Searcharxiv⌕ Search

arXiv subjects

Zongze Li

Publications and source records attributed to Zongze Li.

At least 19 recordsLinked to original sources

RAG in a Trenchcoat: When Minimal Memory Is Enough for Agentic Systems, and When It Isn't

Agentic memory research has accumulated extraction, supersession, reflection, and graph synthesis on top of retrieval-augmented generation. However, we found that almost none of those primitives earn their cost in our tested memory benchmarks: we show this with BRIDGE, a deliberately minimal RAG memory with chained observation notes, hybrid BM25 with cosine retrieval, and chronological rendering, which reaches parity within bootstrap CI with the prior SOTA on LongMemEval-S and BEAM, and beats the best memory baseline on EverMemBench by 5.7-6.3 points at matched reader. It runs at ~5-7K reader-prompt tokens per query, using a single LLM call for each ingestion or query. We argue that this parity is itself the result: standard benchmarks score what memory retrieves, not what memory enables when the agent acts. To test the latter, we introduce five adversarial probes (Rule Application +-, Pattern Induction Style / Procedural, Dynamic Multiturn) judged at generation time by a cross vendor verification chain, where each query can only be solved by applying a remembered rule or pattern it never mentions. Across 5 probes x 20 scenarios, no architecture wins universally: atomic fact extractors (LangMem) win rule application, while BRIDGE's block-summary memory wins pattern induction. This result holds across four domains, three graders, and three reader families. Each design fails where the other wins, for reasons a recall benchmark cannot see: atomic-fact extractors discard the form a demonstrated pattern takes, while block-summary memory retrieves the rule but the reader does not act on it, a gap that presenting the rule more prominently largely closes. What separates memory designs is not what they retrieve but what they enable an agent to do, and that is what agentic memory should be evaluated on.

cs.AI↗

SemICP: Semantic Non-Rigid Point Cloud Registration with Elastic Energy Regularization

Purpose: Accurate point cloud registration is essential in computer-aided interventions (CAI) to align multi-modal medical images for intraoperative guidance. Classical methods, such as Iterative Closest Point (ICP), remain attractive for their explainability and minimal training requirements, but typically ignore anatomical semantics and biomechanical properties during regularization. Methods: We present Semantic ICP (SemICP), a novel non-rigid point cloud registration framework that combines semantically informed point matching with linear elastic energy regularization. Semantic labels are used to improve correspondence matching by constraining correspondences to be anatomically consistent. A novel control-point deformation representation with linear-elastic energy regularization is introduced to encourage biomechanically plausible deformations. SemICP was evaluated on a simulation phantom and four datasets against established baselines. It was also tested with labels from AI-based segmentation in a fully automatic segmentation-registration pipeline. Results: On the finite element phantom, SemICP recovered 86% of the interior displacement from surface observations alone. On the real datasets, SemICP achieved the lowest target registration error on Liver MR-US and the lowest surface distances on TORS, and ranked second on AbdominalMR-CT and HippocampusMR, where NR-ICP achieved lower surface distances. The fully automatic registration pipeline was shown to be effective for US-MR registration and to improve the alignment of expert-annotated structures. Conclusion: SemICP improves interior accuracy at competitive surface accuracy by combining semantic correspondence constraints and linear energy regularization. Combined with AI-based segmentation, SemICP provides an effective pipeline for multi-modal registration in CAI.

cs.CV↗

AlphaDiana: A Measurement System for Agents in Verifiable Reasoning Tasks

Agents pair reasoning models with harnesses that organize model calls, execute tools, load skill instructions, and retain history in memory. Understanding agent performance requires linking execution conditions to outcomes, behavior, and resource use. We introduce AlphaDiana, a measurement system with modular interfaces for datasets, agents, environments, and scorers. It preserves harness-native control loops while controlling file reuse and memory retention separately. AlphaDiana supports macro analyses of complete agents and micro interventions on agent components such as tools, skills, and memory under matched conditions. We study three open-source models and three open-source harnesses on five answer-based benchmarks, with two models also tested on two interactive benchmarks. We find that harnesses do not consistently improve the accuracy of reasoning: gains on questions unsolved by the model alone can be outweighed by losses on questions it answers correctly without tools. This inconsistency extends to component interventions: neither skill guidance nor retained history reliably improves accuracy. Behavior and output-token use must be assessed with correctness: low token entropy can accompany wrong answers, and less output per execution can mean more output tokens per correct answer when accuracy falls. AlphaDiana links answer correctness and execution status to traces and resource records for comparing agents and diagnosing failures under shared task and resource policies.

