arXiv:2603.22306v1 Announce Type: new Abstract: Affective judgment in real interaction is rarely a purely local prediction problem. Emotional meaning often depends on prior trajectory, accumulated context, and multimodal evidence that may be weak, noisy, or incomplete at the current moment. Although multimodal emotion recognition (MER) has improved the integration of text, speech, and visual signals, many existing systems remain optimized for short-range inference and provide limited support for persistent affective memory, long-horizon dependency modeling, and robust interpretation under imperfect input. This technical report presents the Memory Bear AI Memory Science Engine, a memory-centered framework for multimodal affective intelligence. Instead of treating emotion as a transient output label, the framework models affective information as a structured and evolving variable within a memory system. It organizes processing through structured memory formation, working-memory aggregation, long-term consolidation, memory-driven retrieval, dynamic fusion calibration, and continuous memory updating. At its core, multimodal signals are transformed into structured Emotion Memory Units (EMUs), enabling affective information to be preserved, reactivated, and revised across interaction horizons. Experimental results show consistent gains over comparison systems across benchmark and business-grounded settings, with stronger accuracy and robustness, especially under noisy or missing-modality conditions. The framework offers a practical step from local emotion recognition toward more continuous, robust, and deployment-relevant affective intelligence.

arXiv:2603.22312v1 Announce Type: new Abstract: This paper computationally investigates whether thought requires a language-like format, as posited by the Language of Thought (LoT) hypothesis. We introduce the ``AI Private Language'' thought experiment: if two artificial agents develop an efficient, inscrutable communication protocol via multi-agent reinforcement learning (MARL), and their performance declines when forced to use a human-comprehensible language, this Efficiency Attenuation Phenomenon (EAP) challenges the LoT. We formalize this in a cooperative navigation task under partial observability. Results show that agents with an emergent protocol achieve 50.5\% higher efficiency than those using a pre-defined, human-like symbolic protocol, confirming the EAP. This suggests optimal collaborative cognition in these systems is not mediated by symbolic structures but is naturally coupled with sub-symbolic computations. The work bridges philosophy, cognitive science, and AI, arguing for pluralism in cognitive architectures and highlighting implications for AI ethics.

arXiv:2603.22345v1 Announce Type: new Abstract: Multimodal emotion recognition in conversations (MERC) aims to identify and understand the emotions expressed by speakers during utterance interaction from multiple modalities (e.g., text, audio, images, etc.). Existing studies have shown that GCN can improve the performance of MERC by modeling dependencies between speakers. However, existing methods usually use fixed parameters to process multimodal features for different emotion types, ignoring the dynamics of fusion between different modalities, which forces the model to balance performance between multiple emotion categories, thus limiting the model's performance on some specific emotions. To this end, we propose a dynamic fusion-aware graph convolutional neural network (DF-GCN) for robust recognition of multimodal emotion features in conversations. Specifically, DF-GCN integrates ordinary differential equations into graph convolutional networks (GCNs) to {capture} the dynamic nature of emotional dependencies within utterance interaction networks and leverages the prompts generated by the global information vector (GIV) of the utterance to guide the dynamic fusion of multimodal features. This allows our model to dynamically change parameters when processing each utterance feature, so that different network parameters can be equipped for different emotion categories in the inference stage, thereby achieving more flexible emotion classification and enhancing the generalization ability of the model. Comprehensive experiments conducted on two public multimodal conversational datasets {confirm} that the proposed DF-GCN model delivers superior performance, benefiting significantly from the dynamic fusion mechanism introduced.

