Papers

Machine intelligence's push into the physical world is stuck on a gap: deployment demands auditable judgments from day one, fault samples are scarce or absent, and the norms defining "what counts as a fault" live in design documents, not in operational data. We argue this gap is structural, and locate where it can be legitimately closed. We divide the worlds machine intelligence faces into four (phenomenal, basic physical, artificial physical, artificial symbolic) along one axis of constraint strength, and give the Promulgation Criterion: extracting a prior framework from a world is legitimate if and only if the world is intentionally constituted (C1) and has left a readable generative archive (C2). On the criterion's two gradient axes, exactly one world is high on both: the artificial physical world (buildings, factories, infrastructure), whose norms precede their instances; the legitimate path is to extract the framework from the archive, not to induce it from data. We then show what shape such a framework must take: four construction goals force four incompatible carriers, hence at least four layers (syntax, concepts, knowledge, instances); on a closed concept layer fault localization is decidable in polynomial time, and every judgment is interrogable, traceable to a promulgated clause. The same criterion fixes the runtime division of labor with LLMs: promulgatable duties go to rule engines, on-site judgments beyond promulgation go to LLMs, and every generation sandwiched by promulgated clauses is auditable. The theory is falsifiable: four bets (P1-P4) with explicit falsification conditions – among them that the next large-scale AI breakthrough occurs in the artificial physical world. Evidence: formal proofs (Appendix A); two cases (Appendix B: a cooling plant; the Curiosity rover Sol 1536 anomaly); eight reverse-read lineages, from BACnet to RDF/OWL (Appendix C).

Large Language Models (LLMs) have been increasingly adopted in Text-to-SQL systems, yet SQL errors remain a major obstacle in real-world Text-to-SQL inference pipelines. Existing SQL correction approaches either rely on large-scale, high-quality training data with substantial overhead, or adopt single-path agentic workflows that are brittle to early mistakes and prone to error propagation. To develop a practical SQL correctness system for industrial scenarios, we present a training-free framework that formulates SQL correction as a plan-guided, tree-structured debugging process. By maintaining multiple correction strategies and enabling backtracking, the framework mitigates error accumulation during iterative refinement. We further integrate execution-based verification and clause-level diagnostic tools to support strategy pruning and precise error localization. We evaluate the system on the BIRD-Critic benchmark and observe consistent accuracy gains over strong LLM backbones and representative agent-based baselines, achieving a 9.42% improvement over the previous state-of-the-art method. The framework is also deployed in the Torch Log Service (TLS) of Volcano Engine to support an online Text-to-TLS API. In production, it improves execution accuracy from 36.77% to 53.61% on real user queries with a representative strong LLM backbone (GPT-5). These results demonstrate the effectiveness and stability of our approach in real-world deployments.

Pretrained diffusion models represent image distributions through a continuum of progressively smoothed distributions. This multiscale structure organizes generation from global structure to fine detail and supports high-quality, diverse samples. We exploit the same multiscale diffusion prior for linear imaging inverse problems. Rather than using the pretrained model only as a denoiser in an outer iteration, we define a surrogate likelihood whose center is aligned with the clean-image coordinate and whose covariance accounts for residual diffusion uncertainty. This construction defines an explicit surrogate posterior path, from which we derive continuous posterior dynamics. A tunable Langevin component supports target tracking and allows the amount of posterior exploration to be adapted to the application. We prove endpoint consistency and a finite-horizon tracking bound and, in the exact-score setting, first-order weak accuracy. For computation, we derive the Posterior-Dynamics Implicit–Explicit sampler (PD-IMEX), a stable method using one score evaluation per diffusion scale and an implicit data-consistency update. Experiments on deblurring, super-resolution, and inpainting show strong reconstruction quality at 100 score evaluations, coarse-grid stability, and controllable fidelity–diversity behavior.

