Papers

Runtime safety monitors based on Linear Temporal Logic (LTL) and finite automata (FSA) are increasingly deployed to intercept unsafe tool-call sequences in LLM agents. Yet the same monitor achieves 68-75% attack coverage on some model architectures and near-zero on others, with no explanation from capability scores, training data, or prompt design. We provide the missing theory. We prove that the recall of any fixed-invariant FSA monitor is bounded above by the concentration of the attack distribution: the fraction of attacks covered by the k most frequent trigger-completion patterns. When attacks concentrate (low Shannon entropy), a small fixed invariant set achieves high recall; when they disperse across many structurally distinct patterns (high entropy), no fixed invariant set of tractable size can, regardless of how the invariants were derived. We validate this entropy-coverage bound across eight frontier LLM architectures. GPT-class and DeepSeek backends yield highly concentrated attacks (H ~ 0.24 bits; one pattern covers 96%), explaining 68-75% recall; Gemini variants yield high-entropy distributions (H ~ 2.81 bits; 7 clusters each <= 7%), explaining near-zero recall (6-13%), invariant to architecture-matched retraining. Entropy accounts for 76% of variance in coverage (Pearson r = -0.87, p = 0.005, 95% CI [-0.98, -0.78]), holding under leave-one-out (r in [-0.91, -0.82]). We introduce a pre-deployment entropy test that predicts monitor coverage from a small attack sample, enabling architecture-aware monitor selection before deployment. The bound and test are architecture-agnostic and apply to any FSA-based runtime monitor over discrete action sequences.

Large language models can now generate fluent and complete stories, yet many outputs still feel formulaic and unnatural because of cliches, over-explanation, linear causal progression, and stereotyped endings, an immediately recognizable AI flavor. Existing detection and evaluation methods often stop at source labels or holistic scores, while revision methods typically target predefined issues through localized edits, limiting their ability to support multiple plausible revision strategies or guide story-wide changes in information release, causal organization, and ending treatment. We introduce CraftAlign, a framework that aligns AI stories with the craft of human storytelling by both assessing Human/AI writing patterns and providing revision guidance. CraftAlign comprises two learned modules and an inference-time guidance pipeline. A feature estimator built on Qwen3.5-9B predicts 304 explicit writing features spanning style and narrative. A class-conditional energy model scores the resulting feature configuration against Human and AI writing patterns, conditioning on the original writing prompt when available. At inference time, CraftAlign applies schema-valid structured perturbations, selects changes that move the feature configuration toward the Human writing pattern, and converts them into natural-language guidance for a separate editor to rewrite the full story. Experiments show that CraftAlign accurately distinguishes Human and AI writing patterns and that its guidance outperforms revision baselines across editors and in a human study.

Constructing an executable chemical-process model remains manually intensive. Chemical engineers translate underspecified requests into coupled decisions about unit operations, thermodynamics, streams, specifications, degrees of freedom (DoF), initialization, solver repair, and optimization; one error can invalidate the model. CRAFTS mirrors the staged workflow of chemical engineers by decomposing simulation building into bounded subtasks assigned to seven bounded roles, with deterministic IDAES/Pyomo gates between stages. Given a natural-language request, process flowsheet diagram (PFD) evidence, and curated chemical-engineering knowledge, Input Understanding and Intent recover requirements, constraints, and process semantics; visual, topology, and specification specialists translate them into typed simulator contracts; and Debug and Optimization support bounded repair and eligible optimization. Fine-tuning is applied to the three schema-critical visual, topology, and specification roles, while the remaining roles use untuned Qwen. The resulting VisualGraphIR, TopologyIR, SpecIR, BuildPlan, and SolveReport expose unit, port, thermodynamic, numerical, and execution decisions. Compatible constructors, property packages, and runners are attached only after semantic artifacts pass engineering gates. We introduce OpenIDAES-450, a 450-case IDAES process- simulation dataset, and evaluate the complete seven-role LangChain/LangGraph workflow through solve and eligible optimization on its frozen 82-case held-out split. CRAFTS completes the prescribed validation and execution contract for for 91.5% of cases and achieves unit, stream, and directed-connection F1 scores of 0.815, 0.791, and 0.782. These results demonstrate the effectiveness of role specialization, typed intermediate representations, and deterministic engineering gates for reliable automated process-model construction.

