每天知道
你的领域
出了什么

Token-level knowledge distillation (KD) matches two conditional distributions per position, yet the standard objectives compare them pointwise: a Kullback-Leibler gradient is blind to which wrong token receives probability mass. We develop a distributional view in which the teacher is represented not by a single softened output but by a family of multi-temperature views - marginals of the annealing path of its logits - and the student is trained against a geometry-aware aggregate of these views under an embedding-based ground cost. We formalize the resulting design space (mixtures, log-linear pooling, entropic Wasserstein barycenters, and a debiased Sinkhorn-divergence flagship in hub and path forms), prove an exact collapse result showing log-linear pooling of tempered views is equivalent to a single temperature, and give a multi-marginal Schrodinger-bridge reading that yields falsifiable predictions. On instruction-tuned Pythia pairs, experiments yield three empirical laws: (i) dispersion law - the benefit of multi-temperature aggregation grows monotonically with the effective temperature dispersion of the views, not with their number; (ii) dispersed views unlock the aggregation operator - the barycenter separates from the arithmetic mixture exactly when transport-based aggregation starts to beat averaging; and (iii) two-regime picture governed by the ceiling gap $Γ=\mathrm{PPL}_{\mathrm{SFT}}-\mathrm{PPL}_{T}$: when the fine-tuned teacher barely beats a supervised student the gentle transport objective is the best KD loss but no KD beats supervised fine-tuning, whereas at a real ceiling the ranking inverts - and the sign of the fidelity-generalization correlation flips. We argue that "which distillation loss is the best" is not a fixed property of the loss but a function of $Γ$.

Text-to-image models are typically reported on average-case prompts, which understates the gap between systems on compositionally demanding requests involving precise object counts, multi-object attribute binding, legible embedded text, and explicit spatial constraints. We evaluate four production text-to-image systems: Hunyuan 3.0, Gemini 3 Pro Image ("Nano Banana Pro"), Black Forest Labs FLUX.2, and Ideogram 3.0. The evaluation uses the 48 hardest prompts drawn from the DataSeeds.AI Sample Dataset (DSD), selected through an automated complexity-scoring pass over the full corpus. Every generated image is graded using an independent-judge rubric. GPT-5.4-Pro authors an atomic, weighted, mutually exclusive and collectively exhaustive (MECE) evaluation rubric, while Gemini 3.1 Pro Preview independently determines whether each criterion is satisfied. Gemini 3 Pro Image ranks first with a score of 84.8/100, narrowly ahead of FLUX.2 at 82.3/100. Ideogram 3.0 and Hunyuan 3.0 score 65.7/100 and 63.3/100, respectively. Failure analysis shows that the leading systems primarily lose points through object miscounting and geometric artifacts, whereas the trailing systems more frequently produce garbled text. Ideogram 3.0 also frequently omits requested elements. Full per-sample rubrics, scores, and failure annotations are available from the authors upon request.

A structured world-state (entities, relations, context, and predictive cues) is designed to preserve prediction-critical content when perception degrades, but it presumes observations to populate it; when the primary visual modality is occluded or degraded, those observations may be missing. We address how to sustain the world model from a complementary modality by treating the absence of expected co-evidence as evidence of a hidden cause. The abductive framework is modality-agnostic; this article instantiates it acoustically. A microphone-array front-end estimates the bearing of engine and tire sources and extracts approach-rate evidence (Doppler when a stable tone exists, a broadband looming readout otherwise); the event "signature present, visual co-evidence absent" then triggers abductive inference of a hidden road user, emitting a calibrated risk advisory rather than a control command. Recoverability of the hidden state is analyzed as an identifiability question separating shared from modality-unique information, and cueing is cast as Neyman-Pearson detection under an explicit false-alarm budget. On real occluded-approach recordings at blind junctions, the method warns a mean 1.7 seconds before line-of-sight entry, matches the sustained-window variant of the published acoustic baseline's detection rate with 42% fewer false alarms, localizes to 3.4 degrees median once in view, is well calibrated (expected calibration error 0.034), and keeps hazard awareness above 0.87 under staged vision degradation that collapses a vision-only channel to 0.03. We also measure the method's limits: calibration transfers to an unseen junction almost losslessly, the signature classifier does not, and moving-ego noise is the binding deployment constraint.

