arXiv:2606.25008v1 Announce Type: new Abstract: Neural scaling laws describe how pre-training loss decays as power laws with training time, model size, and compute. This position paper argues that the exponents of these power laws are fixed by generic mechanisms: a one-third time scaling due to the strong nonlinearity of Softmax, an inverse width scaling due to representational superposition, and an inverse depth scaling due to ensemble averaging of Transformer layers. These mechanisms are robust to a wide range of data structures and architectural details, placing current large language models in a universality class with fixed exponents. The coefficients, however, are expected to be sensitive to data and architecture details, and directly determine practical quantities such as the optimal model shape and the compute-optimal frontier. We therefore argue that understanding the coefficients is the key to near-term performance improvements, and that a closer examination of the current universality class may reveal pathways to better universality classes.
Science Journals
arXiv:2606.25786v1 Announce Type: new Abstract: This document outlines the changes adopted for ACM HotNets 2026, spanning its scope, review process, and program structure. Rather than isolated adjustments, these changes form a coherent effort to clarify and extend HotNets' role as a venue for agenda-setting research, community discussion, and experimentation with how the networking community evaluates, disseminates, and discusses research. In particular, HotNets 2026 broadens its scope to include perspective and community-facing contributions, introduces distinct evaluation criteria for technical and perspective papers, adopts a more collaborative and discussion-oriented review process, rethinks how accepted work is presented and discussed at the workshop, and explores responsible uses of generative AI (GenAI) in reviewing and research dissemination. We believe these changes will help HotNets continue to serve as a home for ambitious, unconventional, and thought-provoking ideas, while also positioning it as a venue for experimenting with new approaches and formats that larger conferences, e.g., SIGCOMM or NSDI, might later adopt. We use this document to solicit feedback from the community, both on these changes and on how HotNets can best serve the networking community in the future. We plan to collect feedback during and after the event and to prepare a follow-up report summarizing the community's reactions and lessons learned.
arXiv:2606.25828v1 Announce Type: new Abstract: Direct motional coupling of opposite-charge trapped-ion pairs could provide a pathway to extend ultra-low temperatures and quantum control to negative ions that lack the suitable electronic energy structures required for direct laser cooling. Because positive and negative ions cannot be confined within a single electrostatic potential well, I investigate a configuration where single ions are trapped in close proximity within separate potential wells to couple their motion. I analytically and numerically evaluate the electrostatic trapping requirements. As a concrete implementation, I present an optimized segmented surface Paul trap design to couple an antimatter hydrogen molecular ion ($\bar{H}_2^-$) and a beryllium ion ($^9 Be^+$). A motional coupling frequency of 5 kHz can be achieved at an ion-ion separation of $35 \mu m$, with an ion height of $50 \mu m$, axial trap frequencies of 4 MHz, and static trap voltages with a magnitude of $\approx 20 V$. Finally, I outline three applications for this technique: quantum logic spectroscopy of $\bar{H}_2^-$ for matter-antimatter comparisons, the preparation of cold neutral deuterium atoms via near-threshold photo-detachment of $D^-$ for optical trapping, and quantum information processing using equal-mass opposite-charge ion pairs.
arXiv:2606.25852v1 Announce Type: new Abstract: Group-based reinforcement learning effectively post-trains LLM agents for long-horizon, sparse-reward tasks by deriving step-level credit from trajectory outcomes. However, this ties a step's credit to its rollout's final outcome: semantically near-identical intermediate steps receive opposite credit depending on whether their trajectory eventually succeeded or failed. Such semantic credit inconsistency sends conflicting gradients to similar actions and wastes the partially-correct progress inside failed rollouts. Motivated by this, we propose Semantic Consistency Policy Optimization (SCPO), a value-free reward-shaping method that mitigates this inconsistency by recovering step-level credit from successful siblings in the same rollout group. Concretely, SCPO scores each failed step against a successful sibling and adds positive step-level credit for new progress along that sibling. On ALFWorld and WebShop, SCPO matches or exceeds strong group-based baselines, reaching 93.7+/-4.1 percent success on ALFWorld and 74.8+/-2.0 percent on WebShop at 1.5B parameters, with gains concentrated on the hardest multi-step tasks.
