arXiv subjects

Kai Chen

Publications and source records attributed to Kai Chen.

At least 19 recordsLinked to original sources

arXiv2026

The Value of Memory in Language Agents

Memory enables language agents to carry useful distinctions from past experience into future decisions. Its value depends on which distinctions survive, whether a bounded agent can use them, and what can be recovered or revised as demands evolve. We develop a decision-theoretic framework that follows this value from raw history through persistent state, operational use, and costly recovery and revision. Original and retained value are expressed in units of avoided task loss; risk-cost frontiers describe attainable recovery and preservation across uses. Unknown demand yields value profiles, while interactive feedback changes the distribution of future work. A cost-aware simulation condition and five minimal counterexamples expose the roles of information timing, retrieval access, component interaction, feedback and demand uncertainty. In particular, a pointer with zero mutual information about the answer can still reduce the cost of exact recovery. Diagnostics on LoCoMo, a two-project update task and longitudinal cases exhibit failures of preservation, use and measurement. At fixed retrieved evidence, a saved ledger changes subsequent answers; at fixed agent outputs, an evaluator revision changes scores. The framework connects these distinctions to operational risk bounds and testable principles for allocating memory, preserving recovery options and revising evidence dependencies.

cs.AI↗

arXiv2026

A Survey on Long-Term Memory Security in LLM Agents: Attacks, Defenses, and Governance Across the Memory Lifecycle

The emergence of writable, cross-session persistent memory in LLM agents introduces a qualitatively different threat landscape from conventional input-centric security concerns, characterized by three properties: persistence, statefulness, and propagation. To systematically characterize this landscape, we propose a Memory Lifecycle Framework that organizes attacks, defenses, and their cross-phase dependencies along two axes: six lifecycle phases (Write, Store, Retrieve, Execute, Share & Propagate, Forget & Rollback) and four security objectives (Integrity, Confidentiality, Availability, Governance). This analysis in turn exposes the need for explicit security requirements at the system level, motivating Verifiable Memory Governance (VMG), a framework of five architectural primitives for specifying and evaluating auditable, recoverable control over memory state within declared system and threat-model boundaries. Our analysis motivates integrating provenance, versioning, and policy-aware retention with retrieval and execution defenses to support auditable prevention, containment, and recovery across the lifecycle.

cs.CR↗

arXiv2026

Architecture-Aware Reinforcement Learning Makes Sliding-Window Attention Competitive in Math Reasoning

The rapid progress of reasoning and agentic large language models (LLMs) has increased the demand for long-context inference, but self-attention (SA) scales quadratically with context length. To address this, we study SWARR (Sliding-Window Attention with Reinforced Adaptation for Math Reasoning), a practical recipe for adapting SWA models to mathematical reasoning. SWARR has two stages: (1) efficient conversion from a pretrained SA model to SWA with supervised fine-tuning (SFT), which avoids pretraining a new base model, and (2) policy adaptation with reinforcement learning (RL). We find that SWA still underperforms SA after SFT, and we hypothesize that this gap is caused in part by a data-architecture mismatch: most SFT data are prepared for SA models and may contain long-range dependencies that are difficult for SWA to model. Because on-policy RL optimizes self-generated trajectories under the SWA constraint, it can adapt trajectories to better match SWA. Experiments on mathematical reasoning benchmarks show that this recipe substantially narrows the gap between SWA and SA, recovering much of the accuracy lost during SWA conversion while preserving the efficiency benefits of linear-complexity attention. Our central contribution is the empirical finding that RL changes the conclusion one would draw from conversion and SFT alone about SWA's viability for math reasoning.

cs.AI↗

arXiv2026

Concept Acupuncturist: An Efficient and Precise Concept Erasure Method for Diffusion Models via Sparse Autoencoders

Diffusion-based image generation models produce high-quality images but may generate harmful or policy-violating content, motivating machine unlearning (MU) to selectively remove target concepts while preserving general capabilities. However, existing MU approaches often struggle with the distributed and entangled nature of concept representations, leading to inefficient optimization and imprecise edits when applied through large-scale fine-tuning over full diffusion trajectories. To bridge this gap, we propose Concept Acupuncturist (ConceptACP), a novel MU framework that integrates interpretability-guided analysis with inference-free fine-tuning for efficient and precise concept removal. By establishing mappings between semantic concepts and cross-attention blocks in the U-Net denoiser and localizing the most concept-relevant block for targeted optimization, ConceptACP avoids full diffusion trajectory optimization, thereby improving fine-tuning efficiency. To mitigate concept entanglement, we propose a feature-level correction mechanism based on Sparse Autoencoders (SAEs), where SAE-derived sparse, semantically meaningful feature activations are manipulated to construct targeted correction signals, which are then projected back to the output space of the corresponding block and used as supervision for precise concept removal. Consequently, ConceptACP eliminates the need for external inference-time modules while achieving efficient and fine-grained unlearning. Instantiated on SDXL-Turbo, ConceptACP improves the forgetting success rate by 36.34 percentage points in the standard one-step generation setting on average over the best-performing baseline, while preserving image quality and prompt fidelity, and generalizes effectively across other Stable Diffusion backbones. Our code is available at https://github.com/Eran-Xia/ConceptACP.

