Large Language Models (LLMs) have shown remarkable capabilities across tasks, but they still struggle with complex reasoning that requires integrating scattered knowledge. Graph Retrieval-Augmented Generation (GraphRAG) addresses this challenge by organizing fragment knowledge as graphs, while recent Graph Foundation Models (GFMs) further improve GraphRAG with graph-native retrieval and reasoning capabilities. However, existing GraphRAG methods usually adopt decoupled retrieve-then-generate pipelines, which create objective misalignment between retrieval and reasoning, and training GFM retrievers typically depends on expensive human annotations. To address these limitations, we propose \ourmethod, a unified end-to-end GraphRAG framework that tightly couples a GFM retriever with an LLM generator. Given a query, the GFM retriever first reasons over the graph to score and retrieve relevant document nodes. These node embeddings are then passed through a fully differentiable graph bridge, which projects them into soft-weighted graph prefixes according to retrieval scores and combines them with raw document texts to form graph-augmented prompts. In this way, structural information is injected directly into LLM generation, while the final answer loss can be back-propagated through the bridge to jointly optimize retrieval and reasoning. To stabilize training and reduce reliance on costly annotations, we further introduce a semantic-guided distillation module that provides auxiliary supervision using pseudo-labels derived from query-document semantic similarities. Extensive experiments on three multi-hop QA datasets and three out-of-domain GraphRAG benchmarks show that \ourmethod consistently outperforms strong baseline methods and generalizes well, which demonstrates the effectiveness of end-to-end GraphRAG with tightly integrated GFM retrieval and LLM reasoning.
https://arxiv.org/abs/2610.34866
User memory helps personalized systems build on their understanding of users over time. In e-commerce, users convey their preferences and needs through both explicit language and implicit interactions with products. We study memory built from logs of these interactions, examining how it can \emph{remember} concrete evidence and established interests, and use that evidence to \emph{infer} underlying needs and preferences that inform judgments beyond well-observed product domains. Existing memory methods primarily target language-based interactions, and evaluations offer limited real behavioral data and limited separation of these capabilities. This leaves open how to jointly support and separately evaluate remembering and inference from behavioral logs. We formulate the problem of continual behavioral user memory around these joint requirements and develop a dedicated evaluation framework and memory method to address them. \textbf{BehaviorMemBench} is a benchmark built from real e-commerce logs covering 1{,}194 users and 17{,}462 behavioral batches over more than three years, with automatic future-behavior labels and a task-independent memory interface to separately evaluate remembering and inference under a common protocol. \textbf{BehaviorMem} is a memory design that organizes explicit behavioral facts and inferred user characteristics across long-term and short-term scopes in a two-axis state, with evidence-grounded formation, continual maintenance, and query-independent readout. Experiments across two builders and two task readers show that BehaviorMem leads the evaluated agentic memories on remember and infer, and reveal different method rankings across the two capabilities and competitive statistical baselines. Validation on a public dataset further supports its usefulness for downstream tasks based on behavioral histories.
https://arxiv.org/abs/2610.34790
Users are increasingly delegating tasks to AI agents, allowing agents to perform long-running, multi-step workflows on their behalf. However, agents may perform these tasks in ways that depart from users' expectations for how they should be carried out, leading to what we call agentic delegation failures. We leverage a large corpus of 163,487 Reddit posts, describing AI agent use, to develop MANDATE: a taxonomy of 23 failure categories organized into seven groups. A follow-up study with 73 AI agent users examined the relationship between prevalence and perceptions of severity and noticeability, finding that failures rated as more severe are considered harder to notice, and that prior experience with a failure is associated with greater noticeability. Finally, we assess the maturity of research on each failure and identify commonly encountered failures that have received little attention (e.g., agents running without making progress). Our findings inform opportunities to make delegation easier for users to oversee and better studied by the research community.
https://arxiv.org/abs/2610.34748
Aligned audio, performance MIDI, and musical scores provide valuable supervision for music analysis and performance modeling, yet datasets that jointly align all three modalities remain scarce and limited in scale and genre diversity. We introduce Gangqinpu, a multi-genre piano dataset containing 18,786 score-performance pairs and 1,007.1 hours of source audio, with automatic note-level alignments and derived beat/downbeat annotations. It spans diverse genres, including pop, classical, jazz, folk/traditional, rock, and blues. Our automated pipeline transcribes piano audio, aligns the resulting MIDI with musical scores, assigns genre labels, and filters out non-expressive performances with a score-informed binary classifier. We further construct joint splits with ASAP using conservative score-similarity grouping to reduce data leakage. We evaluate Gangqinpu on performance MIDI-to-score transcription, expressive performance rendering, and beat tracking. Downstream experiments demonstrate the value of Gangqinpu as complementary training data, with cross-dataset gains in beat tracking and primarily in-domain benefits for transcription and expressive performance rendering. The code and dataset are publicly available at this https URL.