cs.AI↗

SEIS: Self-Evolving Inference Systems

Inference systems determine how fast and how cheaply language models can be served, so making them faster has direct practical value. However, prior work focuses mostly on optimizing certain parts such as kernels or memory within the large system. In this work, we take a holistic approach and apply agentic self-evolution to optimize the whole system end-to-end. Our SEIS (Self-Evolving Inference Systems) autonomously optimizes the entire mini-sglang engine without human intervention through iterative sessions with inherited experiences and code changes. Serving Qwen3-0.6B on H100, the resulting engine reaches 3.27X the throughput of the original mini-sglang implementation and beats SOTA engines like vLLM, TensorRT-LLM, and SGLang in the single-request workload. The correctness of the optimized inference engine by SEIS is tested in terms of numerical difference and downstream accuracy on math and long-context retrieval tasks. The code and session histories show that the speedup comes from redesigning the whole engine and that building on earlier sessions beats independent attempts. These results suggest that agentic self-evolution can optimize a complex system end-to-end. The evaluation also has to evolve with the engine, and letting agents evolve it is a natural next step.

cs.AI↗

ColoACT: Multi-Cue Action Chunking for Smooth Autonomous Colon Navigation on a Self-Propelled Endoscopic Robot

Autonomous colonoscopic navigation can reduce operator burden and the risk of loop formation or tissue trauma, but remains challenging due to deformable anatomy, weak-texture and specular endoscopic visuals, and contact-rich viscoelastic interactions. Existing methods either rely on geometry-driven pipelines, which are efficient and interpretable yet brittle due to manually engineered features and switching logic, or adopt learning-based policies, whose inferred depth/geometry can become temporally inconsistent or overly smooth under weak texture and specular highlights while simulation-trained variants (e.g., deep reinforcement learning) may further suffer from a sim-to-real gap. We propose ColoACT, an autonomous navigation system that integrates an RGB-D-E based Action Chunking Transformer policy (ColoACT policy) for a compact self-propelled Bevel-Gear-Based Endoscopic Robot (BGER). The ColoACT policy augments RGB with estimated relative depth and a gradient-based pseudo-elevation map to enhance fold-ridge saliency and other high-frequency geometric cues, and enables smooth continuous control of the BGER by predicting overlapping action chunks and fusing them via temporal ensembling. In different \textit{ex-vivo} porcine colons (approximately 60 cm), our system achieves success rates of 85.4\% and 72.5\% in straight and curved segments, respectively, and achieves 70\% success in 90-degree turns and 60\% in double-bend sequences, with feasibility further demonstrated in challenging triple-bend segments. The project page is available at: https://Adamhu1.github.io/ColoACT/.

cs.RO↗

Local LLM Agents as Vulnerable Runtimes:A Source-Code Audit of the Agent Runtime Layer