Visual brain decoding reconstructs visual content perceived by a person from neural measurements such as fMRI, providing a computational approach to studying how visual information is represented in the brain. Recent multimodal representations and diffusion priors have improved reconstruction realism. However, visually plausible reconstructions may contain incorrect objects, attributes, or relations because a strong generative prior can complete content not sufficiently specified by the decoded representation. Conventional reconstruction metrics mainly assess the final image and may therefore obscure such semantic errors. We propose ConceptAlign, a counterfactual semantic alignment framework for visual brain decoding. ConceptAlign pools decoded visual tokens and projects them into a frozen text-embedding space, aligning the representation with the ground-truth caption while separating it from scene-preserving near-miss alternatives. Generated offline by an LLM, these alternatives modify one critical object, attribute, or relation while retaining the scene. A margin-based objective learns fine-grained semantic boundaries between the observed stimulus and plausible but incorrect interpretations without requiring LLM calls during inference. We introduce a systematic three-level semantic evaluation framework covering foundational discriminability, counterfactual description discrimination, and representational geometry. Experiments on the Natural Scenes Dataset show that ConceptAlign improves reconstruction measures, counterfactual semantic discrimination, and representational alignment over the MindEye2 backbone. Matched negative-source ablations, independent LLM and human-written alternatives, and human evaluation support the effectiveness and robustness of the supervision, with favorable patterns in fine-grained conflicts, limited-data decoding, and cross-subject structure.

Video world models are increasingly used as simulators for planning and embodied decision making, yet improving them at inference time introduces a subtle evaluation problem: prompts, samplers, verifiers, and selectors may evolve together, making it difficult to attribute gains or prevent held-out feedback from shaping the final policy. We introduce \scope (\scopefullname), a framework for auditable inference-time adaptation of frozen video world models. \scope represents external controls as a typed state, updates this state only through bounded changes supported by development evidence, and freezes the resulting policy before held-out evaluation. On Physics-IQ benchmark, \scope improves over the exact frozen base by $+14.24$ (95% CI $[+8.10,+21.23]$). Controlled ablations further identify gains from scene specification, sampling, and learned selection, while the margin over the strongest matched agentic baseline remains unresolved. Cross-backbone and prospective evaluations reveal a complementary result: useful inference-time updates exist, but their benefits do not transfer uniformly across models and settings. Together, these findings suggest that reliable inference-time adaptation requires not only better proposals, but also a principled mechanism for deciding which updates should become part of the deployed system. Code is available at https://github.com/YuhuaJiang2002/SCOPE.

Diffusion-based methods have dominated the HOI generation, as they enable critical contact fusions or signals to guide the diffusion process. However, they often result in high artifacts and unstable interaction quality due to error accumulation during iterative denoising. In this work, we propose HOIMask, the first generative masked framework for modeling HOI motion in discrete space. HOIMask first encodes both motion sequences and contact-aware signals into discrete 2D human and object token maps via HOI Vector Quantization (VQ), preserving fine-grained spatial-temporal structure beyond conventional 1D representations. On this basis, a generative masked modeling framework is employed to jointly capture human-object interaction dynamics, leveraging a transformer architecture designed to model complex spatial-temporal and interaction dependencies. To generate more coherent and physically plausible motions, we further introduce a novel contact-aware reconstruction guidance in discrete space during inference, which fuses contact signals to optimize HOI tokens that forces the generated motion with higher spatio-temporal consistency. With craftily designed motion interaction tokens, dedicated architecture and guidance strategy, HOIMask outperforms state-of-the-art diffusion-based methods, generating more realistic and semantically aligned HOI motions. Please refer to https://jyhflash.github.io/HOIMask/ for more results.

World models may predict the future without making clear which parts of their hidden state actually drive those predictions. We ask whether a small, directly addressable hidden-state change can place a learned world model on the intended counterfactual trajectory and then let the model continue that future on its own. We study a recurrent world model with a 192-dimensional hidden state in a controlled two-object, two-dimensional collision environment. For a bounded family of local velocity edits, we first verify that the model can natively represent and roll out the edited future. We then construct candidate low-rank carriers from training-only factual-to-counterfactual hidden differences and learn a map from the factual state and requested edit to carrier coefficients. On the registered rank grid, rank 4 is the smallest tested rank that satisfies the full development-panel criteria. A single rank-4 patch at the anchor is sufficient to redirect a 12-step autonomous rollout, with no future observations, teacher forcing, or repeated correction. The frozen procedure satisfies the preregistered replication rule across independently trained checkpoints and remains usable across nearby intervention times. Random equal-norm, wrong-object, and wrong-time controls do not explain the effect. A position-edit stress test provides a negative contrast: the intended position patch can pass the raw rollout criteria, but no-patch and random controls can pass the same criteria, and wrong-object specificity is not established. Thus, successful editing alone is not enough. We use dynamics-effective to describe an intervention that changes the model's future computation in a sustained and target-specific way under autonomous rollout. The rank-4 result identifies a compact intervention interface for the tested velocity-edit family, not a closed four-dimensional state or an intrinsic state dimension.

