Searcharxiv⌕ Search

arXiv subjects

Jingwei Sun

Publications and source records attributed to Jingwei Sun.

At least 19 recordsLinked to original sources
arXiv2026

Rethinking How to Remember: Beyond Atomic Facts in Lifelong LLM Agent Memory

To enable reliable long-term interaction, LLM agents require a memory system that can faithfully store, efficiently retrieve, and deeply reason over accumulated dialogue history. Most existing methods adopt an extracted fact based paradigm: handcrafted static prompts compress raw dialogues into atomic facts, which are then stored, matched, and injected into downstream reasoning. Nevertheless, such fact-centric designs inevitably discard fine-grained details in original dialogues and fail to support deep reasoning over scattered isolated facts. Moreover, static prompts cannot maintain consistent extraction granularity across diverse dialogue styles. To address these limitations, we propose TriMem, which maintains three coexisting representation granularities, including raw dialogue segments anchored by source identifiers for storage fidelity, extracted atomic facts for efficient memory retrieval, synthesized profiles that aggregate dispersed facts into holistic semantic understanding for deep reasoning. We further adopt TextGrad-based prompt optimization, which iteratively refines extraction and profiling prompts via response quality feedback, achieving lifelong evolution without any parameter updating. Extensive experiments on LoCoMo and PerLTQA across multiple LLM backbones demonstrate that TriMem consistently outperforms strong memory baselines. The code is available at https://TMLR-TriMem.github.io .

cs.CL↗
arXiv2026

Steering Vision-Language Models for Visual-Contextual Privacy in Location Disclosure

Vision-language models (VLMs) can infer sensitive information such as geolocation from subtle visual cues, creating privacy risks when responses disclose more detail than the visual context warrants. We study visual-contextual privacy in geolocation disclosure, where a VLM should reveal location information only at the granularity appropriate to the visual context depicted in the image. We first show that this contextually appropriate granularity is recoverable from pre-generation hidden states, even when generated responses are misaligned, with disclosure levels exhibiting an ordered structure along a common hidden-state direction. Motivated by this, we propose an inference-time route-and-steer framework: an ordinal granularity router maps visual-contextual privacy representations to a disclosure route, and a disclosure-level-conditioned steering module applies route-specific hidden-state interventions during decoding. Across four VLM backbones, our method improves visual-context-aware disclosure alignment, better balances over- and under-disclosure, and incurs only modest changes on general capability benchmarks. It further transfers to another dataset with a distinct data distribution and privacy-risk taxonomy.

cs.CV↗
arXiv2026

Gamma: Grounded Multi-Agent Memory with Mechanical Audit for Memory-Intensive Manipulation

Pretrained vision-language-action (VLA) policies fail on memory tasks, behaviour that depends on events observed long before the current frame, because their training data is overwhelmingly Markov and mapping episodic memory directly to action chunks is hard. Yet the same policy, finetuned once to follow a privileged oracle's grounded-subgoal text, nearly solves these tasks: the difficulty lies entirely in mapping memory to a grounded symbolic subgoal. Existing work entrusts this stage to a single VLM that selects keyframes and reasons over raw frames, and falls far short of the oracle: key identities and dynamics are buried in noisy pixels, the reasoning spans hundreds of frames in an unbounded context, and hallucination corrupts the memory itself. We instead present GAMMA (Grounded multi-Agent Memory with Mechanical Audit), which builds the stage as a collaborating two-agent memory system: a detector tool (SAM-3) for spatial grounding, a VLM writer that distils each replan window into one grounded text line, an append-only text memory bank, and a VLM reasoner that reads the bank and emits the next grounded subgoal at the oracle's cadence, wrapped in a propose-verify harness that admits, defers, corrects, or rejects every agent claim against mechanical visual evidence, at training-corpus construction and at deployment alike. On a sixteen-task memory-manipulation suite GAMMA recovers a large fraction of the oracle ceiling with the policy held fixed, reaching the ceiling on several tasks, and transfers with unchanged contracts to a second benchmark.