arXiv:2603.22347v1 Announce Type: new Abstract: While Landauer's principle establishes the fundamental thermodynamic floor for information erasure and Fisher Information provides a metric for local curvature in parameter space, these classical frameworks function effectively only as approximations within regimes of sparse rule-constraints. They fail to explain the super-linear, and often explosive, computational and energy costs incurred when maintaining symbolic interpretability during the reconfiguration of advanced intelligent systems. This paper introduces the property of intelligence inertia and its underlying physical principles as foundational characteristics for quantifying the computational weight of intelligence. We demonstrate that this phenomenon is not merely an empirical observation but originates from the fundamental non-commutativity between rules and states, a root cause we have formally organized into a rigorous mathematical framework. By analyzing the growing discrepancy between actual adaptation costs and static information-theoretic estimates, we derive a non-linear cost formula that mirrors the Lorentz factor, characterizing a relativistic J-shaped inflation curve -- a "computational wall" that static models are blind to. The validity of these physical principles is examined through a trilogy of decisive experiments: (1) a comparative adjudication of this J-curve inflation against classical Fisher Information models, (2) a geometric analysis of the "Zig-Zag" trajectory of neural architecture evolution, and (3) the implementation of an inertia-aware scheduler wrapper that optimizes the training of deep networks by respecting the agent's physical resistance to change. Our results suggest a unified physical description for the cost of structural adaptation, offering a first-principle explanation for the computational and interpretability-maintenance overhead in intelligent agents.

arXiv:2603.22350v1 Announce Type: new Abstract: Deterministic pre-execution safety gates evaluate whether individual agent actions are compatible with their assigned roles. While effective at per-action authorization, these systems are structurally blind to distributed attacks that decompose harmful intent across multiple individually-compliant steps. This paper introduces Session Risk Memory (SRM), a lightweight deterministic module that extends stateless execution gates with trajectory-level authorization. SRM maintains a compact semantic centroid representing the evolving behavioral profile of an agent session and accumulates a risk signal through exponential moving average over baseline-subtracted gate outputs. It operates on the same semantic vector representation as the underlying gate, requiring no additional model components, training, or probabilistic inference. We evaluate SRM on a multi-turn benchmark of 80 sessions containing slow-burn exfiltration, gradual privilege escalation, and compliance drift scenarios. Results show that ILION+SRM achieves F1 = 1.0000 with 0% false positive rate, compared to stateless ILION at F1 = 0.9756 with 5% FPR, while maintaining 100% detection rate for both systems. Critically, SRM eliminates all false positives with a per-turn overhead under 250 microseconds. The framework introduces a conceptual distinction between spatial authorization consistency (evaluated per action) and temporal authorization consistency (evaluated over trajectory), providing a principled basis for session-level safety in agentic systems.

arXiv:2603.22357v1 Announce Type: new Abstract: Aphasia, a language impairment primarily resulting from stroke or brain injury, profoundly disrupts communication and everyday functioning. Despite advances in speech therapy, barriers such as limited therapist availability and the scarcity of personalized, culturally relevant tools continue to hinder optimal rehabilitation outcomes. This paper reviews recent developments in neurocognitive research and language technologies that contribute to the diagnosis and therapy of aphasia. Drawing on findings from our ethnographic field study, we introduce two digital therapy prototypes designed to reflect local linguistic diversity and enhance patient engagement. We also show how insights from neuroscience and the local context guided the design of these tools to better meet patient and therapist needs. Our work highlights the potential of adaptive, AI-enhanced assistive technologies to complement conventional therapy and broaden access to therapy. We conclude by outlining future research directions for advancing personalized and scalable aphasia rehabilitation.

arXiv:2603.22359v1 Announce Type: new Abstract: Current AI agent frameworks commit early to a single interaction protocol, a fixed tool integration strategy, and static user models, limiting their deployment across diverse interaction paradigms. To address these constraints, we introduce STEM Agent (Self-adapting, Tool-enabled, Extensible, Multi-agent), a modular architecture inspired by biological pluripotency in which an undifferentiated agent core differentiates into specialized protocol handlers, tool bindings, and memory subsystems that compose into a fully functioning AI system. The framework unifies five interoperability protocols (A2A, AG-UI, A2UI, UCP, and AP2) behind a single gateway, introduces a Caller Profiler that continuously learns user preferences across more than twenty behavioral dimensions, externalizes all domain capabilities through the Model Context Protocol (MCP), and implements a biologically inspired skills acquisition system in which recurring interaction patterns crystallize into reusable agent skills through a maturation lifecycle analogous to cell differentiation. Complementing these capabilities, the memory system incorporates consolidation mechanisms, including episodic pruning, semantic deduplication, and pattern extraction, designed for sub-linear growth under sustained interaction. A comprehensive 413-test suite validates protocol handler behavior and component integration across all five architectural layers, completing in under three seconds.

