arXiv:2607.11801v1 Announce Type: new
Abstract: Large audio-language models (LALMs) often underperform on fine-grained, non-semantic attributes of speech, such as a speaker's emotion, despite strong performance on speech content. Improving this without the cost of retraining calls for an effective inference-time intervention, yet most existing methods intervene only after the audio encoder and operate at a relatively coarse granularity. The encoder itself, where acoustic information is first extracted from the waveform, remains largely unexplored, especially at the level of individual neurons. We introduce IAAN, Identifying and Amplifying Acoustic Neurons, a training-free and label-free method that scores each feed-forward neuron in the audio encoder by contrasting its activation on the real waveform with that on a noise reference lacking the real audio's acoustic information. IAAN then amplifies a small set of the highest-scoring neurons at inference. Across ten non-semantic speech attributes, IAAN improves average accuracy by 25.7 points on Audio-Flamingo-3, 21.4 on Qwen2.5-Omni, and 9.7 on Kimi-Audio. It also improves a model already explicitly fine-tuned to prioritize acoustic evidence. In controlled comparisons, both the encoder locus and neuron-level selectivity prove necessary for this gain. Intervening after the encoder, at the decoding side or inside the language model, yields little to no improvement, or even deteriorates accuracy. The improvement also depends on which specific neurons are amplified, not merely on their number, confirming that IAAN's acoustic score succeeds in identifying the neurons that matter. These results show that a small, precisely targeted intervention inside the audio encoder is an effective and largely untapped way to strengthen the acoustic understanding of LALMs, opening a new direction for inference-time methods that improve acoustic perception through neuron-level access to the encoder.
Science Journals
arXiv:2607.10180v1 Announce Type: new
Abstract: We introduce ActiveFly-Bench, the first benchmark to bridge cyberspace reasoning and physical-world interaction for UAV embodied perception. The benchmark decomposes active perception into three hierarchical tasks: Aerial Embodied Question Answering (Air-EQA), Observation Behavior Planning (OBP), and Fine-grained Language-guided UAV Control (FLUC), explicitly connecting high-level task understanding, behavior planning, and low-level control. The datasets are collected from both real-world and simulated outdoor environments for training and evaluation. We further develop ActiveFly, a closed-loop UAV agent that integrates visual-language reasoning with fine-grained control, and deploy it on a physical UAV platform. Experiments with representative VLMs and VLA models show that current UAV agents still struggle with behavior planning, viewpoint adjustment, and robust task completion in active perception. These results establish ActiveFly-Bench as a new testbed for embodied aerial intelligence.
arXiv:2510.22510v3 Announce Type: replace
Abstract: While continuous diffusion has shown remarkable success in continuous domains such as image generation, its direct application to discrete data has underperformed pure discrete formulations. To understand this gap, we introduce token identifiability, an analytical framework characterizing how Gaussian noise corrupts discrete data through two mechanisms: discrete identity corruption and continuous rank degradation. We reveal that these mechanisms scale differently with vocabulary size, creating a temporal dissonance that forces a tradeoff between learning continuous geometry and discrete structure. To address this, we propose CANDI (Continuous ANd DIscrete diffusion), a hybrid framework that decouples discrete and continuous corruption, enabling simultaneous learning of both. This unlocks the benefits of continuous diffusion for discrete spaces: on controlled generation, CANDI enables classifier-based guidance with off-the-shelf classifiers through simple gradient addition; on text generation, CANDI outperforms masked diffusion at low NFE, demonstrating the value of learning continuous gradients for discrete spaces. We include the code on the project page: https://patrickpynadath1.github.io/candi-lander.
Route, Communicate, and Reason: Gated Routing and Adaptive Depth for Efficient Multi-Agent Reasoning
arXiv:2607.10836v1 Announce Type: new
Abstract: Multi-agent ensembling multiplies active parameters and inference cost without answering three basic questions: which agents to consult, how deeply a query should traverse a hierarchy of agents, and when inter-agent communication is worth its cost. We present GRADE (Gated Routing and Adaptive Depth for Efficient Reasoning), a hierarchical multi-agent system in which four lightweight learned gates jointly govern agent selection, hierarchy depth, inter-agent communication, and branch pruning. Training uses CoGRPO (Collaborative Group-Relative Policy Optimization), a novel critic-free recipe that adapts GRPO to multi-agent hierarchies and assigns a shared advantage signal to every gate and agent that participated in a rollout. Agent models are drawn from a hot-swappable Expert Registry; per-agent calibration maps allow experts to be replaced at inference time without retraining. At $\sim$17B average active parameters, GRADE outperforms all baselines on GSM8K, MMLUPro, and GPQA, surpassing the strongest baseline by 4.8 points on MMLUPro at half the active compute. On AIME-2025, where model depth dominates, GRADE remains competitive to existing frameworks. Ablations isolate the hierarchy and masked cross-attention as the largest contributors to accuracy, and show that per-agent calibration is necessary for safe hot-swapping.
