Ahmad Halimi Razlighi, Maximilian H. V. Tillmann, Edgar Beck, Bho Matthiesen, Armin Dekorsy
原文摘要
Cooperative multi-task semantic communication (CMT-SemCom) improves task execution performance by leveraging shared representations. However, as we demonstrated in [1], cooperative multi-tasking can be either constructive or destructive, depending on the semantic relationships among tasks. To ensure constructive cooperation, we propose a semantic-aware task clustering method for CMT-SemCom. We have formulated a sequential multi-stage optimization problem in which semantically aligned tasks are clustered once after a short initial training phase, and then end-to-end (E2E) joint training is conducted exclusively within the discovered groups. Specifically, the problem decomposes into two stages: (i) a semantic clustering problem leveraging hierarchical density-based spatial clustering, and (ii) an intra-cluster E2E CMT-SemCom learning problem. Simulation results demonstrate that the proposed framework effectively mitigates destructive cooperation and negative transfer, yielding accuracy gains compared to unclustered multi-tasking and individual training baselines.
Iñaki Dellibarda Varela, R. Sendra-Arranz, Pablo Romero-Sorozabal, J. M. Valverde-García, Annemarie F. Laudanski, Álvaro Gutiérrez, Eduardo Rocon, Manuel Cebrian
原文摘要
Large language model (LLM)-based multi-agent systems (LLM-MAS) are kept out of safety-critical domains by emergent failures that are hard to detect and to attribute. Current evaluation relies on a centralised evaluator, human or LLM judge, that must master the domain and hold the whole interaction history in one context. We propose that the agents executing a task can themselves attribute its failures. We introduce POIROT (Phased Oversight via Introspection, Review and Opinion Tallying), a modular protocol in which each agent casts a private vote within a hazard space derived from the system architecture, and the votes are aggregated into a single attribution. Before the vote, phases are added that progressively increase the interaction between agents, from independent self-assessment to sharing the self-assessments and peer interrogation. We build the protocol phase by phase, measure the contribution of each phase through an ablation, and compare every configuration with a single-LLM evaluator. We evaluate one proprietary and two open-weight LLMs on the Who&When benchmark and on BLAME (Benchmark for Localising Agent Malfunctions Effectively), a new benchmark for fault attribution in clinical rehabilitation and algorithmic trading systems. Agent-based attribution outperforms the single-LLM evaluator, but the benefit comes from independent self-assessment, not from interaction: sharing the assessments makes agents converge on the same, often wrong, culprit, and interrogation does not recover the loss. Without sharing and interrogation, the protocol outperforms the evaluator in all nine model-benchmark configurations (up to +50 percentage points). POIROT and BLAME are released as open source.
Photonic neural network accelerators have emerged as a promising solution for reducing the power consumption of deep neural network (DNN) inference by executing matrix multiplication on analog photonic hardware. However, their inference accuracy is limited by low-bit quantization required for photonic processing and by analog photonic noise. This paper proposes an approach for improving inference accuracy by exploiting the high similarity among activation vectors in DNNs. Specifically, our approach constructs a representative vector for each group of similar activation vectors, computes its inner product digitally with high precision, and offloads the residual inner products to the photonic accelerator. Since the activation vectors within each group are highly similar, the residual vectors have much smaller norms, thereby reducing the effects of quantization errors and analog photonic noise. Furthermore, our approach preserves the photonic inner-product computation pipeline, allowing seamless integration into existing and future photonic accelerators without hardware modification. Experimental results on six representative DNN models demonstrate that the proposed approach can improve inference accuracy across diverse network architectures and application domains.
We study how the sign operator interacts with gradient noise to shape the learning dynamics and scaling behavior of sign-based optimizers. In power-law kernel regression with diagonal features, we establish a functional scaling law (FSL) for SignSGD under general learning rate schedules. The FSL describes the learning dynamics through two processes: signal learning and noise forgetting. The key mechanism is noise-induced preconditioning: gradient noise smooths the sign operator, inducing an effective preconditioned SGD. Noise-curvature alignment enables this preconditioning to transform the effective Hessian from $\mathbf{H}$ to $\mathbf{H}^{1/2}$. The resulting dynamics exhibit faster signal learning but slower noise forgetting than SGD. This tradeoff yields a regime-dependent comparison with SGD: (i) SignSGD achieves improved data efficiency in the hard-task regime; (ii) both optimizers achieve the same data efficiency in the easy-task regime, but SignSGD maintains this efficiency at larger batch sizes, requiring fewer optimization steps for a fixed data budget. Power-law kernel regression experiments support the proposed mechanism and predicted scaling behavior. LLM pre-training experiments further suggest that the SignSGD FSL better captures the batch-size dependence of loss dynamics for sign-based optimizers than the original SGD FSL.
