arXiv:2607.18101v1 Announce Type: new
Abstract: On-device model adaptation is essential to enable lifelong personalization on resource-constrained hardware, but compute, power, and memory limitations of such devices make end-to-end backpropagation impractical for modern deep neural networks. This work proposes a heterogeneous adaptation pipeline that repurposes a commercial edge AI inference accelerator, Hailo-8L, for frozen-backbone feature extraction during on-device training. The computational graph is partitioned so that the pre-trained backbone is quantized to INT8 and run on the accelerator, while only a lightweight FP32 classification head is fine-tuned on the host CPU, enabling frequent, energy-efficient in-field updates with most weights remaining fixed. Across multiple architectures and datasets, this pipeline achieves up to 15.4x faster wall-clock training time compared to a Raspberry Pi 5 CPU baseline, offers competitive throughput in favorable settings, and consistently reduces energy per sample. Post-training quantization restoration is shown to be crucial for preserving the quality of accelerator-generated features and mitigating accuracy loss in quantization-sensitive architectures. Overall, the results demonstrate a practical approach to efficient on-device adaptation using inference-oriented edge accelerators. The implementation is available at https://github.com/MatPiech/accelerator-training.
Science Journals
arXiv:2607.16866v1 Announce Type: cross
Abstract: Quantum-measurement reduction contains two distinct global-optimization layers: a continuous problem of splitting an observable and allocating shots within a fixed measurement dictionary, and a nonconvex outer problem of designing the dictionary and calibrating its data-driven uncertainty model. We solve the inner layer globally and certifiably as a second-order cone program (SOCP), and use RANGE, a robust adaptive nature-inspired global optimizer, for the combinatorial and statistical outer layer. For any declared set of contexts, per-shot costs, score functions, and covariance model, the SOCP returns the minimum leading shot cost among unbiased linear stratified estimators. The conic dual supplies an independently checkable lower-bound witness; after feasibility repair, an external verifier recomputes $L \le \Phi \le U$ from stored data without trusting the optimizer. Pilot measurements yield simultaneous finite-sample covariance brackets, and the dual becomes a pricing oracle for omitted contexts. Discrete RANGE searches covering sub-dictionaries, Pareto compression fronts, and candidate contexts; continuous RANGE performs an explicitly empirical, coverage-constrained calibration of covariance-radius models, while rigorous certificates retain the proved finite-sample radius. RANGE compresses molecular context dictionaries by 4.3-6.1x at 0.2-2.1% certified-frontier excess. Standard strategies are exactly optimal for H2 yet leave factors of 2.1-7.7 in shots within their own settings by H2O. Adding fully commuting contexts lowers the certified optimum by up to 56%; on 29-35-qubit production f-element Hamiltonians under a declared Hartree-Fock-proxy covariance model, the capped-dictionary enlargement saves 31-70% of the shots, and transformations reducing block-encoding cost need not reduce sampling cost.
arXiv:2607.17933v1 Announce Type: cross
Abstract: We present CARDAL (Curvature-Aware Rank-Adaptive Distributed Augmented Lagrangian), a distributed multi-GPU solver for large-scale semidefinite programs (SDPs) based on a rank-adaptive Burer-Monteiro factorization and an augmented Lagrangian method. At fixed ranks, a matrix-free L-BFGS method with negative-curvature corrections targets an approximate Euclidean second-order stationary point of the factored augmented Lagrangian. A reverse multiplier shift turns a negative dual-slack direction into exact negative curvature after rank expansion, and a small joint rank-lift problem selects a batched low-rank correction. A verified slack lower bound provides an a posteriori approximate KKT certificate. Our analysis establishes generic global-optimality guarantees for heterogeneous products of PSD cones at per-block ranks near the Barvinok-Pataki scale, together with a finite-accuracy counterpart under blockwise cost smoothing. For scalable execution, CARDAL distributes constraint rows, factor columns, and PSD blocks over a Constraint x Rank x Cone device mesh. The primal residual, gradient, Hessian-vector products, and slack matrix-vector products are evaluated using device-local operations and axis-wise collectives. On the Mittelmann benchmark, CARDAL exhibits stronger robustness than existing low-rank GPU approaches under a uniform accuracy standard. Experiments on large-scale SDP relaxations from robotics, electronic structure, and Max-Cut demonstrate the complementary scaling regimes of the three distribution axes, with observed wall-clock speedups of up to 4x on four H100 GPUs.