arXiv:2606.25073v1 Announce Type: new Abstract: In cooperative multi-agent reinforcement learning (MARL), from a deployment perspective, it is challenging and expensive to train agents from scratch for each new environment or task. In this work, we propose GCT-MARL, a transfer learning framework that builds on the multi-view graph contrastive backbone of MAIL and augments it with a per-view, adaptively weighted alignment loss and a two-phase training protocol specifically designed for transfer across populations of varying sizes and compositions. We empirically demonstrate that the proposed framework markedly accelerates convergence on the target task relative to from-scratch training, in both homogeneous (within-faction, varying N) and heterogeneous (cross-faction and mixed unit-type) transfer scenarios. Furthermore, we show that the framework naturally supports continual learning by sequentially chaining the two-phase transfer protocol across a series of related tasks. Overall, this work provides a unified approach to mitigating key limitations in current MARL transfer methods with new insights at both methodological and empirical levels.
arXiv:2606.25854v1 Announce Type: new Abstract: We analyze the decrease of simplex diameters under iterated refinement of spherical Delaunay complexes. Unlike in ordinary subdivision, the refined Delaunay complex need not be a subdivision of the previous one, so mesh contraction is not automatic. We derive explicit contraction bounds for several families of Steiner points, including Delaunay analogues of barycentric and edgewise subdivision. The proof reduces the problem to sharp covering estimates for Euclidean simplices. These estimates are obtained through a strengthening of Maurey's empirical method via pivotal sampling and a dimension-dependent version of the approximate Carath\'eodory theorem. Theoretical results and numerical experiments show that Delaunay refinements achieve stronger contraction than their subdivision counterparts.
arXiv:2606.25855v1 Announce Type: new Abstract: The efficiency of high-power laser systems is limited by wavefront distortions in the beam, particularly non-common path aberrations, which reduce the peak intensity at the focal plane. Compensating for these aberrations requires the calibration of the adaptive optics system. Conventional calibration methods rely on a time-consuming iterative optimization that is highly sensitive to initial conditions. While deep learning-based models offer high speed, they often demonstrate insufficient accuracy. In this work, we present a hybrid wavefront reconstruction method that combines a convolutional neural network to generate an initial estimate of the wavefront distortions, with the L-BFGS (Limited-memory Broyden-Fletcher-Goldfarb-Shanno) algorithm for its subsequent refinement. In numerical simulations, the method achieved an efficiency of $\sim 0.99$ in 80% of the cases for a root-mean-square (RMS) of wavefront distortions ranging from 0 to $1.3\lambda$. In a physical experiment, for initial wavefront distortions with RMS values from 0.15 to $0.6\lambda$, the method achieved an efficiency of $\sim 0.75$. As a result, focusing with a Strehl ratio of $0.96 \pm 0.02$ was attained within 2 to 4 iterations of the algorithm, confirming the applicability of the method for the fast and accurate calibration of adaptive optics systems under real experimental conditions.
arXiv:2606.25095v1 Announce Type: new Abstract: The rapid growth of AI workloads is driving unprecedented increases in data center power demand, current transients, and thermal stress, exposing fundamental limitations in traditional 48 V rack architectures, low-voltage AC distribution, and line-frequency transformer interfaces. This paper reviews the three stages of architectural shifts required to support next-generation AI data centers and identifies three enabling technological building blocks: high-voltage conversion-ratio DC/DC converters, facility-level low-voltage DC distribution, and medium-voltage solid-state transformers. The advantages, technical challenges, and potential solutions associated with each building block are reviewed. Finally, future research directions and open challenges are discussed.