cs.CV↗

arXiv2026

Compiling Autoregressive LLM Control into Policy-Exact Parallel Inference for Hierarchical RAG

Large language models increasingly act as controllers in structured decision pipelines, yet their decisions are still executed by autoregressive text decoding. Hierarchical RAG is one such setting: at each frontier of a document tree, an LLM commits to $b$ of $k$ sections under an input-dependent action grammar, executed as either failure-prone free-form JSON or reliable but strictly serial grammar-constrained generation. We show that this serialization is unnecessary for bounded-horizon control: the reachable future forms a shallow action-prefix tree, which we compile --- one root prefill plus one tree-causal probe evaluate every prefix-conditioned action distribution with serial decoding's tokens, positions, and causal ancestry. A prefix-state lemma makes compiled control policy-exact in exact arithmetic: the same policy with a two-forward critical path and zero output tokens. One further probe pass certifies the exact order-marginalized MAP set. Across four document-disjoint splits and two scales on the primary MetaX device, compiled control matches the serial policy on $84$--$89\%$ of questions --- every instrumented divergence a floating-point near-tie --- and cuts control latency by $35$--$47\%$ versus token-wise serial decoding and $6$--$7\%$ versus an idealized chunked engine; an H800 repetition preserves the ordering.

cs.CL↗

arXiv2026

D-Pingmesh: Dynamic Probing and Fault Localization for Reconfigurable Datacenter Networks

Pingmesh monitors datacenter networks through periodic pings between servers. In reconfigurable datacenter networks, topology engineering (TPE) and traffic engineering (TE) make uniform server-pair probing inefficient: a pinglist wastes budget retesting common paths while missing low-probability interfaces. Frequent control-plane updates further require rapid pinglist refresh, and cause path ambiguity at control-plane boundaries for fault localization. We present D-Pingmesh, a probing and fault-localization system co-designed with the TPE/TE control plane. Its control-plane-aware generator builds task-efficient pinglists by greedily admitting probes with high interface coverage under the current TPE/TE snapshot. An incremental replay-and-repair mechanism updates pinglists across reconfigurations rather than recomputing from scratch. At control-plane boundaries, a Rho-EM localizer jointly infers per-task old/new-path mixture weights and the fault set through EM-style iteration. We evaluate D-Pingmesh on an 8-pod reconfigurable topology under dynamic TPE/TE. At a 0.99 coverage target, D-Pingmesh uses 17.7x fewer tasks than Pingmesh and generates each pinglist about 5.1x faster than RD-Probe. In ns-3 simulation, boundary-aware localization improves micro F1 from 0.871 to 0.953 over the deTector PLL baseline. A testbed deployment demonstrates second-scale fault localization with low host overhead. These results demonstrate that D-Pingmesh is practical for reconfigurable datacenter networks with control-plane-aware monitoring.

cs.NI↗

arXiv2026

MemT-1.0: Towards Memory-Native Evolution

Memory-native evolution requires making memory an intrinsic capability of foundation models, enabling them to continuously accumulate and utilize information. To this end, we introduce MemT-1.0, a new architecture of memory foundation models that internalize memory as part of model computation. Building on Metis, MemT-1.0 employs Gated DeltaNet-2 with a stable token-level memory update mechanism within each turn. We further introduce Mixture-of-Storage to scale native memory capacity. Following a deliberate backbone preparation, we develop a two-stage training pipeline consisting of continued pre-training and post-training, supported by 34B tokens of training data and efficient training and inference infrastructure. We release a model family comprising MemT-1.0-Nano and MemT-1.0-Flash, and conduct a comprehensive evaluation of memory and general capabilities. The results show that they outperform the parametric memory baselines on average and remain competitive with retrieval-augmented methods, while largely preserving general capabilities. At a history length of 128K tokens, MemT-1.0-Nano achieves approximately $1.86\times$ end-to-end inference speedup and a $66\times$ reduction in persistent storage relative to full context KV cache. By internalizing retrieved procedural experience, MemT-1.0 improves task success on agentic tasks, providing a practical step toward memory-native evolution.

