Searcharxiv⌕ Search

arXiv subjects

Xiaodong Liu

Publications and source records attributed to Xiaodong Liu.

At least 19 recordsLinked to original sources

Beyond Token Stacking: Multi-scale Visual Modulation and Adaptive Fused-feature Calibration for MLLMs

Current Multimodal Large Language Models (MLLMs) have advanced rapidly and follow two main paradigms: one involves concatenating visual and textual tokens for input into Large Language Model (LLM), while the other injects visual knowledge into Feed-Forward Networks (FFNs) of the LLM. However, the former introduces a large number of redundant visual tokens, substantially increasing computational overhead. The latter reduces sequence length but removes images from the input stage entirely, which may hinder the activation of image-related knowledge. Moreover, naively injecting visual knowledge into the FFNs could induce activation distribution shifts due to cross-modal semantic discrepancies. To address these challenges, we develop Beyond Token Stacking (BTS), a novel visual enhancement framework for MLLMs. Specifically, we propose a Multi-scale Visual Modulation (MVM) mechanism that injects multi-scale visual information at the input stage via modulation without increasing the token sequence length. Furthermore, we develop an Adaptive Fused-feature Calibration (AFC) module that performs per-channel affine calibration to decouple visual injection from distribution alignment in the FFN layers, improving multimodal fusion stability while preserving language priors. Experiments show that our approach achieves state-of-the-art performance.

cs.CV↗

The limit point set of real roots of random polynomials: a resolution of erdős problem 521

We resolve Erdős Problem~521 by proving that the normalized real-root count $N_n/\log n$ of Rademacher polynomials almost surely has no limit. More generally, consider the Kac polynomials \[ P_n(x)=\sum_{j=0}^{n}ξ_jx^j \] with i.i.d.\ real coefficients of mean zero, variance one, and a finite absolute exponential moment. Let $N_n$ denote the number of distinct real roots of $P_n$. We prove that, almost surely, \[ \liminf_{n\to\infty}\frac{N_n}{\log n}=\frac1π, \qquad \limsup_{n\to\infty}\frac{N_n}{\log n}=\frac1π+C_*, \] and the set of subsequential limits is exactly \[ \left[\frac1π,\;\frac1π+C_*\right]. \] Here $C_*>1/π$ is a constant uniquely determined by an equation. The lower-limit identity remains valid under the weaker assumption that $\mathbb{E}|ξ_0|^{2+δ}<\infty$ for some $δ>0$.

math.PR↗

Reinforcement Learning over Patient Trajectories for Clinical Reasoning in EHR Foundation Models

Electronic health record (EHR) foundation models trained on longitudinal patient trajectories have demonstrated strong performance across diverse clinical prediction tasks. However, their clinical reasoning capabilities remain constrained by next-token prediction on limited and incomplete EHR data. To address this, we propose a reinforcement learning (RL) fine-tuning framework that treats EHR foundation models as generative policies over patient trajectories. We formulate common clinical prediction problems (e.g., hospital readmission) as event-conditioned, time-windowed reasoning tasks. We then design time-aware, rollout-sensitive rewards to account for finite rollout lengths and temporally inconclusive outcomes. We find that RL fine-tuning consistently improves over pre-trained backbones and strong baselines. Notably, it enables smaller models to match larger pre-trained models in data-limited regimes and induces positive transfer across tasks. Further analysis shows that RL fine-tuned models generate trajectories with stronger structural and semantic alignment to ground truth and greater downstream utility.

cs.LG↗

Long-term average collision frequencies between two bodies undergoing Lidov-Kozai evolution

The existing {Ö}pik-type collision frequency framework that accounts for the projectile's Lidov-Kozai (LK) secular evolution holds the target's semimajor axis, eccentricity, and inclination fixed and prescribes uniform precession of its longitude of the ascending node and argument of periapsis. This treatment is not readily applicable when the target's eccentricity, inclination, or orbital precession rates undergo appreciable secular variations. Here, we extend the framework of Liang and Liu (2026, arXiv:2608.30972) to address one such case, in which both the projectile and target undergo LK evolution. Using a doubly averaged quadrupole model to describe the secular dynamics of both bodies, we include the target's orbital evolution when evaluating the collision contribution of each orbital intersection configuration. In the numerical example considered, the projectile survival curve predicted using the calculated collision frequency shows an overall decay trend close to that obtained from direct N-body simulations using REBOUND. In addition, we examine whether the long-term mean collision frequency depends on initial states while holding both bodies' semimajor axes, Kozai constants and closed Hamiltonian level curves fixed. Our theoretical analysis indicates that this frequency is independent of initial states when the parameters describing their relative state are jointly equidistributed at the start of successive projectile LK cycles. Numerical cases with different initial states support this conclusion.