arXiv:2606.25249v1 Announce Type: new Abstract: Modern Information Retrieval (IR) systems typically use a "retrieve-then-rerank" pipeline, where a computationally expensive, pre-determined cross-encoder re-ranks the top results from a fast initial retriever. While effective, this approach often applies heavy re-ranking models regardless of query complexity, resulting in high latency and wasted computational resources on simple queries. We propose Adaptive Re-Ranking, an utility-based labeling framework for cost-aware routing and present empirical evidence (via oracle analysis and a trained baseline router) that per-query routing offers large potential gains but is non-trivial to learn from limited supervision. We train a routing classifier with 3 strategies: sparse retrieval (BM25), dense re-ranking (MiniLM-L6-v2), and heavy neural re-ranking (BGE-v2-m3). Compared to BGE our method achieves 1.15-53x lower median latency and 1.11-5.22x lower mean latency across all datasets we have tested, while delivering -17.5% to +4.0% nDCG@10, which is competitive in some datasets. Our findings show that routing queries based on our novel utility function offers a scalable solution for reducing computational costs and latency in a variety of IR systems.
arXiv:2606.25098v1 Announce Type: new Abstract: The rapid expansion of artificial intelligence (AI) infrastructure is driving unprecedented growth in electricity demand from data centers. Traditional power-system planning treats large computing facilities as inflexible peak loads, leading to costly infrastructure upgrades and long delays in grid interconnection. Recent work has shown that AI clusters can reduce electricity consumption during peak demand through software-based workload orchestration. This article explores how modern GPU-based AI data centers can operate as grid-interactive assets that respond dynamically to power system conditions. We describe an architecture integrating grid signals, workload scheduling, and power telemetry for fine-grained cluster power control. Experimental results from a real-world deployment on a 130 kW GPU cluster demonstrate multiple forms of flexibility, including rapid load reduction, sustained curtailment, and carbon-aware operation while preserving service levels for priority jobs. We further demonstrate performance-aware load shifting across geographically distributed clusters, enabling workloads to migrate toward regions with lower grid stress. Together, these capabilities transform AI infrastructure from static electricity consumers into flexible resources that support grid reliability, accelerate interconnection, and improve computing sustainability.
arXiv:2606.25120v1 Announce Type: new Abstract: Aviation software certification has operationalised three structural requirements for governed software systems since 1992: structured governance linkage between governing specifications and operational evidence, context-bounded validity that triggers revalidation when operational context changes, and an objective evidence architecture that defines what proof means and what makes it sufficient. These requirements appear in DO-178C and DO-330 and are enforced through FAA and EASA certification. No existing framework requires these structural properties as intrinsic properties of individual AI governance documents. A system prompt, an AGENTS.md file, a governance policy, or a task envelope can be deployed without satisfying any of the three requirements aviation has enforced for three decades. Aviation is the most technically rigorous instance: its standard-setting bodies have acknowledged that their frameworks break down for AI systems, yet none requires these properties of individual governance documents. Aviation's structural requirements break down at the system level because AI systems are non-deterministic, but remain transferable at the document level: the governance artifact is a static artifact whose structural properties can be evaluated independently of the stochastic system it governs. The paper maps DO-178C's traceability architecture, DO-330's requalification triggers, and DO-178C's objective evidence requirements onto three structural findings: epoch limits on governance document validity, proof surfaces as the revalidation feedback mechanism, and the absence of structural completeness requirements in AI governance instruments. An empirical companion (arXiv:2604.21090) found that 37% of AI governance documents fall below the structural quality threshold. PromptQ's seven-principle framework operationalises these requirements at the governance document layer.