arXiv:2603.22386v1 Announce Type: new Abstract: Large language model (LLM)-based systems are becoming increasingly popular for solving tasks by constructing executable workflows that interleave LLM calls, information retrieval, tool use, code execution, memory updates, and verification. This survey reviews recent methods for designing and optimizing such workflows, which we treat as agentic computation graphs (ACGs). We organize the literature based on when workflow structure is determined, where structure refers to which components or agents are present, how they depend on each other, and how information flows between them. This lens distinguishes static methods, which fix a reusable workflow scaffold before deployment, from dynamic methods, which select, generate, or revise the workflow for a particular run before or during execution. We further organize prior work along three dimensions: when structure is determined, what part of the workflow is optimized, and which evaluation signals guide optimization (e.g., task metrics, verifier signals, preferences, or trace-derived feedback). We also distinguish reusable workflow templates, run-specific realized graphs, and execution traces, separating reusable design choices from the structures actually deployed in a given run and from realized runtime behavior. Finally, we outline a structure-aware evaluation perspective that complements downstream task metrics with graph-level properties, execution cost, robustness, and structural variation across inputs. Our goal is to provide a clear vocabulary, a unified framework for positioning new methods, a more comparable view of existing body of literature, and a more reproducible evaluation standard for future work in workflow optimizations for LLM agents.

arXiv:2603.22404v1 Announce Type: new Abstract: Consider a market of competing model providers selling query access to models with varying costs and capabilities. Customers submit problem instances and are willing to pay up to a budget for a verifiable solution. An arbitrageur efficiently allocates inference budget across providers to undercut the market, thus creating a competitive offering with no model-development risk. In this work, we initiate the study of arbitrage in AI model markets, empirically demonstrating the viability of arbitrage and illustrating its economic consequences. We conduct an in-depth case study of SWE-bench GitHub issue resolution using two representative models, GPT-5 mini and DeepSeek v3.2. In this verifiable domain, simple arbitrage strategies generate net profit margins of up to 40%. Robust arbitrage strategies that generalize across different domains remain profitable. Distillation further creates strong arbitrage opportunities, potentially at the expense of the teacher model's revenue. Multiple competing arbitrageurs drive down consumer prices, reducing the marginal revenue of model providers. At the same time, arbitrage reduces market segmentation and facilitates market entry for smaller model providers by enabling earlier revenue capture. Our results suggest that arbitrage can be a powerful force in AI model markets with implications for model development, distillation, and deployment.

arXiv:2603.22558v1 Announce Type: new Abstract: Generating synthetic populations from aggregate statistics is a core component of microsimulation, agent-based modeling, policy analysis, and privacy-preserving data release. Beyond classical census marginals, many applications require matching heterogeneous unary, binary, and ternary constraints derived from surveys, expert knowledge, or automatically extracted descriptions. Constructing populations that satisfy such multi-way constraints simultaneously poses a significant computational challenge. We consider populations where each individual is described by categorical attributes and the target is a collection of global frequency constraints over attribute combinations. Exact formulations scale poorly as the number and arity of constraints increase, especially when the constraints are numerous and overlapping. Grounded in methods from statistical physics, we propose a maximum-entropy relaxation of this problem. Multi-way cardinality constraints are matched in expectation rather than exactly, yielding an exponential-family distribution over complete population assignments and a convex optimization problem over Lagrange multipliers. We evaluate the approach on NPORS-derived scaling benchmarks with 4 to 40 attributes and compare it primarily against generalized raking. The results show that MaxEnt becomes increasingly advantageous as the number of attributes and ternary interactions grows, while raking remains competitive on smaller, lower-arity instances.