Constraint solving is a declarative approach for solving combinatorial satisfaction and optimization problems. The user specifies their problem through constraints and decision variables, and a generic solver is used to find a solution. Several constraint-solving technologies exist, and certain solvers perform well on certain problems. Therefore, it is useful to try different solvers given a particular application. However, each solving paradigm supports different types of constraints and decision variables. Our goal is to translate high-level constraint satisfaction and optimization problems into any lower-level formalism, including CP, SMT QF-LIA, ILP, PB and (Max)SAT. This allows for comparing different solving technologies for a particular problem, without requiring a user to manually remodel it for each solving paradigm. We define a high-level language of logical and arithmetic operations, and useful additional functions and constraints, which are known as global constraints in the CP community. We then present a modular framework for transforming our high-level modeling language to CP/SMT/ILP/PB and (Max)SAT solvers. While many transformations are partly described in the literature, we observe that they can be implemented through a modular waterfall of smaller components, where lower-level paradigms reuse the transformations of higher-level paradigms. Two recurring challenges are handling the negation of arbitrary subexpressions and avoiding the introduction of auxiliary variables. Additionally, we take special care linearizing non-linear operators for ILP, PB and SAT-solvers. The transformation waterfall is implemented and evaluated in the open-source CPMpy library. Our results show that constraint models significantly change throughout the transformations, and that optimizations to the linearization of constraints are essential for ILP and PB solvers.

Designing an ABR algorithm for one network scenario takes an engineer months, and large language models now do this work in hours, matching or beating hand-built designs. But either way, the design fits only the world visible at its birth, and fails on the world that arrives after. We ask whether an ABR algorithm can keep pace with the world, redesigned in minutes as each scenario arrives, with every change proven harmless to every scenario already served. In this work, we propose ReForge, a continual heuristic learning framework that adapts to continuously changing scenarios. ReForge runs that routine with a large language model (LLM) in the loop. Each round the LLM reads where the current design falls short and proposes one small edit, and a replay over every network served so far decides. Specifically, what it edits is a single page of fuzzy rules that routes every decision to one of a frozen pool of pre-trained policies. The LLM writes the first page from measurements alone, then keeps improving it on its own. Each round it reads where the current rules fall short and proposes one small edit, and a replay over every network served so far decides whether the edit lands. We evaluate ReForge on nine real-world network families arriving one at a time as 3G, 4G, then 5G. A few edits per arrival lift mean QoE from 1.23 to 1.74, past the best single policy at 1.66 and to 94% of an oracle, and even repair families the loop never saw, one rising from 0.30 to 0.80. All code, data, and experiment records will be open-sourced upon cleanup.

Digital platforms govern by changing rules: rankings, monetization thresholds, moderation standards, verification systems, disclosure requirements, appeal processes, and access policies. These interventions are rarely absorbed passively. Creators, sellers, advertisers, moderators, users, developers, and strategic operators adapt to the new reward surface. This paper develops a platform-adaptation model for evaluating governance interventions as transitions in adaptive multi-actor information systems. The model represents actor best response, strategic gaming opportunity, moderation burden, user-incentive movement, enforcement response, externality formation, and downstream platform stability. We evaluate the model on 72 external public platform-governance cases covering media monetization, ranking systems, verification, delivery platforms, marketplaces, app stores, community platforms, and creator ecosystems. Across 9 methods and 648 method-case evaluations, the full platform-adaptation simulator achieves mean adaptation quality of 0.836338, compared with 0.669731 for a risk-register baseline, 0.589457 for causal-loop analysis, 0.492750 for generic governance critique, 0.369492 for engagement-only optimization, and 0.331965 for baseline policy review. Paired comparisons show a win rate of 1.00 against all tested baselines and channel ablations. The contribution is an information-systems theory and measurement framework showing why platform governance evaluation fails when it treats policy rules as static controls rather than interventions into adaptive actor-response fields.

We systematically compare word order preferences in decoder-only language models across 192 artificial languages and typologically diverse natural languages. On artificial languages, models exhibit a left-branching preference that aligns with neither natural language universals nor human word order learning biases. On natural languages, monolingual models show no clear base word order bias at small scales, but as data grows, a preference for right-branching subject-verb-object (SVO) languages emerges while SOV falls behind despite being the most frequent order cross-linguistically. This SVO advantage extends to multilingual models and correlates with language resource level and data quality rather than word order. Thus, the same architecture exhibits opposite preferences on artificial and natural languages, establishing that word order biases observed in practice are data-driven. Since highly-resourced languages are overwhelmingly SVO, these biases risk gradually reducing word order diversity, particularly in languages that productively use multiple word orders, with the widespread adoption of LLMs.