https://arxiv.org/abs/2610.34694
Background The combination of hematoma location, expansion and final volume influence neurological outcome. The aim of this study was to derive and validate a quantitative and interpretable region-based NCCT image feature model for predicting hematoma expansion. Methods The Ethnic/Racial Variations of Intracerebral Hemorrhage(ERICH) study was used for derivation and the Virtual International Trials in Stroke Archive(VISTA) was used for validation. NCCT images at presentation and 24+/-6hours were pre-processed, segmented and categorized into lobar, basal ganglia or thalamic regions. Hematoma expansion was defined by region and according to its effect on clinical outcome. Transport-based morphometry (TBM) produced a unique transport-based representation from each NCCT hematoma image. Linear classifiers applied to transport representations were used to derive a predictive model. The regional TBM model's performance was compared to alternate methods using AUROC analyses. Results The derivation dataset included 574 patients, comprising 162 lobar, 257 basal ganglia and 155 thalamic hematomas. The validation dataset included 196 patients, comprising 31 lobar, 119 basal ganglia and 46 thalamic hematomas. A combined regional model predicted hematoma expansion in the validation dataset with an AUROC of 0.787. The principal image features which varied by location were size, heterogeneity, eccentricity and density distribution towards the hematoma periphery. The regional TBM model outperformed alternative clinician based and machine-learning methods. Conclusions TBM discriminated and visualized regional features of hematoma expansion from presentation NCCT scans. This provides preliminary evidence for regional radiographic features of hematoma expansion. It has the potential to improve hematoma expansion prediction and understanding of hematoma expansion pathophysiology.
https://arxiv.org/abs/2610.34645
Objective. Deformable image registration remains challenging because corresponding anatomy may exhibit different intensity relationships. Hand-crafted modality-robust descriptors such as Modality-Independent Neighborhood Descriptor with Self-Similarity Context (MIND-SSC) and features from pretrained vision encoders provide distinct representations for correspondence estimation, but how they should be combined remains unclear. We investigate whether tighter integration improves registration over a sequential, scale-separated design. Approach. We introduce FlexiMIND, a training-free and label-free two-stage framework in which frozen FlexiCT-2D encoder features establish contextual correspondence and MIND-SSC refines the deformation within ConvexAdam. Five alternative feature-coupling strategies were evaluated under a protocol across CT-CT (DIR-Lab 4DCT 4DCT, 10 cases), MR-MR (ACDC, 50 patients), and MR-CT (Learn2Reg Abdomen MR-CT, 8 pairs) registration. Performance was assessed using target registration error (TRE), Dice similarity coefficient (Dice), SDlogJ, and local folding. Main results. FlexiMIND achieved a TRE of 1.27 mm on DIR-Lab 4DCT, a DSC of 0.807 on ACDC, and an organ-averaged DSC of 0.825 on Abdomen MR-CT. Across 15 coupling-strategy dataset comparisons, only four outperformed the reference method, with a maximum absolute improvement of 0.0043 in Dice, whereas degradation reached 20.6 % in TRE and 37.9 % in Dice. Changes in registration accuracy were closely associated with changes in deformation-field regularity, and strategies that better preserved regularity generally remained closest to the reference performance. Landmark-level analysis showed that the pretrained encoder stage primarily reduced large correspondence errors rather than uniformly improving local precision. Code is available upon acceptance at this https URL.
https://arxiv.org/abs/2610.34600
Despite advances in automated liver lesion segmentation, a gap remains between image-derived predictions and clinically interpretable radiology reports. Bridging this gap requires quantitative measurements that are both reproducible and traceable. We present an evidence-grounded agentic AI framework that connects liver and lesion segmentation with deterministic quantification and structured reporting. TotalSegmentator-based nnU-Net models delineate the liver, Couinaud segments, and lesions, while computational tools extract volumes, dimensions, counts, and anatomical locations. A large language model (LLM) organizes these measurements into a predefined reporting schema, with numerical consistency checks against source measurements. On 51 LiTS cases, the segmentation stage achieved a liver Dice score of 0.947, lesion sensitivity of 75.6%, and positive predictive value of 68.9%. Exploratory screening of 25,494 MERLIN reports identified 10,208 (40.0%) with liver-related abnormalities. By connecting quantitative imaging evidence with agentic report generation, this work takes a step toward human-centered medical AI that prioritizes traceability and clinical relevance.