arXiv:2607.18115v1 Announce Type: new
Abstract: Physical artificial intelligence (AI) systems involve distributed sensing agents with embedded AI models that must coordinate to perceive, reason, and act in networked environments. Transmitting raw sensor data incurs significant communication overhead, latency, and redundancy. While semantic communication (SC) mitigates these challenges by transmitting task-relevant information, existing deep learning-based joint source-channel coding approaches exhibit limited adaptability, poor out-of-distribution generalization, and scalability challenges. To address these limitations, this paper proposes a framework for compositional semantic communication (CSC), enabling heterogeneous physical AI sources to transmit semantic representations (SRs) that compose meaningfully at a base station (BS) or edge server for remote inference. First, a category-theoretic measure of compositional semantics is developed to quantify each device's contribution to inference tasks beyond mutual information. Second, Grothendieck topologies and presheaves formalize semantic composition across devices, ensuring consistency and task relevance. Building on these foundations, multi-device coordination is formulated as a Stackelberg game in which devices commit to encoding strategies and the BS optimally composes received SRs. An ADMM-based algorithm computes equilibrium signaling strategies. Equilibrium existence is established under mild conditions and is Pareto optimal when compositional information yields increasing collective benefit. Simulation results demonstrate that the proposed approach achieves up to 17% bandwidth reduction and 53% lower end-to-end latency than cooperative multi-agent, distributed gradient descent, and uniform-selection CSC baselines while maintaining 85% inference accuracy across diverse autonomous driving scenarios.
arXiv:2510.22228v2 Announce Type: replace
Abstract: Layer pruning has emerged as a widely adopted technique for improving the efficiency of large language models (LLMs). Although existing methods demonstrate strong performance retention on general knowledge tasks, their effect on long-chain reasoning, a more brittle yet crucial capability, remains largely unexplored. In this work, we study the impact of layer pruning on long-chain reasoning through the lens of test-time scaling, a key mechanism in modern LLMs that enables strong reasoning capacity by allocating more computation at inference time. With extensive experiments, we demonstrate that pruning even one or two layers can severely impair test-time scaling, with performance collapsing drastically on long reasoning benchmarks even when performance on knowledge-intensive and shallow reasoning tasks remains stable. Furthermore, we find that standard supervised fine-tuning remedies fail to recover test-time scaling once it has deteriorated. Through in-depth analyses, we identify the mechanisms underlying this fragility of test-time scaling and highlight the fundamental risks of applying layer pruning to reasoning-intensive LLMs. These findings call for a rethinking of layer pruning strategies and provide insights for developing methods that preserve the robustness of reasoning. We open-source the codebase in \href{https://github.com/keyu-wang-2002/Layer-Pruning-Harms-Inference-Scaling}{https://github.com/keyu-wang-2002/Layer-Pruning-Harms-Inference-Scaling}.
arXiv:2510.23472v2 Announce Type: replace
Abstract: Chip placement is a vital stage in modern chip design, and black-box optimization (BBO) has been applied to it for decades. Early BBO efforts, however, were limited by immature problem formulations and inefficient algorithm designs, leading to worse efficiency, quality, and scalability than mainstream analytical methods. Recent advances in BBO have shown strong potential, but a unified, BBO-specific benchmark for thoroughly assessing various problem formulations and BBO algorithms is lacking. To fill this gap, we propose BBOPlace-Bench, the first benchmark tailored for evaluating and developing BBO algorithms for chip placement. It integrates three BBO problem formulations and offers a modular, flexible framework that enables users to seamlessly implement, test, and compare their own algorithms. It aggregates representative modern chip cases and standardizes their formats, providing uniform and comprehensive information to support BBO optimization. Moreover, it integrates representative BBO algorithm families, including simulated annealing, population-based search (including GA, CMA-ES, and PSO), and Bayesian optimization, and systematically evaluates their performance across different problem formulations using key chip-placement metrics. We position these experiments primarily as illustrative case studies under a shared evaluation protocol, including common benchmark instances, metric definitions, evaluation pipeline, and search budgets. Under this protocol, some BBO configurations (e.g., GA under the mask-guided optimization formulation) are competitive with representative analytical and reinforcement learning baselines. BBOPlace-Bench not only facilitates the development of efficient BBO-driven solutions for chip placement but also broadens the practical application scenarios urgently needed by the BBO community.