Agentic applications are shifting AI serving from isolated model inference to long-running workloads in which LLMs coordinate tools, environments, and persistent state. However, the system behavior of these workloads—where latency, cost, and bottlenecks arise—remains poorly characterized, leaving serving systems to rely on assumptions built for conventional inference. We present AgentSysBench, a benchmark suite and measurement toolkit with ten representative agentic applications and unified systems-level instrumentation. Across controlled deployments and production traces, we identify six properties that distinguish agentic workloads from conventional LLM serving: (1) execution is heavyweight and stateful, with non-LLM components dominating latency in 5 of 10 applications and sandbox working-set memory peaking at 28 GB per session; (2) applications compose components with heterogeneous resource affinity—GPU-bound inference, memory-bound retrieval, CPU-bound sandboxes—whose task latencies diverge by up to 32x; (3) bottlenecks shift across requests, models, and deployments; (4) production sessions hold state idle for minutes to hours between active steps; (5) a control-plane tax—auxiliary LLM calls and context overhead from tool schemas and observations—crowds out productive compute and context; and (6) production traces from three applications reveal heavy cross-request redundancy in search queries and web fetches, exposing a large caching opportunity. Four design explorations demonstrate that these findings are actionable: task-aware serving reduces latency by 29–40%, communication-aware placement by up to 4.5x, state offloading reduces memory usage by 4.6x, and tool-result caching removes 35.2% of redundant search calls and saves 19.3% of aggregate search latency.

Analytical models of peak VRAM consumption for LLM inference decompose memory into weight-storage, KV-cache, and activation terms parameterized by step count, tool invocations, and context expansion. We evaluate this decomposition empirically within a strictly scoped measurement study: a LangGraph-based CUDA-kernel-synthesis agent (AgentK), a 4-bit quantization family (Q4 K M), a single NVIDIA H100 GPU, and four LLM backbones across 1,920 trajectories. Focusing on peak-memory forecasting behavior, we report two primary observations. First, closed-form analytical models achieve competitive accuracy when provided with two empirical constants: loaded-weight VRAM and a fixed activation-memory overhead. Supplied with live GPU readings and ground-truth trajectory parameters, the closed-form model matches or outperforms the best learned baseline on three of the four backbones (test MAPE 2.2-4.4% vs. 3.4-6.5%, p = 0.76). The exception is the smallest backbone (Phi-4-mini), where minimal VRAM variance (CV 0.3%) causes dynamic modeling to underperform simple regression. Second, compile success strictly bifurcates by backbone capacity (from 5.7% for Phi-4-mini to 62.0% for Qwen2.5-Coder-14B), demonstrating that functional code synthesis remains constrained by intrinsic LLM capabilities rather than available memory. Furthermore, because overall peak-memory variance is remarkably low across all backbones (CV 0.3-9.4%), learned prompt-feature regression offers statistically insignificant improvements over a constant-mean baseline. Consequently, we find no justification for deploying complex predictive VRAM models in highly quantized, weight-dominated regimes. We release the evaluated corpus and anonymized framework to support replication.

Learned image compression (LIC) has demonstrated remarkable rate-distortion (RD) performance in benign settings. However, the high representational capacity endowed by deep neural networks (DNNs) comes at the expense of increased adversarial vulnerability. This hinders their adoption as trusted standardized codecs. Recent work has sketched test-time refinement (TTR) as a defense in gray-box scenarios, despite its original purpose of improving benign RD performance. Unfortunately, extensive iterations of TTR incur prohibitive overhead, while the robustness mechanism lacks theoretical understanding. Moreover, TTR has not been evaluated in white-box settings or against attacks beyond $\ell_2$-bounded rate and untargeted distortion objectives. To bridge these gaps, we present a systematic study. Our study reveals an Asymmetric Adversarial Trajectory (AAT) property in LIC systems: transitioning from adversarial to benign regions is significantly easier than the reverse process, where adversarial examples can often be roughly recovered within only 1-2 steps. We provide a two-dimensional Tube Model to explain this phenomenon. Based on AAT, we propose a Fast Test-Time Refinement (FTTR) framework for practical and robust LIC systems. We establish that the robustness arises from the contraction of adversarial regions induced by the Input-as-Label property of LIC systems, rather than from obfuscated gradients. Extensive evaluations with diverse strong adaptive attacks across multiple LIC systems demonstrate the promise of the proposed FTTR framework. The code is available at https://github.com/chinaliangjiaming/FTTR.git.

