arXiv:2606.05899v1 Announce Type: new
Abstract: We develop a high-dimensional statistical theory of low-rank adaptation (LoRA) in attention models, capturing the interplay between pre-training and fine-tuning. We introduce a solvable framework in which a single-head attention layer is first pre-trained on a data-abundant task and subsequently adapted via a rank-one LoRA update on limited data. In the high-dimensional limit, both stages admit a sharp asymptotic characterization in terms of a finite set of order parameters, yielding explicit predictions for test errors and representation alignment. Our analysis shows that the impact of pre-training on LoRA is summarized by an effective noise term, from which we derive prescriptions for the optimal pre-training procedure. We also demonstrate a regime with a mismatch between the value of the test error and representation quality, and propose an application of our theory to active fine-tuning.
Science Journals
arXiv:2606.05903v1 Announce Type: new
Abstract: Performance variations in sensor arrays, caused by intrinsic differences or installation conditions, can lead to inconsistent results during shape sensing. To obtain accurate results, a large amount of data is usually required, and a separate model must be retrained for each sensor array, thereby increasing the cost and time of data acquisition, transmission, and computation. To address this issue, this work proposes an encoder-decoder architecture for surface shape sensing based on sparse strain sensors and further incorporates meta-learning and few-shot adaptation strategies to enable adaptation across different groups of sensor arrays. Experimental results demonstrate that, after the cross-sensor adaptation, a newly deployed sensor array achieves a sensing error of approximately 4.0 mm relying on less than 5.0% newly labeled data and requiring an adaptation time of under 1 second, which represents a substantial improvement from 23.0 mm error without adaptation and 20-minute data collection time required to train a new model. Moreover, the number of points with errors below 5.0 mm increased by more than 65.0%. These results indicate that the proposed method can substantially reduce the cost and training burden of surface shape sensing, and it has broad potential applications in soft robotics and wearable devices.
arXiv:2606.05905v1 Announce Type: new
Abstract: We study an assignment problem where a set of agents and a set of facilities lie on a line metric. The goal is to compute an assignment of agents to facilities to approximately minimize the social cost (the total distance of agents from their assigned facilities) given only partial information regarding the metric. Unlike previous work which focused solely on algorithms with access to the ordinal preferences of the agents over the facilities (ORD), we also consider the value of information regarding approval preferences (APP), and inter-facility distances (DIST). For different combinations of these three information types, we establish tight bounds on the distortion of deterministic algorithms, showing that it is possible to improve over the optimal bound of $3$ that can be achieved using only ORD information. Among other results, we show a tight bound of $1+\sqrt{2}$ for APP+DIST which holds even for general metrics, and a tight bound of $2$ for ORD+APP+DIST.
arXiv:2606.05909v1 Announce Type: new
Abstract: Ambient clinical scribes increasingly combine Automatic Speech Recognition with Large Language Models to automate documentation. However, traditional metrics like Word Error Rate mask systemic safety degradation. We present a paired acoustic stress test to isolate the causal impact of noise on clinical reasoning. For the same dialogues, we inject diverse noise types while keeping the downstream model configuration frozen. Crucially, we uncover a dangerous disconnect between signal fidelity and clinical safety. Stationary ambient noise increased the Word Error Rate by a negligible 0.71 percentage points yet nearly doubled the rate of unsafe outputs. Our analysis reveals that minor acoustic perturbations can invert clinical meaning without substantially inflating error rates. Furthermore, we demonstrate a lightweight mitigation strategy that mitigates safety degradation under noisy conditions without requiring model fine tuning.
arXiv:2606.05914v1 Announce Type: new
Abstract: This paper provides and analyzes a dataset detailing the characteristics and execution data of all jobs submitted to the IN2P3 Computing Center (Villeurbanne, France), a national research and support unit of the CNRS, in 2024. The main additional value of this contribution compared to previously available datasets consists in the combination of an extended time interval considered, the inclusion of memory usage data and its recency, on top on improving the diversity of datasets provenance. This allows researchers to simulate and evaluate scheduling algorithms on a real workload over a large time window. Thus, specificities due to seasonal, monthly, and weekly user behaviors can be taken into account, which is not possible with smaller or synthetic datasets. It is composed of 44M jobs submitted by 1k users running on a cluster of a maximum of 312 machines supporting 46k concurrent threads and providing 105To of RAM.