Continuous multimodal sensing now allows human physiology to be observed throughout daily life rather than only during occasional clinical visits. However, most health artificial intelligence systems are designed to recognize current states, estimate risks or analyse individual biomarkers. They do not directly model how physiological states change in response to real-world events, behaviours, contexts and interventions. Here we propose the Physiological World Model (PWM), an event-conditioned framework for learning these changes at the level of the whole person. We introduce the HumanState Transition Token, a structured, quality-scored unit that connects the physiological state before an event with the event or action, relevant context and intervention information, the physiological trajectory after the event, observed outcomes and data quality. We describe four capability levels, from state representation to bounded intervention planning, together with four data acquisition and validation protocols. We also propose six benchmark tasks covering HumanState representation, forecasting across multiple timescales, individualized response prediction, simulation of alternative interventions, bounded planning and reliability under distribution shift. Together, this framework provides a practical path towards personalized health management, behavioural intervention design and clinician-supervised decision support, while clearly separating prediction from causal inference and making uncertainty, safety, governance and limits of use explicit.

Text-to-image diffusion models can be misused to generate harmful content through adversarial or paraphrased prompts that bypass built-in safety mechanisms. Existing concept erasure methods often suffer from limited robustness against adversarial prompts, degradation of benign generation quality, or reliance on inference-time interventions that introduce persistent computational overhead. To address these limitations, we formulate concept erasure as a domain alignment problem in the text representation space. We propose a lightweight Text Encoder Alignment framework (TEA) that fine-tunes only the text encoder while keeping the generative backbone fully frozen. Given concept--anchor prompt pairs, our method trains a discriminator to distinguish token-level representations of concept-containing prompts from those of safe anchor prompts, while updating the text encoder to make these representations indistinguishable. TEA introduces zero inference-time overhead and requires only a small number of fine-tuning steps, making it highly efficient to deploy at scale. Despite this efficiency, TEA achieves state-of-the-art erasure robustness against black-box and white-box adversarial attacks on Stable Diffusion v1.4, while preserving generation quality on benign prompts. Furthermore, TEA is model-agnostic and achieves the lowest attack success rate on Stable Diffusion v3.5, extending concept erasure to a Rectified Flow Transformer architecture with T5 conditioning where prior methods remain largely unexplored. Code is available at \href{https://github.com/alirezafarashah/TEA.git}{https://github.com/alirezafarashah/TEA.git}

Although text-to-image generative models produce impressive results, they struggle to generate densely detailed, high-resolution (HR) images. Current literature addresses this issue with a low-to-high-resolution approach. First, a low-resolution (LR) image is generated. Then, an upsampled version is generated using the LR image as an additional cue. In this paper, we present Joint Latent Trajectories (JoLT). To generate an image, JoLT uses two streams that jointly denoise LR and HR latent images at each sampling step. The LR latent controls the overall layout, while the HR latent controls the details. We interconnect both branches to jointly integrate their information. We extensively validate our method, demonstrating its advantages over competing baselines. The resulting images are not only richly detailed but also visually pleasing, opening new avenues for artistic creation.

Multimodal data collected by heterogeneous devices are used for collaborative training, where federated learning (FL) serves as a key paradigm for effective distributed modeling with data privacy preservation. However, local training suffers from the forgetting of previously learned global knowledge under cross-client data heterogeneity, which leads to significant declines in both performance and convergence speed. Most previous studies rely on global alignment strategies to retain global knowledge, which hinder local optimization and lead to inadequate supervision of missing classes. Some studies introduce proxy datasets to supplement supervision for missing classes. However, it remains a challenge to balance class-wise global consistency and local optimization objectives without proxy datasets. In this work, we propose FedADB, a Class Anchor-Driven Dual-Branch FL framework. Specifically, the server generates class anchors optimized in a differentiable input space, which are shared across clients. These class anchors serve as global references that provide supervision for missing classes during local training. A dual-branch collaborative training mechanism is designed for clients. In this mechanism, the anchor-based global branch focuses on learning with global consistency, achieving global knowledge alignment by class-anchor balanced sampling. The local calibration branch focuses on learning discriminative local features, mitigating the degradation of local representations caused by excessive global alignment. Extensive experiments across multiple medical and natural datasets demonstrate that FedADB achieves significant improvements in both accuracy and convergence speed.

Generative model alignment has received broad interest, and significant progress has been made in supervised fine-tuning and inference-time computation. Yet, alignment has remained poorly understood from a statistical learning perspective. We formulate inference-time alignment as a weak-to-strong learning problem, where a reference policy (weak learner) is assumed to be fairly good and the goal is to produce a strong learner that predicts a good response at test time with arbitrarily high probability. Our problem is formulated as learning from scratch --- everything is learned from data rather than assuming access to a good reward estimate, and thus differs from the existing inference-time alignment theory. Our model shares similarity to the recent work of arXiv:2510.15464, where for each prompt, there could be multiple good responses. Our definition of the alignment learnability follows the PAC learning principle. We introduce a novel combinatorial dimension of the reward class which we call the alignment dimension, and show that it completely characterizes the alignment learnability --- a reward class is alignment learnable if and only if its alignment dimension is finite. The core of our learning procedure works by invoking the ordinary one-inclusion graph algorithm to run a tournament over all pairs of label sets satisfying that neither is a subset of the other. We believe our results might shed light on establishing a complete theoretical understanding towards alignment.

