored. In this paper, we introduce LoCo, the first model-agnostic framework to localize where and when specific concepts emerge within autoregressive models. Driven by the native coarse-to-fine nature of next-scale generation, our method precisely maps conceptual knowledge across three distinct dimensions: Layer, Scale, and Position. To systematically localize and evaluate concept routing without context bias, we propose LoCoBench, a comprehensive dataset spanning 10 diverse categories. Extensive probing on image and video autoregressive models like Infinity, HunyuanImage-3.0, and InfinityStar shows that the localized positions are both interpretable and causally related to concept emergence. Building on these insights, we apply our localization method to three key applications: concept erasure, model personalization, and adversarial concept injection. Experiments demonstrate that our targeted intervention achieves state-of-the-art performance, substantially reducing computational overhead while preserving benign utility. Overall, our findings offer insights into how conceptual knowledge is routed during autoregressive generation, introducing a practical pathway for more interpretable, efficient, and secure adaptation.

cs.CV↗

Doubly Robust Distributionally Robust Offline Contextual Pricing

Offline contextual pricing often relies on logged observational data, but faces challenges from distributional shifts between training and deployment environments. Distributionally robust optimization (DRO) provides a principled approach to off-policy evaluation and learning (OPE/L). However, existing methods are mostly limited to discrete actions. Recent work has explored DRO for continuous treatments using inverse propensity weighting (IPW), while such IPW-based estimators can be sensitive to the convergence rate of propensity score estimates, particularly when estimated nonparametrically, which may lead to larger estimation errors and regret. In this work, we develop a doubly robust (DR) framework for distributionally robust OPE/L in continuous pricing settings. For evaluation, we propose a localized DR estimator that addresses the computational challenges of worst-case expectations by fitting only a small number of regressions, comparable to standard non-robust DR, while achieving semiparametric efficiency under mild product rate conditions. For learning, we leverage the inherent smoothness of demand noise to handle pricing-specific discontinuities in revenue outcomes (e.g., threshold-based purchase decisions), establishing a finite-sample regret bound of $\tilde{\mathcal{O}}_p(T^{-s/(2s+1)})$ for smoothness orders $s=1,2$. This bound improves upon existing regret rates in existing DRO-based off-policy learning (OPL) for continuous treatments. Extensive experiments under various levels of distribution shift validate our proposed framework.

cs.LG↗

Imaging-anchored multiomics in cardiovascular disease: integrating cardiac imaging, bulk, single-cell, and spatial transcriptomics

Cardiovascular disease arises from interactions between inherited risk, molecular programmes, and tissue-scale remodelling that are observed clinically through imaging. Cardiac MRI (CMR), computed tomography (CT), and echocardiography are integral to routine cardiovascular care, while bulk RNA sequencing, single-cell RNA sequencing, and spatial transcriptomics are providing increasingly detailed molecular characterization of cardiac tissue. Yet, these imaging and molecular data are still analysed in largely separate pipelines. This review examines joint representations that link cardiac imaging phenotypes to transcriptomic and spatially resolved molecular states. An imaging-anchored perspective is adopted in which echocardiography, CMR, and CT define a spatial phenotype of the heart, and bulk, single-cell and spatial transcriptomics provide cell-type- and location-specific molecular context. We define the representation requirements of each modality, compare multimodal fusion strategies, and synthesize integrative pipelines for radiogenomics, spatial alignment, and image-based gene-expression prediction, together with their validation requirements, limitations, and failure modes. Spatial multiomic maps of human myocardium and atherosclerotic plaque, together with single-cell, spatial, and multimodal medical foundation models, are advancing imaging-anchored multiomics; however, cost, scalability, and tissue availability remain substantial barriers to large-scale cardiovascular translation.