arXiv:2606.05916v1 Announce Type: new
Abstract: Open-vocabulary object detection seeks to identify novel object categories that were not part of the training data. Many knowledge distillation-based approaches have shown promising performance by transferring knowledge from pre-trained vision-language models to object detection. However, these methods often overlook structured, image-specific relationships between objects, such as interactions and spatial arrangements. This oversight can significantly restrict the effectiveness of detecting novel categories. To address this issue, we propose a Scene-guided Relational Modeling detection framework. This framework utilizes scene graphs to capture structured semantic and spatial relationships between candidate regions and their contextual objects. It explicitly models interactions among neighboring regions and incorporates a Relation Attention Module to implicitly amplify the key relational cues extracted from the scene graph. Furthermore, we present a scene-based textual alignment branch that distills category knowledge from captions to guide relational alignment. This approach facilitates a seamless integration of visual relations with semantic information for enhanced detection performance. Comprehensive experiments show that our model achieves superior performance compared to other OVOD methods, improving the AP for novel categories on COCO and LVIS datasets.
arXiv:2606.05937v1 Announce Type: new
Abstract: Large Language Models (LLMs) are increasingly used, including in political applications, but their political fairness has been little studied. We assess it using perplexity, posing that a fair model should give equal probability to all political groups. However, we find, across ten LLMs and three datasets covering 37 languages, that LLMs are more perplexed by the texts of far right and nationalist parties than of social-democratic parties. We find this to be consistent with previous work on translation fairness, to the point that perplexity correlates with downstream translation metrics. Our method is applicable to both base LLMs as well as their instruction-tuned counterpart, and we find that both are highly correlated, suggesting that the political fairness of LLMs stems from their pretraining, and is hardly affected by instruction-tuning.
arXiv:2606.05938v1 Announce Type: new
Abstract: High-order thermodynamic nonequilibrium (TNE) in three-dimensional compressible flows reflects the breakdown of low-order kinetic moment closure in strong-gradient regions. Using Chapman-Enskog analysis, we identify the kinetic moment constraints required to describe third-order TNE. The analysis yields the third-order constitutive relations and evolution equations for the viscous stress and heat flux, together with second-order expressions for their associated higher-order fluxes. These constraints enable the construction of a three-dimensional super-Burnett-level discrete Boltzmann model with 91 discrete velocities. The resulting D3V91 model reproduces shock-tube wave structures and resolves high-order TNE contributions that lower-order DBMs do not capture reliably. These results demonstrate that high-order TNE has a multigradient, rather than single-gradient, origin. For the four TNE quantities considered here, odd-order central moments, including the heat flux and the viscous-stress flux , are primarily governed by temperature gradients, whereas even-order central moments, including the viscous stress and the heat-flux-related flux , are dominated by velocity gradients. These leading-gradient dependences are not exclusive; they are substantially modified by density gradients, secondary gradients and transition-layer widths through higher-order derivative terms, gradient products and cross-couplings. When the secondary contributions become comparable to the leading-gradient terms, the nonequilibrium response transitions from a near-linear regime to an approximately exponential regime. This work establishes a super-Burnett-level DBM framework that treats kinetic moment closure and multigradient coupling consistently, providing a basis for resolving and interpreting high-order TNE in three-dimensional compressible flows.
arXiv:2606.05941v1 Announce Type: new
Abstract: The BRUIT-FM Challenge asks participants to reduce ''noise'' on an open dataset of real and synthetic broadband seafloor seismology data. The dataset includes signals from earthquakes, ocean infragravity waves and seafloor currents. It also includes training datasets of synthetic data with solutions. Participants are asked to remove or separate these signals to obtain the best earthquake and seafloor compliance signals. The authors will make a first assessment of the noise reduction, based on noise reduction and ease of use. Participants are invited to a community workshop, in which they will validate the preliminary assessment and collaborate on a community paper describing the results, identifying promising pathways and proposing new directions. The best results will be added to a new version of the dataset, which will persist as a benchmark for future developments.
arXiv:2606.05956v1 Announce Type: new
Abstract: Bidirectional heuristic search can potentially reduce search effort for problems amenable to backward search. Therein, it is well-known that front-to-front heuristics can reduce the number of node expansions, but their overhead is so high that overall runtime almost always increases. We propose BiXDFBnB, a bidirectional depth-first branch-and-bound algorithm that adapts the Single-Frontier Bidirectional Search (SFBDS) framework - originally developed for shortest-path (MIN) problems - to the Generalized Longest Simple Path (GLSP) setting. Because SFBDS inherently operates on paired states, front-to-front (F2F) heuristic evaluation arises naturally and avoids the overhead typically associated with bidirectional frontier management. We show that this adaptation can be successfully applied to maximization (MAX) problems while efficiently handling overlapping constraints. BiXDFBnB is applied to several types of longest-path problems: Longest Simple Path (LSP), Snakes, and Coil-in-the-Box (CIB). Empirical evaluation shows that the new algorithm frequently reduces the number of node expansions and, in some cases, also improves overall runtime.