David Hartman, Jaroslav Hlinka, Dominik Klepl, Anna Pidnebesna, Petra Vidnerová
原文摘要
Symmetry in complex networks is a structural property with broad implications for network function, dynamics, and robustness. While classical approaches rely on exact automorphisms and thus remain highly sensitive to small perturbations, recent work of Liu has introduced approximate formulations that relax the strict invariance condition and recast symmetry detection as an optimization problem over permutations. We investigate how node-level similarity information can be used to bias the search toward structurally plausible vertex correspondences, with network centralities serving as one interpretable instance. Building on this perspective, we adapt simulated annealing, genetic algorithms, tabu search, particle swarm optimization, differential evolution and the quadratic symmetry approximator to this problem, add similarity guidance to each, and introduce Bilinear Approximate Symmetry Inference (BASIN), which learns a soft assignment matrix via bilinear embeddings. Using a unified evaluation pipeline, we compare their guided and unguided variants, BASIN and a random baseline on four synthetic graph families, structural brain connectomes, and three empirical networks with known automorphism groups. Node-similarity guidance is decisive where a planted orbit exists that the unguided search misses, and algorithm-dependent elsewhere, on synthetic and empirical networks alike. Overall, tabu search achieves the strongest performance; its guided variant (SGTS) has the lowest median score among the guided variants in all 44 synthetic conditions, while BASIN is the better of the continuous relaxations on rewired lateral graphs. The effect of node-similarity guidance is therefore regime-dependent rather than uniformly positive.
Simplified swarm optimisation (SSO) updates each variable by one draw that copies the global or personal best, keeps the current value or redraws uniformly, with probabilities $c_g$, $c_p$, $c_w$ and $c_r$. As a finite Markov chain under frozen memory, its coordinate kernel $c_w I + (1 - c_w)Π$ has an explicit stationary law and exact convergence rate $c_w^t$. The elitist chain has a geometric hitting-time bound, from a non-optimal start a fixed $c_r$ needs at least $\exp(nc_r/2)$ expected evaluations to reach a unique optimum, and a coupling under changing memories gives $O(n\log n)$ and $O(n^2)$ expected generations on OneMax and LeadingOnes for every population size with redraw share of order $1/n$ and no inertia, and within a $\log n$ factor for fixed inertia. With a frozen memory, inertia cannot speed up first entry into any set from stationary positions, though it can from others. Where the three reference values agree, only $c_r$ acts; where they disagree, copy branches recombine the memories, forcing a redraw only where no active branch holds the target value. Branch logging and checkpoint replay show inertia slowed the tested OneMax and LeadingOnes runs and suggest memory-mediated gains on wider gaps. On Jump functions SSO needed 3.8--8.9 times fewer evaluations than the (1+1) evolutionary algorithm with rate $1/n$, but no consistent advantage over rate-adjusted mutation or a $(μ+1)$ genetic algorithm with uniform crossover was established. The analytical redraw rule improved six of eight benchmark--setting pairs and redundancy allocation, but was slower than $c_r = 0.1$ on Jump$_2$.
Benjamin Bourel, Jean-Christophe Lombardo, Alexis Joly, Mathias Chouet, Thomas Paillot, Antoine Affouard, Vanessa Hequet, Herv{é} Go{ë}au 等 9 位作者
原文摘要
1. Monitoring plant biodiversity requires identifying large numbers of plants across space and time. Although artificial intelligence (AI) can now identify plant species from images with high accuracy, most existing systems rely on users manually submitting images. As ecological monitoring increasingly generates large image collections from cameras, drones, robots and online platforms, there is a growing need for plant identification services that can be directly integrated into automated data-processing workflows. 2. Here, we present Pl@ntNet's machine-to-machine AI plant identification service. This service uses an application programming interface (API) to provide programmatic access to Pl@ntNet's plant recognition models and the associated taxonomic processing. This service enables the implementation of automated identification workflows without human intervention and can thus directly link large image collections and automated observation systems to an AI-based plant identification system. 3. We establish that this service, which covers 85,000 species from around the world, (1) has a species prediction accuracy of 0.89, (2) process images up to 666 times faster than workflows requiring human intervention, (3) and has processed over 84 million identification requests since its launch, including more than 40 million in 2025 alone. We demonstrate how it has already enabled the implementation of a new generation of ecological workflows: identifying plants from images of insects to reconstruct the plant-pollinator interactions, analysing geolocated images from social media to study plant distribution, detecting invasive plants using images taken at the roadside, and identifying weeds using agricultural robots. Finally, we illustrate its application using a reproducible example: identifying plants from insect photos on iNaturalist. In approximately 30 min on a standard laptop, 2,424 images were processed to generate a dataset of 708 plant-pollinator interactions, with a weighted precision of 0.98 for flower detection and an accuracy of 0.88 for genus plant identification. 4. In conclusion, this machine-to-machine species identification is already operating on a considerable scale and meets a genuine need. All of this paves the way for a new paradigm in biodiversity monitoring, in which automated sensors and image collections can be directly connected to AI-based identification services.