In social learning, agents learn not only from direct evidence but also through interactions with their peers. We investigate the role of imprecision in such interactions and ask whether it can improve the effectiveness of the collective learning process. To that end we propose a model of social learning where beliefs are equivalent to formulas in a propositional language, and where agents learn from each other by combining their beliefs according to a fusion operator. The latter is parametrised so as to allow for different levels of imprecision, where a more imprecise fusion operator tends to generates a more imprecise fused belief when the two combined beliefs differ. In this context we describe both difference equation models and agent-based simulations of social learning under a variety of conditions and with different initial biases. The results presented suggest that for populations with a strong initial bias towards incorrect beliefs some level of imprecision in fusion can improve learning accuracy across a range of learning conditions. Furthermore, such benefits of imprecision are consistent with a stability analysis of the fixed points of the proposed difference equation models.

Large language models (LLMs) are integral to complex intellectual tasks, yet output quality remains constrained by user-provided prompts. Iterative multi-turn prompting often leads to context degradation and diminishing cognitive returns. We present PAWNI (Prompt Architecture Wizard using Neural Intelligence), an agentic conversational interface of eight agents that transforms unstructured queries into structured prompts through guided question-and-answer dialogue informed by a self-evolving knowledge base. Rather than optimising the model's response, PAWNI optimises the question itself by front-loading intent clarification. We also propose a three-tier framework of 18 prompt elements across Essential, Enhancement, and Elevation categories. To evaluate system behaviour and validate a measurement protocol, we conducted an exploratory within-subjects study (N=4) across four complex tasks, integrating 32-channel EEG, NASA-TLX workload, and behavioural metrics. Participants produced more structurally complete prompts with PAWNI (42% to 91% of assessed elements), rated LLM outputs higher across all quality dimensions, and reported lower workload (39.6 vs. 21.7 NASA-TLX). Every participant reached satisfactory output in a single turn, compared to 1-12 turns unaided. While effect sizes are unstable due to sample size, direction consistency supports the hypothesis that optimising prompt formulation front-end is a critical lever for human-AI collaboration.

High-stakes decisions under uncertainty, such as medical emergency triage, require more than accurate predictions. They depend on estimating the likelihood of alternative outcomes while explicitly weighing the consequences of different actions, principles that have long formed the foundation of medical diagnosis and decision making. Yet language models are increasingly used for high-stakes clinical recommendations without explicit specification of the utilities governing these decisions. Here we show that emergency triage with language models can be understood within a probabilistic decision framework, providing a case study of a broader decision-analytic paradigm for steering, evaluating, and deploying language models in high-stakes settings. Using clinical vignettes from a structured evaluation of a consumer triage system, we analyze recommendations for treatment under alternative utility functions that specify the relative costs of missed emergencies and unnecessary escalation. We find that capable language models adjust recommendations in response to stated utilities, revealing that the same underlying predictions can support markedly different decision policies. These findings show that effective deployment depends not only on improving predictions but also on making decision objectives explicit. More broadly, they suggest that language models for high-stakes applications should be understood and evaluated as probabilistic decision systems whose recommendations depend jointly on predictive performance and explicit utilities.

Estimating causal effects from real-world spatiotemporal data is challenging due to hidden confounders and interference. Standard causal identification methods assume conditional exchangeability given observed covariates, which fails whenever hidden confounders affect both treatment and outcomes - a common setting in domains such as climate, environmental policy, epidemiology, and regional economics. In this paper, we propose a novel spatiotemporal proximal causal inference framework that extends proximal identification theory to spatiotemporal settings. The proposed method jointly captures local and neighborhood-level confounding information by introducing treatment- and outcome-inducing proxies, and we derive a spatiotemporal outcome confounding bridge function that identifies the potential outcome without requiring direct recovery of the hidden confounder. We establish the identifiability of this bridge function under proxy exclusion restrictions and a spatiotemporal completeness condition, and show that the resulting estimator recovers the outcome through a proximal generalization of the g-computation formula. To operationalize this identification result, we propose a neural architecture that learns proxies via transformer-based spatiotemporal encoders - coupled with a conditional mutual information critic to enforce exclusion restrictions and a moment-matching network to guarantee that the learned bridge function satisfies the underlying identifying equation. We further introduce a stabilized weighting scheme to address treatment support imbalance. Experiments on synthetic datasets demonstrate that our approach achieves comparable performance to baseline causal inference methods, while providing, to our knowledge, the first theoretically grounded outcomes for the hidden confounding in the presence of spatiotemporal interference through a proximal causal inference framework.