arXiv:2606.05958v1 Announce Type: new
Abstract: Activation steering has become a popular way to control Large Language Model (LLM) behavior without fine-tuning. Since the technique is plug-and-play, users share datasets and precomputed vectors to steer model activations. However, we show that a \emph{stealth data poisoning attack} silently compromises this pipeline. By substituting $4{-}6\%$ of tokens in the steering dataset, an attacker can silently align the resulting vector with an anti-refusal direction. This jailbreaks the target model while preserving the intended steering effect on benign prompts. Under this threat model, a malicious actor can distribute an apparently safe bundle containing texts, vectors, and weights, alongside an equivalence certificate that the end-user can verify. We test the attack on two open-weight model families and eight model-attribute combinations, observing that poisoned vectors reach an absolute attack success rate (ASR) of $20{-}55\%$, $+19\%$ to $+51\%$ over a clean reference. Finally, we find that a refusal-direction orthogonalization defense can recover ${\approx}82\%$ of the ASR gap without harming benign behavior.
arXiv:2606.05960v1 Announce Type: new
Abstract: Embodied intelligence is moving from laboratory demonstrations toward industrial deployment, with the logistics industry serving as a key application scenario. Learning-based policies offer a promising path beyond traditional perception-planning-control pipelines, but their scalability depends on how embodied data can be collected, organized, and reused. This research studies a data-centric framework for industrial embodied intelligence by constructing a logistics data flywheel. Our framework converts daily operations into reusable data assets, uses World Models to generate reliable supervision for long-tail parcel manipulation, and feeds deployment feedback back into policy improvement. As an initial result, \textit{WM-DAgger} introduces a World-Model-based data aggregation framework that synthesizes out-of-distribution recovery data for robust imitation learning. Building on this result, ongoing work explores how large-scale in-the-wild multimodal data, including labeled human demonstrations, unlabeled operational videos, and system-level robot logs, can be aligned for policy learning and transformed into feedback for continual system improvement.
arXiv:2606.05961v1 Announce Type: new
Abstract: Large Language Models (LLMs) are increasingly employed as proxies for human behavior in computational social science. However, their tendency to internalize biases from training data raises concerns about their reliability in politically sensitive domains, specifically in regard to their susceptibility to persuasive language. In this work, we examine whether LLMs endorse persuasion-infused messages and whether partisan persona prompting modulates such endorsement. We evaluate six LLMs from different geographic regions on content annotated with persuasion techniques drawn from real-world media sources, measuring the likelihood of endorsement using a five-point Likert scale. The models are prompted as either a neutral social media user or as a user with left- or right-leaning political views. Results show that without political conditioning, LLMs generally do not endorse messages containing persuasion techniques, though model-level differences emerge, and that partisan persona prompting increases polarization of endorsement, particularly for persuasion-infused content. Endorsement further varies by persuasion technique and topic. These findings raise concerns about agentic LLM deployments in politically sensitive environments and complicate their use as reliable simulators of human political cognition.
arXiv:2606.05972v1 Announce Type: new
Abstract: Causal graphs provide a high-level language for making mechanisms transparent. Recent work uses Large Language Models (LLMs) to recover causal graphs of external-world processes. Instead, in this paper, we use causal graphs to model LLM inference itself, providing stakeholders with a transparent view of how the model perceives and organizes high-level concepts to produce a prediction. We propose a four-phase method for constructing such graphs. Given a target LLM and a set of textual examples, our method discovers class-discriminative, human-interpretable concepts and maps each input to LLM-perceived concept states. We then introduce an MCMC-inspired counterfactual augmentation procedure that expands the sparse observational data through chains of counterfactuals. This enables stable causal discovery with $\sigma$-CG, yielding informative, interpretable graphs. We apply our method to three LLMs across disease diagnosis, sentiment analysis, and LLM-as-a-judge classification tasks. We evaluate the learned graphs for predictive fidelity and structural stability, and the MCMC-inspired augmentation for convergence and downstream utility. Our results show that the discovered causal graphs capture meaningful dependencies consistent with LLMs' reasoning. Together, this paper provides a foundation for concept-level explainability of LLMs.