https://arxiv.org/abs/2610.34599
Agent harnesses enable large language models to act through tools, observe the results, and iteratively work toward a task goal. Verification checks and revises agents' work to improve output quality. Yet there is limited guidance on how harnesses should organize this effort. We study verification as an axis of test-time scaling by examining when to intervene, who checks the work, and how checking leads to revision. Our framework spans four layers of the agent lifecycle: initial instructions, mid-turn reminders, completion gates, and independent review. We compare 17 designs across seven frontier coding models on Terminal-Bench (TB) and SWE-bench Verified (SWE), and introduce VisualBench for expert evaluation of visual and interactive artifacts. Initial instructions and execution-time reminders yield modest, model-dependent average gains on the coding benchmarks, without substantially increasing average output token use. Completion gates can yield larger gains but may substantially increase cost. Independent review by a critic offers a more favorable accuracy-cost tradeoff, improving average accuracy by 5.9-9.5 percentage points on TB and 3.9-6.6 on SWE at 1.13-1.25 and 1.00-1.04 times baseline cost, respectively. On VisualBench, independent review achieves net preferences of 13.2-23.8 points over the baseline across its four designs, exceeding the evaluated L0-L2 designs. Additional review rounds or independent critics further improve average TB accuracy by up to 3.5 percentage points. Together, these results provide empirical guidance for verification design in agent harnesses and support independent review as a practical axis of test-time scaling.
https://arxiv.org/abs/2610.34577
Multimodal humor detection requires interpreting language, vocal delivery, and visual behavior jointly. Expressive fusion must also regulate how interactions of different orders enter the final judgment. We propose Möbius Interaction Filtering (MoIF), which learns a modality-subset prediction function and imposes order-selective spectral regularization through its aggregation structure. Private temporal encoders and balanced reciprocal interaction construct supervised unimodal, bimodal, and trimodal predictors. Their uniform aggregate is exactly a Möbius-domain filter with a geometrically decreasing interaction-order response. This formulation retains higher-order representations while regulating their contribution to prediction. Controlled UR-FUNNY perturbations reveal a pronounced decline in label-aligned signal-to-noise ratio across interaction orders. The induced filter improves the final ratio by 1.62-2.79 times relative to the co-trained trimodal predictor while preserving its own clean class-aligned signal under perturbation. Five-seed experiments achieve 70.6% accuracy and weighted F1 on UR-FUNNY, exceeding thirteen reproduced methods on these detection metrics. CMU-MOSI and CMU-MOSEI evaluations establish applicability to sentiment prediction, and cross-seed analyses show more stable estimates on all three datasets. MoIF connects the organization of multimodal interactions to their statistical reliability through an executable spectral constraint. Code is available at this https URL.
https://arxiv.org/abs/2611.12913
Decentralized multi-agent systems (MAS) enable large language model agents to interact without centralized control. Agents need to adapt to other agents at test time and reuse accumulated experience to refine their interaction strategies for future interactions across cooperative, mixed-motive, and adversarial settings. However, existing approaches often develop adaptation mechanisms for particular interaction settings. This raises a central question: how can decentralized MAS support test-time adaptation and reliable experience reuse across interaction settings? We propose EvolveMAS, a unified self-evolution framework that supports experience-driven adaptation within and across episodes. EvolveMAS pairs each main agent with a dedicated MemoryAgent for test-time memory governance, which maintains isolated Structured Memory outside the main agent's reasoning context. EvolveMAS organizes adaptation into two complementary processes: Within-Episode Online Adaptation and Cross-Episode Knowledge Evolution. During Within-Episode Online Adaptation, MemoryAgent updates episodic, semantic, and working memory from interaction evidence to inform subsequent actions. During Cross-Episode Knowledge Evolution, the main agent uses memory and episode outcomes to distill reusable interaction knowledge comprising partner profiles and interaction strategies for subsequent episodes. Across five benchmarks and six evaluation configurations spanning the three interaction settings, EvolveMAS achieves the highest normalized average score of 77.2 among the compared methods. Ablations show that test-time memory governance and cross-episode knowledge reuse both improve performance. Interaction traces further illustrate how updated memory and distilled knowledge guide subsequent interactions. These results support experience-driven adaptation in decentralized MAS without updating model parameters.