Local LLM agents such as OpenClaw and Nanobot run on end-user machines and act on host resources - the shell, filesystem, browser, stored credentials, and messaging applications - through natural-language goals. These agents have become privileged software runtimes that mediate between user intent, model outputs, and host-level actions. Existing research characterizes the landscape through prompt injection, malicious skills, marketplace risks, or black-box evaluation of agents. But the implementation layer that performs this mediation, the prompt builder, parser, tool dispatcher, skill loader, memory writer, network client, and permission gate, has remained an unexamined safety boundary. To our knowledge, no prior work has examined the agent's source tree to audit these components for implementation-level security weaknesses. We present CLAWAUDIT, a static auditing framework for measuring vulnerability exposure in local LLM agent runtimes. CLAWAUDIT derives a five-category vulnerability taxonomy from STRIDE and develops custom static-analysis rules that target agent-specific patterns absent from established rule sets for vulnerability analysis. We instantiate the taxonomy in two backends, 47 Semgrep YAML rules and 30 CodeQL queries, and evaluate on OPENCLAWBENCH, a benchmark of 446 source-code-level advisories from the OpenClaw repository and split temporally into 229 rule-derivation (train) and 217 held-out (test) advisories. On the held-out test, CLAWAUDIT raises Semgrep recall from 21.7% (Pro baseline) to 66.8%, and CodeQL recall from 13.8% (security-extended) to 75.1%. Train/test gaps remain within 4 percentage points for all four configurations, indicating that the rules generalize to vulnerabilities unseen during rule writing. A preliminary live-code audit shows that these recall-oriented rules require manual triage, motivating semantic filtering before production deployment.

cs.CR↗

What Do Evolutionary Coding Agents Evolve?

Recent work pairs LLMs with evolutionary search to iteratively generate, modify, and select code using task-specific feedback. These systems have produced strong results in mathematical discovery and algorithm design, yet a fundamental question remains: what do they actually evolve? Progress is typically summarized by the best score a run reaches under a task-specific evaluator, but that score can reflect several different mechanisms: new algorithmic structure, re-tuning an existing strategy, recombining ideas already in the model's internal knowledge, or overfitting to the evaluator. Distinguishing these mechanisms requires inspecting the search process itself, not only its final outcome. We introduce EvoTrace, a dataset of evolutionary coding traces spanning four evolutionary frameworks, reasoning and non-reasoning models, and 16 tasks across mathematics and algorithm design. To analyze these traces, we develop EvoReplay, a replay-based methodology that reconstructs the local search states behind high-scoring solutions and tests controlled interventions, including adjusting constants, removing program components and substituting models or prompting contexts. We annotate every code edit in EvoTrace with one of nine recurring edit types using an LLM-as-judge pipeline validated against blind human re-annotation. Across EvoTrace, most score gains come from a small subset of these edit types. We further find a deterministic cycling pattern: about 30% of code lines added during search are byte-identical re-introductions of previously-deleted lines, present throughout nearly every run. These results show that benchmark gains in evolutionary coding agents can arise from qualitatively different mechanisms, only some of which correspond to new algorithmic structure. EvoTrace enables more diagnostic evaluation of evolutionary coding agents beyond final benchmark scores.

cs.NE↗

Not All Prefills Are Equal: PPD Disaggregation for Multi-turn LLM Serving

Prefill-Decode (PD) disaggregation has become the standard architecture for modern LLM inference engines, which alleviates the interference of two distinctive workloads. With the growing demand for multi-turn interactions in chatbots and agentic systems, we re-examined PD in this case and found two fundamental inefficiencies: (1) every turn requires prefilling the new prompt and response from the last turn, and (2) repeated KV transfers between prefill and decode nodes saturate the bandwidth, leading to high latency and even service degradation. Our key insight is that not all prefill operations are equally disruptive: append-prefill, which processes only the new input tokens while reusing cached KV states, incurs an order-of-magnitude smaller decoding slowdown than full prefill. This motivates routing append-prefill to decode nodes locally. However, through comprehensive analysis, we show that no single fixed routing strategy satisfies all Service Level Objectives (SLOs) simultaneously. Based on this insight, we propose Prefill Prefill-capable Decode (PPD) disaggregation, a dynamic routing system that decides when to process Turn 2+ requests locally on decode nodes using cached KV states. PPD adapts to varying SLOs via configurable weights and seamlessly integrates with traditional PD deployments. With extensive evaluations, we show that PPD reduces Turn 2+ time-to-first-token (TTFT) by $\sim$68\% while maintaining competitive time-per-output-token (TPOT), effectively alleviating KV transfer congestion under high load. PPD provides a flexible and efficient paradigm for multi-turn LLM serving.