We present a suite of 50 unpublished research-level theoretical physics problems with verifiable answers. Without AI assistance, a typical problem in our collection could be solved by a world expert in about a day or less but would take considerably longer for experts in adjacent specialties. We test the ability of off-the-shelf AI agents to tackle these problems autonomously, without human hints. When given access to tools available in a realistic setting, including the ability to search the literature and write scripts for numerical and symbolic work, the best-performing agent, GPT-6-Astra harnessed in Codex, obtains the final answers with a near-perfect accuracy of 98% at an average cost under USD 3 per problem. On average, agents show only small changes in accuracy when denied web search and access to background research papers, suggesting significant memorization of the literature. These results suggest that state-of-the-art agents can accurately perform concrete calculations required across theoretical physics research. We comment on the broader implications of these findings and discuss possible explanations for the current lack of agent-led breakthroughs in theoretical physics despite their outstanding problem-solving abilities.
Xiaokai Li, Ziyu Zhou, Qibin Cheng, Yangcheng Wu, An Zeng, Ying Fan
原文摘要
As large language models increasingly generate mathematical proofs, whether they develop a proving style distinct from that of human mathematicians remains an open question. This question is particularly pressing as mathematicians increasingly voice concerns that proofs generated by large language models are difficult to interpret. Here we ask three large language models to independently reprove all 465 propositions of Euclid's \textit{Elements} and compare their citation choices against the human-proof baseline through Jaccard similarity and complex network analysis. We find that AI-generated proofs differ from human proofs, and this difference is stable and reproducible: mean pairwise Jaccard similarity among AI runs (~0.68) is nearly twice that between AI and human proofs (~0.37), and cross-run classification consistency is also high. By examining citation structure in detail, we identify the concrete signature of this difference: approximately 40% of AI citations point to foundational elements (definitions, common notions, and postulates), compared with only 13\% in human proofs; even among theorem-to-theorem citations, AI tends to cite more distant results. Roughly half of AI's extra citations trace to the upstream dependencies of the human proof. Together, these findings suggest that the AI proving style is a tendency to construct arguments on its own from foundational elements rather than directly citing later, more established results, which we term "building from the ground up". Our findings offer a new perspective on how AI systems organise and invoke mathematical knowledge, with implications for understanding the interpretability of AI-generated proofs.
Yinghao Zhang, Yalin Liu, Hong-Ning Dai, Yan Zhang, Qubeijian Wang, Bishenghui Tao, Yaohan Li
原文摘要
Victim search in wilderness areas is time-critical. Drones enable rapid area search, but scheduling them is challenging because victims may continue moving, detection results are uncertain, and effective decisions must combine structured mission information with unstructured contextual clues and domain knowledge. To address this, we propose an agent-aided drone scheduling system. The system employs an agent at the ground control station (GCS) to coordinate multiple drones over successive search cycles. In each cycle, the agent applies a scheduling policy that combines a large language model (LLM) for reasoning about where to search with an optimizer for determining how to dispatch the drones under energy constraints. Simulation results show that the proposed system consistently completes the victim-search mission, whereas none of the six non-LLM baselines does so and direct LLM scheduling is considerably less reliable. The source code is publicly available at https://github.com/1643204431/llm-uav-victim-search.
Predicting every observation in a high-rate stream requires a model to finish before the next arrival. A large model may offer better predictions but miss that deadline. We propose RESYNC: the large model runs asynchronously, while a fast path updates its latest exact but stale latent state using the observations received since that state was computed and applies its unchanged prediction head. The updater learns from the frozen model's representations without task labels. Across one real and two synthetic streams and five architecture families, RESYNC beats the best compact or distilled alternative with a paired 95% interval above zero in all twelve dataset--architecture pairs with a material teacher advantage at anchor ages of 4--12 observations, and in eleven pairs at older tested ages up to 32. The share of the teacher's gain preserved remains complete on the real stream but is lower and generally declines with anchor age on the synthetic streams. In live replay with 1 ms arrivals, the synchronous large models cannot deliver any predictions on time while RESYNC serves at least 99.4% of arrivals within the interval, with 0.35--0.62 ms median latency on one CPU thread: around the same speed as or faster than the compact model, and 7.9--17.7$\times$ faster than synchronous large-model inference. RESYNC targets settings with background compute available but a strict per-observation response deadline.