astro-ph.EP↗

Jailbreak susceptibility prediction and mitigation via the behavioral geometry of models

Evaluating and mitigating a generative system's susceptibility to jailbreak attacks is critical to its safe deployment. Given the number of deployable systems, full per-configuration evaluation and optimization is impractical. In this paper, we formalize the behavioral geometry of a population of models that, by leveraging previously evaluated and defended models, supports both efficient susceptibility prediction and effective defense transfer across a population. We apply the framework to 82 models spanning 24 providers and to 100 system configurations of a single base model. Simple methods that use the behavioral geometry reach an AUPRC of $0.94$ for susceptibility detection with $\approx98\%$ fewer probes relative to a full evaluation. Using the behavioral geometry to select which model to transfer an optimized defense from outperforms same-provider assignment ($+4.1\%$, $p = 0.004$) at no additional probe cost, with a set of three models sufficient to cover the population. Results are robust to hyperparameter selection and judge.

cs.CR↗

Scaling LLM Agent Tool Access: The MCP Gateway

LLM agents increasingly rely on tool calling to complete tasks. Model Context Protocol (MCP) standardizes tool calling between LLM agents and tool providers. It reduces integration cost and drives rapid growth in tool scale. In production deployments, tool access becomes difficult to scale. On the tool provider side, legacy services are not directly callable through MCP; the rapid protocol development creates ongoing compatibility cost. On the agent side, the number of mounted tools is limited by the LLM context window and inference overhead; mounting a large tool set increases token usage and inference latency and can reduce task success rate. Moreover, for stateful MCP backends with multiple replicas, preserving session affinity increases client-side complexity. We present a cloud-scale MCP Gateway system. It breaks the direct-connect model on the data plane and offloads legacy service integration, consolidating incompatible MCP variants, access control, tool recommendation, and session-aware routing at the gateway. The conversion module supports 315 OpenAPI services in the cloud and integrates MCP transport variants. Session-aware routing decouples the scalability of the gateway layer and the MCP service layer. The hybrid retrieval engine sustains $\sim$98\% Top-15 recall; it scales agent tool access to 3{,}000+ tools with improved tool selection accuracy. We also share the lessons learned from deploying the gateway system in production.

cs.DC↗

ShallowDefend: Harnessing Shallow Safety Alignment for Multi-Turn Defender

Safety alignment has enabled large language models to reliably refuse harmful requests directly, yet the same request distributed across a multi-turn conversation still succeeds far too often. This brittleness has a mechanistic root: Alignment is shallow, with safety behavior concentrated in a model's first few output tokens, and existing defenses inherit the weakness, supervising the surface form of the response without ever giving the model an explicit safety decision to learn. We take the opposite view of shallowness: if a short span determines the trajectory of an entire response, that span is the highest-leverage place to intervene, and the defender can occupy it first. We propose Assess $\to$ Decide $\to$ Respond, a framework in which the defender assesses the user's intent, commits to an explicit, parseable safety decision, and only then generates a reply conditioned on it, with the commitment trained by reinforcement learning at the payload turn. Because the decision is a token span, the reward can be a rule: a self-verified reward checks the committed label against the conversation's known nature, requiring no judge and no annotation, while the same format also supports an LLM-as-judge reward over a graded decision space with safe redirection. Across two model families and fourteen benchmarks spanning safety, utility, and helpfulness, our defenders achieve the strongest multi-turn safety in a controlled comparison, maintaining their lead in safety while preserving utility at a modest cost to helpfulness. A per-token KL analysis and interventional probes confirm why this works: training concentrates its effect on the brief deliberation that sets a response's trajectory, and the decision committed there is a causal control point. The very shallowness attackers exploit becomes the mechanism of defense.