cs.NI↗

Unified framework for outage-constrained rate maximization in secure ISAC under various sensing metrics

Integrated sensing and communication (ISAC) is poised to redefine the landscape of wireless networks by seamlessly combining data transmission and environmental sensing. However, ISAC systems remain susceptible to eavesdropping, especially under uncertainty in eavesdroppers' channel state information, which can lead to secrecy outages. On the other hand, diverse and complex sensing performance requirements further complicate resource optimization, often requiring custom solutions for each scenario. To this end, this paper introduces a unified optimization framework that holistically addresses both the worst-case user secrecy rate and the sum secrecy rate across multiple users. Besides putting the two commonly used objectives into a single but flexible objective function, the framework accurately controls secrecy outage probabilities while accommodating a broad spectrum of sensing constraints. To solve such a general problem, we integrate the sensing requirements into the objective function through an auxiliary variable. This enables efficient alternating optimization and the proposed approach is theoretically guaranteed to converge to at least a stationary point of the original problem. Extensive simulation results show that the proposed framework consistently achieves higher optimized secrecy rates under various sensing constraints compared to existing methods. These results underscore the proposed unified framework's superiority and versatility in secure ISAC systems.

eess.SP↗

AlphaApollo: A System for Deep Agentic Reasoning

We present AlphaApollo, an agentic reasoning system that targets two bottlenecks in foundation-model reasoning: (1) limited reasoning capacity for complex, long-horizon problem solving and (2) unreliable test-time evolution without trustworthy verification. AlphaApollo orchestrates models and tools via three components: (i) multi-turn agentic reasoning, which formalizes model-environment interaction with structured tool calls and responses; (ii) multi-turn agentic learning, which applies turn-level reinforcement learning to optimize tool-use reasoning while decoupling actions from tool responses for stable training; and (iii) multi-round agentic evolution, which refines solutions through a propose-judge-update loop with tool-assisted verifications and long-horizon memory. Across seven math reasoning benchmarks and multiple model scales, AlphaApollo improves performance through reliable tool use (> 85% tool-call success), substantial gains from multi-turn RL (Avg@32: Qwen2.5-1.5B-Instruct 1.07% -> 9.64%, Qwen2.5-7B-Instruct 8.77% -> 20.35%), and improvements from evolution (e.g., Qwen2.5-3B-Instruct 5.27% -> 7.70%, Qwen2.5-14B-Instruct 16.53% -> 21.08%). This project is still ongoing. We welcome feedback from the community and will frequently update the source code and technical report.

cs.AI↗

Reference-guided Policy Optimization for Molecular Optimization via LLM Reasoning

Large language models (LLMs) benefit substantially from supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR) in reasoning tasks. However, these recipes perform poorly in instruction-based molecular optimization, where each data point typically provides only a single optimized reference molecule and no step-by-step optimization trajectory. We reveal that answer-only SFT on the reference molecules collapses reasoning, and RLVR provides sparse feedback under similarity constraints due to the model's lack of effective exploration, which slows learning and limits optimization. To encourage the exploration of new molecules while balancing the exploitation of the reference molecules, we introduce Reference-guided Policy Optimization (RePO), an optimization approach that learns from reference molecules without requiring trajectory data. At each update, RePO samples candidate molecules with their intermediate reasoning trajectories from the model and trains the model using verifiable rewards that measure property satisfaction under similarity constraints in an RL manner. Meanwhile, it applies reference guidance by keeping the policy's intermediate reasoning trajectory as context and training only the answer in a supervised manner. Together, the RL term promotes exploration, while the guidance term mitigates reward sparsity and stabilizes training by grounding outputs to references when many valid molecular edits exist. Across molecular optimization benchmarks, RePO consistently outperforms SFT and RLVR baselines (e.g., GRPO), achieving improvements on the optimization metric (Success Rate $\times$ Similarity), improving balance across competing objectives, and generalizing better to unseen instruction styles. Our code is publicly available at https://github.com/tmlr-group/RePO.