arXiv:2603.22169v3 Announce Type: replace
Abstract: We introduce VersualRL, a closed-loop framework for task-level robot planning that uses visual execution feedback to iteratively refine executable Behavior Trees through structured natural-language critique. VersualRL combines a vision-language model critic, which analyzes visual observations and Behavior Tree execution traces, with a large language model actor that performs discrete and interpretable policy updates. During physical deployment, both foundation models remain frozen, while the executed Behavior Tree is updated at the symbolic level without online gradient-based policy optimization. This enables transparent reasoning, explicit causal feedback, and human-interpretable policy evolution. We validate VersualRL on a real mobile robot performing a multi-stage navigation and manipulation task under execution uncertainty. Experimental results show that the framework supports explainable policy improvements, closed-loop adaptation to execution failures, and successful deployment on physical robotic systems.
arXiv:2603.22590v2 Announce Type: replace
Abstract: With the increasing deployment of automated and agentic systems, ensuring the adversarial robustness of automatic speech recognition (ASR) models has become highly relevant. We observe that changing the precision of an ASR model during inference reduces the likelihood of adversarial attacks to succeed. We take advantage of this fact to make models more robust simply by randomly sampling the precision during prediction. Moreover, this insight can be turned into an adversarial example detection strategy by implementing a simple Gaussian classifier that thresholds the differences between outputs of models run with different precision. To further enhance security boundaries, we combine the approach with an existing uncertainty-based defense mechanism, which forces adaptive adversaries to introduce highly perceptible noise to bypass detection. An experimental analysis across various ASR models, languages, and attack types demonstrates a significant increase in adversarial robustness, competitive detection capabilities, and resistance to adaptive threats.
arXiv:2603.27752v2 Announce Type: replace
Abstract: Large language models can still hallucinate in retrieval-augmented generation (RAG), producing claims that are unsupported by or conflict with the retrieved context. Detecting such errors remains challenging when faithfulness is judged solely against the retrieved context: many existing detectors return holistic answer-level scores, while others target open-domain factuality or fail to provide evidence-grounded diagnostics.
We present RT4CHART, a retromorphic testing framework for context-faithfulness assessment. RT4CHART decomposes an answer into independently verifiable claims, performs hierarchical local-to-global verification against the retrieved context, and assigns each claim one of three labels: entailed, contradicted, or baseless. It further maps these claim-level decisions back to specific answer spans and returns explicit context-side evidence, enabling fine-grained auditing rather than opaque scoring.
We evaluate RT4CHART on RAGTruth++ (408 samples) and our re-annotated RAGTruth-Enhance (2,675 samples). RT4CHART achieves the best answer-level hallucination-detection F1 score among the evaluated baselines. On RAGTruth++, it attains a precision of 0.845, a recall of 0.718, and an F1 score of 0.776, representing an 83% relative improvement over the strongest baseline. It also achieves a span-level F1 score of 47.5% on RAGTruth-Enhance.