cs.RO↗
arXiv2026

Rethinking Faithfulness in LLMs: A Pairwise Context-Sensitive Perspective

Large language models (LLMs) are expected to answer questions faithfully based on the provided context, abstaining when the context information is insufficient to answer the questions. Existing faithfulness evaluations typically assess each question-context instance in isolation; however, such instance-level evaluation fails to capture a fundamental requirement of faithful behavior: the ability to adapt model responses to changes in available contexts. In particular, a model should provide correct answers when sufficient evidence is present and abstain when it is not. In this work, we propose a Pairwise Faithfulness Benchmark (PFaithBench) that evaluates whether a model can switch between answering and abstaining for the same question under supporting versus non-supporting contexts. Our evaluations across thirty-nine models with seven model families demonstrate that faithfulness fundamentally involves a trade-off between answering and abstaining, and that most current models exhibit a strong bias toward answering, with most faithfulness errors arising from over-answering, i.e., models tend to fabricate a response even when the provided context is insufficient. We further conduct a series of studies on faithfulness training under different data constructions. Our results show that training outcomes are highly sensitive to the specific composition of answering and abstaining data. Constructing answering and abstaining data from mismatched sources can cause models to rely on dataset-specific shortcuts rather than actual context sufficiency. Moreover, increasing answer-supervised data improves answering performance but exacerbates over-answering, while increasing abstaining data reduces hallucination but leads to over-abstention. The code and data are released at https://github.com/tmlr-group/PFaithBench.

cs.CL↗
arXiv2026

Test-time Calibration Learning for Large Language Model Reasoning

Reliable large language models (LLMs) must not only produce accurate answers but also express confidence that faithfully reflects their probability of being correct. Such calibration is essential for identifying uncertain predictions and supporting reliable decision-making in real-world deployment. Recent studies incorporate calibration learning into reinforcement learning (RL), jointly optimizing answer correctness and verbalized confidence using ground-truth correctness supervision. However, their reliance on labeled data limits their applicability in practical test-time settings, where ground-truth labels are unavailable and calibration may need to adapt to newly encountered target tasks. To address this challenge, we propose Test-Time Calibration Learning (TTCL), a label-free framework that jointly adapts reasoning accuracy and verbalized confidence directly on unlabeled target-task data. Specifically, TTCL derives self-supervision signals for both correctness and calibration from multiple model-generated responses, enabling calibration learning at test time without ground-truth labels. Theoretical analysis further establishes TTCL as a bounded surrogate for the ideal calibration objective. Extensive experiments on mathematical reasoning and factual question answering demonstrate that TTCL consistently improves both accuracy and calibration across diverse models and tasks. On base models, TTCL achieves an average relative accuracy improvement of +40.13% and an ECE reduction of +70.80% across eight benchmarks. Moreover, TTCL can further improve both accuracy and calibration for already calibrated models under domain shift, particularly when source-domain calibration transfers poorly to target tasks. In the math-to-factQA setting, TTCL achieves an average relative accuracy gain of +20.35% and reduces ECE by +53.83%. The source code is released at https://github.com/tmlr-group/TTCL.

cs.LG↗
arXiv2026

LatentSift: Policy-State Filtering for Token-Efficient Verification of Software Engineering Agents

Test-time scaling improves software engineering agents by generating multiple candidate trajectories and selecting the best one. Verifying and selecting among these long interactions can consume as many tokens as generation itself. Existing hybrid workflows first apply an LLM-based execution-free (EF) verifier to filter candidates before running tests, which adds another model pass over every trajectory. We introduce LatentSift, a token-free and execution-free filter that replaces this first stage with hidden states the policy already produces while generating the candidates. It represents each candidate through its reasoning, observation, and function-call states, compares them with positive and negative banks of such states collected from successful and unsuccessful trajectories during policy training, and fuses the resulting distance scores with a learned linear score to retain promising candidates for the execution-based stages. On SWE-bench Verified, across three agents and two policy sizes, LatentSift cuts EF-verifier tokens by 66.6--81.0% and total verification tokens, which include test generation, by 49.1--62.1% at K=16, while hybrid Best@16 matches or improves on each agent's reference workflow, rising from 59.26% to 60.06% on DeepSWE-Preview.

cs.SE↗
arXiv2026

PrivAct: Internalizing Contextual Privacy Preservation via Multi-Agent Preference Training

Large language model (LLM) agents are increasingly deployed in personalized tasks involving sensitive, context-dependent information, where privacy violations may arise in agents' action due to the implicitness of contextual privacy. Existing approaches rely on external, inference-time interventions which are brittle, scenario-specific, and may expand the privacy attack surface. We propose PrivAct, a contextual privacy-aware multi-agent learning framework that internalizes contextual privacy preservation directly into models' generation behavior for privacy-compliant agentic actions. By embedding privacy preferences into each agent, PrivAct enhances system-wide contextual integrity while achieving a more favorable privacy-helpfulness tradeoff. Experiments across multiple LLM backbones and benchmarks demonstrate consistent improvements in contextual privacy preservation, reducing leakage rates by up to 12.32% while maintaining comparable helpfulness, as well as zero-shot generalization and robustness across diverse multi-agent topologies. Code is available at https://github.com/chengyh23/PrivAct.