arXiv:2606.25121v1 Announce Type: new Abstract: We study the logical relation of the P-NP separation conjecture in the Blum-Shub-Smale-model over the complex numbers with the P-NP separation conjecture in Valiant's algebraic model. This amounts to comparing Hilbert's Nullstellensatz Problem, that is, deciding feasibility of a given system of polynomial equations over the complex numbers, with the problem of evaluating the permanent of a given complex matrix. We compare the respective uniform models of computations and prove that $P_C\ne NP_C$ in the Blum-Shub-Smale-model over $C$ implies the separation $VP^0(u)\ne VNP^0(u)$ of the uniform versions of Valiant's constant-free complexity classes over $C$. For the nonuniform models we show the analogous implication: the separation $P^0_C(nu)\ne NP^0_C(nu)$ of the nonuniform, constant-free Blum-Shub-Smale classes over $C$ implies the separation $VP^0\ne VNP^0$ of Valiant's constant-free complexity classes over $C$. In the reverse direction, we conjecture that $VNP_C\not\subseteq\overline{VP}_C$ implies that $P_C(nu)\ne NP_C(nu)$.
arXiv:2606.25130v1 Announce Type: new Abstract: Building on the null-infinity-layer construction, we develop an H1-conforming finite-element formulation of hyperboloidal compactification for the exterior Helmholtz equation. A change of coordinates maps infinity to a finite outer boundary, and a rescaling removes the leading oscillatory decay. We derive the transformed equation and a global sesquilinear weak formulation with bounded coefficients. The compactified boundary contributes an explicit boundary mass term, and its trace gives the far-field pattern up to a known normalization. We compare the resulting method with finite-element discretizations using perfectly matched layers (PML) and report benchmark results in two and three dimensions. Numerical experiments include scattering by a unit disk, resonance in a trapping geometry, a manufactured benchmark in three dimensions, and a submarine benchmark.
arXiv:2606.25887v1 Announce Type: new Abstract: Several fundamental problems in computational geometry admit algorithms with running time $f(d) \cdot n^{2-\Theta(1/d)}$ for $n$ points in $d$ dimensions, making them among the most prominent examples of barely subquadratic computation. Notable members of this class include Furthest Pair, Bichromatic Closest Pair, (Bichromatic) Maximum Innter Product, and Hopcroft's Problem. Chen [Theory Comput. 2020] proved that, assuming the Strong Exponential Time Hypothesis (SETH), these problems require $n^{2-o(1)}$ time when the dimension satisfies $d=2^{\Theta(\log^* n)}$. We extend this lower bound to all efficiently constructible dimensions $d=\omega(1)$. Thus, assuming SETH, the dependence of the best known algorithms on the dimension is essentially unavoidable. The proof utilizes techniques in OpenAI's recent disproof of the Erdos unit distance conjecture. The proof was initially discovered by ChatGPT 5.5 Pro. The authors have validated and substantially edited the proof to improve the presentation.
arXiv:2606.25134v1 Announce Type: new Abstract: Safe control has been widely studied in various safety-critical applications, for instance, autonomous driving. In order to ensure the autonomous vehicle does not collide with other vehicles, it is essential to obtain an accurate expectation of surrounding vehicles' behavior and react adaptively. Instead of assuming fully cooperative and homogeneous vehicles using the same safety-critical controllers, recent works have been exploring different data-driven approaches to model the neighboring vehicles' underlying controllers with observed data. However, existing works either suffer from 1) the inter-vehicle influence during the multi-vehicle interaction, which makes it hard to determine the causality of surrounding vehicles' behavior in controller modeling, or 2) being dominated by the worst-case analysis, which may lead to overly conservative behavior. In this paper, we extend the prior work on Parametric-Control Barrier Function (Parametric-CBF) to multi-robot interactions with embedded causality inference to explicitly reason over the inter-vehicle influence. Given the learned Causality-based Parametric-CBF, we present an adaptive safety-critical controller that allows the ego vehicle to safely react to surrounding vehicles with the learned expectation. We demonstrate that by leveraging the motion flexibility among multi-vehicle systems, task efficiency can be greatly improved in various interaction-intensive scenarios.