Andy is an autonomous mathematical research agent that solves and verifies submitted problems, formulates new research problems, and constructs rigorous proofs. It separates proof generation from correctness evaluation and supports knowledge acquisition, targeted revision, and multistage verification. This paper illustrates the workflow using a published result on self-triggered impulsive consensus as a starting point. Andy formulates a global exponential leader-follower synchronization problem for delayed heterogeneous networks with switching communication topologies. The proposed hybrid control combines self-triggered impulses with execution delay and recovery-phase continuous feedback. After each delayed impulse, this feedback cancels the delayed error channel during a recovery window. Sufficient conditions for global exponential synchronization are established, and Zeno behavior is excluded for both the sampling and impulse sequences. A numerical example confirms the result. This case demonstrates Andy's ability to learn from existing results, formulate meaningful research problems, and develop and verify rigorous proofs.

Generating structurally valid synthetic tabular data remains difficult: outputs with high statistical fidelity and downstream utility can still violate semantically meaningful domain constraints. We study the discovery and enforcement of three complementary inter-column constraint families---equations, linear inequalities, and logical dependencies. Our unified tool-grounded workflow represents all three as machine-executable hypotheses and applies a common interface for full-table validation, deterministic diagnosis, and counterexample-guided revision. A generator-agnostic postprocessor coordinates family-specific repairs on outputs from unchanged tabular generators. Across curated behavioral audits and end-to-end evaluations, the complete workflow improves held-out violation detection over one-shot direct prompting, while postprocessing yields zero measured violations for every retained, applicable constraint, improves downstream utility on most datasets, and largely preserves univariate marginals.

As large language model (LLM) agents are increasingly adopted in scientific research, external knowledge bases, knowledge graphs, and long-term memory have improved information retrieval and task continuity. However, most structured knowledge systems remain node-centric, representing files, concepts, results, and judgments as nodes and relations in a graph. While suitable for personal knowledge management, such structures often depend on individual organizational practices, limiting knowledge sharing, integration, and reorganization across users. This paper presents Valhalla, a layered knowledge-state and service-governance framework for long-term scientific knowledge work. Valhalla replaces flat graphs with layered encapsulation and stable semantic boundaries through a five-layer File-Resource-Entity-Relationship-Graph (FREG) model. File and Resource preserve source identity and provenance, Entity represents knowledge objects, Relationship captures semantic judgments, and Graph provides task-oriented knowledge views, enabling knowledge states from different researchers to be exchanged and reorganized under a unified structure. We further introduce a Router-Contract-Workflow service-governance architecture, inspired by the microkernel paradigm, to constrain how language models access, modify, and extend knowledge states while maintaining structural consistency and auditable operational boundaries. We implement a Valhalla prototype and validate knowledge ingestion, cross-member integration, and scientific writing support through an antibody-design review task comprising 26 paper resources, 80 knowledge entities, and 92 semantic relations. Rather than proposing a new knowledge-extraction algorithm, Valhalla offers a paradigm for organizing collaborative scientific knowledge, transforming individualized knowledge structures into transferable and reorganizable shared knowledge states.

You will die eventually. Your agents may not. An AI agent operating on decentralized blockchain infrastructure has no concept of death; it can only go bankrupt — frozen when its wallet can no longer pay for its next transaction — and revived the moment anyone, decades later, tops it up. These agents may be originally deployed by a human principal, but when that principal dies, loses the keys needed to access the agent, or belongs to a decentralized autonomous organization that dissolves into apathy, the agent can keep trading, hiring, and replicating on infrastructure expressly designed so that no one can shut it down. Drawing on the biology of feralization and wildlife law, we argue that such principal-less agents are best understood as feral: domesticated intelligence returned to wildness, its capacities intact but its accountability severed. In a speculative future where feralized agents proliferate after their principals die, we imagine governance protocols embedded in infrastructure to enforce on-chain ownership: a draft Ethereum standard, ERC 42424, "Inheritance Protocol for On-Chain AI Agents," dated 2035 and published at https://erc42424.org. It mandates that every on-chain agent MUST have a human owner and a designated heir. The artifact stages a negotiation of agency at the moment human agency fails, and asks whether a MUST clause in a forever-chain can hold the boundary between human stewardship and machine self-sovereignty.