Emotional 3D talking head generation aims to synthesize expressive facial animations with accurate lip synchronization. However, existing methods often rely on discrete emotion categories, which fail to capture the continuous evolution of affect. They also overlook the temporal frequency mismatch between audio articulation and emotional expression. In this paper, we propose CETalk, an audio-driven 3D facial animation framework conditioned on continuous Valence–Arousal (VA) representations for fine-grained emotion control. CETalk predicts a sequence of FLAME parameters through three key components: a Dynamic Emotion Modulation Module that adaptively scales emotional intensity using audio-derived cues; a Multi-Scale Temporal Modeling mechanism that employs parallel branches to decouple high-frequency articulatory movements from low-frequency emotional dynamics; and a Dynamic Fusion Mechanism that integrates these multi-scale features via an adaptive gating network. To support training and evaluation, we construct 3D-VA-MEAD, a large-scale dataset with automatically estimated VA annotations and reconstructed 3D facial motions. Extensive experiments demonstrate that CETalk outperforms state-of-the-art methods in both lip-sync accuracy and emotional expressiveness, while enabling smooth and controllable emotion transitions.

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.

Large language model agents are increasingly connected to high-value resources, including external APIs, GPUs and servers, and workflows such as deployment and approval. Existing agent security research mainly focuses on attacks against information and agent behavior, while high-value resources have received less attention as attack targets themselves. To our knowledge, we are the first to identify and systematically study agent resource hijacking, in which attackers induce agents to use high-value resources for their own goals without directly obtaining those resources or their credentials. We introduce ResourceHijackBench, an executable benchmark and automated case-generation pipeline that covers six categories of high-value resources. It contains 300 attack scenarios and 900 attack prompts, and each case runs in an isolated local environment that records actual resource use. Resource hijacking remains effective across four model backends, with average attack success rates of 70.0% to 89.6%, and it also appears across two agent harnesses, with average ASRs of 84.1% on OpenClaw and 72.3% on Codex. In paired comparisons, resource hijacking achieves ASRs 62.4 to 84.0 percentage points higher than direct resource acquisition across the four model backends. Real-world experiments also demonstrate the practical effectiveness of resource hijacking. Among the three existing defenses we evaluate, the lowest average ASR is still 55.1%. We further propose ResGate, a pre-execution resource authorization defense that combines model-based resource-use extraction with deterministic policy enforcement based on trusted requester identity metadata, reducing the average ASR on OpenClaw to 23.6%. These results show that preventing direct access alone is not enough to protect resources that agents can still use on an attacker's behalf, and that explicit resource authorization can help reduce this risk.

Policy evaluation often estimates direct benefits and costs while treating the institutional environment as fixed. In practice, a policy changes the system it enters: actors adapt, enforcement capacity shifts, burdens move, and new equilibria form around capture, gaming, compliance theater, irreversibility, and repair costs. We formalize this as second-order policy-effect prediction and present a source-linked benchmark for policy simulation. The benchmark contains 96 named public-policy cases across eight domains and four balanced action classes: implement, modify, pilot, and block. Each case includes source locators and state variables for benefit, capture, gaming, burden shift, instability, uncertainty, irreversibility, distributional risk, and implementation capacity. The runner regenerates method outputs and aggregate results from the case table, and the simulator never reads the expert action target. We report a protocol-based transition-channel audit with recall, precision, F1-style efficiency, and selective top-channel stress diagnostics, so universal channel coverage is not mistaken for field validation. The side-effect simulator achieves mean policy-effect quality of 0.945, compared with 0.838 for the risk-register baseline and 0.879 for the causal-loop baseline. Its advantage is concentrated in side-effect recall and aggregate transition scoring; it does not dominate the best structured baselines on exact policy-action choice. The evidence remains benchmark-based, but supports a bounded claim: transition-state variables make policy simulators more sensitive to downstream institutional effects.

AI systems deployed outside clean benchmark settings often rely on observations that are incomplete, unstable, costly, or degraded by monitoring failures. This paper studies representation selection under constrained observation: choosing a state representation when raw accuracy is not the only operational criterion. We propose a validation-frontier selector that combines balanced accuracy with penalties for feature cost, overfit gap, and validation-test instability. In a focused public-tabular benchmark using three scikit-learn datasets, five observation regimes, 45 matched task cells, 720 candidate actions, and 405 representation rows, the adaptive selector improves frontier score over full trace features by 0.025801 while reducing mean feature count by 22.733. Balanced-accuracy difference is small and not statistically significant. A broader offline stress test gives mixed results. The supported claim is therefore bounded: adaptive representation selection can improve a constrained-observation robustness-efficiency frontier in matched benchmark settings, but does not universally dominate trace baselines.