cs.CL↗

arXiv2026

Agentic Physics-Grounded Video Generation via Memory-Augmented In-Context Reasoning

Physical plausibility is essential for realistic video generation. Existing methods impose physical guidance before synthesis via simulations or refined prompts, but this can limit dynamic diversity, miss errors that emerge only after generation, and require repeated regeneration when inconsistencies remain. We therefore study a generate-diagnose-repair paradigm that preserves the original generation capability and selectively repairs physical inconsistencies afterward, raising two key questions: what to repair and how to repair? To address them, we introduce PhysRepair, an agentic video generation framework that combines causal reasoning over the observed physical failure with prior repair experience to enable localized and adaptive correction. Specifically, Spatiotemporal Causal Constraints identify the first invalid physical state and its causally affected states, producing a repair specification that guides physically plausible local regeneration while preserving unaffected content. Experience-augmented In-context Learning combines the ongoing repair context with relevant experience from previous repairs, allowing the agent to adapt its repair decisions across diverse physical failures and accumulate completed trajectories for future cases. Experiments on Physics-IQ Verified across solid mechanics, fluid dynamics, optics, thermodynamics, and magnetism demonstrate that PhysRepair consistently improves the physical plausibility of generated videos over existing methods. Our project page is available at https://xxkkxxx.github.io/PhysRepair-Website/.

cs.CV↗

arXiv2026

PrefixFlow: Training-Time KV Caching for Schedule-Level Prefix Reuse in LLM RL Training

In reinforcement learning (RL) post-training, one prompt fans out into $N$ trajectories, yet standard trainers replay the shared prefix's forward and backward pass $N$ times; in long-context retrieval-augmented generation (RAG) and multimodal workloads, this repetition dominates training compute. Existing designs reuse the prefix within one computation graph (a packed sequence, fused kernel, or tree traversal); under the trainer's memory limits and parallelism, their compute savings do not reliably become shorter steps. We present PrefixFlow, which takes the prefix out of that graph and into the training schedule through a training-time KV cache: Phase A runs one prefix forward and caches its keys and values ($K/V$); Phase B trains suffixes as ordinary microbatches that read only $K/V$ and write back only their gradients ($gK/gV$); Phase C runs one delayed prefix backward on the accumulated $gK/gV$. Memory and parallelism then become manageable: dormant prefix activations are offloaded during Phase B to free high-bandwidth memory (HBM) for suffixes, and context, pipeline, and data parallelism each schedule the phases through a local rule while tensor and expert parallelism need no change, so saved compute becomes shorter steps under any microbatch split. On dense models, and on mixture-of-experts (MoE) models under Rollout Routing Replay, PrefixFlow matches full-sequence training within bfloat16 (BF16) tolerance. Against an identically configured baseline, it speeds up end-to-end verl RL training on a long-prefix RAG workload by $1.90\times$ ($6.1\times$ on the actor update). In controlled sweeps, actor-side speedup grows with prefix ratio and group size, reaching $4.4\times$ at an 83% prefix ratio, while suffix-dominated splits gain little; offloading dormant prefix activations lowers Phase-B peak HBM by up to 59% and extends trainable sequences by $1.66\times$.

cs.DC↗

arXiv2026

SHIP: Structured and Hierarchical Insight with Privacy

Extracting insights from textual data while protecting user privacy is increasingly important. Existing methods produce private insights in diverse forms for different analytical purposes, making their utility difficult to assess and compare. In this work, we formulate differentially private (DP) insights as reusable representations for answering analytical queries, providing a common basis for utility assessment. Answering these queries, however, requires fine-grained details and cross-attribute statistics that existing representations may omit. To capture this information, we introduce Structured and Hierarchical Insight with Privacy (SHIP), which automatically maps queries to attributes, builds a hierarchical hybrid taxonomy, and releases a DP synthetic table for further analysis. We also develop an evaluation protocol based primarily on QA performance and structured fidelity. Our experiments demonstrate SHIP's strong performance on both measures relative to existing private insight baselines, while identifying queries that remain challenging for structured representations.

cs.CL↗

arXiv2026

Functional Subnetwork Localization for Efficient Continual Learning in Large Language Models