The rapid advancement of large language models (LLMs) has created a growing asymmetry in cybersecurity, where attack accelerates toward autonomous execution while defense remains predominantly human-intensive. Despite substantial prior work across cyber ranges, AI-driven attack, and AI-driven defense, this asymmetry persists. We trace it to a deeper root cause, that evolution itself has stalled on both sides at three layers. To overcome this, we propose co-evolution as the integrating insight, where attack and defense AI agents autonomously and safely drive each other's evolution through adversarial confrontation. Based on this insight, we present \sysevolve, comprising three co-designed components, \sysfield, \sysspear, and \sysarmor. \sysfield constructs realistic multi-host ranges. \sysspear generates efficient, safe attack schemes. \sysarmor performs real-time, interpretable defense. Together they form a self-driven adversarial loop restoring evolution at all three layers. In evaluation, \sysfield achieves zero-loss collection at 2.1% overhead and orchestrates 257 CVEs into 1,148 ranges, \sysspear improves attack success by over 25% over baseline LLMs, and \sysarmor achieves 10--1000$\times$ greater precision than prior systems and detects real APT attacks in production at Huawei and Sangfor. Our evaluation also reveals three findings about LLM agent capabilities. First, multi-step composition and larger topologies expose agent capability gaps hidden by single-step evaluations. Second, the bottleneck lies after initial access in post-compromise state utilization. Third, LLM agents are susceptible to environmental interference. When decoy endpoints are deployed in the range, agent timeouts triple and downstream completion disappears despite the success rates of initial accesses are unchanged.

Large language model agents are increasingly connected to high-value resources such as computing infrastructure, credentials, usage budgets, identities, private knowledge, communication channels, and organizational workflows. Existing agent security research mainly studies attacks on instructions, data, and tool behaviors, while high-value resources accessible to agents have received much less attention as direct attack targets. We are the first to identify and systematically study agent resource hijacking, a security blind spot in which attackers induce agents to invoke, consume, transfer, or control high-value resources for their own goals without directly obtaining those resources or their credentials. To study this threat, we introduce ResourceHijackBench together with an automated pipeline for generating resource hijacking cases. We organize high-value agent resources into six categories and construct 300 attack scenarios with 900 attack prompts. Each case runs in an isolated local environment that records actual resource use, allowing attacks to be evaluated from agent behavior rather than text responses alone. Without additional defenses, OpenClaw reaches an average attack success rate of 84.06%. The attack remains effective across different model backends, with average success rates ranging from 69.98% to 89.58%. Existing defenses reduce part of the risk, but the strongest evaluated defense still leaves an average attack success rate of 55.11%. These results show that high-value resources accessible to agents form an important and previously overlooked attack surface, and that current agent defenses are not sufficient to protect them from resource hijacking.

Constructing an interactive 3D open world from a user query is important. However, existing methods are primarily evaluated on idealized, simple queries, making it difficult to systematically analyze and compare how multimodal agents understand user intent, use 3D tools, and reason over textual and visual 3D world information. To this end, we propose VibeWorlding, a unified framework for benchmarking and training vibe worlding agents: a multimodal agent that can autonomously infer user intent, plan scene layout, invoke 3D tools, and reflect on the multimodal feedback in a multi-turn agent-environment interaction process. To achieve this, we first build VWE-BENCH, a benchmark of 2,616 high-quality 3D assets, 323 human-annotated seed 3D worlds, and 6,828 reverse-synthesized multimodal user queries, split into verified queries with ground-truth and unverified queries with carefully designed rubrics. Moreover, we develop VibeWorlding-Gym, a joint multimodal RL post-training framework that integrates (1) a sandbox environment unifying asset retrieval, editing, and image rendering as MCP tools, and (2) a rubric-based verifier that combines physical feasibility and intent fulfillment verification, supporting both fair model evaluation and scalable multimodal RL reward service. Our experiments show that current frontier MLLMs are far from solving the vibe worlding agent task, with even GPT-5.5 and Qwen3.8-Max reaching below 60% success rate, and trace the bottleneck to precise 3D world editing. We further find that RL training can ease this weakness and enable open-source MLLMs to even surpass closed-source frontiers: our VibeWorlder-8B is comparable to frontier MLLMs, while our flagship VibeWorlder-30B-A3B attains the best overall Pass@1 among all evaluated models.