Photonic neural accelerators offer ultra-low latency linear algebra, yet their usable numerical precision, sustainable throughput, and application accuracy remain tightly bounded by surrounding electro-optical interfaces and physical non-idealities. Pre-fabrication design-space exploration (DSE) requires a cross-layer electronic design automation (EDA) framework capable of evaluating disparate optical topologies under identical neural workloads and physical assumptions. We present \textbf{PISim}, a unified Photonic Inference Simulator that couples pretrained deep neural network execution with physics-based models of six accelerator architecture families, six electro-optic modulator technologies, four optical source organizations, and direct or coherent readout schemes. PISim embeds stage-resolved impairment operators---covering control and analog-to-digital quantization, finite extinction ratio, shot noise, correlated relative-intensity noise (RIN), thermal drift, propagation loss, and channel crosstalk---alongside an analytical cycle-level interface timing model to expose electronic-optical bottlenecks. Using simulated MNIST multilayer perceptron (MLP) benchmarks, we demonstrate how PISim identifies optimal precision operating points under multi-impairment stress (peaking at 6 bits for Mach--Zehnder and microring modulators), quantifies a 52.5 percentage-point accuracy gap between VCSEL-array and parallel-fanout laser architectures arising, under the modeled conditions, from power partitioning and correlated source noise, and models interface-limited throughput scaling across emitter-array widths from 32 to 128 channels. This comprehensive framework establishes an extensible, reproducible foundation for accuracy-constrained Pareto optimization prior to photonic tape-out.
Jiancheng Pan, Yuzheng Liu, Hengyuan Cao, Qiyuan Zhang, Yangyan Li, Biao Gong, Zheng Zhang, Ka Long Cheng 等 10 位作者
原文摘要
We present UniSpark-9B, a unified multimodal model that investigates hybrid attention for visual understanding and generation. Built on a Mixture-of-Transformer-Experts backbone with approximately 9B parameters in each branch, UniSpark combines full attention for explicit token interactions with GatedDeltaNet linear attention for compact context propagation. Autoregressive text prediction and visual flow matching jointly support visual question answering, text-to-image generation, instruction-guided editing, and interleaved text-image prediction. In interleaved generation, noisy visual targets are isolated from subsequent predictions, while clean visual representations provide context for later outputs. A blockwise protocol enforces this distinction through attention masks in full-attention layers and recurrent-state retention rules in linear-attention layers. The selected training sources contain 159.7M generation samples, 213.4M understanding samples, and 624.8K interleaved sequences. A four-stage curriculum jointly trains multimodal understanding and visual generation, refining answer accuracy and instruction following while extending image synthesis to 2K. UniSpark scores 2470.3 on MME-S and 0.83 on GenEval. Qualitative examples further illustrate its capabilities in interleaved image generation, video synthesis, and video editing. Together, these results demonstrate that hybrid attention provides an effective backbone for unified multimodal understanding and generation.
Low-light image and video enhancement has received increasing attention in recent years. However, most existing enhancement methods are developed and evaluated on still images, while progress in low-light video enhancement remains limited by the lack of real-world paired video datasets. Acquiring paired low-light and normal-light video sequences is challenging because it requires spatially and temporally aligned captures while avoiding motion blur, exposure mismatch, and correspondence errors in dynamic scenes. To address this gap, we introduce PiLLVD, the Pi-based Low-Light Video Dataset, a paired low-light video dataset captured using an accessible Raspberry Pi-based multi-camera system with off-the-shelf cameras. PiLLVD provides paired low-light and aligned normal-light reference sequences for dynamic indoor and outdoor scenes, including diverse human and object motion under varying illumination conditions. Unlike existing datasets that are often image-based, static, synthetic, or limited to simple motion patterns, PiLLVD focuses on motion-rich scenes where frame-to-frame stability and motion-related artefacts are important for evaluation. In addition to presenting the dataset and acquisition pipeline, we benchmark representative pre-trained low-light video enhancement methods on PiLLVD under a zero-shot setting. The results show that existing methods do not generalise uniformly across the proposed sequences, highlighting the challenges posed by the proposed dataset and its value for both benchmarking existing methods and supporting the development of future low-light video enhancement algorithms.
Shujie Duan, Jia Hong Puah, Haocheng Wu, Yifei Mao, Hongting Zhou, Chenxuan Li, Zitao Fang, Yang Lu 等 9 位作者
原文摘要
Time series forecasting aims to predict the future evolution of complex systems by learning their underlying dynamics from past observations. Despite recent advances in deep learning, Transformers in particular still struggle to capture the temporal evolution of system dynamics. Existing approaches have been largely driven by architectural innovation, with comparatively little attention paid to training strategies that explicitly exploit the temporal structure governing system dynamics. In this work, we introduce DeltaTSF, a dynamics-aware forecasting framework based on discrete derivative representations of time series. Specifically, DeltaTSF jointly learns the series and their discrete temporal derivatives across both input and forecast horizons within a unified, model-agnostic framework. By coupling these representations over time, the model is encouraged to maintain temporal consistency, a property we show to induce a form of discrete Taylor consistency during training. This formulation enables the model to better reflect the underlying system dynamics, resulting in more stable and physically coherent forecasts. Notably, DeltaTSF does not require explicit derivative computation and introduces no additional computational overhead at inference time. Extensive experiments across diverse benchmarks and backbone architectures demonstrate that DeltaTSF consistently improves forecasting performance and yields more dynamically coherent predictions than strong SOTA methods. These results highlight the effectiveness of enforcing dynamic coherence through derivative representations as a model-agnostic paradigm that operates alongside and complements architectural design. Source code will be made publicly available upon acceptance.