https://arxiv.org/abs/2610.34450
Traditional autoencoders rely on static network architectures to learn latent representations. Recently we suggested a dynamic architecture inspired by Hebbian plasticity, enabling topological self-organization among convolutional feature maps (R. Koç et al., Proceedings of the SIU, 2026). Early implementations of this paradigm infer relationships among feature maps using global average pooling and model their interactions through scalar-valued connections. This approach falls short to fully exploit the underlying spatial information, embedded in input images. Furthermore, abrupt transitions between different loss functions and training dynamics can introduce representation shocks, leading to unstable feature learning. In this paper, we introduce a Spatially-Aware Self-Organizing Graph Autoencoder (S-LooM) to address these structural bottlenecks. By using spatial cross-correlation, and spatial kernels, S-LooM establishes dynamic connections among feature maps. We replace static scalar message passing of autoencoders with learnable spatial kernels and eliminate high-frequency deconvolutional artifacts observed in decoders using nearest-neighbor upsampling. Via a soft-blending training curriculum, the latent topology aligns smoothly during training. Our experiments show that S-LooM yields a more stable topology, compared to the state of the art dynamic architectures. Meaning, the graph structure that emerge has more resistance to feature entanglement and noise. It forms functional communities responsible for visual attributes, such as illumination, shape and texture even under high connection densities.
https://arxiv.org/abs/2610.34346
Extracting salient information from long-horizon dialogue histories remains a significant challenge for Large Language Models (LLMs) due to the restricted context windows and efficient retrieval challenges. Current methods typically rely on history organization or local token-level optimization, suffering from information loss, retrieval redundancy, and myopic attention. Inspired by the view in cognitive science that dialogue constitutes sense-making and memory association, we propose CoLink, a framework that enhances LLMs' global context awareness in domain-specific multi-turn dialogues through information linking and compression. CoLink first clusters historical contexts into topic graphs to establish cross-semantic associations with external knowledge. Guided by the adaptive control of thought, it then retrieves and activates knowledge via synergizing semantic relevance and graph-structural topology. Crucially, to achieve effective compression of history and knowledge, we ground multi-turn dialogues in information theory and derive a principled optimization objective for extracting the core sense-making information. Coupled with a competitive slot mechanism, CoLink condenses the global context into dedicated memory and knowledge slots, efficiently distilling lengthy, redundant inputs into an extremely compact latent representation. Extensive empirical results on both public and our MMD datasets validate CoLink's superiority and generalizability in knowledge linking and sense modeling.
https://arxiv.org/abs/2610.34335
Text-to-motion models generate for the skeletons they were trained on, while a new character arrives as a skeleton file and nothing else. We study text-to-motion for skeletons that no training clip has used, with no clips of the character, no template, no retargeting and no fine-tuning. We present TopX, a 73M-parameter flow-matching transformer that generates joint rotations and a root trajectory for any kinematic tree from its tree, rest pose and joint names. Motion is normalised by the rest pose and the size of the skeleton, and the rest pose enters as a world-frame descriptor that does not depend on how the rig's author oriented each joint, with rest conventions randomised in training. Attention is factorised over the joints and the frames of a motion and carries a spatial--temporal encoding designed for skeleton graphs of any topology: queries and keys are rotated across joints by a heat-kernel coordinate of the skeleton graph, which is defined for every tree and does not change with the order of its joints or the eigenbasis of its Laplacian, and across frames by a rotary encoding of frame order. To measure what this buys, we build SkelShift: four motion corpora under one codec, 143,334 captioned clips on skeletons of 4 to 142 joints, from which 4,232 clips on 2,043 skeletons are held out by clip, rig, tree class or species and organised along topology, joint count, body plan, species and rig source, and scored by motion quality: Fréchet distances on public kinetic features and on kinematic descriptors, foot skating and ground contact. Against UniMate trained on the same clips, on test rigs without a training clip, TopX lowers the descriptor FID from 5.73 to 4.58 when UniMate's guidance is tuned for the kinetic FID, and the kinetic FID from 1.48 to 0.55 when it is tuned for the descriptor FID, also on motions no training skeleton performs; fewer of its frames sink below the ground.