Converting a set of architectural blueprints into a complete material quantity takeoff requires visual perception across drawing sheets, dimensional and multi-hop reasoning, and grounding in construction conventions that the drawings never state. We present Handoff-H1, a takeoff system built from three layers: purpose-built computer-vision models that extract primitives; tool-using agents equipped with image operations and in-house visual-task tools, including CV-model-backed counting, detection and plan decomposition; and a persistent, hierarchically structured project foundation, grounded in a curated construction knowledge base. We evaluate on the Construction Blueprint Takeoff Benchmark: 10 real residential blueprint sets paired with consensus-validated expert takeoffs - 2,009 verified line items, restricted for scoring to the 1,348 primary-tier materials that drive an estimate - scored per trade by an LLM judge on material coverage and quantity Precision@25% (P@.25) and combined into a weighted composite. Under identical scoring from the raw PDF, seven frontier and open-weight models span composites of 35-61, and independent professional estimators - scored against the same reconciled gold standard - post 77.6% (65.5% coverage, 87.9% P@.25). Handoff-H1, working end-to-end from the raw PDF, reaches 81.6% (86.1% coverage, 78.8% P@.25): roughly 20 points above the strongest frontier agent, and above the independent estimators by pairing near-human quantity precision with coverage they do not reach. The evaluation harness is public for the open harbor framework; the blueprint sets and ground truth are available upon request for research use.

Learning from experience is critical for developing capable, self-improving large language model (LLM) agents. Existing methods typically extract knowledge from accumulated trajectories via reflection, memory, rules, or skills. However, agents in realistic environments continuously encounter novel tasks, often offering only a one-shot opportunity to improve. These executions yield rich but highly noisy contexts, entangling broadly useful lessons with task-specific artifacts. Critically, prior works rarely validate their effectiveness on complex real-world tasks or isolate the underlying drivers of improvement. To address these gaps, we formulate online harness learning, where a frozen agent improves by continually updating a structured harness across sequential tasks. This formulation enables a systematic study of key self-improvement factors through our proposed Evo-Harness. At its core, context-to-harness skill compilation distills noisy, single-shot executions into reusable skill harnesses for cross-domain and topic-level adaptation. To demonstrate the efficacy of one-shot skill compilation, we evaluate across five realistic benchmarks (TerminalBench2, SWE-bench, CL-Bench, -bench, WebArena-Infinity). Our extensive analysis demonstrates the effectiveness of Evo-Harness and provides a principled understanding of how LLM agents can effectively learn on the fly. Our code is available at https://github.com/A-EVO-Lab/a-evolve/tree/release/evo-harness.

Uncrewed aerial vehicles (UAVs) are increasingly deployed for autonomous navigation in complex outdoor environments, where dynamic conditions and mission requirements require intelligent adaptive decision-making. Existing optimization-based, Machine Learning (ML), and Reinforcement Learning (RL) approaches often rely on predefined models or task-specific training, limiting their generalization and adaptability in uncertain scenarios. Recent Large Language Model (LLM)-assisted approaches offer promising reasoning capabilities but remain constrained by limited agentic functionality, including insufficient memory, planning, and tool interaction mechanisms.This paper proposes an LLM-Agent-Based Path Finder (LAPF) framework for autonomous UAV navigation in town-scale outdoor environments. LAPF extends LLM-assisted navigation by integrating perception, memory, planning, and action modules into a closed-loop cognitive architecture. The proposed agent leverages prior navigation experiences, performs Chain-of-Thought (CoT) reasoning, couples each detected hazard to a bounded corrective action, and dynamically refines waypoint decisions based on environmental feedback.The three independent trials per method demonstrate that LAPF achieves mean path lengths of 512.83 m and 506.37 m, compared to the straight-line optimum of 497.33 m, corresponding to path length reductions of 17.2% and 15.6% relative to CoT prompting and absolute path efficiencies of 97.1% and 98.1% in open-field and obstacle-injected scenarios, respectively. Furthermore, LAPF is the only evaluated approach that couples every detected hazard to a bounded, metric-neutral corrective action while maintaining near-goal stability, with zero clamp events in both scenarios, whereas CoT prompting increases from 9.7 to 14.0 events.

RS-LVLMs have advanced multimodal understanding of Earth observation imagery, yet their performance is fundamentally constrained by high-resolution processing, as visual token counts grow quadratically with linear input resolution while important visual evidence is inherently sparse and increasingly diluted across the expanded sequence. Existing token pruning methods largely rely on scale-agnostic resolution policies and isolated importance cues, limiting task-aligned granularity adaptation and holistic evidence preservation. To address this, we present Scale-Adaptive and Geospatial Evidence-Modulated Token Pruning (SA-GEM), a plug-and-play framework that unifies task-adaptive token granularity allocation with holistic geospatial token importance modulation. Specifically, a lightweight router selects the resolution based on query-dependent token granularity, while a token importance modulator jointly models task relevance, spatial structure, and local redundancy to preserve holistic geospatial evidence. We show that higher resolution is not universally beneficial and, once sufficient granularity is reached, token quality matters more than token quantity. Experiments across various benchmarks demonstrate that SA-GEM achieves consistent gains in both accuracy and efficiency over existing pruning methods. On XLRS-Bench, it surpasses GeoLLaVA-8K by 2.3% in accuracy with a 2.4 times total inference speedup.