Biological intelligence continually adapts to new information while retaining prior skills through localized plasticity. In contrast, continual learning in large language models (LLMs) remains vulnerable to catastrophic forgetting. A promising direction is to identify functionally specialized subnetworks within LLMs that enable targeted and resource-efficient adaptation. Using gradient-attribution profiling, we localize cognition-associated functional subnetworks exhibiting extremely low mutual overlap. Building on this finding, we propose Function-isolated Memory-Efficient Optimization (CiMEO), a continual learning framework that routes each training example to its corresponding functional subnetwork, restricting parameter updates and optimizer states to selected regions. Across 1.4 million sequential instructions, CiMEO achieves the highest average performance among the compared methods on five generative benchmarks spanning programming, examination-style tasks, and mathematical reasoning, while mitigating catastrophic forgetting. Subspace-overlap and principal-angle analyses further indicate that localized updates help preserve pre-trained representations, highlighting functional subnetwork localization as a practical approach to efficient and stable continual learning in LLMs.

cs.AI↗

arXiv2026

EvoSpatial: Experience-Driven Skill and Tool Evolution for Spatial Reasoning Agents

Spatial reasoning agents can inspect tool outputs and repair errors, but reusing these lessons requires preserving both the computation and the conditions that make it valid. We introduce EvoSpatial, a framework that learns from execution experience by updating a library of linked skills and tools. Skills describe procedures, geometric assumptions, and verification steps; tools implement the operations those procedures require. When solving a task, the agent retrieves relevant skills and synthesizes reusable computational functions if existing tools are insufficient. After each task, execution feedback and evidence checks guide skill revision and tool consolidation, without ground-truth answers or benchmark scores. Across 20 benchmarks, EvoSpatial averages 63.9, 4.0 percentage points above the average reported by SpatialClaw. The evolved library also improves held-out and cross-benchmark performance when updates are disabled, and trajectories generated with evolved skills improve supervised fine-tuning, even without external skills at evaluation. Keeping checked procedures on record, together with the tools that support them, lets an agent carry spatial reasoning experience from one task to the next.

cs.CV↗

arXiv2026

HyrCap: Hybrid Rank-Calibration of Action Proposals for Temporal Event Understanding

Temporal action detection in long untrimmed videos suffers from inaccurate localization and low-quality proposals, especially with sparse annotations and ambiguous boundaries. For open-ended video understanding and LLM-driven event analysis, reliable proposals connect temporal localization, event annotation, and semantic understanding. However, conventional temporal action detection (TAD) pipelines lack explicit modeling of proposal localization quality, cross-representation support, and inter-proposal relations. We propose Hybrid Rank-Calibration of Action Proposals (HyrCap), a proposal-centric query-based TAD framework for temporal event understanding. Hybrid combines complementary video features with different localization cues, while Rank-Calibration adjusts proposal ranking confidence to reflect localization reliability and candidate competition. HyrCap treats dense query-based predictions as an over-complete temporal event hypothesis space. Cross-Representation Consensus Calibration, Localization Quality Estimation, and Proposal Redundancy Suppression improve proposal ranking reliability, localization quality, and candidate discrimination. Inspired by LLMs' capabilities in fine-grained semantic recognition, event description, and reasoning, we introduce HyrCap for Semantic Action Understanding, using calibrated proposals for standard TAD and as inputs to LLM-based semantic understanding. We define a staged evaluation protocol to separately measure event localization, segment-level semantic understanding, and combined event understanding performance. Experiments on public benchmarks show that HyrCap achieves state-of-the-art performance across multiple metrics and provides a finer-grained paradigm and evaluation protocol for query-based event segmentation and annotation.

cs.CV↗

arXiv2026

EgoWholeView: Camera-Only Head-Mounted Capture of Body and Bimanual Hand Motion for Embodied Demonstrations

Head-mounted body and bimanual motion capture supports self-avatars, teleoperation, and embodied demonstrations. However, a single egocentric camera is vulnerable to a limited field of view, scale ambiguity, and hand-object occlusion. We present EgoWholeView, a wearable capture system of six synchronized wide-angle RGB cameras with a role-specialized pipeline. The outward views localize the rig in a world frame, while an overlapping downward pair independently predicts SMPL-X body and MANO hand parameters. Both predictors are adapted to egocentric fisheye data. To combine the two, reliability cues measured from the predictions themselves weight the fusion of body and hand joint rotations, while shape is aggregated across views and frames. The pair's baseline fixes metric scale, and triangulation recovers root translation. Relative to monocular prediction, fusion reduces PA-MPJPE by 3.7% for the body and 11.2% for the hands, and temporal refinement lowers angular-acceleration magnitude by 32.2% and 31.0%, respectively. On our six-camera recording, the pipeline produces body and hand states in a shared world frame using benchmark-fitted gates without retuning, and temporal refinement lowers hand acceleration magnitude by 41.08%. Using only RGB cameras and no third-person view, EgoWholeView offers a low-cost, portable route to egocentric interaction data.