arXiv:2606.25893v1 Announce Type: new Abstract: This paper introduced a Hierarchical Generative Density Model (HGDM) for the long-term propagation of orbital fragmentation clouds. Validation against high-fidelity Monte Carlo simulations showed that the proposed surrogate accurately reproduces the dominant multidimensional structures of propagated clouds while consistently outperforming classical band-formation approximations based on independent angular variables. Accurate cloud reconstructions were obtained using only a few hundred to a few thousand propagated fragments, yielding reductions exceeding two orders of magnitude in computational cost and three orders of magnitude in storage requirements; future work will investigate its application to large-scale debris-environment evolution and collision-cascade simulations associated with the Kessler syndrome.
arXiv:2606.25136v1 Announce Type: new Abstract: General-purpose robots operating in partially observable environments, such as homes, require memory to support autonomy. They must recall diverse information from the past, such as where objects were placed, which tasks a human partner has completed, and when an appliance was turned on. Achieving this versatility requires a general memory retrieval mechanism. Transformer architectures that use attention over long contexts for memory retrieval provide a promising approach, as they learn retrieval from data rather than relying on task-specific or hand-designed rules. However, directly incorporating them into imitation learning from offline data introduces two key challenges: (1) the policy may learn spurious correlations between past information and predicted actions, and (2) errors accumulate in memory due to prediction inaccuracies and their compounding interactions with the environment, causing model drift and cascading failures. To address both challenges, we introduce HALO, a visuomotor policy with an attention-based memory retrieval mechanism for long-horizon control. First, to suppress spurious correlations, HALO distills vision-language model (VLM) priors into the policy. It generates memory-dependent question--answer pairs from demonstration trajectories and trains jointly with a video question--answering objective, steering retrieval toward task-relevant information. Second, to reduce the impact of accumulated errors in memory during closed-loop control, HALO uses sparse attention that restricts retrieval to only the most relevant parts of the history. Together, these components enable more reliable long-horizon control by guiding the policy to retrieve task-relevant information from up to eight minutes of past experience. Project website: https://robin-lab.cs.utexas.edu/HALO
arXiv:2606.25139v1 Announce Type: new Abstract: Agentic AI offers significant potential to automate complex building-engineering workflows. However, most existing applications remain isolated proof-of-concept demonstrations and lack reusable domain capabilities, human-verified evaluation cases, and standardized benchmarking infrastructure. This study presents Buildrix, an open, community-driven platform for developing, sharing, executing, and evaluating agentic AI skills for building engineering. Buildrix integrates three components: a Python command-line package for developing, validating, publishing, installing, and managing skills and test cases; a web-based Hub for organizing open challenges, reusable skills, test cases, reviews, and benchmark results; and a local agent harness that supports skill discovery, external toolchain provisioning, progressive context loading, and multi-step workflow execution. Buildrix skills are organized as standardized, self-contained packages containing task instructions, executable scripts, dependencies, and supporting resources. Quantitative test cases can be verified by domain experts and promoted to golden test cases for reproducible benchmark evaluation. Buildrix provides an open foundation for reusable capability development, transparent evaluation, and community-driven advancement of agentic AI in building engineering.
arXiv:2606.25140v1 Announce Type: new Abstract: Following the successful completion of the Muon g-2 experiment run at Fermilab, the Muon Campus facility has been reconfigured from delivering 3 GeV muon beams to the g-2 storage ring to providing slow-extracted 8 GeV proton beam spills for the Mu2e experiment. The first full-scale commissioning run with slow extraction was conducted during the 2025 run, followed by the second run in early 2026. We present the results and current status of this commissioning campaign.