cs.CL↗
arXiv2026

FlashQuant: Sparse-Dense Fusion for Memory-Efficient Outlier-Aware LLM Inference

Low-bit quantization reduces the memory footprint and computational cost of large language model (LLM) inference. However, high-magnitude outlier weights can induce substantial quantization errors and degrade model accuracy. Outlier-aware quantization addresses this issue by retaining outliers in high precision while quantizing the remaining weights, resulting in a low-bit dense GEMM path and a high-precision sparse SpMM path. Existing implementations execute these paths in separate GPU kernels, despite their shared activations and outputs, thereby missing opportunities for intra-operator reuse and incurring redundant global-memory accesses. This inefficiency is particularly pronounced in memory-bound decoding workloads. We propose FlashQuant, a content-sharing execution framework for outlier-aware W4A16 decoding. FlashQuant fuses the dense GEMM and sparse outlier SpMM paths into a single GPU kernel, enabling on-chip reuse of activation and output tiles across heterogeneous computations. It introduces three key techniques: sparse-dense tiling, which aligns outlier processing with dense GEMM tiles; Tile-COO outlier encoding, which enables efficient sparse access and reduces shared-memory bank conflicts; and pipelined scheduling, which overlaps computation with data movement. Experiments show that FlashQuant reduces outlier-processing overhead, achieving $2.74\times - 4.18\times$ speedup over cuBLAS BF16 and up to $1.53\times$ speedup over the strongest unfused outlier-aware baseline.

cs.DC↗
arXiv2026

DeltaLog: Deferred Materialization of Recurrent States for Linear Attention Decoding

Linear attention models eliminate the quadratic prefix computation and context-growing KV cache of softmax attention by replacing pairwise token interactions with recurrent state updates. However, existing decoding implementations often materialize and write back the full recurrent state after every generated token, making state maintenance a major source of memory traffic, especially for models with large states and many heads. This paper presents DeltaLog, a recurrent-state decoding scheme that reduces this overhead without changing the model semantics. Specifically, DeltaLog represents the recurrent state as a dense base state together with a bounded log of recent compact updates. Most decode steps append only compact update factors to this log, while periodic merge steps fold the accumulated updates back into the dense base state. Thus, the model observes the same dense state as in eager decoding, but most full-state write-backs are replaced by lightweight append operations. We implement DeltaLog for GDN, KDA, and RWKV6 and integrate it into a prototype serving stack. Across these models, DeltaLog accelerates the recurrent-state update kernel by up to $1.86\times$, reduces profiled recurrent-state write traffic by up to $7.83\times$, and achieves $1.05$--$1.20\times$ end-to-end serving speedups over dense recurrent baselines.

cs.DC↗
arXiv2026

EMO: Energy Efficiency Modeling and Optimization for AI Workloads

The massive energy consumption of GPU-accelerated AI workloads challenges sustainable computing. We observe that execution asynchrony (e.g., CPU-GPU, concurrent streams, multi-GPU) creates slack, allowing non-critical kernels to run at lower frequencies to save energy without impacting end-to-end latency. However, existing approaches fail to simultaneously achieve workload generality and fine-grained slack discovery, while high-fidelity modeling incurs prohibitive overhead. We present EMO, a lightweight framework exploiting these fine-grained opportunities. First, to identify where to optimize, EMO constructs a low-level dependency graph capturing asynchrony and performs what-if timing analysis to precisely identify slack windows. Second, to determine how to optimize, EMO introduces dependency-aware kernel packing. It aggregates kernels to preserve critical paths while collapsing redundant details, enabling high-fidelity latency-energy modeling with minimal profiling cost. Finally, EMO combines graph analysis and pack-level models to formulate energy optimization as a constrained combinatorial problem, efficiently solving for optimal frequency policies under given latency targets. Evaluations show EMO reduces energy consumption by 15%--28% with only 2%--5% performance loss and negligible overhead.

cs.PF↗
arXiv2026

AgentHijack: Benchmarking Computer Use Agent Robustness to Common Environment Corruptions