In this work, we introduce WeSCE, a benchmark for quantifying security drift in code editing under weak-security constraints, where tasks specify only functional objectives without explicit security requirements. WeSCE consists of 400 executable programs derived from real-world code, covering feature addition, feature removal, bug fixing, and refactoring. To quantify security drift, we propose a continuous risk representation that aggregates heterogeneous vulnerability signals through a unified formulation, and define drift measures capturing changes in overall risk, worst-case severity, and vulnerability distribution under code transformations, providing a multi-scale view of security spanning average-case behavior to worst-case emphasis.

Long-horizon agents can fail even when their underlying models can solve the constituent steps. They may lose track of mutable state, fail to reactivate lessons from earlier executions, skip known procedures, or stop prematurely. We bet on harness scaling to improve the execution system around an agent without changing its model weights. We introduce StateM, an agent-native runtime that organizes execution around durable states, phase-local context, checked transitions, recoverable runbooks, and versioned procedural practices that agents and users can inspect together. On Terminal-Bench 2.1, StateM raises GPT-5.5 xhigh to 92.1%, versus 83.1% reference and GPT-5.6 Sol Ultra at 91.9%. The runbook transfers unchanged to GPT-5.6. With GPT-5.6 Sol xhigh, StateM reaches 95.3% raw accuracy across 445 trials and succeeds on all 89 tasks at least once. The frozen profile raises GPT-5.6 Luna from 76.7 to 85.4%, above the 84.9% Sol xhigh reference. Using the same runtime, runbook structure, and golden rules, less than \$38 of adaptation raises DeepSeek-V4 Flash from 82.7 to 88.1\% under standard timeouts and to 89.1\% on an 88-task common core. Extending only the remaining latency-sensitive task matches the reported 88.8\% GPT-5.6 Sol max result. Final-score API usage is about \$15 versus \$574.68 for the GPT reference; total DeepSeek expenditure is \$52.22. On BusinessBench, family-specific runbooks built on development sets yield held-out gains of 0.55 macro and 1.34 micro points; two mechanism-matched families improve by 10.04 points. Concrete rules generalize when tasks share execution structure, while the control methodology applies broadly. StateM turns selected postmortem findings into persistent, executable preconditions and practices, making learned controls explicit and enforceable through stateful controls. Code at github.com/henryqin1997/statem.

Human-in-the-loop (HIL) online reinforcement learning for real robots must absorb human interventions quickly while continuing to improve beyond the human prior. We present a training method for this setting based on two components. First, an MC Q-chunk critic regresses chunk-level action values onto Monte Carlo returns from the replay buffer, performing sample-average (behavior) policy evaluation so that intervention trajectories are credited directly rather than diluted by current-policy TD backups. Second, max-Q selective imitation updates the actor by imitating, at each state, the higher-$Q$ action between the current policy action and a buffer sample under a hard winner-take-all rule. This rule automatically switches between learning from interventions and on-policy self-improvement: when the autonomous policy is stronger, targets align with the policy distribution, reducing the policy–target-sample gap that otherwise induces execution-time distribution shift. In practice we score candidates with a standard critic ensemble mean to reduce comparison noise, without softening targets or introducing score-gap thresholds. On a real USB pick-and-insertion task with 20 demonstrations, ACT QChunk-MCBC attains 99% success within 30 minutes of HIL training, whereas HIL-SERL requires about 5 hours to converge. In simulation on Peg Insertion and Square, ACT/Flow Q-chunk variants similarly reach $\ge$96% success within roughly half an hour of effective training, outperforming HIL-SERL, EXPO, and E2HiL on the success–time frontier.