arXiv:2607.11106v1 Announce Type: new
Abstract: Recent multimodal large language models (MLLMs) have made remarkable progress on fine-grained perception tasks under the "Thinking with Images" (TwI) paradigm by iteratively performing various visual tool operations. However, this paradigm relies heavily on frequent external tool calls and repeated image re-encoding, which leads to substantial computational overhead and inference latency. To address these issues, we propose Beyond the Eye (BEE), a novel implicit visual tool paradigm centered on self-regulated capability. BEE directly incorporates visual tool invocation behaviors into the training objective and encourages the model to develop a self-regulated invocation mechanism. This design enables the model to adaptively balance internal knowledge and implicit tools, avoiding redundant tool usage while substantially reducing inference latency. Specifically, BEE involves a two-stage training process: (1) Formalized Chain-of-Thought (CoT) Supervised Fine-tuning (SFT). We construct CoT trajectories with structured tool slots and mixed invocation states. This stage activates the model's implicit tool representations and adaptive switching capability. (2) Self-regulated Reward-Driven Alignment. To address redundant tool usage caused by ambiguous cognitive boundaries, we first introduce the Net Tool Gain (NTG) metric to quantify this phenomenon. Based on this observation, we further propose a self-regulated reward mechanism. This mechanism penalizes ineffective tool dependency and encourages the model to perform knowledge routing, ensuring that implicit tools are invoked only when the model's internal knowledge is insufficient. BEE achieves state-of-the-art performance in fine-grained visual perception while remaining competitive in general reasoning tasks and achieving substantial gains in inference efficiency.
arXiv:2607.11083v1 Announce Type: new
Abstract: Real-world target speaker extraction (TSE) remains challenging because target speech, interference, and enrollment are recorded under mismatched acoustic conditions with reverberation, noise, and irregular conversational overlap. This paper describes the SonicAGI submission to the REAL-TSE Challenge (IEEE SLT 2026). We take a data-centric approach that combines fully simulated mixtures from clean speech with real meeting overlaps, and use a frozen offline enhancer to provide a denoised mirror of real targets for auxiliary supervision. For the online track, we introduce SwiftNet-Lookahead, which inserts a single bounded-lookahead module before a strictly causal iterative separator and keeps the total system latency at 96 ms. For the offline track, we use a frame-level enrollment cross-attention USEF-TFGridNet with a magnitude-domain fusion stage that trades off perceptual quality and speaker fidelity. In the official evaluation, SwiftNet-Lookahead ranks second in Track~1 and USEF-TFGridNet ranks fifth in Track~2, both exceeding the challenge baselines. These results suggest that real-data-oriented training and track-specific modeling are effective for conversational TSE.
Rethinking MCP Security: A Large-Scale Study of Runtime MCP Servers and Security Scanner Reliability
arXiv:2607.11086v1 Announce Type: new
Abstract: The Model Context Protocol (MCP) has rapidly established itself as a standard interface for enabling LLM-based agents to interact with external tools and services. As MCP servers are increasingly entrusted with security-sensitive operations, understanding their real-world risks has become critical. In practice, due to the absence of large-scale runtime MCP servers, such understanding largely relies on security scanners applied to a small number of cases, yet the reliability of these assessments remains unclear.