Ablation studies show that claim-based local processing drives most of the observed improvement, while global verification provides selective benefits across datasets. Finally, our re-annotation identifies 1.68X more hallucination cases than the original labels, suggesting that commonly used benchmarks substantially underestimate the prevalence of hallucination.
arXiv:2603.27958v2 Announce Type: replace
Abstract: Analogical reasoning tests a fundamental aspect of human cognition: mapping the relation from one pair of objects to another. Existing evaluations of this ability in multimodal large language models (MLLMs) overlook the ability to compose rules from multiple sources, a critical component of higher-order intelligence. To close this gap, we introduce CARV (Compositional Analogical Reasoning in Vision), a novel task together with a 5,500-sample dataset as the first diagnostic benchmark. We extend the analogy from a single pair to multiple pairs, which requires MLLMs to extract symbolic rules from each pair and compose new transformations. Evaluation on the state-of-the-art MLLMs reveals a striking performance gap: even Gemini-2.5 Pro achieving only 40.4% accuracy, far below human-level performance of 100%. Diagnostic analysis shows two consistent failure modes: (1) decomposing visual changes into symbolic rules, and (2) maintaining robustness under diverse or complex settings, highlighting the limitations of current MLLMs on this task.
arXiv:2603.29219v2 Announce Type: replace
Abstract: Sign language is the primary approach of communication for the Deaf and Hard-of-Hearing (DHH) community. While there are numerous benchmarks for high-resource sign languages, low-resource languages like Arabic remain underrepresented. Currently, there is no publicly available dataset for Syrian Arabic Sign Language (SyArSL). To overcome this gap, we introduce SyriSign, a dataset comprising 1500 video samples across 150 unique lexical signs, designed for text-to-SyArSL translation tasks. This work aims to reduce communication barriers in Syria, as most news are delivered in spoken or written Arabic, which is often inaccessible to the deaf community. We evaluated SyriSign using three deep learning architectures: MotionCLIP for semantic motion generation, T2M-GPT for text-conditioned motion synthesis, and SignCLIP for bilingual embedding alignment. Experimental results indicate that while generative approaches show strong potential for sign representation, the limited dataset size constrains generalization performance. We will release SyriSign publicly, hoping it serves as an initial benchmark.
arXiv:2607.18181v1 Announce Type: new
Abstract: Battery-free Internet of Things (IoT) requires iterative design of vibration energy harvesters (VEHs) under coupled physical constraints, while LLMs are emerging as interface layers for engineering workflows. However, existing engineering benchmarks primarily assess final artifact validity, offering limited insights into how LLMs behave across different stages of coupled physical design. We introduce VEHBench, an engineering-native diagnostic benchmark for LLM-assisted VEH design, featuring 763 literature-grounded tasks scored by an analytical physical oracle. VEHBench evaluates four design roles: specification triage, verifier-guided search, corrupted-state recovery, and policy-conditioned selection. Experimental results reveal that LLM capability is strongly stage-dependent: no single model consistently dominates the entire workflow, and response-control profiles expose distinct behavioral patterns across design roles. VEHBench thus provides a stage-aware foundation for evaluating, selecting, routing, and improving verifier-grounded engineering LLMs. The benchmark artifact is available at https://huggingface.co/datasets/AnonymousVehbench/vehbench
arXiv:2607.18184v1 Announce Type: new
Abstract: We present anyakrakusuma, an open-source Python library that solves the discrete static Schr\"{o}dinger bridge problem, the entropically regularized counterpart of optimal transport, through a log-domain Sinkhorn--Knopp iteration and reconstructs the entropic interpolation between two empirical point clouds. The solver is paired with a diagnostic pipeline that characterizes the optimal coupling and the intermediate distributions through information-theoretic and geometric measures. We exercise the library on four idealized planar cases spanning a circle-to-circle dilation, a spiral-to-mixture fragmentation, a rigid reorientation of two moons, and a Lissajous-to-trefoil deformation. The log-domain formulation is necessary rather than merely convenient at the parameters studied, where the cost-to-regularization ratio reaches four hundred and the Gibbs kernel underflows double precision across most of its range; the iteration nonetheless attains a marginal residual of $10^{-9}$ and unit marginal fidelity in every case. Residual histories decay geometrically over approximately eight decades at per-iteration contraction factors between $0.966$ and $0.976$, which are local rates near the fixed point that lie many orders of magnitude below the worst-case Hilbert-metric bound. The covariance analysis recovers an imposed ninety-degree reorientation to within $0.07^\circ$, roughly forty times smaller than its uncertainty, across a masked interval of near-isotropy on which the principal axis is unobservable. The diagnostics are reported with explicit attention to the regimes in which each is well defined, including the differential entropy, which is meaningful only on the open interpolation interval. The presented cases are constructed rather than measured; quantitative application to empirical point clouds requires further study.