Stage-one stellarator design searches a high-dimensional family of three-dimensional plasma boundaries and fixed-boundary MHD equilibria for configurations that jointly meet requirements on confinement, field-line topology, force balance, stability proxies, and geometry. These specifications do not provide a general constructive map to a validated finite-beta equilibrium. High-quality targets are commonly developed through iterative numerical optimization whose outcome depends on the initial configuration, active Fourier resolution, objective priorities, and local solver budget. Coordinating this process is computationally costly and expert-intensive, limiting both design throughput and the production of consistently evaluated data. We present a proof of concept for agentic stage-one optimization. A bounded language-model agent diagnoses the current equilibrium and selects the next local optimization experiment, while deterministic DESC execution owns prescribed profiles and flux, symmetry, metric evaluation, solver validity, and acceptance. On a common-budget subset from an expanding finite-beta campaign, the number of gate-valid configurations increases from five inputs to nineteen outputs; median Boozer QS RMS decreases from $2.39\times10^{-4}$ to $1.07\times10^{-4}$, and median maximum principal curvature decreases from $62.56$ to $33.00\,\mathrm{m}^{-1}$. A complementary long route achieves a $9.10\times$ QS reduction while repairing magnetic-well and curvature defects. The system also records every attempted local action as transition evidence, yielding 734 structured parent–action–outcome records in the reported experiments. These results show that agentic outer-loop control can sustain finite-beta, multi-objective search and turn repeated optimization into a scalable source of improved equilibria and reusable decision data.

The emergence of transformer-based deep learning models has brought unprecedented performance across various domains, particularly in natural language processing and computer vision. However, deploying these models, especially on resource-constrained devices, poses significant challenges due to their high computational complexity and large memory size and bandwidth requirements. This complexity has led researchers to use low-bit model weights to reduce memory usage and improve efficiency. In addition to reducing processing and memory demands, quantization introduces another useful property: value locality, where the extremely large number of parameters are restricted to a limited range of values. To fully take advantage of this locality, this paper presents DeVIT, an acceleration method for vision transformers that leverages differential computation to enable multiplier-less matrix multiplication.

Programmable-payment protocols such as x402 enable per-request micropayments, but they do not determine which payable service an autonomous agent should buy under a finite wallet. We formulate this buyer-side problem as agent-native payment decision-making: contextual provider selection under wallet pressure, chosen-only paid feedback, and changing market conditions. We propose 402Pilot, a protocol-agnostic buyer-side decision layer between autonomous agents and payment execution that implements purchasing policies for selecting among payable providers. We instantiate it with PA-DCT, a payment-aware discounted contextual Thompson-sampling policy that adapts purchasing decisions under wallet pressure while learning from post-payment feedback. To evaluate buyer-side payment policies, we introduce 402Pilot-Bench, a frozen-replay benchmark spanning 823 tasks, five heterogeneous provider pipelines, and three market regimes, each evaluated over 30 paired seeds. PA-DCT achieves the strongest fixed-wallet adaptive trade-off among non-oracle policies: it maintains competitive service quality while spending only 39 to 43 percent of the wallet and reallocates spending as market conditions change. It attains the best non-oracle PA-gap/T under the price shock and the best mean and worst-case ranks across the nine scenario-metric combinations of quality, ROI, and PA-gap/T. Comparisons with learning baselines and component ablations further support the effectiveness and design of the proposed decision policy. These results suggest that programmable payment must be complemented by buyer-side decision-making capable of learning service value and adapting purchasing decisions accordingly.

Multimodal large language models (MLLMs) are rapidly evolving with expanded context windows and stronger reasoning capabilities, enabling multi-chart understanding and multi-step inference. These abilities are increasingly important as MLLMs are adopted in complex agentic tasks. However, existing benchmarks largely emphasize single-chart perception, while simple chart-to-chart connections are insufficient to evaluate these capabilities. To capture multi-chart complexity while ensuring consistency and validity, we design a synthesis pipeline supported by latent graphs. Building on this pipeline, we introduce LongChart, a benchmark whose VQA sets contain an average of 6.5 images and 31.2 questions. We evaluate 10 state-of-the-art MLLMs and examine three factors that influence performance: reasoning patterns, auxiliary tools, and robustness to image perturbations. Our results show that MLLM accuracy decreases and varies substantially as computational complexity increases, highlighting directions for future research in multi-chart reasoning.

Large Language Models (LLMs) have a bounded context window. The context window is the maximum input size an LLM can consume for a single inference. AI agents rely on a process called context compaction to fit their state within the context window when calling an LLM. Despite its ubiquity, context compaction has received essentially no formal analysis. In this paper, we initiate a formal study of context compaction. We first introduce a framework consisting of two games that capture the two algorithmic strategies for context compaction used by contemporary AI agents in practice. The Context Selection Game models context compaction algorithms that select a subset of an agent's accumulated state to retain. The Context Generation Game models context compaction algorithms that summarize an agent's state by an arbitrary message of bounded length. We then prove an equivalence between the Context Generation Game and one-way communication complexity. The minimum context compaction budget for answering a set of queries within a target error is equal to the one-way communication complexity of the induced communication problem at the same error. Known bounds from communication complexity therefore transfer directly to context compaction. We also show that the Context Selection Game corresponds to a restricted class of one-way communication protocols. Any gap between selection and generation is therefore a gap between two classes of communication protocols. We prove that there exists a set of queries for which generation needs strictly less budget than selection. The equivalence between the Context Generation Game and one-way communication also lets us measure how well a deployed context compaction algorithm performs on a query relative to the optimal strategy. As an example, we present a case study that evaluates Anthropic's context compaction endpoint on set membership queries.