In this study, we revisit how MCP security is measured. We present MCPZoo, the largest collection of MCP servers for dynamic analysis to date. MCPZoo is constructed through a multi-agent framework for transforming in-the-wild static repositories into dynamic services. The framework emulates how human experts build, diagnose, and iteratively repair deployment and runtime defects by combining environment inference with feedback-driven refinement. To ensure practical interactivity at runtime, the servers are validated via real protocol interactions. As a result, MCPZoo contains 64,611 unique MCP servers (113,927 in total), with more than 37,288 supporting dynamic analysis. Leveraging MCPZoo, we conduct the first ecosystem-scale measurement of MCP servers and the scanners that analyze them. While existing scanners report that 96.89% of servers are risky, we find that these signals are unreliable. In particular, manual validation shows that less than 50% of sampled alerts are true positives, and scanner outputs exhibit clear inconsistency across scanners. Overall, MCPZoo enables large-scale, reproducible measurement of MCP server security and exposes limitations of current scanning practices. We further release a public query interface to support practical risk assessment of MCP servers.
arXiv:2607.11564v1 Announce Type: new
Abstract: Researchers organize the papers they collect into personal folder hierarchies in reference managers, and route each new paper into the folder where it belongs. This task differs from standard hierarchical text classification. A user's folder hierarchy is not a fixed, shared taxonomy but a private and evolving folksonomy whose folder meanings may be topical, shorthand, venue-based, or process-oriented, and are often defined by the papers already stored inside them. We formalize this setting as personalized hierarchical paper routing (PHPR): assigning an incoming paper to folders in a user-specific hierarchy without per-user training. We propose PaperRouter-Agent, a training-free LLM agent that grounds routing decisions in folder members rather than folder names alone. The agent first narrows the candidate hierarchy, retrieves folder-specific evidence, verifies fit by inspecting member papers, and incorporates similarity-gated feedback from past user rejections. A formative study on real personal libraries shows that PaperRouter-Agent raises overall Recall@1 from 0.39 to 0.61 and Recall@3 from 0.57 to 0.83, with the largest gains on organizational folders defined by metadata such as venue or year, where single-shot methods collapses (Recall@1 0.09 to 0.50). On the public LaMP-2 benchmark, the same approach improves accuracy from 44.5% to 51.5% (+9.0 macro-F1) over a single-shot baseline, while remaining low-cost for practical use.
arXiv:2509.12090v3 Announce Type: replace
Abstract: Reconstructing cardiac motion from CMR sequences is critical for diagnosis, prognosis, and intervention. Existing methods rely on complete CMR stacks to infer full heart motion, limiting their applicability during intervention when only sparse observations are available. We present TetHeart, the first end-to-end framework for unified 4D heart mesh recovery from both offline full-stack and intra-procedural sparse-slice observations. Our method leverages deformable tetrahedra to capture shape and motion in a coherent space shared across cardiac structures. Before a procedure, it initializes detailed, patient-specific heart meshes from high-quality full stacks, which can then be updated using whatever slices can be obtained in real-time, down to a single one during the procedure. TetHeart incorporates several key innovations: (i) an attentive slice-adaptive 2D-3D feature assembly mechanism that integrates information from arbitrary numbers of slices at any position; (ii) a distillation strategy to ensure accurate reconstruction under extreme sparsity; and (iii) a weakly supervised motion learning scheme requiring annotations only at keyframes, such as the end-diastolic and end-systolic phases. Trained and validated on three large public datasets and evaluated zero-shot on additional private interventional and public datasets without retraining, TetHeart achieves state-of-the-art accuracy and strong generalization in both pre- and intra-procedural settings. Code and dataset is available at https://github.com/Scalsol/TetHeart.
arXiv:2509.20114v3 Announce Type: replace
Abstract: We study \emph{online episodic Constrained Markov Decision Processes} (CMDPs) under both stochastic and adversarial constraints. We provide a novel algorithm whose guarantees greatly improve those of the state-of-the-art best-of-both-worlds algorithm introduced by Stradi et al. (2025). In the stochastic regime, \emph{i.e.}, when the constraints are sampled from fixed but unknown distributions, our method achieves $\widetilde{\mathcal{O}}(\sqrt{T})$ regret and constraint violation without relying on Slater's condition, thereby handling settings where no strictly feasible solution exists. Moreover, we provide guarantees on the stronger notion of \emph{positive} constraint violation, which does not allow to recover from large violation in the early episodes by playing strictly safe policies. In the adversarial regime, \emph{i.e.}, when the constraints may change arbitrarily between episodes, our algorithm ensures sublinear constraint violation without Slater's condition, and achieves sublinear $\alpha$-regret with respect to the \emph{unconstrained} optimum, where $\alpha$ is a suitably defined multiplicative approximation factor. We further validate our results through synthetic experiments, showing the practical effectiveness of our algorithm.