https://arxiv.org/abs/2610.34321
Training a spatial VLM to answer which side an object lies on does not require it to preserve the metric displacement that determines that answer, while supervision for a distance estimate may leave categorical relations largely unconstrained. When post-training emphasizes one readout, improvements can therefore concentrate on that interface while transfer to others remains weak or even degrades; we call this observed specialization interface lock-in. ChronoReplay addresses it by replaying heterogeneous spatial supervision as chronologically ordered episodes, retaining source-native targets, and balancing supervision across families. Under the same observation access, a single ChronoReplay checkpoint improves OVO-S-Bench, OST-Bench, and UCS-Bench together, including a 13.41-point gain on OST-Bench, and achieves the highest three-benchmark mean among the supervision settings compared here. Same-data controls further separate the ingredients of this transfer: answer serialization redistributes gains across OST and UCS, family-balanced learning improves over ordinary joint SFT and natural turn weighting, and chronological replay exceeds a shifted-order control by about 2.41 OST points across two paired seeds. Adaptive frame selection then changes evidence access while keeping the trained model fixed, providing an additional inference-time gain at lower visual cost. Streaming spatial performance depends on what distinctions supervision preserves, how heterogeneous histories are organized during learning, and what causal evidence inference makes available at test time.
https://arxiv.org/abs/2610.34283
Skill documents are becoming the default adaptation surface for frozen large language model (LLM) agents, but current skill generators face two key limitations. First, humans and different LLMs differ in how they organize procedural knowledge and approach task solving. Skills authored from a human or single-model perspective may therefore fail to align with the reasoning behaviors of other backbones, limiting cross-model generalization. Second, sequential skill descriptions inadequately represent procedures with parallel branches and explicit dependencies, while existing skill optimizers primarily operate on prose and lack explicit support for graph-structured skill evolution. We introduce MultiEvoSkill, a skill-optimization framework comprising three core components: 1) Skill compiler casts the skill as a typed graph with tailored operators, producing a compact, informative representation with explicit procedural topology; 2) Heterogeneous agent system proposes revision advices using collective intelligence, where minimal and informative issues are identified based on multiple sets of experiences. 3) Structure-aware debate and evolution mechanism organizes candidate edits into a heterogeneous graph, resolves them under dependency and importance constraints, and distills the collective proposal into the artifact. Multi-dimensional gating ensures the information gain and topological correctness of each evolution. Optimal skills are then obtained when the system converges at a consistent state. Various experiments and ablations against SOTA methods validate the effectiveness and generalization of MultiEvoSkill.
https://arxiv.org/abs/2610.34260
Materials research relies on scattered and growing data, knowledge, and simulation tools, and a single study often spans multi-day workflows across literature analysis, computation, and experiment. Large language model agents can now retrieve materials data, operate simulation codes, and close design-synthesis-characterization loops with autonomous laboratories. However, extending these agents with new capabilities, keeping long-running tasks on track, and letting them learn from past tasks and user feedback remain difficult. Here we present MatOS, an agent operating system for materials research with three core components: (1) a layered scientific asset graph that organizes data, knowledge, tools, skills, roles, and workflows as reusable assets; (2) a stateful adaptive workflow that runs tasks as explicit execution graphs with scientific checks, so that they can be resumed and traced; and (3) evidence-driven asset evolution, which revises these assets from execution records and user feedback. On ResearchClawBench, MatOS improves on the previously best-performing agent by 11.15% with the same base model and extends the cost-score Pareto frontier. With a common base model, it outperforms Codex CLI and Claude Code, exceeding Codex CLI by 2.13% on MatTools and 4.17% on SciAgentGym. In eight research cases, MatOS narrowed more than 30 million candidate structures to 97 solid-electrolyte candidates, identified 89 halide variants with higher computed conductivity than their parent structures, linked the force errors of 11 pretrained interatomic potentials to local atomic environments, and supported two new methods that reduce errors by 3.1% and 13.6% relative to their baselines. These results show that organizing research capabilities as reusable, evolving assets enables agents to carry out long, multistage materials studies. Code is available at this https URL.