arXiv:2606.25143v1 Announce Type: new Abstract: Objectives: To develop a codebook for self-stigma across cognitive, affective, and behavioral domains, and to estimate the prevalence, co-occurrence, and temporal patterns of these indicators in Reddit posts by people who use drugs. Methods: We developed a ten-indicator codebook through consensus-based abductive coding spanning cognitive (self-labeling, pessimism/self-defeatism, deservingness/worthlessness), affective (shame, guilt/self-blame, despair/hopelessness), and behavioral (concealment, anticipated rejection, desire to quit, ambivalence) domains; two coders reached substantial agreement (Cohen's k = 0.72). We then scaled classification with a large language model validated against expert coding (k = 0.73, F1 = 0.80), analyzing 72,115 thread-initiating posts from 1,660 English-language users (2006-2025). Results: 3,838 posts (5.3%) from 1,228 users (74.0%) contained self-stigma; all ten indicators discriminated self-stigma posts (RR 3.6 to 86.2), led by self-labeling (56.0%) and despair/hopelessness (48.5%). Self-stigma was integrated: core and behavioral indicators were strongly associated at the user level (OR = 4.65, 95% CI 3.12-6.94, p < 0.001), and 87.0% of posts with behavioral indicators also contained a core indicator. Contrary to progressive models, behavioral indicators emerged earlier than core ones (desire to quit at median position 0.08 vs. shame at 0.38). Nine of ten indicators were stable across posting trajectories; only pessimism increased (OR = 1.62, 95% CI 1.25-2.10). Conclusion: Among people who use drugs online, self-stigma is an integrated phenomenon in which behavioral indicators rarely appear without internalized ones and often precede them. Most expressions remain stable over time, but pessimism about change deepens, marking a target for early digital intervention and showing that progressive stage models do not map directly onto textual disclosure.
arXiv:2606.25146v1 Announce Type: new Abstract: The initial development phase of UAV swarms largely depends on simulation for experimental design and validation, yet existing open-source tools are often unmaintained, have steep learning curves, or are built around a single fixed scenario. The need for a comprehensive, modular simulation platform is a recognized research gap. This paper presents SwarmFly, a MATLAB-based simulation and test platform for multi- UAV swarms that addresses these gaps. SwarmFly combines a real-time operational map, four swarm coordination modes (leader-follower, decentralized, heterogeneous relay, and heterogeneous speed), simulated IMU telemetry, and IP-based geolocation with a plugin architecture that lets researchers add behaviors, fault models, and analysis tools without touching the core code. Eight bundled plugins extend the base simulator into a full test harness. The SwarmFly platform exposes multi-agent aerial swarms to a wide range of internal and external disruptions, enabling observation and quantification of underlying swarm control and behavioral mechanisms. This study verifies and characterizes each subsystem through eight experiments that measure formation accuracy, wind tolerance, fault recovery, energy endurance, and airspace compliance. The platform runs entirely in MATLAB. Its modular design supports straightforward extension toward hardware-in-the-loop testing, larger swarms, and higher-fidelity dynamics. An open-source release is available at [https://github.com/abhishekphadke/SwarmFly.git]
arXiv:2606.25149v1 Announce Type: new Abstract: The last few years have seen a significant rise in interest in highly autonomous and proactive systems, fueled by advances in AI. Systems that anticipate user needs, take initiative, and act without explicit user input. Such systems span a wide range of applications, from smart lighting that adapts to user activity to assistive robots that plan actions in advance to intelligent thermostats that learn routines and adjust environments proactively. Despite this breadth, the concept of proactivity remains loosely defined and inconsistently applied across research and practice. Current usage of the term often conflates fundamentally different system behaviors. For instance, simple reminders or recommendation systems are frequently labeled as proactive, even though underlying mechanisms and intentions differ significantly. This conceptual ambiguity limits our ability to systematically design, compare, and evaluate proactive systems. Moreover, existing methodologies for design and evaluation are largely rooted in reactive interaction paradigms, failing to address the unique challenges posed by proactive behavior, including timing, appropriateness, user control, transparency, and trust. This multidisciplinary workshop aims to establish a clearer and more rigorous foundation for understanding proactive systems. We bring together researchers and practitioners from Human-Computer Interaction, AI, and related fields to (1) develop a shared conceptualization of proactivity, (2) identify gaps and limitations in current design and evaluation approaches, and (3) co-create human-centered guidelines and research directions for future systems. Through interactive discussions and collaborative activities, the workshop seeks to map key challenges and opportunities, ultimately advancing robust and consistent frameworks for designing and evaluating proactive technologies.