Deep search has become a fundamental capability of large language models (LLMs) for solving open-domain complex tasks. However, existing approaches typically rely on linear sequential reasoning for both trajectory generation and inference, making it difficult to consistently preserve intermediate states and constraints throughout long-horizon multi-hop search. Consequently, they often suffer from context forgetting, search drift, and inefficient exploration. To address these limitations, we propose $\textbf{G-ReAct}$, a reasoning framework for deep search that organizes reasoning as $\textbf{state evolution over a fixed-topology query graph}$. The evolving graph state explicitly tracks search progress and guides subsequent decisions, transforming exploratory search driven by textual history into graph-guided reasoning under explicit constraints. G-ReAct supports both training and inference: it generates high-quality deep-search trajectories for supervised fine-tuning and provides structured guidance for inference-time search without additional fine-tuning. Experiments demonstrate that with only 1.9K generated trajectories for fine-tuning, Qwen3-30B-A3B-Thinking-2507 achieves $52.6\%$ accuracy on BrowseComp-ZH and $79.0\%$ on XBench, outperforming comparable open-source methods trained on substantially larger datasets, including RL-enhanced methods. Furthermore, when applied at inference time, G-ReAct consistently improves the performance of existing strong LLMs on deep-search tasks. We will publicly release all code and model weights.

Language generation in the limit is a theoretical framework for studying how a generator can learn to produce new valid strings from a stream of positive examples. In this model, an adversary chooses an unknown language from a countable family and enumerates its elements in an arbitrary order, while the generator must eventually output only elements of the language that have not yet appeared in the enumeration. Reliable generation is thus formalized through two eventual guarantees: validity and novelty relative to the observed data. To further quantify the breadth of the generator's outputs, Kleinberg and Wei (FOCS 2025, STOC 2026) introduced lower density as a measure of output coverage. Given an order representing the importance or relevance of possible outputs, lower density is the asymptotic lower bound, as $n$ grows, on the fraction of the first $n$ elements of the target language that the generator outputs before they appear in the data. Kleinberg and Wei showed that $1/2$ is the optimal lower-density guarantee for deterministic algorithms. We develop a simple and unified framework for obtaining optimal lower-density guarantees. We first give a deterministic algorithm that recovers the optimal guarantee of $1/2$ with a significantly simpler analysis than prior work. We then demonstrate the flexibility of our framework through two extensions. First, against an oblivious adversary, randomization raises the optimal guarantee to $1-1/e$. Second, for any finite collection of orders, the optimal deterministic and randomized guarantees can be achieved simultaneously with respect to every order, so accommodating multiple notions of importance or relevance entails no loss in the optimal guarantee.

Recent meta-reasoning frameworks improve LLM reasoning by wrapping chain-of-thought generation in an iterative control loop, allowing more effective backtracking, termination of reasoning loops, and injection of promising reasoning patterns, among other strategy adjustments. Despite promising results, methods often rely on backward-looking reward functions, utilize coarse search actions, or require additional reasoning controller training requiring many-shot supervision. We introduce Cognitive Demand Steering (CDS), a training-free meta-reasoning framework equipped with residual demand assessment: at each step, an LLM-based progress evaluator characterizes the residual reasoning required to arrive at a solution rather than merely evaluating the previous step. This allows a meta-controller to select reasoning interventions comprising both general-purpose exemplars and actions (e.g., general guidance for quantitative reasoning) that directly tackle this forward-looking demand signal. This shift eliminates the need for any trained component while enabling zero-shot transfer across models and tasks with no adaptation. Rather than relying on coarse characterizations, we employ cognitive scales to both design interventions as well as profile initial problem complexity and residual demand signal over 16 dimensions motivated by cognitive science (e.g., attention and scan, learning and abstraction, spatio-physical reasoning), giving the controller a fine-grained vocabulary for diagnosing. Averaged across three frontier LLMs and six reasoning benchmarks, CDS improves accuracy by $21.9\%$ over direct calls and $9\%$ over standard CoT reasoning, with the largest gains on difficult mathematics and coding tasks.