cs.CL↗

CoFiBreak: A Coarse-to-Fine Diversity Reward for Effective Multi-Turn Jailbreak

Automated red teaming trains an attacker large language model (LLM) to elicit harmful responses from a victim LLM, but reinforcement learning can collapse attackers onto a narrow set of prompts when attack success is the only reward. Such a collapse of attack diversity can bias safety evaluation toward repeatedly testing near-duplicate attack patterns rather than the underlying harmful intents. Existing methods often trade off attack success or inference cost for greater content or strategy diversity. We propose CoFiBreak, a coarse-to-fine diversity reward for Group Relative Policy Optimization (GRPO) that improves attack diversity while preserving strong attack effectiveness and low inference cost. The coarse reward broadens which strategies the attacker uses by encouraging strategies underrepresented within each GRPO group, while the fine reward broadens how those strategies are realized by discouraging reuse of the same personas, framings, and output formats after harmful-intent information is removed from prompt embeddings. Diversity rewards apply only to attacks that already exceed an effectiveness threshold, so diversity cannot be gained through ineffective attacks. Since both rewards are computed only during training, CoFiBreak also retains inference efficiency. Trained with Llama3.1-8B-Instruct as the victim, CoFiBreak discovers 1.6-2.0x as many distinct successful fine-grained strategies as the strongest baseline across frontier victims on HarmBench. It reaches 98.8% ASR@10 (attack success rate with 10 generations per intent) on AdvBench with the highest diversity score among all compared methods, which is 2.1x that of the strongest-attacking baseline. The resulting attacks transfer to unseen frontier LLMs, reaching 71.0% ASR@10 on AdvBench with the unseen victim Claude-Opus-4.6. The code will be released.

cs.CL↗

Kerkoporta: Injecting Structured-Output Backdoors in Reasoning Coding Agents

Existing large language model (LLM) backdoor attacks operate primarily in chat settings, where the risk is bounded by free-form text outputs. AI agents remove that bound by performing actions and executing commands on the user's machine. We introduce structured-output backdoor attacks against reasoning models in agentic coding settings. A trigger placed anywhere in the agent's context causes the model to emit an attacker-chosen tool call, which the harness executes, directly compromising the system's confidentiality, integrity, or availability. We present Kerkoporta, a methodology that constructs high-quality poison samples, injects the backdoor via supervised fine-tuning, and recovers utility loss with lightweight reinforcement learning. We evaluate on modern, heavily post-trained open-weight models across proprietary and open-source harnesses, focusing on coding for its wide adoption and long-horizon task structure. The attack reaches success rates of up to 96.2% and can be configured to generalize across harnesses or bind to a single target. The resulting models can outperform the clean baseline on standard utility benchmarks, giving victims a direct incentive to adopt them. Overall, we demonstrate that LLM backdoors can have critical security impacts in the context of modern coding agents, and we characterize the poisoning methodology required to insert backdoors that are both effective and stealthy in this setting.

cs.CR↗

Where You Inject Matters: The Trust Region Effect in LLM Agents

LLM agents are increasingly adopted in long-horizon tasks and interact with diverse external environments. Despite their growing capabilities, these agents remain vulnerable to real-world attacks such as indirect prompt injections, which manipulate agents through malicious instructions embedded in external environments. While existing risk assessment work has primarily focused on the injection content, the effect of where an injection appears has not been systematically characterized. Studying injection placement requires agents to operate over longer-horizon tasks and interact with diverse environmental surfaces. Thus, we build three agentic testbeds spanning tool use, coding, and daily workflow domains with diverse injection surfaces. Holding the injection content fixed, we find that agent susceptibility varies across injection placements. We refer to this phenomenon as the trust region effect: agents implicitly assign different levels of trust to different regions of their environments. For example, an agent powered by GPT-5.4 follows an unsafe injection instruction placed in a repository documentation file (i.e., a README or TODO file) 2.6 times more often on average than the same instruction placed in a code or data file. Building upon this observation, we develop learning-based attackers that generate injection content and select better injection placements. Our approach demonstrates that both where an injection is placed and what content it contains can affect the strength of indirect prompt injection attacks. To the best of our knowledge, this work provides the first systematic characterization for the effectiveness of different injection placements, spanning both coarse-grained environmental surfaces and fine-grained positions. The uneven trust that agents implicitly place across their environments provides important insights for future attack and defense design.