cs.CV↗

arXiv2026

Decoupling Trap States in Amorphous Oxide Semiconductor Transistor Using a Conductive-Semiconducting-Conductive MOS Structure

We propose a novel amorphous oxide semiconductor (AOS)-based conductive-semiconducting-conductive (CSC) MOS test structure to effectively sense, discriminate, and analyze distinct trap-induced conductance losses in AOS transistors. In conventional test structures, non-ideal metal-AOS contacts convolute the conductance loss generated by various trap states. The CSC architecture mitigates these contact-related artifacts, enabling the decoupling of interface states and shallow states. Furthermore, the contributions from leakage, series resistance, and device-geometry effects are identified and calibrated. Our results yield bias-resolved trap profiles and reveal three distinct regimes: (1) In the depletion region, interface states dominate the conductance loss. These states exhibit sub-microsecond time constants and their density increases slightly with deeper depletion, which is associated with donor-like states. (2) From weak to moderate accumulation, shallow states become the primary contributor, displaying a Gaussian-like volumetric density-of-states distribution with response times spanning from microseconds to nanoseconds. (3) In the strong accumulation regime, interface states regain dominance with an increased state quantity, while their time constants remain invariant. These bias-resolved profiles clarify how different trap populations impact AOS transistors under distinct operating conditions, providing intuitive metrics for understanding reliability degradation. Consequently, the CSC structure offers a simple, scalable, and robust platform for trap characterization in AOS transistors.

cond-mat.mtrl-sci↗

arXiv2026

Gestalt: Large Multimodal Interplay Model

In this paper, we propose Gestalt, a new paradigm of large multimodal model built around multimodal interplay. Despite rapid advances, large multimodal models are reaching a bottleneck: existing approaches focus primarily on accommodating additional modalities while overlooking the distinct characteristics of each modality and the relations among them. Motivated by the multistage property of human multisensory perception, we propose a multimodal interplay pyramid that organizes multimodal modeling as a progression from modality-specific processing, through cross-modal alignment, to deeper multimodal integration. Guided by this pyramid, Gestalt adopts a unified discrete diffusion framework and an interplay-partitioned architecture, with learnable interplay tokens mediating cross-modal exchange and integration. The pyramid also structures its data organization and training strategy. Strong performance across image generation, multimodal understanding, and text-only evaluation shows that Gestalt significantly improves cross-modal integration while preserving modality-specific information, effectively harnessing the strengths of diffusion-based multimodal models and offering a promising path toward unified multimodal intelligence. Project page: https://GeWu-Lab.github.io/Gestalt.

cs.CV↗

arXiv2026

ResOPD: Tail Residualization for Sparse On-Policy Distillation

On-policy distillation (OPD) trains a student model on self-generated trajectories, but transmitting dense teacher distributions across long reasoning traces creates prohibitive communication and memory bottlenecks. Practical systems therefore rely on sparse teacher interfaces, typically transmitting either the sampled-token score or a small Top-$k$ distribution. However, this sparse setting faces a fundamental dilemma: sampled-token estimators are unbiased but suffer from severe gradient variance, whereas directly optimizing Top-$k$ objectives introduces systematic bias. To improve this trade-off, we propose ResOPD (On-Policy Distillation with Tail Residualization), which provides unbiased full-vocabulary reverse KL gradient estimation under on-policy sampling, with substantial variance reduction under the same sparse payload in the evaluated settings. ResOPD aggregates the unobserved vocabulary into an observable coarse tail event, computes its exact aggregate gradient, and samples only the fine-grained within-tail residual, which requires no additional teacher queries or forward passes. Extensive experiments demonstrate that ResOPD substantially reduces gradient variance, stabilizes online training dynamics, and improves downstream performance across the evaluated settings. These results establish ResOPD as an efficient, plug-and-play variance reduction primitive for sparse on-policy distillation.

cs.LG↗

arXiv2026

Equivariant Visual-Tactile Diffusion Policy for Contact-Rich Manipulation

Imitation learning for contact-rich manipulation requires high-quality expert data that is expensive to obtain. This makes learning a sample-efficient policy a key issue. To address this, we propose VISTA, a workspace-level equivariant visuotactile diffusion policy for data-efficient contact-rich imitation learning. VISTA projects visual and tactile observations into spherical tokens, injects tactile contact cues into visual spherical directions through permutation-equivariant spherical fusion, and rotates the fused harmonic representation using the end-effector orientation. The resulting representation conditions an equivariant diffusion policy to predict spatially consistent actions. Extensive experiments in both simulation and real-world robotic settings show that VISTA substantially improves data efficiency over strong visuotactile imitation learning baselines. Project website: https://vista-paper.github.io/

cs.RO↗