cs.LG↗

System Prompt Poisoning: Persistent Attacks on Large Language Models Beyond User Injection

Large language models (LLMs) have gained widespread adoption across diverse applications due to their impressive generative capabilities. Their plug-and-play nature enables both developers and end users to interact with these models through simple prompts. However, as LLMs become more integrated into various systems in diverse domains, concerns around their security are growing. Existing studies mainly focus on threats arising from user prompts (e.g. prompt injection attack) and model output (e.g. model inversion attack), while the security of system prompts remains largely overlooked. This work bridges the critical gap. We introduce system prompt poisoning, a new attack vector against LLMs that, unlike traditional user prompt injection, poisons system prompts hence persistently impacts all subsequent user interactions and model responses. We systematically investigate four practical attack strategies in various poisoning scenarios. Through demonstration on both generative and reasoning LLMs, we show that system prompt poisoning is highly feasible without requiring jailbreak techniques, and effective across a wide range of tasks, including those in mathematics, coding, logical reasoning, and natural language processing. Importantly, our findings reveal that the attack remains effective even when user prompts employ advanced prompting techniques like chain-of-thought (CoT). We also show that such techniques, including CoT and retrieval-augmentation-generation (RAG), which are proven to be effective for improving LLM performance in a wide range of tasks, are significantly weakened in their effectiveness by system prompt poisoning.

cs.CR↗

Communication Efficient Robotic Mixed Reality with Gaussian Splatting Cross-Layer Optimization

Realizing low-cost communication in robotic mixed reality (RoboMR) systems presents a challenge, due to the necessity of uploading high-resolution images through wireless channels. This paper proposes Gaussian splatting (GS) RoboMR (GSMR), which enables the simulator to opportunistically render a photo-realistic view from the robot's pose by calling ``memory'' from a GS model, thus reducing the need for excessive image uploads. However, the GS model may involve discrepancies compared to the actual environments. To this end, a GS cross-layer optimization (GSCLO) framework is further proposed, which jointly optimizes content switching (i.e., deciding whether to upload image or not) and power allocation (i.e., adjusting to content profiles) across different frames by minimizing a newly derived GSMR loss function. The GSCLO problem is addressed by an accelerated penalty optimization (APO) algorithm that reduces computational complexity by over $10$x compared to traditional branch-and-bound and search algorithms. Moreover, variants of GSCLO are presented to achieve robust, low-power, and multi-robot GSMR. Extensive experiments demonstrate that the proposed GSMR paradigm and GSCLO method achieve significant improvements over existing benchmarks on both wheeled and legged robots in terms of diverse metrics in various scenarios. For the first time, it is found that RoboMR can be achieved with ultra-low communication costs, and mixture of data is useful for enhancing GS performance in dynamic scenarios.

cs.RO↗

The 1st International Workshop on Disentangled Representation Learning for Controllable Generation (DRL4Real): Methods and Results

This paper reviews the 1st International Workshop on Disentangled Representation Learning for Controllable Generation (DRL4Real), held in conjunction with ICCV 2025. The workshop aimed to bridge the gap between the theoretical promise of Disentangled Representation Learning (DRL) and its application in realistic scenarios, moving beyond synthetic benchmarks. DRL4Real focused on evaluating DRL methods in practical applications such as controllable generation, exploring advancements in model robustness, interpretability, and generalization. The workshop accepted 9 papers covering a broad range of topics, including the integration of novel inductive biases (e.g., language), the application of diffusion models to DRL, 3D-aware disentanglement, and the expansion of DRL into specialized domains like autonomous driving and EEG analysis. This summary details the workshop's objectives, the themes of the accepted papers, and provides an overview of the methodologies proposed by the authors.

cs.LG↗

Balancing Latency and Model Accuracy for Fluid Antenna-Assisted LM-Embedded MIMO Network