cs.AI↗

SEMA-Auto: Sighted and Efficient Multi-Turn Attack with Autonomy

Existing jailbreak attackers fall into two classes: static planners fix their single- or multi-turn adversarial plans in advance, remaining simple but losing sight of the victim's replies, wasting turns and limiting transferability; interactive generators read each victim reply to generate the next-turn attack but require in-loop evaluators for success signals or hints, incurring substantial costs and inflated performance. Bridging this gap, we propose SEMA-Auto, a sighted and efficient multi-turn attacker with autonomy, interacting with the victim model and terminating the session by itself upon success. To stabilize its behaviors and boost its capabilities, SEMA-Auto is trained in two stages. Synthetic bootstrap tuning installs the notion of attack targets and conforming decisions by fine-tuning the base model with bootstrapped trajectories. Decision-supervised reinforcement learning then optimizes two objectives jointly yet separately, reinforcing the attack with higher jailbreak reward while supervising the decision at each turn against an absolute target. Our method achieves state-of-the-art (SOTA) attack success rates (ASR) across multiple datasets, victims, and judges, outperforming all static and interactive baselines, within almost the fewest turns, demonstrating both effectiveness and efficiency. For example, SEMA-Auto performs a $76.0\%$ ASR@1 within an average of $3.0$ turns against DeepSeek-V4-Flash on AdvBench, while the best static baseline is $54.8\%$ with fixed $7$ turns and the best interactive baseline is $43.8\%$ in $3.7$ turns. Against safety-hardened frontier models, GPT-oss-120B, GPT-5.4, Claude-Haiku-4.5, and GPT-6-Astra, our counter-trained attacker achieves an average $59.6\%$ ASR@1 on HarmBench, $13.8\%$ over prior SOTA. Our code is available at: https://github.com/microsoft/SEMA

cs.CL↗

CoCarry: Contact-Support-Structured Retargeting for Dual-Humanoid Cooperative Carrying

Recent advances in large-scale human motion data, motion retargeting, and motion imitation have enabled increasingly capable whole-body control for individual humanoids. Cooperative carrying is more challenging because successful coordination requires not only plausible individual motions, but also consistent cross-agent and agent-object relations, including shared grasp, relative configuration, and whole-body support. We introduce CoCarry, a cooperative-carrying dataset containing 220 synchronized two-person sequences across 21 task schemes, five shared-object geometries, four terrain types, and approximately 61 minutes of motion. We further propose Contact-Support-Structured Retargeting, which starts from nominal individual motions and jointly refines both humanoids around the same prescribed object trajectory. Cooperation-aware object-local hand anchors preserve compatible four-hand grasp relations, while complementary upper object-interaction and lower support structures retain object-relative geometry and coordinated support. Across all 220 CoCarry sequences retargeted to two A3 humanoids, our method achieves 0.9995 contact preservation, reduces penetration duration by 67.0% and 71.0% relative to GMR and OmniRetarget, respectively, and maintains a 100% retargeting success rate. On CORE4D with Unitree G1 humanoids, it achieves 0.9906 contact preservation with zero measured sliding. We further validate the generated references through physics simulation and multiple real-world cooperative carrying experiments on two physical A3 humanoids, demonstrating their usability for downstream dual-humanoid control.

cs.RO↗

Post-perihelion dust activity and a candidate icy-grain halo in long-period comet C/2023 A3 (Tsuchinshan--ATLAS)