Modern sensing, communication, and learning systems generate heterogeneous network signals, with local data differing in dimension, modality, and geometric structure. Processing such data requires a mathematical framework capable of simultaneously modeling heterogeneous local signal spaces and the transformations relating them. Network sheaves provide such a framework by associating local vector spaces with network entities and linear restriction maps with their interactions. This is the first paper to develop a unified sheaf signal processing (SSP) framework on network sheaves, extending the fundamental operations of signal processing, namely spectral analysis, filtering, and sampling, to heterogeneous local spaces. Unlike graph and topological signal processing, where signals are modeled over a common vector space, SSP jointly models heterogeneous local signal spaces and the linear transformations relating neighboring spaces through restriction maps. We define the Sheaf Fourier Transform (SFT), whose frequencies quantify signal inconsistency induced by the network topology, the restriction maps, and the local geometry. Building on this representation, we develop polynomial sheaf filters and formulate sampling as the joint selection of network nodes and intra-node components. We derive perfect recovery conditions for bandlimited sheaf signals and propose a greedy sampling-set design algorithm. To incorporate application-dependent signal models, including different bases, dictionaries, and learned embeddings, we introduce representation sheaves and characterize the natural transformations that preserve spectral properties and guarantee interoperability across representations. Experiments on synthetic, motion-capture, and financial datasets validate the proposed framework and demonstrate consistent improvements over canonical graph signal processing baselines.

Human pairwise comparison provides a direct basis for perceptual infrared-visible image fusion assessment, but dense annotation becomes costly as method pools grow. We present the Learned Perceptual Image Fusion Measure (LPIFM), among the earliest learned fusion assessors trained directly on dense human A/B/Tie comparisons. LPIFM jointly examines both source images and both fused candidates, combining a shared hierarchical encoder, triadic interaction, and a tie-aware objective to predict comparative preference and perceptual indifference. We construct and publicly release all 6,300 unordered comparisons among 25 methods on 21 VIFB scenes, collected through blinded, randomized annotation and expert adjudication. Across four VIFB evaluation settings, LPIFM achieves 79.2-84.0% agreement with human pairwise judgments and Spearman correlations of 0.941-0.977 with human-derived method rankings. On full method pools, accuracy exceeds the strongest of 19 conventional metrics by 16.3-21.1 pp. Consistency diagnostics show 99.98-100% candidate-swap agreement and no observed decisive preference cycles. External experiments on EVAFusion further demonstrate rapid adaptation to a different fusion-evaluation preference protocol. After only three epochs of fine-tuning, LPIFM surpasses all 19 conventional metrics in accuracy, macro-F1, and ranking correlation. LPIFM provides a scalable instrument for human-aligned fusion assessment, with the preference corpus, model weights, and code publicly available.

Diffusion Transformers (DiTs) have emerged as a core architecture in generative modeling due to their scalability and adaptability to multimodal tasks. DiTs comprise isotropic transformer blocks, and learn representations progressively across depth, where the denoising objective drives later layers to focus on fine-detail reconstruction. This results in degraded representation quality and an imbalanced encoder-decoder behavior. Prior approaches such as representation alignment (REPA) mitigate this by encouraging stronger early representations via training regularization. Alternatively, U-Net-style DiT architectures introduce explicit multi-scale encoder-decoder structures for improved convergence. But they build on standard U-Net wisdom via learnable operators for spatial downsampling, which are not well-suited to transformer architectures, introducing inefficiencies and compatibility issues with components such as cross-attention and representation regularization. In this work, we propose UDT, a U-Net diffusion transformer that combines the representation power of DiTs with the encoding-decoding benefits of U-Nets, through data-adaptive token merging for downsampling and upsampling, while preserving the DiT token dimension. Our baseline UDT architecture outperforms existing U-Net DiTs and achieves performance comparable to REPA across all model sizes. Furthermore, using architectural optimization and REPA, UDT outperforms SiT's 7.9 FID at 1400 epochs (w/o CFG) within 40 epochs (~ 40x faster convergence) for XL model size on 256x256 ImageNet. Finally, it achieves strong image generation performance with CFG, reaching FID of 1.38 (320 epochs) with SD-VAE and 1.35 (500 epochs) with VA-VAE, providing a new backbone for DiTs with strong empirical benefits.

The continued development of LLMs toward persistent and adaptive intelligence increasingly requires long-term memory mechanisms that preserve and reuse information across interactions. Existing memory systems either compress and structure histories for efficient access or perform deep research over broader trajectories. The former lowers online cost but may omit temporal, causal, or cross-step dependencies, while the latter improves evidence coverage at substantial latency and inference cost. This raises a key question: can a memory system achieve strong answer quality while maintaining low online latency? We introduce Router-Mem, an evidence-conditioned progressive execution framework for long-horizon agent memory. Router-Mem first applies a shared low-cost retrieval prefix to obtain evidence. A lightweight sufficiency router then predicts whether the context supports early termination, which enable a single-token decision at inference time. It is trained with evidence-level supervision and rationale-conditioned representation distillation. When evidence is insufficient, Router-Mem reuses retrieval hits to expand memory blocks and perform deeper analysis and aggregation. Experiments on AMA-Bench and BEAM show that Router-Mem achieves 55.17% and 38.77% score while reducing average inference time by 27.3% and 25.5% compared with full memory execution.