Autonomous computer use agents that powered by multimodal large language models (MLLMs) are emerging as capable assistants for completing complex digital workflows. However, real-world execution environments are far from ideal: pop-ups, resolution changes, and competing applications frequently interfere with agent perception and control. We introduce AgentHijack, a benchmark designed to evaluate the robustness of computer-use agents under common corruptions, where the uncertainties in dynamic environment disrupt the execution flow without direct adversarial intent. Specifically, AgentHijack introduces 9 configurable common corruptions to replicate realistic imperfect scenarios. We evaluate a variety of desktop tasks that utilize MLLM-based agents and discover that even minor instances of corruption can result in substantial performance degradation, which emphasizes the fragility of agents and underscores the necessity of robustness evaluation. Afterward, we propose AgentHijack-Agent, a framework that integrates an action generator with enhanced grounding capabilities and an onlooker responsible for behavior summarization and environment checking. Extensive experiments validate its effectiveness. Our code, environment, baseline models and data are publicly available at: https://AgentHijack.github.io.

cs.AI↗
arXiv2026

EgoIntrospect: An Egocentric Dataset and Benchmark for User-Centric Internal State Reasoning

Despite extensive efforts on egocentric video datasets and benchmarks, understanding users' internal states, which is crucial for enabling seamless AI assistant experiences, remains largely overlooked. In this work, we introduce EgoIntrospect, the first egocentric dataset captured in user-driven scenarios with self-annotations that explicitly reveal users' interactive intentions with AI assistants. EgoIntrospect was collected using a cross-device setup, providing synchronized video, audio, gaze, motion, and physiological signals. It consists of 180 hours of recordings from 60 subjects, with an average recording duration of 3 hours per subject. Leveraging EgoIntrospect, we formalize a suite of tasks centered on user internal states, including affective experience, interactive intent, and cognitive memory. We further process the annotations to construct benchmarks that evaluate the ability of modern multimodal large language models to reason about users' internal states from egocentric observations. Experiments on our benchmark suggest that existing multimodal large language models struggle to effectively leverage multimodal signals to infer users' subjective internal states. The dataset and annotations will be made publicly available to advance research in egocentric vision and wearable AI assistants. Project page: https://ego-introspect.github.io/

cs.CV↗
arXiv2026

CATS: Cascaded Adaptive Tree Speculation for Memory-Limited LLM Inference Acceleration

Auto-regressive decoding in Large Language Models (LLMs) is inherently memory-bound: every generation step requires loading the model weights and intermediate results from memory (e.g., High-Bandwidth Memory (HBM) for GPU servers), making throughput bottlenecked by memory bandwidth rather than compute. Speculative decoding addresses this by enabling parallel verification of multiple draft tokens, effectively amortizing the cost of each target-model call. However, existing speculative decoding methods are designed under the assumption that HBM is sufficiently large to hold both the target model and an auxiliary draft model simultaneously -- an assumption that breaks down on memory-constrained devices such as edge platforms with limited DRAM. We analyze the inference bottleneck in this memory-limited regime and propose CATS, a self-speculative decoding framework that conducts cascaded verification and correction based on the memory budget and parameter offloading patterns on memory-limited devices. This design maximizes token acceptance rate and end-to-end speedup while keeping the peak memory footprint on the device equal to that of the target model alone. We evaluate CATS on different models across five benchmarks on real edge devices. CATS can achieve a wall-clock speedup of up to 5.08x with no degradation in generation quality, outperforming the SOTA method by up to 1.45x under edge memory constraints.

cs.LG↗
arXiv2026

Latent Bridge: Feature Delta Prediction for Efficient Dual-System Vision-Language-Action Model Inference

Dual-system Vision-Language-Action (VLA) models achieve state-of-the-art robotic manipulation but are bottlenecked by the VLM backbone, which must execute at every control step while producing temporally redundant features. We propose Latent Bridge, a lightweight model that predicts VLM output deltas between timesteps, enabling the action head to operate on predicted outputs while the expensive VLM backbone is called only periodically. We instantiate Latent Bridge on two architecturally distinct VLAs: GR00T-N1.6 (feature-space bridge) and π0.5 (KV-cache bridge), demonstrating that the approach generalizes across VLA designs. Our task-agnostic DAgger training pipeline transfers across benchmarks without modification. Across four LIBERO suites, 24 RoboCasa kitchen tasks, and the ALOHA sim transfer-cube task, Latent Bridge achieves 95-100% performance retention while reducing VLM calls by 50-75%, yielding 1.65-1.73x net per-episode speedup.