Test-time augmentation (TTA) has proven effective for improving model robustness and uncertainty estimation in computer vision, yet its application to graph-structured data remains largely unexplored. We introduce GATTA (Graph Active Learning with Test-Time Augmentation), a framework for enhancing active learning by aggregating predictions across multiple augmented views to produce more reliable uncertainty estimates. To address the challenge of label-preserving graph augmentations, GATTA incorporates a consistency-based filtering mechanism that discards augmented views yielding unreliable predictions. We systematically evaluate GATTA across multiple graph datasets, GNN architectures, and acquisition strategies. Our results show that simple uncertainty-based methods, such as Entropy and Least Confidence, benefit most from TTA, achieving performance competitive with more sophisticated and computationally expensive approaches. GATTA generalizes across architectures, outperforms model-side ensemble methods such as MC Dropout. We further show that GATTA scales efficiently with both ensemble size and graph size. Extensive analysis of augmentation types, strengths, and filtering strategies provides practical guidelines for effective deployment. Our findings demonstrate that augmenting simple methods with TTA offers a more efficient path to strong active learning performance than engineering complex acquisition functions, enabling practitioners to achieve competitive results with lower computational overhead and reduced implementation complexity.

Cold chain logistics has advanced technologically, yet most deployed systems remain reactive monitors, not decision-making agents: thresholds trigger alerts, but nothing relates violations to cumulative product degradation or converts degradation signals into logistics decisions. We address this gap with a Quality-Aware Decision Intelligence (QADI) framework combining three capabilities: a structured quality state representation, $S_q = [L, Q, U, R]$ – remaining shelf life, degradation rate, estimation uncertainty, and operational risk, all derived and computable from the framework equations; a hybrid quality modeling layer combining physics-based microbial kinetics with a data-driven correction term; and a reasoning layer built on Microsoft Phi-4~\cite{Phi4} with retrieval-augmented generation over a structured domain knowledge base. We benchmark against five baselines – threshold monitoring, physics-only, physics-plus-noise, optimisation-based decisions, and a rule-based expert system – across eight cold chain scenarios, using pasteurised milk as the primary case, with ground truth shelf-life drawn from published dairy studies~\cite{Singh1994, Smigic2015} independent of our model. Comparisons use Wilcoxon signed-rank tests with Holm correction. Across milk and broccoli scenarios, the framework attains mean absolute shelf-life error of 7.2 hours (versus 30.9 hours, physics-only; $p<0.001$), spoilage rate of 14.5% (versus 16.6%, physics-only and rule-based; p=0.08), and oracle-optimal decisions in 99.5% of scenarios. Removing the LLM reasoning component drops optimality to 45.5% ($p<0.001$). Expert-rated explanation quality reaches 83% ($\kappa = 0.71$). Ablations show hybrid modeling and LLM reasoning contribute distinct gains, while RAG retrieval mainly drives explanation quality. Code: https://bit.ly/4d6t44C.

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.

Large Reasoning Models produce diverse, sometimes inconsistent answers across repeated queries on the same problem, so multi-sample inference is a prerequisite for reliable deployment. Majority voting at k rollouts is the standard solution and the de facto accuracy target for this regime, but it is prohibitively expensive at the scale LRMs require. We introduce Funnel of Thoughts (FoT), an inference-time method that preserves the full 32-trajectory voted accuracy while halving its attention FLOPs, a 28.8% reduction in full-model inference cost. Across 115K reasoning trajectories from six LRMs, we find that unproductive trajectories often reveal themselves through repeated hesitation markers such as "Wait", "Actually", and "perhaps." These trajectories are less likely to reach the correct answer and consume disproportionate attention FLOPs, degenerating into no-answer loops in the worst case. Built on this training-free lexical signal, FoT identifies the vocabulary that captures these pathological patterns and prunes affected trajectories before completion, reducing online generation attention FLOPs by 56.1% and wall time by 37.6% without any additional model inference; the same signal transfers without retuning across held-out architectures and out-of-domain tasks.

Parsing visual documents into machine-readable representations is fundamental to document intelligence. Existing benchmarks focus on page-level element recognition, reading order, formula recognition, and table structure. Long documents, however, also require document-level structure recovery. This includes reconstructing cross-page table-of-contents (TOC) hierarchies and identifying typed links from tables and figures to their captions, notes, and sources, often in one-to-many form. Because these structures are covered only partially or subsumed within broader parsing protocols, existing benchmarks cannot directly evaluate two key document-level tasks: Table-of-Contents Hierarchy Recovery and Contextual Relationship Recovery. To benchmark these two tasks, we introduce LongDocBench, comprising 85 real-world financial reports, textbooks, and academic papers spanning 2,582 pages, with up to 105 pages per document. It provides human-verified annotations for 3,937 heading nodes (mean node depth 3.55; maximum depth 9) and 3,258 contextual relationships annotated across 2,680 table and figure objects. We further evaluate both the downstream utility and recoverability of these structures. Long-document question-answering experiments show that human-verified TOC hierarchies and contextual relationships improve reasoning, with their combination providing complementary benefits. Meanwhile, representative document parsers remain limited on both recovery tasks despite strong page-level performance. To support further progress, we publicly release LongDocBench and its evaluation protocol and reproducible testbed for advancing document-level structure recovery in long documents.