The normalized Transformer (nGPT) realizes hyperspherical representation learning by constraining model parameter vectors and activation vectors to the unit hypersphere. In this paper, we describe a practical training recipe for nGPT and evaluate it on modern hybrid Mamba-2–Transformer Mixture-of-Experts (MoE) models. The recipe introduces Logit Gradient Preconditioning, Logarithmic Learning Rate Decay, GatedAdamW, angular update control, and optional exploration mechanisms. Compared with an unnormalized model of the same hybrid MoE architecture trained with AdamW, the 30B-total-parameter nGPT model reaches the same validation loss using approximately half as many training tokens. The recipe scales across the models considered, which contain up to 30B total parameters.

All Transformer-based large language models compute attention via the Euclidean inner product, an architectural choice that Dong et al. (2021) proved causes representational rank to decay doubly exponentially with depth in pure self-attention stacks. We develop a theoretical framework that targets this structural limitation at the mathematical level by replacing the flat Euclidean metric with learned per-token Riemannian metrics. Our contributions are threefold. (1) We prove that Riemannian attention scores with heterogeneous per-token metrics are non-Gram—they cannot be factorized as QK^T with factorization dimension O(d). We are explicit that this is a structural observation, not a proof of rank preservation. (2) We establish that low-rank metric factors render all geometric operations tractable: geodesic distance in O(d*r) per token and metric inversion in O(d*r^2) via the Woodbury identity—both far below the O(d^3) cost of a general matrix—making Riemannian attention feasible at billion-parameter scale with negligible overhead. (3) We present the Fiber Bundle Transformer, a complete architecture specification in which each token position carries its own Riemannian metric, attention is geodesic distance computation, feed-forward updates use metric-preconditioned steps, and the connection carries explicit curvature and torsion proxies. We derive formal predictions about correctly implemented geometric architectures and identify the central open problem: proving or disproving that heterogeneous Riemannian metrics prevent the rank collapse that row-stochastic attention matrices otherwise cause. This paper presents theoretical analysis and architectural design; empirical validation is the subject of future work.

Video large language models (Video-LLMs) represent videos as dense sequences of visual tokens, whose length grows with the temporal and spatial extent of the input. These tokens often contain substantial redundancy arising from repeated visual patterns, leading to unnecessary computation in the subsequent language-model processing. Existing token compression methods, including pruning and merging, perform compression online during inference, repeatedly incurring additional computation for each input video and often relying on model-specific designs that limit their generality, we instead rethink this paradigm by shifting the costly compression process offline. We propose ONCE, a plug-in video token compression framework that introduces an offline-to-online paradigm: a frequency-aware global codebook is learned once in the visual feature space and reused for lightweight online compression through codebook lookup and aggregation, reducing repeated per-video computation and the need for model-specific compression designs. Extensive experiments across multiple video understanding benchmarks and against diverse compression baselines demonstrate that our approach achieves a strong accuracy-efficiency trade-off, maintaining competitive performance while achieving the lowest inference latency among compared methods.

Hierarchical Graph Retrieval-Augmented Generation (GraphRAG) organizes corpus knowledge at multiple levels of granularity, yet fixed context construction may fail to translate these multi-resolution representations into a context suited to the current query. We identify this mismatch as the representation–inference gap. We propose Agentic Context Engineering for Hierarchical GraphRAG (ACE-GraphRAG), an inference-time context policy layer that supplements and adapts the initial context for generation. ACE-GraphRAG formulates context construction as a policy over gap-aware refinement, retrieval branches, and task-conditioned adaptation. Parallel Differential Retrieval acquires supplementary evidence from depth-oriented factual and breadth-oriented semantic branches. These evidence increments are consolidated with the initial context while preserving provenance and abstraction levels. Full-ACE applies the full policy uniformly within each task family, whereas Adaptive-ACE selects task- and topology-specific policies for individual queries. We evaluate ACE-GraphRAG on HotpotQA, 2WikiMultiHopQA, and four UltraDomain subsets across multi-hop QA and query-focused summarization. Full-ACE outperforms the evaluated RAG and GraphRAG baselines across both task families, while Adaptive-ACE further improves multi-hop QA and is preferred over Full-ACE on all four UltraDomain subsets. Ablation and topology analyses support treating context construction as a query- and task-dependent inference policy rather than a fixed procedure.