arXiv:2505.02722v2 Announce Type: replace
Abstract: Although large language models (LLMs) have demonstrated impressive reasoning capabilities across general domains, their effectiveness in real-world clinical practice remains limited. This is likely due to their insufficient exposure to real-world clinical data during training, as such data is typically not included due to privacy concerns. To address this, we propose enhancing the clinical reasoning capabilities of LLMs by leveraging real-world clinical data. We constructed reasoning-intensive questions from a nationwide sepsis registry and fine-tuned Phi-4 on these questions using reinforcement learning, resulting in C-Reason. C-Reason exhibited strong clinical reasoning capabilities on the in-domain test set, as evidenced by both quantitative metrics and expert evaluations. Furthermore, its enhanced reasoning capabilities generalized to a sepsis dataset involving different tasks and patient cohorts, an open-ended consultations on antibiotics use task, and other diseases. Future research should focus on training LLMs with large-scale, multi-disease clinical datasets to develop more powerful, general-purpose clinical reasoning models.
arXiv:2604.27967v2 Announce Type: replace
Abstract: Background: We introduce StructGP, a continuous-time multi-task Gaussian process that couples process convolutions with differentiable structure learning to uncover a sparse, ordered directed acyclic graph (DAG) of inter-variable dependencies while preserving principled uncertainty. We further propose LP-StructGP, which augments StructGP with latent pathways-shared, temporally shifted trajectories inferred via subject-specific coupling filters and a softmax gating mechanism-to capture cross-patient progression patterns. Both models are trained under sparsity and acyclicity constraints using scalable low-rank updates using likelihood-based objectives. Results: In simulations, graph recovery improved with cohort size, with the median Structural Hamming Distance reaching zero at the largest cohort size, while pathway assignments showed high Adjusted Rand Index. Our analysis establishes that the ordered StructGP graph is identifiable from the population marginal likelihood. On a MIMIC-IV septic shock cohort (n=1,008; norepinephrine, creatinine, mean blood pressure), StructGP improves short-horizon (6 h) forecasting over independent-task baselines (average RMSE 0.68 [95% CI: 0.63-0.74] vs. 0.88 [0.83-0.94]) and, with 15 additional inputs, markedly outperforms unstructured kernels (0.63 [0.58-0.69] vs. 3.02 [2.85-3.18]) with superior calibration (coverage 0.96 vs. 0.84). For long horizons (up to 6 days), LP-StructGP further reduces error for creatinine (RMSE 0.95 [0.88-1.03] vs. 1.17 [1.08-1.25]) and improves overall coverage (0.93 [0.93-0.94] vs. 0.91 [0.91-0.92]). On the PhysioNet Challenge, StructGP attains competitive accuracy (MAE 3.72e-2) relative to a strong published graph neural model. Conclusion: These results show that structured process convolutions with latent pathways deliver interpretable, scalable, and well-calibrated forecasting for irregular clinical time series.
arXiv:2606.14409v2 Announce Type: replace
Abstract: In this report, we present Hy-Embodied-0.5-VLA, abbreviated as HyVLA-0.5, an end-to-end system that spans the full robot learning stack: data collection, model design, continued pre-training and supervised fine-tuning, RL post-training, and real-world deployment. Each component serves a distinct role in this stack.