q-bio.QM↗

StepAudio 3 Realtime Technical Report

Realtime spoken interaction demands deep reasoning, prompt responses, and fluid turn-taking. We present StepAudio 3 Realtime, an audio-language foundation model organized around a continuous listen-converse-think-act loop. Deep Perception captures rich acoustic cues to interpret user intent, while Seamless Duplex models synchronized audio streams to handle pauses, backchannels, and interruptions naturally. Crucially, we resolve the tension between deep deliberation and latency via Think-While-Speaking, executing private reasoning in parallel with spoken delivery. In reasoning mode, StepAudio 3 reaches a 73.0 macro average on StepAudioChat. With Think-While-Speaking, it achieves dialogue and reasoning performance comparable to dedicated reasoning models while speaking in real time. Furthermore, an integrated Voice Agent handles asynchronous tool execution without disrupting the dialogue flow. StepAudio 3 Realtime achieves top-tier performance across key dimensions: an exceptional 90.6 on the MMSU benchmark, 98.9 Overall on the Artificial Analysis Full-Duplex Bench, and a 56.0% macro task-success rate on $τ$-Voice.

cs.SD↗

Noise Space Learning for VLA Adaptation

Pretrained vision-language-action (VLA) models encode broad manipulation priors, yet adapting them to specific tasks requires balancing limited robot interaction, computational resources, and safety constraints. Noise space learning (NSL) adapts flow-matching-based VLAs by learning a context-conditioned distribution over noise while keeping the pretrained VLA fixed. Building on prior noise-steering methods, we formulate NSL as an interface for VLA adaptation that accommodates different learning algorithms and instantiate it with PPO, SAC, and AWR. Theoretically, we provide an information-theoretic analysis of how NSL enables efficient adaptation. Empirically, we study these NSL algorithms on LIBERO, RoboTwin, and four real-robot manipulation tasks under interaction, computation, and execution constraints. Across these settings, we develop a suite of effective NSL variants and derive practical recommendations for choosing among them under different resource and deployment regimes.

cs.RO↗

Estimation of the sub-Gaussian Parameter

The sub-Gaussian parameter (also called the variance proxy) of a mean-zero random variable $X$ is defined as $ξ^2_\star = \sup_{λ\in \mathbb{R}} L(λ)$ where $L(λ) = \frac{2}{λ^2} \log \mathbb{E} e^{λX}$ is a weighted cumulant generating function. We study the estimation of $ξ^2_\star$ and prove that the minimax risk is governed by a non-increasing function $δ_P(C) = \sup_{|λ| \geq C} L(λ) - \sup_{|λ| \leq C} L(λ)$ which captures the influence of the tail behavior of the distribution $P$. Over the class of distributions with $δ_P \leq r$ for a non-increasing function $r$, the minimax risk is, up to a multiplicative constant, lower bounded by $r(\sqrt{\log n}) + n^{-1/2}$ and upper bounded by $r((\log n)^{1/2-\varepsilon}) + n^{-1/2 + \varepsilon}$ for any $\varepsilon > 0$. Our estimator for the upper bound is based on constrained maximization of the empirical analogue of $L$. In addition to being almost minimax optimal and adaptive, we further prove that the estimator is asymptotic normal under suitable conditions and that, if the underlying distribution is not sub-Gaussian, the estimator diverges with a rate determined by the heaviness of the distributional tail.

math.ST↗

Is the regression $F$-test doubly robust?

We study the robustness of the $F$-test in random design linear models, and reach a somewhat nuanced conclusion. On the positive side, one of our main results is that the size of the test is close to its nominal level as soon as either the distribution of the normalised error vector is close to uniform on the unit sphere, or the design matrix, after applying a suitable column space-preserving orthogonalisation scheme, is close to being uniformly distributed. This provides a sense in which the $F$-test is doubly robust. Our conclusion is reached by establishing a Kolmogorov to Wasserstein distance Hölder continuity property controlling the departure of the $F$-statistic from its notional $F$-distribution under the null. Writing $n$, $p$ and $p_0$ for the sample size and the dimensions of the full and null models respectively, we prove that the Hölder exponent is $1/3$ when $\min(p-p_0,n-p) = 1$ and $1/2$ when $\min(p-p_0,n-p) \geq 2$. On the other hand, these exponents are relatively small and cannot be improved in general, revealing that the size of the test may depart from its nominal level quite quickly as we move away from settings where the test is exact. In some cases, our conclusions may be improved by working with a local Kolmogorov distance that focuses on discrepancies between distribution functions in the right tail.

math.ST↗