Attributing early Chinese Buddhist translations requires evidence that scholars can trace back to the texts. We develop an interpretable stylometric framework that uses features defined or reviewed by Buddhist philologists, including function characters, sentence formulas, sentence-final characters, sentence length, and part-of-speech n-grams. We construct stylistic profiles from 31 authenticated translations by seven translators from the Eastern Han through the Western Jin, then compare eight disputed texts with An Shigao's profile. Two texts, T1557 and T0101, show the strongest stylistic similarity to the authenticated An Shigao corpus, with scores of 0.783 and 0.687, respectively. The attributions agree with those reported by a published BERT-based approach on the same set of disputed texts. The framework requires no task-specific training and allows scholars to inspect which linguistic features support each proposed attribution.
Julien Vanderheyden, Guillaume Drion, Pierre Sacré, Fulvio Forni
原文摘要
Low-impedance grippers make the dynamics of a grasp matter. Current grasp metrics, however, are static: they describe how a held object resists constant loads, not how it responds to time-varying ones. To quantify this, we rely on small-signal analysis to characterize a grasp by its admittance, the transfer matrix relating load perturbations to object displacements around an equilibrium. This transfer matrix recovers the stiffness and stability of the grasp at zero frequency and captures its natural frequencies and peak gain under time-varying loads. We apply this characterization to a planar two-finger gripper under proportional-derivative control. We show that two grasps of the same object in the same position and with the same contact forces can differ in stiffness by two orders of magnitude. Likewise, two grasps that share the same stiffness can exhibit very different resonances. Nonlinear simulations confirm the effectiveness of our small-signal analysis in capturing the dynamic behavior of grasping.
While quantization compresses LLMs and accelerates inference, outlier activations impede efficient low-bit representation. In particular, massive outliers cause substantial performance degradation. Although rotation-based methods that rely on randomized Hadamard transforms for online inference aim to mitigate this, they can produce fragmented distributions with empty bins when a massive outlier dominates a token's activation. This fragmentation severely undermines uniform quantization performance. Moreover, massive outliers concentrate in the down-projection input, where the preceding nonlinearity requires the rotation to be applied at inference time. This constraint prevents training-based approaches such as SpinQuant from learning the online rotation at this position, as a learned dense matrix cannot use the Fast Walsh-Hadamard Transform and would require a full matrix multiplication during inference. To address this, we propose SignRot (Sign-adjusted Rotation), which identifies massive outlier channels and constructs a sign adjustment vector that unifies the signs of the corresponding Hadamard row, collapsing the bimodal distribution into a unimodal profile amenable to quantization. Since the aligned outlier shifts the activation mean away from zero, we pair SignRot with asymmetric activation quantization, whose zero-point absorbs this shift. As a lightweight, drop-in improvement compatible with any Hadamard-based quantization method, SignRot requires only a single calibration forward pass and adds only element-wise sign flips at inference, without an additional matrix multiplication. Applying SignRot to QuaRot achieves a WikiText2 perplexity of 5.88 and an average zero-shot reasoning accuracy of 72.73, surpassing SpinQuant on the LLaMA3-70B model. Code is available at https://github.com/HYU-ARK-Lab/SignRot.
We present EvenVLM, the first text and multimodal (VQA) dataset for Even (Lamut), an endangered Northern Tungusic language spoken by approximately 6,800 people in Siberia. Our dataset comprises 50,521 structured Even-Russian entries extracted from dictionary and phrasebook sources via GPT-4o Mini-based parsing, from which we synthetically generate 107,543 training examples spanning translation, grammar, reading comprehension, and semantic classification. To address tokenization inefficiency on Even's agglutinative morphology, we propose a three-step tokenizer expansion algorithm that adds Even-specific graphemes (eng, oe, o-umlaut), whitelisted vocabulary, and BPE merges, reducing average tokens per word by 27% (4.99 to 3.64). Using LoRA fine-tuning of Qwen3-VL-4B, we adapt the model for text tasks (0.53% trainable parameters, rank 32) and separately for visual question answering (rank 64, frozen vision encoder). Our best configuration achieves 11.4% exact match and a 13.6 weighted score on text benchmarks, 0.1283 chrF for sentence translation, and 65% accuracy on VQA (up from 0% for the base model and text-only fine-tuning). We release our datasets, evaluation benchmarks, and training pipeline to support further research on Even and other low-resource languages.