Language models (LMs) are increasingly used in recommender systems to generate recommendations or coordinate recommendation tools. However, does their strong performance imply that serving-time LM inference is necessary? We challenge this seemingly reasonable but unverified assumption. To investigate this question, we introduce recovery-based mechanism auditing, which tests whether independent alternatives can achieve the accuracy of LM-based recommender systems without the audited inference mechanisms under controlled information and evaluation conditions. For generative recommendation, we propose LIME-Rec, which combines sequential, collaborative, and offline semantic information without autoregressive LM decoding. On three Amazon Reviews datasets (Beauty, Toys, and Sports), LIME-Rec achieves the highest reported mean in all 12 dataset-metric combinations, with up to a 15.2% relative improvement in Recall@10 over GRAM, an LM-based generative recommender. For agentic recommendation, we propose DART-Rec, which replaces LM-based tool control with deterministic fusion and fixed expert selection. On 1,000 Beauty users, given the same candidate items and expert scores, deterministic fusion achieves a 16.7% relative improvement in Recall@10 over a fine-tuned LM controller. With at most two tool calls, fixed expert selection matches the LM controllers' Hit@10 outcomes for every user in 17 of 18 comparisons. These findings demonstrate that strong recommendation performance does not necessarily require serving-time LM inference: under the evaluated task and information conditions, alternatives without the specified inference mechanisms can match or exceed the accuracy of their LM-based counterparts. To support reproducibility, code is available at https://github.com/Double-wk/LIME-Rec.

Catastrophic forgetting is a major problem in task-incremental learning, where neural networks tend to overwrite previously learned knowledge when trained on new tasks. A number of architecture-based approaches have been proposed to address this problem. However, the architecture-based approaches suffer from another problem related to network capacity when the networks learn long task sequences: As a network is trained on an increasing number of new tasks in a long task sequence, a growing proportion of active parameters becomes static to prevent forgetting of previously learned knowledge. In this paper, we propose Adaptive Hard Attention to the Task (AdaHAT) with an adaptive attention mechanism which allows adaptive updates to static parameters by taking into account the information about previous tasks on both the importance of these parameters to previous tasks and the current network capacity. Based on this idea, we develop a new neural network architecture incorporating our proposed AdaHAT mechanism. AdaHAT extends an existing architecture-based approach, Hard Attention to the Task (HAT), to better support task-incremental learning over long task sequences. We conduct experiments on a number of datasets and compare AdaHAT with task-incremental learning baselines including HAT. Our experimental results show that AdaHAT achieves better average performance across tasks than these baselines, especially on long task sequences, demonstrating the benefits from balancing the trade-off between stability and plasticity of a network when learning such sequences of tasks, alleviating the network capacity problem. Our code is available at pengxiang-wang.com/projects/continual-learning-arena.

Large language model agents increasingly operate in dynamic environments where tool interfaces, APIs, and user requirements change after deployment. Existing self-evolution methods mainly follow two paradigms: harness-based approaches, which externalize feedback into editable memories or skills for rapid adaptation, and parameter-based approaches, which internalize experience into model parameters for deeper capability improvement. However, using either mechanism alone creates a trade-off between flexibility and performance. This paper asks how an agent can coordinate both channels to achieve robust self-evolution. We present COVE, a unified agent self-evolution framework that combines harness-based and parameter-based learning through task-aware routing, stage-aware scheduling, and knowledge optimization. Through this design, COVE treats self-evolution not as indiscriminate accumulation of experience, but as a coordinated process that matches tasks and knowledge types to appropriate learning mechanisms. Experiments across multiple task categories show that COVE outperforms single-channel evolution strategies, demonstrating more robust and efficient improvement under changing environments.

Heterogeneous computed tomography (CT) acquisitions and diverse downstream task requirements limit the transferability of fixed data preparation workflows across data sources and tasks. Existing approaches typically rely on manually designed or dataset-specific rules, making it difficult to accommodate changes in acquisition conditions and analytical objectives without manual intervention. Large language model (LLM)-based agents have shown promise for automating medical workflows, yet their potential for adaptive CT data preparation remains largely unexplored. To bridge this gap, we propose CT-PrepAgent, which enables adaptive CT data preparation through a bounded policy and controlled deterministic execution. Deterministic inspection constructs structured data–task profiles, from which a policy decides an eligible DICOM series or predefined preprocessing profile, while the controlled execution flow guards, resolves, executes, and verifies the decision with bounded recovery when enabled and safe quarantine otherwise. Across three public CT segmentation tasks, CT-PrepAgent derived data-task adaptive preprocessing decisions and achieved the highest macro-average Dice. On two private raw-DICOM cohorts, CT-PrepAgent increased verified output yield from 61.7% to 70.0% and yielded similar registration metrics on common verified outputs. Controlled fault and replay tests validate bounded recovery, safe quarantine, and policy-free replay under tested fault and drift settings.