https://arxiv.org/abs/2610.34203
Optical character recognition (OCR) systems depend strongly on the scale and diversity of their training data. One approach to document OCR first analyzes the page layout and extracts text-line images, which are then transcribed by a recognition model. Multilingual line-level recognition involving Persian, Arabic, and English presents additional challenges because different writing directions, fonts, and styles may coexist within a single line. This paper introduces MuSTEL, a large-scale synthetic multilingual OCR dataset containing more than 14.5 million annotated text-line images, organized into a training partition and predefined small and large test sets. Samples are generated from natural-language corpora, including multilingual content already present in the source texts. The generation framework incorporates capability-aware font selection, within-line font and style changes, mixed-direction text rendering, highlighting, and underlining. A validated repository of 2,009 font files from 711 families supports broad typographic coverage, while varied backgrounds, blur, geometric distortion, skew, stroke width, and image save quality provide diverse document-like appearances. Statistical analysis characterizes the dataset's language composition, transcription lengths, lexical and character-sequence diversity, image dimensions, typography, and rendering conditions. MuSTEL provides a resource for training and evaluating multilingual line-level OCR systems and investigating recognition robustness across languages, writing directions, and visual conditions.
https://arxiv.org/abs/2610.34148
Long-context language models can increasingly process entire documents, making full-context prompting a practical alternative to passage retrieval. However, greater capacity determines how much information a model can receive, but not which information best supports a query. This distinction motivates rethinking retrieval: when the full document already fits, can selecting and organizing its evidence improve answering? The challenge is to retain supporting facts and relations while keeping the context focused on the question. To reconcile these goals, we introduce \ours, a training-free framework that uses semantic abstraction to guide selection while retaining original text for answering. \ours constructs a reusable semantic hierarchy from the full document in a single generation, linking node descriptions to source spans. Given a query, structural--semantic scoring and fine-grained lexical attribution nominate shared source nodes; their union with ancestor context preserves structural relations and complete local text. Across four frontier answer models, \ours improves average QA-F1 by 1.72 points over full-context prompting on three LongBench tasks, surpassing the evaluated baselines on the four-model average while reducing answer-context length by 58.55\%. Our code is available at the $\href{this https URL}{repository}$.
https://arxiv.org/abs/2610.34146
Semantic indexing must preserve reusable source information without committing too early to one business schema, while providing checkable premises for reasoning. We propose 5W1H+Which, a design that separates source-grounded content extraction from progressive ontology binding. The 5W1H questions organize independently addressable content units; Which records links to versioned ontology elements, mapping relations, applicability scopes, and validation states. Time, location, system environment, and participant roles constrain when facts, bindings, and rules may be used. Unbound content remains searchable, whereas formal reasoning requires explicit premise checks. The design distinguishes business valid time, system knowledge time, and operational traces, and tracks dependencies for revalidation and incremental maintenance. Worked examples illustrate cloud migration and temporal policy validity. The resulting linked corpus can support downstream training and evaluation organized around benchmarks, positive and negative examples, trajectories, and costs. Trajectories preserve steps, evidence-backed checkpoints, recovery, and local-to-global failure impact; reusable skills remain downstream abstractions requiring cross-episode validation. We specify falsifiable hypotheses and capability-matched baselines to test cross-task evidence coverage, contextual misuse, and update cost under equal information and budgets. The contribution is a proposed indexing and maintenance contract, not a new universal ontology or post-training algorithm. No experiments have yet been conducted, and no performance improvement is claimed.
https://arxiv.org/abs/2609.35184
Memory is increasingly used in long-horizon agents, yet the net value of building memory over direct retrieval from raw history remains largely unclear. Through a large-scale evaluation across multiple benchmarks and model backbones, we find many dedicated memory systems underperform naive RAG in many cases, indicating a negative net value. Taking RAG as a reference, we dissect the net value of memory into the gains and losses over direct retrieval, termed memory benefit and memory tax. These gains and losses arise because memory process changes how the raw information is organized and accessed. We view such changes as the double-edged deviations from the raw history. They are originally designed and introduced to benefit the task, for example by reorganizing facts or distilling preferences, but some may also remove or obscure evidence, giving rise to memory tax. Since memory benefit reflects the expected gains of its design, we then focus more on the overlooked memory tax, study how it arises from the system and evaluation side, and generalize the concept to more application scenarios of memory. We hope the notion of memory tax can provide a new lens and shift more attention from standalone performance to the net value of memory, informing more reliable memory designs and evaluations. We will open-source the code and results to the community. Project Page: this https URL
https://arxiv.org/abs/2610.34073