cs.RO↗
arXiv2026

ZEUS: Accelerating Diffusion Models with Only Second-Order Predictor

Denoising generative models deliver high-fidelity generation but remain bottlenecked by inference latency due to the many iterative denoiser calls required during sampling. Training-free acceleration methods reduce latency by either sparsifying the model architecture or shortening the sampling trajectory. Current training-free acceleration methods are more complex than necessary: higher-order predictors amplify error under aggressive speedups, and architectural modifications hinder deployment. Beyond 2x acceleration, step skipping creates structural scarcity -- at most one fresh evaluation per local window -- leaving the computed output and its backward difference as the only causally grounded information. Based on this, we propose ZEUS, an acceleration method that predicts reduced denoiser evaluations using a second-order predictor, and stabilizes aggressive consecutive skipping with an interleaved scheme that avoids back-to-back extrapolations. ZEUS adds essentially zero overhead, no feature caches, and no architectural modifications, and it is compatible with different backbones, prediction objectives, and solver choices. Across image and video generation, ZEUS consistently improves the speed-fidelity performance over recent training-free baselines, achieving up to 3.2x end-to-end speedup while maintaining perceptual quality. Our code is available at: https://github.com/Ting-Justin-Jiang/ZEUS.

cs.LG↗
arXiv2026

GCL-Sampler: Discovering Kernel Similarity for Sampled GPU Simulation via Graph Contrastive Learning

GPU architectural simulation is orders of magnitude slower than native execution, necessitating workload sampling for practical speedups. Existing methods rely on hand-crafted features with limited expressiveness, yielding either aggressive sampling with high errors or conservative sampling with constrained speedups. To address these issues, we propose GCL-Sampler, a sampling framework that leverages Relational Graph Convolutional Networks with contrastive learning to automatically discover high-dimensional kernel similarities from trace graphs. By encoding instruction sequences and data dependencies into graph embeddings, GCL-Sampler captures rich structural and semantic properties of program execution, enabling both high fidelity and substantial speedup. Evaluations on extensive benchmarks show that GCL-Sampler achieves 258.94x average speedup against full workload with 0.37% error, outperforming state-of-the-art methods, PKA (129.23x, 20.90%), Sieve (94.90x, 4.10%) and STEM+ROOT (56.57x, 0.38%).

cs.PF↗
arXiv2026

MNO: Multiscale Neural Operator for 3D Computational Fluid Dynamics

Neural operators have emerged as a powerful data-driven paradigm for solving partial differential equations (PDEs), while their accuracy and scalability are still limited, particularly on irregular domains where fluid flows exhibit rich multiscale structures. In this work, we introduce the Multiscale Neural Operator (MNO), a new architecture for computational fluid dynamics (CFD) on 3D unstructured point clouds. MNO explicitly decomposes information across three scales: a global dimension-shrinkage attention module for long-range dependencies, a local graph attention module for neighborhood-level interactions, and a micro point-wise attention module for fine-grained details. This design preserves multiscale inductive biases while remaining computationally efficient. We evaluate MNO on diverse benchmarks, covering steady-state and unsteady flow scenarios with up to 300k points. Across all tasks, MNO consistently outperforms state-of-the-art baselines, reducing prediction errors by 5% to 50%. The results highlight the importance of explicit multiscale design for neural operators and establish MNO as a scalable framework for learning complex fluid dynamics on irregular domains.

cs.LG↗
arXiv2026

AutoEDA: Enabling EDA Flow Automation through Microservice-Based LLM Agents

Electronic Design Automation (EDA) remains heavily reliant on tool command language (Tcl) scripting to drive complex RTL-to-GDSII flows. This scripting-based paradigm is labor-intensive, error-prone, and difficult to scale across large design projects. Recent advances in large language models (LLMs) suggest a new paradigm of natural language-driven automation. However, existing EDA efforts remain limited and face key challenges, including the absence of standardized interaction protocols and dependence on external APIs that introduce privacy risks. We present AutoEDA, a framework that leverages the Model Context Protocol (MCP) to enable end-to-end natural language control of RTL-to-GDSII design flows. AutoEDA introduces MCP-based servers for task decomposition, tool selection, and automated error handling, ensuring robust interaction between LLM agents and EDA tools. To enhance reliability and confidentiality, we integrate locally fine-tuned LLM agents. We further contribute a benchmark generation pipeline for diverse EDA scenarios and extend CodeBLEU with Tcl-specific enhancements for domain-aware evaluation. Together, these contributions establish a comprehensive framework for LLM-driven EDA automation, bridging natural language interfaces with modern chip design flows. Empirical results show that AutoEDA achieves up to 9.9 times higher accuracy than naive approaches while reducing token usage by approximately 97% compared to in-context learning.

cs.AI↗