arXiv:2604.00594v2 Announce Type: replace
Abstract: As the focus in LLM-based coding shifts from static single-step code generation to multi-step agentic interaction with tools and environments, understanding which tasks will challenge agents and why becomes increasingly difficult. This is compounded by current practice: agent performance is typically measured by aggregate pass rates on benchmarks, but single-number metrics obscure the diversity of tasks within a benchmark. We present a framework for predicting success or failure on individual tasks tailored to the agentic coding regime. Our approach augments Item Response Theory (IRT) with rich features extracted from tasks, including issue statements, repository contexts, solutions, and test cases, and introduces a novel decomposition of agent ability into LLM and scaffold ability components. This parameterization enables us to aggregate evaluation data across heterogeneous leaderboards and accurately predict task-level performance for unseen benchmarks, as well as unseen LLM-scaffold combinations. Our methods have practical utility for benchmark designers, who can better calibrate the difficulty of their new tasks without running computationally expensive agent evaluations.
arXiv:2604.01457v2 Announce Type: replace
Abstract: Large language models are often not just wrong, but \emph{confidently wrong}: when they produce factually incorrect answers, they tend to verbalize overly high confidence rather than signal uncertainty. Such verbalized overconfidence can mislead users and weaken confidence scores as a reliable uncertainty signal, yet its internal mechanisms remain poorly understood. We present a circuit-level mechanistic analysis of this inflated verbalized confidence in LLMs, organized around three axes: capturing verbalized confidence as a differentiable internal signal, identifying the circuits that causally inflate it, and leveraging these insights for targeted inference-time recalibration. Across two instruction-tuned LLMs on three datasets, we find that a compact set of MLP blocks and attention heads, concentrated in middle-to-late layers, consistently writes the confidence-inflation signal at the final token position. We further show that targeted inference-time interventions on these circuits substantially improve calibration. Together, our results suggest that verbalized overconfidence in LLMs is driven by identifiable internal circuits and can be mitigated through targeted intervention.
arXiv:2604.01700v2 Announce Type: replace
Abstract: Video frame interpolation aims to synthesize realistic intermediate frames between given endpoints while adhering to specific motion semantics. While recent generative models have improved visual fidelity, they predominantly operate in a unidirectional manner, lacking mechanisms to self-verify temporal consistency. This often leads to motion drift, directional ambiguity, and boundary misalignment, especially in long-range sequences. Inspired by the principle of temporal cycle-consistency in self-supervised learning, we propose a novel bidirectional framework that enforces symmetry between forward and backward generation trajectories. Our approach introduces learnable directional tokens to explicitly condition a shared backbone on temporal orientation, enabling the model to jointly optimize forward synthesis and backward reconstruction within a single unified architecture. This cycle-consistent supervision acts as a powerful regularizer, ensuring that generated motion paths are logically reversible. Furthermore, we employ a curriculum learning strategy that progressively trains the model from short to long sequences, stabilizing dynamics across varying durations. Crucially, our cyclic constraints are applied only during training; inference requires a single forward pass, maintaining the high efficiency of the base model. Extensive experiments show that our method achieves state-of-the-art performance in imaging quality, motion smoothness, and dynamic control on both 37-frame and 73-frame tasks, outperforming strong baselines while incurring no additional computational overhead.
arXiv:2604.02142v2 Announce Type: replace
Abstract: We consider energy-aware planning for an unmanned aerial vehicle (UAV) and unmanned ground vehicle (UGV) team operating in a stochastic environment. The UAV must visit a set of air points in minimum time while respecting energy constraints, relying on the UGV as a mobile charging station. Unlike prior work that assumed deterministic travel times or used fixed robustness margins, we model travel times as random variables and bound the probability of failure (energy depletion) across the entire mission to a user-specified risk level. We formulate the problem as a Mixed-Integer Program and propose PRO-SPECT, a polynomial-time algorithm that generates risk-bounded plans. The algorithm supports both offline planning and online re-planning, enabling the team to adapt to disturbances while preserving the risk bound. We provide theoretical results on solution feasibility and time complexity. We also demonstrate the performance of our method via numerical comparisons and simulations.