arXiv:2607.11078v1 Announce Type: new
Abstract: Can a Video Large Language Model (Video-LLM) follow one person through a long video, keeping track of who they are well enough to report, in order, how their outfit changes across a full TV episode? Benchmarks increasingly score this kind of task, and the strongest open-source 7--8B models now reach 37--38% on InfiniBench's global appearance task, which asks exactly that. But does that score come from tracking the named character, or from something easier? We test this with a nine-condition diagnostic protocol applied to three architecturally distinct open-source Video-LLMs, with Gemini~2.5~Flash as a frontier reference, and find the accuracy does not come from character tracking. When we change the character named in the question to a different cast member, leaving the video and answer options untouched, the models change their answer only 4--31% of the time, so they are largely ignoring who the question asks about. Breaking that test down by the gender of the swapped name shows why: the models react more when the name is changed to a different-gender character than to a same-gender one (a 13--28 point gap), picking up coarse gender cues but unable to tell same-gender individuals apart. This shallow processing surfaces again when we drop the multiple-choice options and ask the same questions open-endedly: open-source accuracy drops 18--25 points, with none of 151 answers fully correct, versus a 12-point drop for Gemini. Further checks rule out the obvious innocent explanations, adding subtitles, using the most informative frames, or doubling the number of frames all leave character tracking unimproved, so the bottleneck is not how much video the model sees but how it ties that video to the person the question names. We release a diagnostic toolkit for auditing what such benchmark scores actually measure.
arXiv:2607.11600v1 Announce Type: new
Abstract: We propose a collaborative meta-learning framework for distributed Bayesian optimization matching centralized performance without raw-data exchange. We show gradient sharing leaks client observations, with leakage worsening as the search converges and queries concentrate near the optimum. We evaluate a differentially private defense and characterize its privacy-utility trade-off.
arXiv:2509.24653v2 Announce Type: replace
Abstract: Large Language Models (LLMs) excel at multi-hop reasoning in distribution, yet fail on unseen compositions, a phenomenon known as the curse of two-hop reasoning. In this work, we argue that this phenomenon can be attributed to a missing supervision on the bridge entity. We formalize this gap by introducing identity bridge, a minimal supervision that enforces a identity mapping on bridge tokens. Under this supervision, even a one-layer transformer with uniform attention (Emb-MLP) can achieve out-of-distribution (OOD) two-hop generalization. We provide a theoretical analysis demonstrating that identity bridge induces an implicit regularization effect, leading the model to establish a direct subject-to-answer association. From an empirical perspective, the performance of standard GPT-2 models aligns closely with simple Emb--MLP models across varying levels of problem complexity. Finally, analyses of fine-tuned mainstream LLMs indicate that correct two-hop predictions consistently coincide with the establishment of a subject-to-answer relationship, extending our findings to realistic settings.
arXiv:2607.09696v1 Announce Type: new
Abstract: Deep neural models for click-through rate prediction often exhibit a sharp decline in validation performance immediately after the first training epoch despite continued improvement in training loss. This instability restricts effective learning and limits model performance. In this study, we analyze this behavior using large-scale industrial datasets and evaluate practical mitigation strategies. While reducing the learning rate provides only incremental gains, controlling feature sparsity yields substantial improvements. Removing highly sparse features and aggregating infrequent feature values stabilizes training, extends useful learning beyond a single epoch, and improves both offline evaluation metrics and online system performance.
arXiv:2607.11362v1 Announce Type: new
Abstract: We equipped an LLM-based search agent with access to a Boolean retrieval engine to search the MS MARCO V2.1 deduped segment collection used by the TREC 2024 RAG track. Over a standard track subset of 86 topics, and operating under a budget of 100 model calls/topic, the agent achieved an NDCG@10 of 0.6863, which would place it above many dense, sparse, and learned-sparse first-stage retrievers. Ranking is based solely on the density of corpus substrings matching a query, with no requirement for supervised learning, global statistics, or term weights. Formally, the query language expresses a strict subset of the regular languages, with a document's score based on the number and length of matches it contains. Although the results are more exploratory than definitive, because they are based on a single test collection that was publicly available during model training, they suggest that simple pattern matching may be sufficient for agentic search.