arXiv:2606.25152v1 Announce Type: new Abstract: Deployed approaches for AI text detection often rely on training-time access to labeled datasets of both human-written and AI-generated text. This approach is vulnerable to three types of distribution shifts that occur continually post-deployment, and for which labeled data is often unavailable: adversarial humanization, new LLMs being released, and temporal drift in human writing. Simultaneously, existing approaches do not leverage a key signal of LLM usage: inference-time homogeneity. We propose a test-time adaptation (TTA) approach, using semi-supervised learning, that adapts to distribution shifts by leveraging homogeneity among unlabeled samples observed at inference time. Empirically, we find that state-of-the-art supervised detectors systematically fail when they encounter distribution shifts in AI-generated and human writing, both adversarial and natural, while test-time adaptation with semi-supervised learning is largely robust; e.g., the commercial model Pangram detects just 24.1% of our adversarial AI-generated text, compared to 90.5% for our test-time approach. We establish that test-time adaptation is a promising framework for AI text detection in the wild. We publicly release our code (which includes code for model training, evaluation, and plots) at https://github.com/kkr36/llm_detection.
arXiv:2606.25894v1 Announce Type: new Abstract: Medical vision-language models typically generate diagnoses through single-pass inference without indicating which image regions support their conclusions. This lack of spatial grounding limits clinical utility: outputs cannot be audited, and models may hallucinate findings on normal scans. We present BrReMark (Brain Rethink via ROI Marking), a framework that introduces explicit region marking into brain MRI diagnosis. The model first generates hypotheses about potential abnormalities and grounds them through explicit bounding box marking, then verifies conclusions by re-examining the marked evidence. Training combines supervised fine-tuning on structured reasoning trajectories with reinforcement learning using a composite reward over localization accuracy and diagnostic reasoning. Furthermore, we integrate a domain randomization-based pathology synthesis augmentation strategy to improve the model's generalizability to out-of-distribution (OOD) data. On internal benchmark, BrReMark improves mAP50 from 0.74% to 37.54% compared to the base model, while achieving 21.57% Clinical F1 and 45.26% diagnostic accuracy. On NOVA OOD benchmark, it also achieves competitive overall performance with a 45.7% reduction in false positives compared to the state-of-the-art, indicating reduced hallucination on rare pathologies. These findings suggest that explicit hypothesis-verification grounding is a practical path toward trustworthy open-ended brain MRI diagnosis across both in-distribution and OOD settings.
arXiv:2606.25161v1 Announce Type: new Abstract: Large language model (LLM) agents rely on long-term memory to support extended interactions and personalized assistance beyond finite context windows. Existing memory agents actively update external memory through generated write, revise, and delete operations, but these updates may omit important information, corrupt existing memory, or introduce unsupported hallucinated content. Once stored, such errors become persistent system-state failures that can affect future reasoning and generation. In this paper, we propose TrustMem, a framework designed to improve the trustworthiness of memory consolidation. TrustMem relies on a Memory Transition Verifier to evaluate the transition process of memory updates in terms of coverage, preservation, and faithfulness. It further constructs preference pairs among candidate updates under the same memory state, enabling preference-guided reinforcement learning to directly optimize memory updating behaviors. Extensive experiments demonstrate that TrustMem improves both memory utility and reliability: it achieves state-of-the-art results across MemoryAgentBench, HaluMem, and the Mem-alpha validation set, improves HaluMem memory extraction by 12.14 F1 points, and reduces transition-level omission, corruption, and hallucination by 40.1\%, 79.1\%, and 50.0\%, respectively, compared with the strongest baseline for each error type.