This paper addresses the challenge of large model (LM)-embedded wireless network for handling the trade-off problem of model accuracy and network latency. To guarantee a high-quality of users' service, the network latency should be minimized while maintaining an acceptable inference accuracy. To meet this requirement, LM quantization is proposed to reduce the latency. However, the excessive quantization may destroy the accuracy of LM inference. To this end, a promising fluid antenna (FA) technology is investigated for enhancing the transmission capacity, leading to a lower network latency in the LM-embedded multiple-input multiple-output (MIMO) network. To design the FA-assisted LM-embedded network with the lower latency and higher accuracy requirements, the latency and peak signal to noise ratio (PSNR) are considered in the objective function. Then, an efficient optimization algorithm is proposed under the block coordinate descent framework. Simulation results are provided to show the convergence behavior of the proposed algorithm, and the performance gains from the proposed FA-assisted LMembedded network over the other benchmark networks in terms of network latency and PSNR.

eess.SP↗

Packet Header Recognition Utilizing an All-Optical Reservoir Based on Reinforcement-Learning-Optimized Double-Ring Resonator

Optical packet header recognition is an important signal processing task of optical communication networks. In this work, we propose an all-optical reservoir, consisting of integrated double-ring resonators (DRRs) as nodes, for fast and accurate optical packet header recognition. As the delay-bandwidth product (DBP) of the node is a key figure-of-merit in the reservoir, we adopt a deep reinforcement learning algorithm to maximize the DBPs for various types of DRRs, which has the advantage of full parameter space optimization and fast convergence speed. Intriguingly, the optimized DBPs of the DRRs in cascaded, parallel, and embedded configurations reach the same maximum value, which is believed to be the global maximum. Finally, 3-bit and 6-bit packet header recognition tasks are performed with the all-optical reservoir consisting of the optimized cascaded rings, which have greatly reduced chip size and the desired "flat-top" delay spectra. Using this optical computing scheme, word-error rates as low as 5*10-4 and 9*10-4 are achieved for 3-bit and 6-bit packet header recognition tasks, respectively, which are one order of magnitude better than the previously reported values.

eess.SP↗

Comparison Research of Millimeter-Wave/Infrared Co-aperture Reflector Antenna Systems Based on a Specialized Film

This paper presents a novel co-aperture reflector antenna operating in millimeter-wave (MMW) and infrared (IR) for cloud detection radar. The proposed design combines a back-fed dual-reflector antenna, an IR optical reflection system, and a specialize thin film with IR-reflective/MMW-transmissive properties. Simulations demonstrate a gain exceeding 50 dBi within 94 GHz plush or minus 500 MHz bandwidth, with 0.46° beamwidth in both azimuth (E-plane) and elevation (H-plane) and sidelobe levels below -25 dB. This co-aperture architecture addresses the limitations of standalone MMW and IR radars, enabling high-resolution cloud microphysical parameter retrieval while minimizing system footprint. The design lays a foundation for airborne/spaceborne multi-mode detection systems.

eess.SP↗

Green Robotic Mixed Reality with Gaussian Splatting

Realizing green communication in robotic mixed reality (RoboMR) systems presents a challenge, due to the necessity of uploading high-resolution images at high frequencies through wireless channels. This paper proposes Gaussian splatting (GS) RoboMR (GSRMR), which achieves a lower energy consumption and makes a concrete step towards green RoboMR. The crux to GSRMR is to build a GS model which enables the simulator to opportunistically render a photo-realistic view from the robot's pose, thereby reducing the need for excessive image uploads. Since the GS model may involve discrepancies compared to the actual environments, a GS cross-layer optimization (GSCLO) framework is further proposed, which jointly optimizes content switching (i.e., deciding whether to upload image or not) and power allocation across different frames. The GSCLO problem is solved by an accelerated penalty optimization (APO) algorithm. Experiments demonstrate that the proposed GSRMR reduces the communication energy by over 10x compared with RoboMR. Furthermore, the proposed GSRMR with APO outperforms extensive baseline schemes, in terms of peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM).

cs.RO↗