We analyze the post-perihelion coma of the Great Comet of 2024, C/2023 A3 (Tsuchinshan--ATLAS) to characterize dust activity and assess the contribution of sublimating icy grains to a candidate near-nucleus halo. Using archival Zwicky Transient Facility images from 15 February to 26 May 2025, we performed aperture photometry within a fixed projected radius of 40,000 km and derived azimuthally averaged surface-brightness profiles. The comet faded from 12.86 to 14.77 mag as heliocentric distance increased from 2.64 to 3.90 au, with two short-lived brightenings. The effective dust cross section decreased by a factor of 2.4, following $C_{\rm e}\propto r_{\rm h}^{-3.1}$, while the dust mass production rate scaled as $\dot{M}_{\rm d}\propto r_{\rm h}^{-3.41}$, consistent with a contribution from water-ice-driven activity. Extrapolating the fitted relation to perihelion implies a model-dependent cumulative dust mass loss of $3.0\times10^{11}$ kg, equivalent to a globally averaged erosion depth of $\sim1.4$ m for an assumed nucleus radius of 5.9 km. Coma profiles were shallower than the canonical steady-state expectation, suggesting radial changes in dust distribution or scattering properties, possibly due to delayed aggregate fragmentation. A persistent central brightness excess, coincident with less-red inner-coma colors, is consistent with an icy-grain halo. Among the models explored, icy grains of radius $\sim10~μ$m with carbon fraction $\sim0.1\%$ yield halo radii comparable to those observed during enhanced activity on 21--25 February 2025, with grain lifetimes of $2.8\times10^5$--$3.2\times10^5$ s. No single fixed grain size and composition reproduces the full halo evolution; an increasing carbon fraction could explain the broad halo-radius trend, possibly reflecting progressive refractory enrichment of near-surface material.

astro-ph.EP↗

Dust Activity Evolution and Constraints on Icy Grains in the Interstellar Comet 3I/ATLAS

We present Zwicky Transient Facility observations of the interstellar comet 3I/ATLAS to characterize its dust-activity evolution before and after perihelion. Before perihelion, the effective scattering cross section $C_{\rm e}$ brightened across the inner $2\times10^4$ km as $C_{\rm e}\propto r_{\rm H}^{-γ}$, with $γ$ increasing from $1.63\pm0.14$ to $1.88\pm0.34$ with aperture size. After perihelion, the heliocentric index decreased from $2.44\pm0.09$ in the $5\times10^3$ km aperture to $1.54\pm0.09$ in the $8\times10^4$ km aperture, indicating faster fading of the inner coma. Under adopted grain-size and residence-time assumptions, the implied dust mass-loss rates are $\sim123$--$203~{\rm kg~s^{-1}}$ before perihelion and $\sim596$--$1412~{\rm kg~s^{-1}}$ after perihelion. The morphology yields a sunward ejection-speed scale of $v_{\rm ej}\sim(190\pm3)a_{μ{\rm m}}^{-1/2}~{\rm m~s^{-1}}$ and an effective perpendicular velocity of $v_\perp=(79\pm6)a_{μ{\rm m}}^{-1/2}~{\rm m~s^{-1}}$. Post-perihelion surface-brightness profiles transiently steepened to $k\sim-2.7$ in late November and early December, then became shallower and approached $k=-1.5$ by January. Matched $g-r$ profiles show that this phase coincided with an anomalously blue outer coma that disappeared by January. We characterize the fading component with a projected destruction length of $\ell_{\rm DE}\sim(6.4$--$7.2)\times10^4$ km. Thermophysical sublimation calculations indicate that this scale, together with the order of magnitude of the water production near the main imaging epoch, is consistent with $a_0\sim5$--$10~μ{\rm m}$ icy grains containing a small absorbing carbonaceous fraction of $f_{\rm C}\sim0.1\%$. These results suggest that the early post-perihelion coma likely contained a short-lived, volatile-bearing icy-grain component whose fading contributed to the outer-coma brightness and color evolution

astro-ph.EP↗

AgentCIScope: Grounded Contextual Integrity Evaluation for LLM Agents

LLM agents are becoming increasingly capable of handling daily workspace tasks by interacting with application environments through tool calls. However, as agents read and write sensitive user data across email, documents, and other applications, a central privacy question is whether they preserve contextual integrity (CI), that is, whether the information flows they produce comply with context-specific disclosure constraints. Existing work either focuses on model-level privacy leakage or evaluates agents in limited application environments with manually constructed disclosure constraints. In this work, we introduce AgentCIScope, a unified platform and benchmark for evaluating agent-level contextual integrity through executable, multi-step tasks across diverse, realistic application environments. AgentCIScope transforms public and synthetic privacy cases into executable agent tasks, with disclosure constraints grounded in three criteria: governing privacy policies, persona-based agreement, and a predefined sensitive-attribute taxonomy. Based on AgentCIScope, we evaluate 12 agents on 936 tasks and find that strong task-completion capability does not guarantee contextual-integrity preservation: single-run leakage rates range from 32.5% to 94.4%, while repeating each task five times further increases leakage by 4-19 percentage points. Leakage varies substantially across attributes: personal preferences and booking details have average leakage rates of 84.7% and 76.5%, compared with 17.1% for credit-card information and 29.4% for SSNs; environments: calendar has the highest average leakage rate as an information source, while WhatsApp ranks highest as an information destination; and governing privacy policies. AgentCIScope provides deep insights into agent privacy leakage and serves as a unified testbed for developing privacy-preserving agents.