arXiv:2606.05994v1 Announce Type: new
Abstract: Medical knowledge graphs (MKGs) infused with clinical knowledge have been increasingly used to model electronic health records (EHRs) to support interpretable predictions in healthcare domain. However, existing MKG-based approaches are limited in capturing pairwise relations between clinical concepts (e.g., conditions, procedures, and medications), and restricts their ability to model higher-order interactions among co-occurring or semantically related concepts. In addition, most representation learning methods that leverage MKGs either collapse temporal information across visits or lack an explicit mechanism for modeling long-range temporal dependencies, which is critical for clinical tasks such as mortality prediction. To mitigate these limitations, we propose HoT-SSM, a parameter efficient and higher-order temporal graph reasoning with state space models. For each visit, HoT-SSM constructs hypergraphs by grouping semantically related clinical concepts into hyperedges using domain knowledge, thereby preserving visit-level clinical context. Further, to model the temporal dynamics while learning the representations, we introduce a novel dynamic hypergraph-based state space model that explicitly captures patients latent state evolution over time while preserving long-range information. The learned representations are used for downstream clinical prediction and reasoning. Experiments on MIMIC-III and MIMIC-IV datasets shows significant performance improvement over the current state-of-the-art models, demonstrating the effectiveness of jointly modeling higher-order clinical interactions and long-range temporal dependencies.
arXiv:2606.05997v1 Announce Type: new
Abstract: We present the AILS-NTUA submission to the EXIST 2026 Lab at CLEF, addressing multimodal sexism identification and characterization in memes (Task 2) and short-form videos (Task 3). Our system follows a feature-engineered late-fusion pipeline built around gradient-boosted regression models and hierarchical post-processing. For memes, we combine visual, textual, demographic, biometric, and LLM-derived semantic indicators designed to capture high-level cues such as stereotyping, objectification, irony, and misogyny. For videos, we investigate the effect of feature selection, frame-based visual representations, OCR-based textual features, acoustic descriptors, and sensor-derived metadata. Development results show that focused LLM-derived semantic cues improve meme sexism identification, while video performance is highly sensitive to feature dimensionality and cross-modal noise. For videos, development results favor compact feature selection, but official test results show that this conclusion does not fully transfer to unseen data, where the unfiltered representation generalizes better. Overall, our findings highlight the usefulness of targeted semantic feature engineering for static memes and the need for more robust temporal modeling in noisy short-form video settings.
arXiv:2606.05628v1 Announce Type: new
Abstract: Focused ultrasound (FUS) offers a promising, non-invasive method for modulating neural activity and delivering therapies deep within the brain with immense clinical potential. However, progress in developing transcranial ultrasound (TUS) for clinical applications has been hindered by several factors. The complexity of the human skull causes focal aberrations and attenuation, thereby presenting a major obstacle to the precise targeting of ultrasound waves. Although phased arrays can correct for these aberrations, their high cost and continuous reliance on magnetic resonance imaging (MRI) pose significant obstacles for widespread academic research and clinical translation. To address these challenges, this thesis proposes an innovative framework for the design, registration, and clinical application of acoustic holograms. First, we introduce a novel frequency-domain topology optimization method that overcomes the breakdown of traditional phase-only designs in the megahertz regime by accounting for volumetric wave-propagation effects, thereby achieving high-fidelity focusing. Second, we present a non-invasive registration strategy that utilizes the nonlinear parametric array (PA) effect to enable precise lens alignment without requiring any imaging modalities, such as MRI. Finally, we demonstrate the utility of this nonlinear parametric array (PA) effect as a tool for monitoring ventricular dilation as a non-invasive proxy for intracranial pressure changes in hydrocephalus. Collectively, these developments provide a path toward accessible, high-precision transcranial ultrasound systems for research and clinical use. In addition, we demonstrate a novel platform for in vitro focused ultrasound neuromodulation that leverages acoustics to advance therapeutic discovery.