A typed-decision model such as Laya reads a text and a list of options that the caller defines, for example ticket categories, and returns a probability for each option. Code around it compares the top probability with a threshold to answer, abstain or escalate to a person. Earlier work reports, for a Turkish model, that temperature scaling can raise held-out calibration error (ECE). This tutorial gives a five-step recipe (measure, diagnose, fit, validate, deploy) in 130 lines of numpy with standard methods, none new. On 1,065 synthetic, LLM-labelled items the released Laya checkpoint has a pooled ECE of 0.073 but a bucket-macro ECE of 0.143 (0.127-0.184), the mean over buckets (questions grouped by type and option count): the 12-option bucket is over-confident by +0.22, the three-to-five-option bucket under-confident by 0.05, and they cancel. One global temperature makes pooled ECE worse (0.173); a cross-fitted temperature per bucket brings macro-ECE to 0.073 (0.078 over 10 fold assignments) and reaches 0.089 with 10 labelled items per bucket, but holding out one of seven evaluation sets gives 0.099. After the refit, a gate aimed at 10% error answers 46.3% of items (10.5% wrong) against 6.9% on the published confidences, but one cut-off per bucket on the published confidences answers 45.4%, so most of the gain comes from separating buckets. Confidence-only maps (Platt, histogram, isotonic) need 20 or more labelled items per bucket to approach the temperature; with 320 per bucket the lowest macro-ECE belongs to Platt. In these files the retrained checkpoints store no per-bucket table and the stored 12-option value 0.1006 is clamped to 0.5 at runtime; the refit helps 31 of 33 checkpoints on random folds and 21 with a set held out. Fit and test share one generator and LLM labels, so these are findings about these checkpoints and this data.
Zipeng Ye, Shenyu Dai, Yong Xie, Kun Wang, Jiyang Guan, Zhikai Chen, Jiaqi Weng, Jialing Tao 等 9 位作者
原文摘要
As agents perform tasks autonomously in local and cloud environments, auditing their execution trajectories can help service providers identify risks and improve agent harnesses and foundation models. Our analysis of production and red-team trajectories shows that some risks are repeatedly missed across auditing configurations. Overall recall obscures which risks are missed and whether higher recall comes with more false positives on legitimate behavior. Existing benchmarks offer limited insight into these failures. Generated trajectories can contain inconsistent tool calls and responses, conspicuous risky actions, or repetitive attack patterns. Grouping results by risk type also makes it difficult to tell whether auditors overlooked evidence, accepted an unsupported conclusion, or lacked information needed to make a judgment. We introduce ReA-Bench, a diagnostic benchmark for auditing responsible agent behavior. It organizes auditing challenges into six tiers: internal safety knowledge, contradictions within trajectories, evidence retrieval in long contexts, verification of sources and reasoning, indirect steering, and missing information. It includes challenging benign cases in which actions resemble risky ones but are consistent with the task and its authorization. It also provides paired trajectory versions to test whether auditors detect the same risks more reliably when the trajectory is shortened or missing information is supplied. Most samples are drawn from real production trajectories, with scenarios manually selected and adapted to preserve realistic tasks and risks that could plausibly arise during execution.
A hosted score in the Gemma 4 Developer Agent competition is one noisy number, so development needs a local evaluator that is fast, safe for agent-written code, and matching the hosted scorer's prompts. drydock-eval evaluates a small coding agent (Gemma 4 E4B) at about 250 runs per hour on one RTX 5090. Contain runs each job in a fresh container without network and passed 200 scripted escape tests; conform provides eight checks that compare a local stack's requests with the scorer's configuration, down to rendered tokens; compare provides paired statistics that respect repository clusters. AI agents built and attacked the kit; a human owner approved machine use and release. This extended version repeats the short paper and adds the full record of the experiments made with and around the kit: 9,660 rows of agent runs, 2,494 grading controls and 72 serving measurements, all released. On four L4 GPUs the 31B model decodes 38.1 tokens per second in one stream and E4B 105.8. Six configurations of a 31B agent solved between 37 and 44 of 70 development tasks, and no pairwise difference was detected. A pre-registered five-arm experiment (231 tasks, 4,620 jobs) measured four prompt divergences: no E4B solve-rate change was detected (largest estimate -2.1 points, 95% interval -4.4 to +0.4). Three changes of the E4B agent led their screens by +1.0 to +6.2 points; each was then tested once on new tasks under a frozen pass rule, and none was confirmed (gains -0.35, -0.39 and +4.10 points). The unchanged baseline, run again on the same 128 tasks, moved by 1.17 points, and 65% of these tasks were solved in none of eight runs. Prompts and agent bundles are withheld; every run-level result is public.
We study continuous-time mean--variance portfolio selection in markets where stock prices are diffusion processes driven by observable factors that are also diffusion processes, yet the coefficients of these processes are unknown. Based on the recently developed reinforcement learning (RL) theory for diffusion processes, we present a general data-driven RL approach that learns the pre-committed investment strategy directly without attempting to learn or estimate the market coefficients. For multi-stock Black--Scholes markets without factors, we further devise an algorithm and prove its performance guarantee by deriving a sublinear regret bound in terms of the Sharpe ratio. We then carry out an extensive empirical study implementing this algorithm to compare its performance and trading characteristics, evaluated under a host of common metrics, with a large number of widely employed portfolio allocation strategies on S\&P 500 constituents. The results demonstrate that the proposed continuous-time RL strategy is consistently among the best, especially in a volatile bear market, and decisively outperforms the model-based continuous-time counterparts by significant margins.