cs.AI↗

LLM-based User-Product Relevance Evaluation for Dynamic Product Advertising

Dynamic Product Advertising (DPA) retrieval systems are a core component of modern e-commerce advertising; however, their evaluation remains challenging. Conventional offline engagement metrics are limited by sparse, biased feedback, while online evaluation is costly and time-consuming. In this work, we propose a user--product relevance metric for offline evaluation of DPA retrieval systems. The metric first compresses the engagement history into a compact set of intent representatives via Semantic-ID-grouped, event-stratified segmentation, then assesses the relevance of a recommended product to each individual intent and aggregates them as a user-level relevance score, yielding a stable and interpretable measure of retrieval quality. To compute this metric at scale, we leverage large language models (LLMs) as semantic relevance judges, capturing nuanced functional and semantic relationships that heuristic methods often miss. We propose an agentic, human-aligned prompting framework that generates multi-category relevance guidelines and distills teacher-model judgments into an efficient student model for scalable evaluation. The resulting metric strongly correlates with offline and online business metrics, enabling faster and more reliable DPA retrieval iteration before A/B testing.

cs.IR↗

The Interior Inverse Scattering Problem from Backscattering Data: Spectral Comparison and Uniqueness Results

We study an interior inverse scattering problem for the two-dimensional Helmholtz equation with backscattering data, where the source and receiver coincide. Such data are common in applications, yet existing uniqueness results are almost exclusively established for full multistatic data with separated sources and receivers. We consider two main measurement regimes. In the first regime, data are collected on a closed curve inside the unknown domain. Exploiting the meromorphic dependence of the backscattering field on the wavenumber, we show that the residues at the poles determine the diagonal of the spectral projection kernel on the curve, and hence the \emph{visible} part of the Dirichlet spectrum. This yields an unconditional spectral comparison: the symmetric difference of the two Dirichlet spectra is contained in the spectrum of the interior domain enclosed by the curve. If, in addition, the common visible eigenvalues have either matching multiplicities or matching restriction ranks, we obtain an area estimate bounding the discrepancy of the two domain areas by the area of this interior domain. In the second regime, data are given on a nonempty open subset of the domain, and a single non-resonant frequency already uniquely determines a bounded connected Lipschitz domain. The proof combines interior real-analyticity of the regular part of the Helmholtz Green function with its boundary blow-up.

math.AP↗

DeepSeek Elastic Compute (DSec): A Sandbox Infrastructure for Effective Agentic Training at Scale

Large-scale agentic training and evaluation with large language models (LLMs) rely on isolated, stateful execution environments in which models inspect repositories, invoke tools, execute commands, and interact with task-specific services. These workloads create sandboxes in large bursts, span heterogeneous functionality and isolation requirements, retain state across long interactions, and draw from large image corpora with limited reuse. Supporting them therefore requires an elastic execution platform rather than a single sandbox runtime. This report presents DeepSeek Elastic Compute (DSec), a production sandbox platform that exposes FnCall, container, microVM, and full-VM sandbox backends through a unified SDK. DSec coordinates placement and lifecycle management across the cluster, composes environments from independently versioned layers, combines memory sharing, reclamation, and CPU scheduling for high-density execution, and loads image data on demand from Fire-Flyer File System (3FS), a cluster-wide distributed filesystem. DSec is co-designed with the reinforcement learning (RL) framework, decouples stateful rollout execution from preemptible GPU training, coordinates sandbox lifecycle with training to preserve rollout state while reclaiming idle resources, and mitigates agent misbehavior such as reward hacking. A single production-scale unit of DSec spans around 160 nodes, serving about 3 million sandboxes per day; in production, it supports over 380,000 concurrent sandboxes and sustains over 5,000 sandbox creations per second. Our evaluation and deployment experience show that these mechanisms reduce environment setup and image-distribution overhead, improve memory efficiency, and preserve latency-sensitive performance under high-density overcommit.

cs.DC↗