Tsallis statistics generalizes Boltzmann-Gibbs statistical mechanics through a single real parameter $q$ that controls the weight assigned to rare and frequent events. Originally proposed to describe physical systems with long-range correlations, multifractal geometry, and heavy-tailed fluctuations, the framework has become a recurring ingredient in modern artificial intelligence (AI): it underlies sparse attention mechanisms (sparsemax and $\alpha$-entmax), maximum-entropy reinforcement learning with controllable exploration, robust and heavy-tailed probabilistic models, and a family of generalized loss functions and regularizers. This paper offers a structured perspective on where Tsallis statistics meets AI. We first review the mathematical core: $q$-entropy and its variational (maximum-entropy) foundation, the $q$-exponential and $q$-logarithm, the $q$-central limit theorem, $q$-Gaussian distributions, and their dynamical origin in superstatistics, emphasizing the properties that matter for machine learning. We then survey applications across softmax generalization, reinforcement learning, sequential and graph neural models, generative and probabilistic modeling, loss design, and optimization, extracting the recurring design pattern in each case: a tunable interpolation between dense/uniform and sparse/peaked behavior governed by $q$. We further argue that the heavy-tailed weight spectra and gradient-noise statistics empirically observed in deep networks are themselves nonextensive signatures, placing modern learning dynamics within the scope of $q$-statistics. Finally, we discuss methodological pitfalls, the relationship to information geometry and $q$-exponential families, and open directions, arguing that $q$ should be treated as a learnable inductive bias rather than a fixed hyperparameter.

Test-time scaling lifts large language model reasoning by sampling many candidate solutions and selecting among them, yet the same recipe transfers poorly to vision-language models (VLMs): recent work shows that simple majority voting beats selection methods built on the model's own self-verification, apparently because at the selection layer an image-grounded answer and a confident guess from the language prior look the same. A natural fix is to make the selection signal one that cannot be computed without the image. We study Perturbation Grounded Selection (Pgs), a label-free, training-free rule that scores each candidate by whether the model re-derives it under label-preserving perturbations of the input (cropping, background masking, mild photometric or geometric jitter); Pgs recovers majority voting when the perturbation set is empty. The decisive question is not whether Pgs beats chain-of-thought only majority voting, but whether the perturbation term adds anything once decoding format and budget are controlled. We therefore introduce a format-matched control (MatchedCtrl): the same short, no-CoT draws spent on the original image. Across TextVQA, MATH-Vision, MMMU, and ViLP, with a Qwen headline (three-seed means) and LLaVA-OneVision coverage in matched-budget selector tables, Pgs appears to beat plain majority voting by up to +31.8 points on TextVQA (Qwen), but MatchedCtrl tracks or exceeds Pgs within noise on every benchmark, including the vision-required ViLP; no Qwen category shows a significant gain over this control. The stability gap is real and image-dependent (up to +0.48), yet does not predict per-instance wins. The result is negative and diagnostic: perturbation consistency is at best a partial diagnostic of visual dependence and, on its own, not a usable selection signal once format is controlled; gains reported against CoT-only majority voting overstate such methods.

Recent offline reinforcement learning methods increasingly rely on expressive generative policies and specialized value guidance. We ask how far systematic modernization of a conventional behavior-regularized actor-critic can go while preserving algorithmic simplicity. ReBRAC-v2 directly trains an exact-likelihood normalizing-flow actor with likelihood, MSE, and MAE behavior regularization, classification-based residual critics, staged optimization, and multi-sample inference. We develop one shared configuration through roughly 600 Bayesian proposals on six OGBench tasks, then adapt only two behavior-regularization coefficients using a 16-point grid for the main comparisons. Across ten common OGBench categories, ReBRAC-v2 averages 74.8 versus 52.3 for the next-best method and ranks first in eight categories. Without structural changes, it also obtains the strongest compared averages on D4RL AntMaze (90.2) and Adroit (33.6). Motivated by these results, we extend evaluation to all 235 state-based single-task problems across the 47 dataset variants in the original OGBench suite. To our knowledge, this is the first complete evaluation in this setting; no compared baseline reports results across the full suite. Per-task scores and configurations establish ReBRAC-v2 as a reference baseline for future work beyond the common task subset. The complete cloning objective retains the highest average after coefficient retuning of its alternatives, while further ablations characterize training, capacity, and inference choices. These results demonstrate the strength of systematic engineering within a minimalist offline RL foundation.