arXiv:2607.11606v1 Announce Type: new
Abstract: We study how little extra information is needed to make adversarial language learning possible. In Gold's model of language identification in the limit, a learner is given an enumeration of the strings from an unknown language chosen from a countable language collection. The learner guesses the identity of the language over the course of the enumeration, and it succeeds if, eventually, all of its guesses are the correct language. Classical results of Gold and Angluin show that many natural collections cannot be learned in this way. Recent work on trace colorings, motivated by the success of thinking-trace strategies in language learning, overcomes this obstruction by annotating every symbol of every string with a color. We ask whether the learner really needs this whole sequence of colors, or whether one color at the end of each string (a terminal coloring) is enough for language identification.
We show that just one terminal bit per string is enough for every countable collection of infinite languages. In fact, the colorings can be chosen collection-independently: there is a single assignment of a two-color terminal coloring to every infinite language such that the same preassigned colorings identify every countable subcollection. Thus, in this model, an entire color trace can be compressed to one bit attached to the end of each example.
Our global construction uses transfinite recursion, and we prove that this kind of nonconstructivity is unavoidable for any bounded number of colors. As a notion of constructivity, we use the formalism of Borel maps (a regularity condition satisfied by natural explicit constructions); we show that no global terminal coloring with a finite number of colors defined by a Borel map can identify all countable subcollections. By contrast, known trace-coloring constructions are Borel when encoded as terminal colorings, but require infinitely many colors.
arXiv:2601.06896v2 Announce Type: replace-cross
Abstract: We present TagSpeech, a unified LLM-based framework that utilizes Temporal Anchor Grounding for joint multi-speaker ASR and diarization. The framework is built on two key designs: (1) decoupled semantic and speaker streams fine-tuned via Serialized Output Training (SOT) to learn turn-taking dynamics; and (2) an interleaved time anchor mechanism that not only supports fine-grained timestamp prediction but also acts as a synchronization signal between semantic understanding and speaker tracking. Compared to previous works that primarily focus on speaker-attributed ASR or implicit diarization, TagSpeech addresses the challenge of fine-grained speaker-content alignment and explicitly models "who spoke what and when" in an end-to-end manner. Experiments on AMI and AliMeeting benchmarks demonstrate that our method achieves consistent improvements in Diarization Error Rate (DER) over strong end-to-end baselines, including Qwen-Omni and Gemini, particularly in handling complex speech overlaps. Moreover, TagSpeech employs a parameter-efficient training paradigm in which the LLM backbone is frozen and only lightweight projectors are trained, resulting in strong performance with low computational cost.
arXiv:2601.11878v2 Announce Type: replace-cross
Abstract: To develop a deep-learning method for achieving fast high-resolution MR elastography from highly undersampled data without the need of high-quality training dataset. We first framed the deep neural network representation as a nonlinear extension of the linear subspace model, then used it to represent and reconstruct MRE image repetitions from undersampled k-space data. The network weights were learned using a multi-level k-space consistent loss. To further enhance reconstruction quality, phase-contrast specific magnitude and phase priors were incorporated, including the similarity of anatomical structures and smoothness of wave-induced harmonic displacement. Experiments were conducted using both 3D gradient-echo spiral and multi-slice spin-echo spiral MRE datasets. Compared to the conventional linear subspace-based approaches, the nonlinear network representation method was able to produce superior image reconstruction with suppressed noise and artifacts from a single in-plane spiral arm per MRE repetition (e.g., 2mm isotropic resolution in 1 min with a total R=10), yielding comparable stiffness estimation to the fully sampled data. This work demonstrated the feasibility of using deep network representations to model and reconstruct MRE images from highly-undersampled data, a nonlinear extension of the subspace-based approaches.