arXiv:2604.03614v2 Announce Type: replace
Abstract: Global optimization of black-box functions from noisy samples is a fundamental challenge in machine learning and scientific computing. Traditional methods such as Bayesian Optimization often converge to local minima on multi-modal functions, while gradient-free methods require many function evaluations. We present a novel neural approach that learns to find global minima through iterative refinement. Our model takes noisy function samples and their fitted spline representation as input, then iteratively refines an initial guess toward the true global minimum. Trained on randomly generated functions with ground truth global minima obtained via exhaustive search, our method achieves a mean error of 8.05 percent on challenging multi-modal test functions, compared to 36.24 percent for the spline initialization, a 28.18 percent improvement. The model successfully finds global minima in 72 percent of test cases with error below 10 percent, demonstrating learned optimization principles rather than mere curve fitting. Our architecture combines encoding of multiple modalities including function values, derivatives, and spline coefficients with iterative position updates, enabling robust global optimization without requiring derivative information or multiple restarts.
arXiv:2604.03829v2 Announce Type: replace
Abstract: Mamba is an emerging, complex workload with various short-range and long-range dependencies, nonlinearities, and elementwise computations that are unable to run at near-peak speeds on modern hardware. Specifically, Mamba's complex dependency graph makes fusion across its full operator cascade difficult, leaving substantial inter-operator memory traffic on the table. To address these challenges, we propose Mambalaya, a novel reconfigurable accelerator that leverages fusion to overcome the limitations of Mamba. We use the recently proposed cascade-of-Einsums abstraction to characterize Mamba's full computational structure, then apply the extended Einsum framework to systematically explore inter-Einsum fusion opportunities. This principled approach yields a series of fusion mappings that reduce off-chip inter-Einsum traffic. These mappings are supported by the underlying Mambalaya architecture. Mambalaya achieves a layer performance speedup of 4.9$\times$ for prefill and 1.9$\times$ for generation over MARCA. In prefill-dominated scenarios, it achieves up to 1.5$\times$ over a recent fine-grained, memory-aware fusion accelerator for Mamba.
arXiv:2603.07598v2 Announce Type: replace
Abstract: Chain-of-thought (CoT) reasoning improves problem solving, but long think traces increase inference cost. Existing CoT compression methods usually optimize completion-level length. For structured thinking models, however, a completion contains both a think segment and an answer segment, so completion-level compression can save tokens by compressing not only the CoT but also the answer. We call this failure mode answer drift. We propose Segment-wise CoT Compression with Answer Alignment (SCA), an answer-preserving think-compression method. SCA parses completions into functional segments, routes compression rewards only to successful think tokens, and protects answer tokens through length and distribution alignment to a frozen base model. Experiments show that, across datasets from multiple domains, SCA achieves state-of-the-art-level chain-of-thought compression while preserving the base model's performance and answer alignment. Training data and code are included in the supplementary code and data package.
arXiv:2603.08173v2 Announce Type: replace
Abstract: Quantization has become essential for the efficient deployment of speech processing systems. Although widely studied, most existing quantization methods were developed for vision and NLP architectures, while the specific challenges of audio signals remain largely overlooked. In particular, we show that audio activations can exhibit large calibration ranges, leading to significant information loss when standard calibration techniques are applied. To address this, we propose ESC, an Evolution Strategy-based Calibration method that formulates activation scaling as an optimization problem and solves it using a two-step local-global scheme driven by an evolution strategy. ESC enables unaltered performance under full INT8 quantization and is the first calibration method to achieve near-lossless performance for full INT4 quantization across multiple speech tasks. Integrating ESC with PTQ methods further reduces performance loss, achieving a 1% relative accuracy degradation on the AST model.
arXiv:2604.04788v2 Announce Type: replace
Abstract: Large language models (LLMs) could produce systematically misaligned output, from hallucinated citations to strategic deception of evaluators, yet these phenomena are studied by separate communities with incompatible terminology. We propose a unified taxonomy organized along three complementary dimensions: degree of goal-directedness (behavioral to strategic deception), object of deception, and mechanism (fabrication, omission, or pragmatic distortion). Applying this taxonomy to 50 existing benchmarks reveals that every benchmark tests fabrication while pragmatic distortion, attribution, and capability self-knowledge remain critically under-covered, and strategic deception benchmarks are nascent. We offer concrete recommendations for developers and regulators, including a minimal reporting template for positioning future work within our framework.