arXiv:2606.06014v1 Announce Type: new
Abstract: Latent world models (LWMs) have strengthened end-to-end autonomous driving by forecasting compact scene dynamics for downstream planning. However, existing LWM-based planners usually generate trajectories directly from entangled latent representations. This compact latent-to-planner pathway lacks explicit modeling of risk, drivability, and diverse style preferences, making driving-style dynamics difficult to supervise, inspect, or modulate before a final trajectory is selected. We propose PLAN-S (PLANning with latent Style dynamics), a planner-facing bridge that addresses this compactness-controllability dilemma by decoding a style-conditioned, four-channel semantic cost map from the latent representation. The cost map is conditioned on ego state and driving style and is consumed up-stream of the planning decision through two host-side interfaces: attention-level fusion for regression planners and reward-level fusion for anchor-score planners. We validate PLAN-S on two architecturally distinct hosts, ResWorld on nuScenes and WoTE on NAVSIM, while keeping the host backbones frozen to isolate the contribution of the proposed bridge. On nuScenes, PLAN-S reduces L2 at every horizon over the baseline, with 0.55 m average L2 and a 42% relative reduction in the 3 s collision rate. On NAVSIM, the rule-cost variant reaches 89.4 Predictive Driver Model Score (PDMS), while the learned cost variant provides complementary gains on baseline-challenging scenes. Ablations show that the cost pathway contributes most directly to safer trajectory selection. Qualitative results further show that PLAN-S can produce diverse cost maps, with spatially consistent variations aligned to different driving styles.
arXiv:2606.06015v1 Announce Type: new
Abstract: We present a multilevel embedding framework for quantum chemistry calculations on a quantum computer. In our framework, a quantum algorithm treats the strongly correlated active space, while a high-level wave-function method such as coupled cluster theory or multireference perturbation theory recovers the remaining correlation in the surrounding region. A sampling-based quantum algorithm, quantum-selected configuration interaction, bridges the quantum and classical treatments. The entire calculation is embedded in a low-cost density functional theory description of the surrounding environment using Manby's projection technique. We apply the framework to organic, metal-organic, and metallic systems, computing bond dissociation energies, adsorption energies, and reaction barriers using only the subset of qubits of a 144-qubit superconducting quantum computer at the University of Osaka and achieving $\sim$1 kcal/mol agreement with classical references for a Menshutkin $\mathrm{S_N2}$ reaction inside a carbon nanotube. Our results may open the way to quantitatively reliable quantum-classical hybrid calculations for large-scale chemical systems.
arXiv:2606.06025v1 Announce Type: new
Abstract: Scientific peer review generation has attracted increasing attention for reducing reviewing burdens and providing timely feedback. However, existing Large Language Model (LLM)-based methods often produce generic comments with insufficient evidence support and weak source traceability, while complex multi-agent systems incur high inference costs. To address these challenges, we propose EGTR-Review, an Evidence-Grounded and Traceable Review Generation framework via Multi-Agent Teacher Distillation. EGTR-Review first constructs a multi-agent teacher that performs structure-aware paper decomposition, key-element extraction, external scholarly evidence retrieval, evidence-state labeling, verification reasoning, and review synthesis. It then distills both intermediate reasoning trajectories and final review comments into a lightweight student model through task-prefix-driven multi-task learning. An evidence-weighted objective further reduces the influence of weak, missing, or non-verifiable supervision. Experiments on public peer-review datasets show that EGTR-Review (Student) outperforms strong prompt-based, fine-tuned, and structured/agentic baselines across automatic metrics, LLM-as-Judge evaluation, and human evaluation, while maintaining strong factual grounding and source traceability with substantially lower token consumption and inference time. Our code, prompts, configurations, and sample data are available on GitHub.
arXiv:2606.05632v1 Announce Type: new
Abstract: Within the past few years, the ability of Large Language Models (LLMs) to generate formal mathematical proofs has improved drastically. We provide a comparison of various LLMs' effectiveness in producing formal proofs in Lean 4 with the goal of assisting those seeking to use LLMs to support their own projects. We utilize both pass@$k$ and refine@$k$ metrics as the benchmark for our comparison and evaluate on subsets of both miniF2F and miniCTX datasets. Our testing shows that overall, Gemini 3.1 Pro and Claude Opus 4.7 perform best. Gemini 3.1 Pro achieved a 92\% success rate on miniF2F via refine@32 whereas Opus 4.7 achieved a 86\% success rate on miniCTX via refine@32. When taking cost into account, NVIDIA Nemotron 3 Super and GPT-OSS 120B were the most efficient, with competitive accuracies and average costs of $<\$0.01$ per correct proof.