arXiv:2603.22561v1 Announce Type: new Abstract: This paper asks whether a bounded neural architecture can exhibit a meaningful division of labor between intuition and deliberation on a classic 64-item syllogistic reasoning benchmark. More broadly, the benchmark is relevant to ongoing debates about world models and multi-stage reasoning in AI. It provides a controlled setting for testing whether a learned system can develop structured internal computation rather than only one-shot associative prediction. Experiment 1 evaluates a direct neural baseline for predicting full 9-way human response distributions under 5-fold cross-validation. Experiment 2 introduces a bounded dual-path architecture with separate intuition and deliberation pathways, motivated by computational mental-model theory (Khemlani & Johnson-Laird, 2022). Under cross-validation, bounded intuition reaches an aggregate correlation of r = 0.7272, whereas bounded deliberation reaches r = 0.8152, and the deliberation advantage is significant across folds (p = 0.0101). The largest held-out gains occur for NVC, Eca, and Oca, suggesting improved handling of rejection responses and c-a conclusions. A canonical 80:20 interpretability run and a five-seed stability sweep further indicate that the deliberation pathway develops sparse, differentiated internal structure, including an Oac-leaning state, a dominant workhorse state, and several weakly used or unused states whose exact indices vary across runs. These findings are consistent with reasoning-like internal organization under bounded conditions, while stopping short of any claim that the model reproduces full sequential processes of model construction, counterexample search, and conclusion revision.

arXiv:2603.22608v1 Announce Type: new Abstract: Users often rely on Large Language Models (LLMs) for processing multiple documents or performing analysis over a number of instances. For example, analysing the overall sentiment of a number of movie reviews requires an LLM to process the sentiment of each review individually in order to provide a final aggregated answer. While LLM performance on such individual tasks is generally high, there has been little research on how LLMs perform when dealing with multi-instance inputs. In this paper, we perform a comprehensive evaluation of the multi-instance processing (MIP) ability of LLMs for tasks in which they excel individually. The results show that all LLMs follow a pattern of slight performance degradation for small numbers of instances (approximately 20-100), followed by a performance collapse on larger instance counts. Crucially, our analysis shows that while context length is associated with this degradation, the number of instances has a stronger effect on the final results. This finding suggests that when optimising LLM performance for MIP, attention should be paid to both context length and, in particular, instance count.

arXiv:2603.22619v1 Announce Type: new Abstract: LLMs often generate seemingly valid answers to flawed or ill-posed inputs. This is not due to missing knowledge: under discriminative prompting, the same models can mostly identify such issues, yet fail to reflect this in standard generative responses. This reveals a fundamental know-act gap between discriminative recognition and generative behavior. Prior work largely characterizes this issue in narrow settings, such as math word problems or question answering, with limited focus on how to integrate these two modes. In this work, we present a comprehensive analysis using FaultyScience, a newly constructed large-scale, cross-disciplinary benchmark of faulty scientific questions. We show that the gap is pervasive and stems from token-level autoregression, which entangles task selection (validate vs. answer) with content generation, preventing discriminative knowledge from being utilized. To address this, we propose DeIllusionLLM, a task-level autoregressive framework that explicitly models this decision. Through self-distillation, the model unifies discriminative judgment and generative reasoning within a single backbone. Empirically, DeIllusionLLM substantially reduces answer-despite-error failures under natural prompting while maintaining general reasoning performance, demonstrating that self-distillation is an effective and scalable solution for bridging the discriminative-generative know-act gap

arXiv:2603.22633v1 Announce Type: new Abstract: Retrieval-Augmented Generation (RAG) systems for biomedical literature are typically evaluated using ranking metrics like Mean Reciprocal Rank (MRR), which measure how well the system identifies the single most relevant chunk. We argue that for full-text scientific documents, this paradigm is incomplete: it rewards retrieval precision while ignoring retrieval breadth -- the ability to surface evidence from across a document's structural sections. We propose GraLC-RAG, a framework that unifies late chunking with graph-aware structural intelligence, introducing structure-aware chunk boundary detection, UMLS knowledge graph infusion, and graph-guided hybrid retrieval. We evaluate six strategies on 2,359 IMRaD-filtered PubMed Central articles using 2,033 cross-section questions and two metric families: standard ranking metrics (MRR, Recall@k) and structural coverage metrics (SecCov@k, CS Recall). Our results expose a sharp divergence: content-similarity methods achieve the highest MRR (0.517) but always retrieve from a single section, while structure-aware methods retrieve from up to 15.6x more sections. Generation experiments show that KG-infused retrieval narrows the answer-quality gap to delta-F1 = 0.009 while maintaining 4.6x section diversity. These findings demonstrate that standard metrics systematically undervalue structural retrieval and that closing the multi-section synthesis gap is a key open problem for biomedical RAG.