Financial diffusion samplers are often reused with a small number of function evaluations (NFE) inside scenario and risk pipelines, where numerical time-grid error can distort terminal quantities such as covariances, tails, and drawdown proxies. We study offline grid construction for a frozen probability-flow ODE: the score network, conditioning, Euler solver, report normalization, and deployment NFE remain fixed. TAPES (Terminal Adjoint-Propagated Error Scheduling) estimates signed local defects by Richardson comparisons, transports them through the remaining discrete flow and a terminal-report adjoint, and optimizes the squared norm of their sum before a capped direct refinement. We show that this transported signed sum is the first-order terminal-report error and quantify, under an aligned-defect model, the inefficiency of diagonal criteria. Across six budgets on a regime-switching simulation and chronologically held-out Mag7 contexts, TAPES gives the lowest design- and audit-report errors in all 24 dataset--budget--report rows within the reported five-pipeline panel. Cross-budget mean relative audit-error reductions are about 10--11% on Regime-SV and about 1% on Mag7, with Bonferroni intervals excluding zero; all displayed winners persist under a doubled-resolution reference, and a small Regime-SV Heun probe preserves the TAPES-first audit ordering. TAPES changes only the offline schedule and retains the prescribed deployment NFE; claims concern numerical fidelity to frozen generators.
We consider time discretization for score-based diffusion models to generate samples from a learned reverse-time dynamic on a finite grid. Uniform and hand-crafted grids can be suboptimal given a budget on the number of time steps. We introduce Adaptive Reparameterized Time (ART), which controls the clock speed of a reparameterized time variable to redistribute computation along the sampling trajectory while preserving the terminal time, with the objective of minimizing the aggregate Euler discretization error. We derive a randomized companion ART-RL that recasts ART as a continuous-time reinforcement learning problem with Gaussian policies, and prove a two-directional bridge between the two: the deterministic ART optimum lifts to an optimal Gaussian policy, and conversely any optimal Gaussian policy must recover the ART control through its mean. This bridge turns continuous-time actor--critic learning into a principled, rather than heuristic, route to the deterministic timestep optimum. Within the official EDM pipeline, ART-RL improves FID on CIFAR--10 across a wide range of budgets; after one-time offline training, the distilled deterministic schedule transfers without retraining to AFHQv2, FFHQ, and ImageNet at no extra inference cost.
Qiyuan Xu, Joshua Ong Jun Leang, Renxi Wang, Wenda Li, Haonan Li, Luke Ong, Conrad Watt
原文摘要
Interactive theorem proving (ITP) underpins program verification and formalized mathematics, but its manual effort limits scalability. LLM-based proof agents promise to ease this effort, but their heavy token consumption and API cost remain a major obstacle. We trace this cost to a shared root: current agents operate on serialized concrete syntax, emitting proofs as source text and recovering proof states through separate, line-number-based queries, so every edit shifts later lines and forces repeated relocation of errors and states. This same dependence on concrete syntax also blocks adoption of Minilang, a recent proof language that reaches SOTA on LLM-based proving but is too new for LLMs' training corpora. We address both problems by lifting the agent off source text and onto the abstract syntax tree (AST): the model supplies proofs as JSON representations of Minilang's AST -- native to tool-calling LLMs -- and drives the prover through a tree-edit model that fuses proof operations and states into one proof tree, so each operation carries its own subgoal's state, readable directly off the tree. We realize this design in \emph{Agent over AST} (AoA). Against Amazon's Isabelle Agent on miniF2F and NTP4VC-Pearl common success sets, AoA cuts API cost by 2.3--4.7x (normalized input-cache accounting), uses 2.9--6.9x fewer tokens and 3.9--8.9x fewer tool calls, and finishes 1.4--2.0x faster -- while also solving far more problems on the harder verification benchmark.
Local language models often exist at several precisions, and quantization is known to change individual answers. This study asks whether the disagreement between two precisions of the same weights, an 8-bit copy that answers and a lower-precision copy that reads the same prompt, works as a per-question error score for abstention, escalation and re-asking. The cost is one extra forward pass over fewer bytes, with no sampling and no training. We simulate round-to-nearest quantization with per-group scales (3 to 6 bits, groups of 32 to 256) on Qwen3-0.6B and Qwen3-1.7B and score 8,398 multiple-choice questions from ARC-Easy, ARC-Challenge, MMLU and HellaSwag subsets. Error detection is measured by AUROC, risk-coverage curves and ECE on a held-out half, with paired bootstrap intervals over questions. For Qwen3-0.6B, the entropy of the 8-bit answer reaches AUROC 0.682; the best disagreement score, selected on a development half, reaches 0.684 (+0.001 [-0.009, +0.011]), and combining it with the confidence features adds +0.009 [+0.002, +0.015], mostly on ARC. On Qwen3-1.7B the selected score is -0.001 [-0.009, +0.006] from entropy and the combination +0.003 [-0.001, +0.006] on the test half. Gaussian weight noise at matched error energy is about as informative as rounding (0.6B only), so we find no evidence that the signal is specific to quantization, and shuffling the features within task and entropy cells removes the gain. One extra 8-bit pass that shifts the option order does better (+0.049 [+0.031, +0.067] over entropy) at about a third more weight traffic; routing to a larger model on free entropy beats both extra-pass scores at equal total traffic. After logistic recalibration the 0.6B confidences are calibrated within a few points of ECE. A deployment recipe and the cases where the method does not help are given.