arXiv:2606.05639v1 Announce Type: new
Abstract: Knowledge Graph Completion (KGC) aims at predicting missing triplets from incomplete knowledge graphs, which is crucial for downstream applications. Recently, Graph Neural Network (GNN)-based methods have achieved remarkable success by performing message passing over query-centered local subgraphs. However, in practice, a query is jointly defined by both the entity and the relation, with both carrying information indispensable for reasoning, yet these methods rely solely on the query relation as the guiding signal, while the information inherent in the query entity is not leveraged to guide inference - the entity serves merely as a structural anchor for subgraph extraction. To this end, we incorporate query entity information into the reasoning process from two perspectives: the first is structural context, i.e., the neighboring structure and relation patterns around the entity, which is encoded by a dedicated context encoder and used to modulate messages; the second is semantic type of the entity, inferred by a large language model, which is incorporated into attention computation and final scoring to provide type-level prior constraints. Together, these two sources of information enable the reasoning process to be guided by both the query relation and the query entity. Experimental results on standard benchmarks demonstrate the effectiveness of the proposed Q-GNN.
arXiv:2606.05644v1 Announce Type: new
Abstract: When retrieved evidence contradicts parametric memory, language models frequently ignore context and default to memorized priors -- a failure that undermines the core purpose of retrieval augmentation. Contrastive decoding amplifies the context-conditioned output to suppress parametric bias, but existing methods rest on an implicit assumption that this bias is uniform across tokens. A single global contrastive weight over-penalizes safe tokens while leaving genuinely conflicted ones insufficiently corrected. We identify token-level conflict concentration: retrieval-memory tension is sharply heterogeneous, concentrated on a small fraction of answer-critical decoding steps. This reframes contrastive decoding from how much contrast to apply to where to apply it. We propose FIDES (Faithful Inference via Deep Evidence Signals), a training-free decoder that reads three internal signals probing retrieval-memory conflict at complementary depths -- output surface, hidden representations, and prediction trajectory -- and fuses them to govern intervention strength at each decoding step. Across three benchmarks and six backbones -- four primary 7B/8B models and two scaling backbones up to 70B -- FIDES achieves the best context fidelity in all 18 settings, outperforming the strongest training-free baseline by +3 to +13 points. On the 70B scale, fidelity reaches 92-94% while F1 surges to 62-63%, demonstrating that token-level selectivity unlocks generation capability that coarse contrastive rules suppress.
arXiv:2606.06031v1 Announce Type: new
Abstract: Masked diffusion language models generate text by iteratively unmasking many tokens in parallel, but this speed comes with a correction problem: tokens generated in the same step are predicted from marginal distributions, and early local dependency errors can later contaminate the context. PRISM addresses this by learning token-level quality scores and remasking unreliable tokens, but its inference rule is coupled: the same forward pass both detects low-quality tokens and computes logits for their replacements, so the erroneous tokens still condition regeneration. We propose NAVIRA, an inference-time decoding policy that separates these two operations and samples remasking positions stochastically. A first forward pass scores tokens; selected tokens are masked; a second forward pass regenerates from the cleaned context. Temperature-controlled remasking reduces repeated correction of the same positions and balances fluency against diversity. In controlled experiments with a 170M masked diffusion language model, decoupling improves fluency, while scheduled stochastic remasking preserves entropy and achieves stronger LLM-judge scores under larger forward-pass budgets. These results show that remasking policy, not only the learned quality signal, is central to reliable masked-diffusion text generation.
arXiv:2606.06036v1 Announce Type: new
Abstract: Despite recent progress, LLM agents still struggle with reasoning over long interaction histories. While current memory-augmented agents rely on a static retrieve-then-reason paradigm, this rigid pipeline design prevents them from dynamically adapting memory access to intermediate evidence discovered during inference. To bridge this gap, we propose MRAgent, a framework that combines an associative memory graph with an active reconstruction mechanism. We represent memory as a Cue-Tag-Content graph, where associative tags serve as semantic bridges connecting fine-grained cues to memory contents. Operating on this structure, our active reconstruction mechanism integrates LLM reasoning directly into memory access, allowing the agent to iteratively explore and prune retrieval paths based on accumulated evidence. This ensures that memory retrieval is dynamically adapted to the reasoning context while avoiding combinatorial explosion caused by unconstrained expansion. Experiments on the LoCoMo benchmark and LongMemEval benchmark demonstrate significant improvements over strong baselines (up to 23%), while substantially reducing token and runtime cost, highlighting the effectiveness of active and associative reconstruction for long-horizon memory reasoning.