arXiv:2603.22651v1 Announce Type: new Abstract: The adoption of large language models (LLMs) for structured information extraction from financial documents has accelerated rapidly, yet production deployments face fundamental architectural decisions with limited empirical guidance. We present a systematic benchmark comparing four multi-agent orchestration architectures: sequential pipeline, parallel fan-out with merge, hierarchical supervisor-worker and reflexive self-correcting loop. These are evaluated across five frontier and open-weight LLMs on a corpus of 10,000 SEC filings (10-K, 10-Q and 8-K forms). Our evaluation spans 25 extraction field types covering governance structures, executive compensation and financial metrics, measured along five axes: field-level F1, document-level accuracy, end-to-end latency, cost per document and token efficiency. We find that reflexive architectures achieve the highest field-level F1 (0.943) but at 2.3x the cost of sequential baselines, while hierarchical architectures occupy the most favorable position on the cost-accuracy Pareto frontier (F1 0.921 at 1.4x cost). We further present ablation studies on semantic caching, model routing and adaptive retry strategies, demonstrating that hybrid configurations can recover 89\% of the reflexive architecture's accuracy gains at only 1.15x baseline cost. Our scaling analysis from 1K to 100K documents per day reveals non-obvious throughput-accuracy degradation curves that inform capacity planning. These findings provide actionable guidance for practitioners deploying multi-agent LLM systems in regulated financial environments.

arXiv:2603.22677v1 Announce Type: new Abstract: Distributional metrics such as Fr\'echet Audio Distance cannot score individual music clips and correlate poorly with human judgments, while the only per-sample learned metric achieving high human correlation is closed-source. We introduce MUQ-EVAL, an open-source per-sample quality metric for AIgenerated music built by training lightweight prediction heads on frozen MuQ-310M features using MusicEval, a dataset of generated clips from 31 text-to-music systems with expert quality ratings. Our simplest model, frozen features with attention pooling and a two-layer MLP, achieves system-level SRCC = 0.957 and utterance-level SRCC = 0.838 with human mean opinion scores. A systematic ablation over training objectives and adaptation strategies shows that no addition meaningfully improves the frozen baseline, indicating that frozen MuQ representations already capture quality-relevant information. Encoder choice is the dominant design factor, outweighing all architectural and training decisions. LoRA-adapted models trained on as few as 150 clips already achieve usable correlation, enabling personalized quality evaluators from individual listener annotations. A controlled degradation analysis reveals selective sensitivity to signal-level artifacts but insensitivity to musical-structural distortions. Our metric, MUQ-EVAL, is fully open-source, outperforms existing open per-sample metrics, and runs in real time on a single consumer GPU. Code, model weights, and evaluation scripts are available at https://github.com/dgtql/MuQ-Eval.

arXiv:2603.22721v1 Announce Type: new Abstract: Recent progress in artificial intelligence has encouraged numerous attempts to understand and decode human visual system from brain signals. These prior works typically align neural activity independently with semantic and perceptual features extracted from images using pre-trained vision models. However, they fail to account for two key challenges: (1) the modality gap arising from the natural difference in the information level of representation between brain signals and images, and (2) the fact that semantic and perceptual features are highly entangled within neural activity. To address these issues, we utilize hyperbolic space, which is well-suited for considering differences in the amount of information and has the geometric property that geodesics between two points naturally bend toward the origin, where the representational capacity is lower. Leveraging these properties, we propose a novel framework, Hyperbolic Feature Interpolation (HyFI), which interpolates between semantic and perceptual visual features along hyperbolic geodesics. This enables both the fusion and compression of perceptual and semantic information, effectively reflecting the limited expressiveness of brain signals and the entangled nature of these features. As a result, it facilitates better alignment between brain and visual features. We demonstrate that HyFI achieves state-of-the-art performance in zero-shot brain-to-image retrieval, outperforming prior methods with Top-1 accuracy improvements of up to +17.3% on THINGS-EEG and +9.1% on THINGS-MEG.