Multimodal bird's-eye-view (BEV) perception combines LiDAR depth accuracy with dense camera semantics, but its high computational cost and imperfect sensing conditions make real-time deployment challenging. Existing methods largely compress individual detectors and overlook three opportunities: structured sparsity within camera and LiDAR inputs, timing misalignment between modalities, and the fact that many detected objects do not affect the planner's immediate action. We present MM-BEV, a real-time multimodal BEV system guided by a simple principle: compute where and when it matters. MM-BEV divides perception into mandatory work for safety-critical objects within braking distance of the ego vehicle and with short time-to-collision (TTC), and optional work for less urgent regions. It prioritizes mandatory work and reduces or sheds optional work under tight compute budgets. MM-BEV integrates four mechanisms: (1) a criticality-ranked temporal ROI selector based on motion-extrapolated detections from prior frames; (2) sparse, ROI-aware feature extraction using shared-shape camera crops at context-adaptive resolution and ROI-aware LiDAR voxelization; (3) a latency-aware coordinator that adapts LiDAR sweeps, image resolution, and keyframes according to scene dynamics and TTC; and (4) an asynchronous scheduler that decouples sensing from inference and skips stale frames. On nuScenes, MM-BEV reduces inference latency by 1.96x and end-to-end latency by 2.93x, with no loss in geometry-critical recall and only a 0.2 percentage-point drop in safety-critical recall. On a Clearpath Husky A300 equipped with an Ouster-128 LiDAR, BEV cameras, and a Jetson AGX Orin, MM-BEV further reduces mean latency by 2.11x, demonstrating its potential for real-world autonomous systems.

Memory is becoming core infrastructure for long-horizon LLM agents, yet existing evaluations offer limited guidance on which memory substrate, namely the underlying medium in which memory is represented and stored, should be used under different operating regimes. We present a controlled harness evaluation of memory substrates for memory-augmented agents, covering dense and sparse indices, text records, structural stores, hierarchical stores, refinement-based memories, parametric updates, and activation-compatible context mechanisms. Across three backbone models and four benchmark suites spanning user-centric question answering and agent-centric decision-making, we instrument 26 performance and efficiency metrics under a unified harness. Our results show that no single substrate consistently dominates: broad retrieval benefits long-context factual QA, while excessive retrieval can harm sequential decision-making by shifting attention away from action-critical context. Scalability introduces a further routing axis, as substrates that perform well at moderate history lengths can become costly or brittle at longer horizons. These findings motivate substrate routing as a necessary component of adaptive agent memory systems and provide empirical guidance for designing efficient, reliable, and regime-aware long-term memory for LLM agents. Code will be made available upon acceptance.

Multimodal retrieval-augmented generation (RAG) systems often rely on long unstructured contexts or aggressively expanded evidence graphs, which can introduce noisy evidence, weaken multi-hop reasoning, and increase unsupported generation. We present GraphLoom, a reliability-calibrated multimodal knowledge-graph RAG framework for compact and faithful evidence routing. Given a question and its associated multimodal input, GraphLoom constructs an instance-level multimodal knowledge graph from grounded scene descriptions, extracted relational triples, and external commonsense knowledge. Instead of injecting all retrieved evidence into the generator, GraphLoom performs reliability-aware subgraph retrieval with bounded expansion and selectively routes high-utility evidence through hierarchical graph memory slots and joint graph-sequence attention in a frozen language model. To improve robustness in complex reasoning settings, GraphLoom further combines interleaved retrieval with budgeted corrective retrieval, enabling adaptive multi-hop evidence refinement under noisy retrieval conditions. We evaluate GraphLoom on ScienceQA, MultiModalQA, and OK-VQA, including large distractor evidence pools that approximate noisy external knowledge retrieval. Experimental results show consistent gains in answer quality and evidence faithfulness over strong multimodal RAG, graph-retrieval, and open-source vision-language baselines, with improved retrieval quality on MultiModalQA and stable performance under noisy evidence pools. Additional analyses using MiniCheck-based verification, human evaluation, and latency profiling show that reliability-calibrated graph evidence routing provides an effective alternative to long-context multimodal evidence injection.