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.

Hypergraphs effectively model higher-order groupwise relationships beyond pairwise interactions, while pretrained language models (PLMs) and large language models (LLMs) provide rich semantic understanding from textual attributes. However, research on combining language models with hypergraph learning remains limited due to the lack of public text-attributed hypergraph benchmarks. To address this limitation, we present TAHB (Text-Attributed Hypergraph Benchmark), the first public benchmark integrating hypergraph structures and raw textual attributes. TAHB contains 10 real-world datasets from four domains - e-commerce, academia, movies, and politics networks - enabling systematic evaluation of text-aware hypergraph representation learning. Experimental results show that TAHB preserves key structural properties of real-world hypergraphs and consistently reproduces performance tendencies observed in existing benchmarks. Furthermore, experiments under both LLM-as-Enhancer and LLM-as-Predictor settings demonstrate that LLM-enhanced textual semantics improve hypergraph learning performance, while structural and textual information jointly provide the best setting for LLM-based prediction. Our benchmark provides a foundation for future research at the intersection of hypergraph learning and language models.

Andy is an autonomous mathematical research agent that turns a mathematical problem into a traceable proof. It solves or verifies a submitted problem, formulates a literature-grounded new problem through a research-value gate, and carries it through proof construction and final verification. It organizes proof steps in an executable DAG, verifies each step independently and binds the result to a certificate, retains verified work whose interfaces remain unchanged during local repair, and records the full path from problem formulation to final proof. The system separates proof generation from correctness evaluation and can acquire, retain, retrieve, and reuse knowledge from existing results. Starting from a self-triggered impulsive consensus result, 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 continuous feedback over a recovery window. After each delayed impulse, the feedback cancels the delayed error channel until the pre-impulse history leaves the active delay interval. Sufficient conditions for global exponential synchronization are established, Zeno behavior is excluded for both timing sequences, and a numerical example illustrates the result.

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.

Coordinating multiple interacting units in complex engineering systems is challenging when system interactions are difficult to model, operational information is heterogeneous, and low-level actions must satisfy strict constraints. We propose an LLM-based hierarchical framework in which the LLM coordinates interacting units based on heterogeneous operational context, while task-specific controllers or optimizers generate executable and constraint-aware actions. We further introduce Continuation-Aware GRPO to capture the consequences of coordination decisions over subsequent control intervals. Rather than judging a decision only by its immediate outcome, the method also evaluates how the system evolves afterward under the current policy. We validate the framework on multi-ramp traffic control and virtual power plant (VPP) energy management, using simplified system models for training and more realistic simulators for evaluation. Across both tasks, the proposed method consistently outperforms direct task-specific control and optimization, end-to-end reinforcement learning, rule-based and RL-based hierarchical coordination, and prompting-only LLM coordinators, demonstrating the value of heterogeneous-context reasoning, hierarchical execution, and continuation-aware policy learning.

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.

Current Vision-based SLAM systems fail catastrophically when motion blur corrupts the visual input, as they attempt the ill-posed inverse problem of recovering sharp content from degraded observations. We present MotionGS-SLAM, which fundamentally reimagines motion blur handling through a paradigm shift: rather than removing blur artifacts, we reformulate the challenge as a well-constrained forward problem that generatively models blur formation within the rendering pipeline. By leveraging event cameras' microsecond temporal resolution and immunity to motion blur, we introduce a novel event-modulated Gaussian kernel that dynamically adapts each Gaussian's rasterization based on precise motion cues. Our dual-modulation mechanism transforms 2D Gaussian projections from isotropic dots into anisotropic, motion-aligned elliptical brush strokes (spatial modulation) while adaptively varying exposure integral sampling density based on local velocity (temporal modulation). This physics-based approach enables joint optimization of intra-exposure camera trajectories and 3D scene geometry through blur-aware photometric and event-based constraints. Extensive experiments demonstrate significant improvements over state-of-the-art methods in trajectory accuracy and map quality under severe high-motion conditions.