arXiv:2603.22744v1 Announce Type: new Abstract: Large language models excel on objectively verifiable tasks such as math and programming, where evaluation reduces to unit tests or a single correct answer. In contrast, real-world enterprise work is often subjective and context-dependent: success hinges on organizational goals, user intent, and the quality of intermediate artifacts produced across long, multi-tool workflows. We introduce LH-Bench, a three-pillar evaluation design that moves beyond binary correctness to score autonomous, long-horizon execution on subjective enterprise tasks. The pillars are: (i) expert-grounded rubrics that give LLM judges the domain context needed to score subjective work, (ii) curated ground-truth artifacts that enable stepwise reward signals (e.g., chapter-level annotation for content tasks), and (iii) pairwise human preference evaluation for convergent validation. We show that domain-authored rubrics provide substantially more reliable evaluation signals than LLM-authored rubrics (kappa = 0.60 vs. 0.46), and that human preference judgments confirm the same top-tier separation (p

arXiv:2603.22752v1 Announce Type: new Abstract: Automated classification of clinical transcriptions into medical specialties is essential for routing, coding, and clinical decision support, yet prior work on the widely used MTSamples benchmark suffers from severe data leakage caused by applying SMOTE oversampling before train test splitting. We first document this methodological flaw and establish a leakage free benchmark across 40 medical specialties (4966 records), revealing that the true task difficulty is substantially higher than previously reported. We then introduce CLiGNet (Clinical Label Interaction Graph Network), a neural architecture that combines a Bio ClinicalBERT text encoder with a two layer Graph Convolutional Network operating on a specialty label graph constructed from semantic similarity and ICD 10 chapter priors. Per label attention gates fuse document and label graph representations, trained with focal binary cross entropy loss to handle extreme class imbalance (181 to 1 ratio). Across seven baselines ranging from TF IDF classifiers to Clinical Longformer, CLiGNet without calibration achieves the highest macro F1 of 0.279, with an ablation study confirming that the GCN label graph provides the single largest component gain (increase of 0.066 macro F1). Adding per label Platt scaling calibration yields an expected calibration error of 0.007, demonstrating a principled trade off between ranking performance and probability reliability. We provide comprehensive failure analysis covering pairwise specialty confusions, rare class behaviour, document length effects, and token level Integrated Gradients attribution, offering actionable insights for clinical NLP system deployment.

arXiv:2603.22767v1 Announce Type: new Abstract: Observational studies can yield clinically actionable evidence at scale, but executing them on real-world databases is open-ended and requires coherent decisions across cohort construction, analysis, and reporting. Prior evaluations of LLM agents emphasize isolated steps or single answers, missing the integrity and internal structure of the resulting evidence bundle. To address this gap, we introduce RWE-bench, a benchmark grounded in MIMIC-IV and derived from peer-reviewed observational studies. Each task provides the corresponding study protocol as the reference standard, requiring agents to execute experiments in a real database and iteratively generate tree-structured evidence bundles. We evaluate six LLMs (three open-source, three closed-source) under three agent scaffolds using both question-level correctness and end-to-end task metrics. Across 162 tasks, task success is low: the best agent reaches 39.9%, and the best open-source model reaches 30.4%. Agent scaffolds also matter substantially, causing over 30% variation in performance metrics. Furthermore, we implement an automated cohort evaluation method to rapidly localize errors and identify agent failure modes. Overall, the results highlight persistent limitations in agents' ability to produce end-to-end evidence bundles, and efficient validation remains an important direction for future work. Code and data are available at https://github.com/somewordstoolate/RWE-bench.