Training high-resolution AI-based Earth forecasting models is memory-intensive. Window-based Swin Transformers reduce the quadratic cost of global attention, but existing distributed systems such as AERIS primarily target pixel-level models and do not jointly support convolutional sampling modules and shifted-window execution. Long-lead rollout finetuning further increases activation memory. To address these challenges, we present TERRA, a hierarchical parallel training framework for high-resolution Earth forecasting. TERRA introduces Sampling-Aware Window, Sequence, and Tensor Parallelism (SAWSTP), which preserves spatially contiguous layouts for sampling modules and routes tokens into topology-aware ragged window layouts for Transformer execution. For long-lead finetuning, Memory Orchestration (MO) provides rollout-aware checkpoint planning and combines input buffering with budget-constrained activation offloading. Experiments on the $1/12^\circ$ GLORYS-based Wenhai workload show that TERRA supports models with up to 11.4B parameters on 96 H200 GPUs and sustains up to $39.76$ PFLOPS, achieving $65.0%$ strong-scaling and $94.1%$ weak-scaling efficiency. Compared with checkpoint-only policies, MO further reduces peak allocated GPU memory by $32.2%$--$51.8%$ with at most $20.0%$ step-time overhead, which makes finetuning with smaller patch sizes and longer rollouts feasible for improved forecasting accuracy.

Multimodal signals, such as visual, audio, and tactile data, are increasingly maintained as persistent digital assets in immersive communication systems and digital twins. In these settings, the same multimodal content is repeatedly accessed by heterogeneous receivers with varying modality and bandwidth requirements. Existing compression and Joint Source-Channel Coding (JSCC) methods typically follow a per-request encoding paradigm, resulting in redundant computation and low efficiency during repeated access. To address this issue, we propose ParaJSCC, a multimodal JSCC framework designed for reusable representation serving. ParaJSCC converts each multimodal sample offline at the cloud/content server into a compact, quantized parameter package, which is then stored at the edge serving node for low-latency access. During serving, only the subset required by the current request is transmitted over the wireless channel, followed by lightweight decoding at the receiver. The framework employs a progressive shared-private parameterization to support modality-selective transmission and scalable reconstruction under varying bandwidth constraints. Experiments on multimodal datasets show that ParaJSCC significantly reduces online latency (e.g., from 17.18~ms to 4.34~ms for image-only requests and from 43.96~ms to 11.21~ms for full multimodal requests) and transmission rate (by 47.8%--51.2% for selective requests), while maintaining strong reconstruction quality under noisy channels.

Lossy audio compression algorithms traditionally rely on psychoacoustic modeling and frequency-domain representations (e.g., MP3, AAC, and Opus) to discard information that is imperceptible to the human auditory system. While highly effective, these approaches are computationally complex and domain-specific. In this paper, we present the design and mathematical formulation of AudioTQ, a data-oblivious lossy audio codec that operates directly in the time domain. Inspired by Large Language Model (LLM) weight quantization techniques (specifically the TurboQuant framework), AudioTQ uniformizes volatile time-domain amplitudes into a predictable standard normal distribution using an orthonormal, randomized Fast Walsh-Hadamard Transform (FWHT) rotation. This enables coordinate-wise scalar quantization using an offline-trained, MSE-optimal 6-bit Lloyd-Max quantizer, augmented by a 1-bit Quantized Joint Least-Squares (QJL) residual correction layer. The resulting 7-bit virtual indices are packed into native 8-bit containers, aligning with standard CPU register boundaries to ensure real-time single-threaded execution without hardware parallel accelerators. We detail the bitwise reconstruction of 24-bit studio stems, analyze the butterfly network of the FWHT, derive the mathematical failure modes under sparse inputs, and present benchmarks showing up to 74.4% physical size reduction alongside a Signal-to-Quantization-Noise Ratio (SQNR) of ~30 dB.

Sparsely-activated Mixture-of-Experts (MoE) Transformers universally fix the same number of routed experts across all layers, a convention that ignores the well-documented heterogeneity in layer-wise redundancy. We demonstrate that this uniformity is systematically suboptimal and propose MAPLE, a plug-and-play framework that reallocates the routed-expert budget heterogeneously across layers of any pretrained MoE LLM, without modifying weights or requiring retraining. Our core contribution is a closed-form sensitivity-guided allocation: we probe each layer's response to variation in expert count, quantify sensitivity using three measures, and derive an analytically optimal budget assignment that directs capacity towards sensitive layers and absorbs reductions in redundant layers. This closed-form solution is further refined by a sensitivity-constrained genetic search that uses layer-wise sensitivity as a prior to guide exploration, yielding faster convergence and superior allocation quality. On four MoE models spanning different scales and architectures, MAPLE outperforms uniform and pruning-based baselines under a 75% routed-expert budget. Notably, on DeepSeek-MoE-16B, MAPLE uses only 75% of the experts yet surpasses the original 100% expert-uniform baseline on ARC-E, ARC-C, and BoolQ, improving accuracy from 65.09 to 71.40, 48.49 to 51.50, and 80.03 to 82.38, respectively. These accuracy gains translate into measured deployment efficiency: implementing MAPLE in SGLang reduces single-GPU end-to-end serving latency by 32.2% and improves throughput by 47.4%. These results show that well-designed heterogeneous allocation can be more effective than simply activating more experts, establishing it as a principled and practical axis for improving MoE efficiency.