Laya is a small model that answers a multiple-choice question about a text in one pass: the question, the options (each with a name and a description) and the text form one sequence of at most 512 tokens, the window, in the English release. A longer text is silently cut at the end, and we measure what that costs. Test documents place a held-out row's text, the evidence, at a chosen token offset among distractor text; 177 items, five tasks, three checkpoints (released, short-row fine-tune, long-document fine-tune), four window sizes and 93,275 scored requests. When the evidence lies beyond the window, accuracy falls to 26.4% for the released checkpoint (27.2% for both fine-tunes), near the 27.0% of guessing and below the 35.6% of always giving the commonest label. Confidence gives no warning (AUROC 0.50 to 0.55 for telling cut from read requests; 0.5 = chance), and the runtime's truncation flag also fires on requests whose evidence was visible. The released checkpoint reads visible evidence badly once text precedes it: 38.5% pooled against 60% when the evidence comes first, so a larger window alone gains +6.2 points. Fine-tuning on long documents adds +27.6 points on visible evidence over the short-row fine-tune (95% interval +22.3 to +33.1) and nothing on cut evidence. That is an in-family result: training rows and test documents share generator and question types, and the main filler is of the training kind. With encyclopedic filler never seen in training the gain is +13.2 (+9.9 to +16.6), with rows of other tasks as filler -2.0. Among training-free rearrangements a 312-token window scan helps most (+11.9 points over the default for the released checkpoint) at six to eleven forward passes. Documents are synthetic, labels are LLM-made and one backbone is studied, so these diagnose this model, not long documents in general.
Many learned predictive models represent their uncertainty as a probability distribution, and planners and policies use this distribution to judge risk. However, the fact that a model outputs a distribution does not tell us whether that distribution captures the uncertainty in the environment which is relevant for a decision. In this work we study this question in distributional reinforcement learning, where the learned return distribution P_θ(Z | s,a) can be compared directly with an environment-grounded distribution P_env(Z | s,a) estimated by snapshot-restart Monte Carlo. Our audit first screens for the claims that a risk-sensitive user would act upon, using the excess Wasserstein gap between the top two actions (which is equal to the mass by which first-order stochastic dominance is violated). Each claim is then tested against ground truth using permutation nulls, bootstrap refutation and FDR control. Across QR-DQN, C51 and IQN on MinAtar (33 runs), 40-95% of the strongest claimed risk trade-offs are refuted at 95% confidence, their placement is statistically indistinguishable from a truth-blind placement, and almost none of them can be confirmed. A similar pattern is seen for a pretrained full-Atari QR-DQN. On positive controls of known magnitude the same audit confirms 96-100% of real claims (correlation 0.89-0.92), so it is able to detect real risk when it is present. Acting on the learned CVaR advice at the most-flagged states gives results ranging from beneficial to significantly worse than chance. Ensembling, risk-sensitive training and recalibration do not fix the problem, and the recalibration experiment indicates that the learned spread has no usable information about the true risk. Based on these results, we suggest that uncertainty estimates, including those of learned world models, should be checked against realized outcomes at the decisions where they are used..
Efficiently scaling industrial Click-Through Rate (CTR) prediction has recently attracted significant research attention. Existing approaches typically employ early aggregation of user behaviors to maintain efficiency. However, such non-unified or partially unified modeling creates an information bottleneck by discarding fine-grained, token-level signals essential for unlocking scaling gains. In this work, we revisit the fundamental distinctions between CTR prediction and Large Language Models (LLMs), identifying two critical properties: the asymmetry in information density between behavioral and non-behavioral features, and the modality-specific priors of content-rich signals. Accordingly, we propose the Efficiently Scalable Transformer (EST), which achieves fully unified modeling by processing all raw inputs in a single sequence without lossy aggregation. EST integrates two modules: Lightweight Cross-Attention (LCA), which prunes redundant self-interactions to focus on high-impact cross-feature dependencies, and Content Sparse Attention (CSA), which utilizes content similarity to dynamically select high-signal behaviors. Extensive experiments show that EST exhibits a stable and efficient power-law scaling relationship, enabling predictable performance gains with model scale. Deployed on Taobao's display advertising platform, EST significantly outperforms production baselines, delivering a 3.27\% RPM (Revenue Per Mile) increase and a 1.22\% CTR lift, establishing a practical pathway for scalable industrial CTR prediction models.