arXiv:2607.10604v1 Announce Type: new
Abstract: Large language models (LLMs) are increasingly used to generate long-form answers for knowledge-intensive tasks, but users often struggle to decide which parts of a response deserve scrutiny, why they may be unreliable, and what to do next. Prior work on uncertainty communication has largely focused on making uncertainty visible through cues such as confidence scores, leaving less support for the broader process of managing uncertainty distributed across a long response. Through a formative study, we examine how users manage such uncertainty across three stages: interpretation, evaluation, and decision. Based on these insights, we derive design guidelines that address both stage-specific and cross-stage needs: uncertainty target representation, evaluative explanation, response guidance, and interactive presentation. We instantiate these guidelines in U-Lens, an uncertainty-management support system that organizes uncertain information in long-form responses into contextual inspection targets, prioritizes them for attention, and connects each target with evaluative context and response options. We evaluated U-Lens in a controlled within-subjects study with 18 participants, comparing it against a confidence-cue baseline. Our results show that U-Lens improved verification efficiency and effort allocation, lowered perceived workload, and strengthened perceived support across interpretation, evaluation, and decision stages. This work reframes uncertainty support for generative AI from presenting isolated, text-centered cues toward supporting the user-centered process of interpreting, evaluating, and acting on uncertain information.
Mizzle: A Complete Concurrent Incorrectness Logic for Preventing False Alarms in Agentic Bug Finding
arXiv:2607.11611v1 Announce Type: new
Abstract: Large language models are increasingly used to find bugs in real-world programs, but they also produce a flood of false alarms that waste developers' time. We propose a method to prevent these false alarms by requiring an LLM to accompany each bug report with a machine-checked proof, in a program logic, that the reported bug is real. We follow the approach of incorrectness logics, whose under-approximate reasoning establishes that a claimed behavior is genuinely reachable, and hence a true positive. In our case, however, the logic must model a realistic programming language, have a mechanization so that proofs can be checked, and be complete, so that no real bug is ruled out for want of a derivation.
We present Mizzle, an incorrectness separation logic for concurrent programs written in a substantial subset of OCaml, parametric in the notion of incorrectness. We mechanize Mizzle in the Rocq proof assistant on top of the Iris framework, and we prove that it is both sound (that is, it never justifies a false alarm) and complete (that is, every incorrect execution admits a derivation). We instantiate Mizzle with three notions of incorrectness: stuckness (triggering undefined behavior), the non-linearizability of a data structure, and the presence of a race. As a proof of concept, we illustrate how an LLM can use Mizzle in order to certify the existence of a bug.
arXiv:2607.09722v1 Announce Type: new
Abstract: We investigate a fractional gravity model in which both the Hubble parameter and the gravitational constant evolve dynamically due to fractional renormalization-group effects. The model incorporates a scalar field coupled to a time-varying $G$, generating nonlocal corrections characteristic of fractional--action cosmology. Analytical and numerical solutions reveal oscillatory regimes, cyclic phases, and rapid variations with implications for BBN and early-universe evolution. A robust numerical framework is developed to integrate the regularized system and compare the resulting $H(z)$ evolution with observational data from the Hubble parameter, baryon acoustic oscillations, type Ia supernovae, gravitational lensing, and black hole shadows, thereby enabling a consistent reconstruction of cosmographic quantities. A Bayesian analysis shows that the Fractional model with $\mu=0$ is the only statistically viable variant. The inferred Hubble parameter is stable across models ($h\simeq 0.72$), while the fractional parameters are significantly better constrained in the $\mu=0$ case ($\alpha=1.20^{+0.25}_{-0.14}$, $\zeta=0.43^{+0.39}_{-0.29}$). The dynamical sector yields $m=30.8^{+28.0}_{-20.9}$ and $\Gamma=108.3\pm1.1$, leading to a positive discriminant and a well-determined relaxation timescale $\tau_{\rm rel}\simeq 9$ Gyr, confirming an overdamped regime. Although the $\mu=0$ model attains a slightly lower $\chi^2_{\min}$ than $\Lambda$CDM, the BIC strongly favors $\Lambda$CDM due to its smaller parameter space. Overall, the model reproduces late-time acceleration and mimics $\Lambda$CDM while introducing distinctive cosmographic signatures. The dynamical systems analysis clarifies the stability structure and parameter dependence, indicating that fractional nonlocal corrections may offer new pathways toward addressing the $H_0$ and $S_8$ tensions.
arXiv:2607.11136v1 Announce Type: new
Abstract: Diffusion models are increasingly deployed as production visual-generation services, where serving high-resolution image and long video generation is often limited by GPU memory. Popular memory-saving techniques such as weight offloading, sharding, and VAE slicing are often not practical because they tend to introduce significant performance overhead. In this paper, we present Xema, a memory-efficient diffusion serving system that exploits predictable tensor lifetimes for trace-guided memory optimization. For each request template, Xema derives an offline memory trace to identify short memory-pressure intervals and applies memory mitigation only within these intervals and only by the amount needed to fit the target GPU budget. Xema further constructs a static memory layout for tensors with predictable lifetimes, reducing fragmentation-induced reserved memory and making offline memory reasoning reliable at runtime. Built on this memory optimization layer, Xema introduces an offline planner that jointly selects parallelism, concurrency, and memory control under GPU memory and SLO constraints. The selected plan is stored in a plan table and directly used by the online serving runtime. We implement Xema on production diffusion pipelines and evaluate it with Flux.2, CogVideoX-5B, and LTX-2. Compared with existing serving configurations, Xema improves SLO attainment by up to 3.7x and reduces planning cost from 6.3 hours to 197 seconds compared with grid search.
arXiv:2607.11633v1 Announce Type: new
Abstract: Service robots in retail stores increasingly rely on cascaded speech pipelines (STT-LLM-TTS), yet many customer-robot interactions are initiated or guided by nonverbal behaviors such as approaching, waving, pointing, or showing items. This paper studies such cues in a real-world store deployment with a teleoperated humanoid robot and shows that a non-negligible portion of robot turns are triggered by nonverbal behaviors rather than spoken input, revealing a limitation of audio-only dialogue systems. In a 6-day in-the-wild deployment, 15.3\% of robot utterances were initiated by users' nonverbal behaviors rather than spoken input. Based on an analysis of observed customer behaviors, we define a set of frequent, service-relevant nonverbal cues and develop a real-time multi-person, multi-label recognizer that runs online from video. We then propose a dialogue framework that conditions LLM-based utterance generation on recognized nonverbal cue tokens, and optionally leverages a vision-language model when items are shown, enabling proactive robot responses without hand-crafted rules. We evaluate the approach offline on nonverbal-triggered turns and demonstrate an online prototype that reacts to users' nonverbal cues in real time.
arXiv:2607.11818v1 Announce Type: new
Abstract: We introduce MM-ToolSandBox, a benchmark and evaluation framework for visually grounded tool-calling agents. The framework provides a stateful execution environment spanning 500+ tools across 16 application domains, supporting multi-image, multi-turn tasks where agents must ground progressively arriving visual inputs into executable tool calls while handling realistic conversational phenomena (goal revisions, error corrections, state mutations). An automated scenario generation pipeline produces diverse, visually grounded scenarios through information-flow-guided planning and multi-stage quality filtering, yielding 258 human-verified nominal scenarios and 50 variants targeting interactive UI applications. Evaluating 12 state-of-the-art models, from 4B open-weight to frontier proprietary systems, shows that current models still lack robust visual tool-calling capability: even the best model achieves below 50% success rate. Our failure analysis further reveals that visual precision, not only planning, is a primary bottleneck for capable models: 53% of failures stem from incorrect information extraction from images despite otherwise correct task workflows. A planning-to-precision crossover emerges with scale: smaller models fail at deciding what to do, while larger models fail at perceiving what they see, suggesting fundamentally different research directions for improving models at different capability levels. The framework and the benchmark are publicly available at https://github.com/apple/ml-mmtoolsandbox
arXiv:2510.08233v3 Announce Type: replace
Abstract: Diffusion large language models (dLLMs) are promising alternatives to autoregressive large language models (AR-LLMs), as they potentially allow higher inference throughput. Reinforcement learning (RL) is crucial to enabling dLLMs to achieve performance comparable to that of AR-LLMs on important tasks, such as reasoning. However, RL algorithms well-suited to dLLMs' unique characteristics have yet to be developed. This paper proposes Distribution Matching Policy Optimization (DMPO), a principled and theoretically grounded RL fine-tuning method specifically designed to enhance the reasoning capabilities of dLLMs by matching the dLLM policy distribution to the optimal, reward-tilted one through cross-entropy optimization. We identify a key implementation challenge with small training batch sizes and propose several effective solutions based on a novel weight baseline subtraction technique. DMPO exhibits superior performance on multiple reasoning benchmarks without supervised fine-tuning, achieving up to a $39.63$ percentage-point improvement in accuracy over prior non-DMPO RL baselines and $67.97$ percentage points over the base model, underscoring the effectiveness of the distribution-matching framework. Our code is available at https://github.com/yuchen-zhu-zyc/DMPO.