This study addresses on-device inference bottlenecks of Transformer models on Tenstorrent's Tensix architecture and proposes an operator fusion strategy that enhances data locality. RMSNorm is fused with matrix multiplication in self-attention and in the FFN, enabling back-to-back execution of memory-bound and compute-bound operators in on-chip SRAM to significantly reduce DRAM reads/writes of intermediate results and scheduling overhead. To support multi-core parallelism, a NoC-based multicast mechanism is leveraged in which row/column master nodes efficiently distribute inputs and weights across the core mesh, alleviating DRAM bandwidth contention. Experiments on the Wormhole platform with Qwen2.5-0.5B, Qwen3-0.6B, and Qwen3-4B show up to 37.44% latency reduction for attention and 15.89% for MLP, with up to 7.91% reduction per decoder layer, while Pearson Correlation Coefficient (PCC) remains above 98.75%, confirming significant end-to-end efficiency gains under numerical consistency.

Standard benchmarks report aggregate accuracy, but practitioners need to know which specific capabilities a model lacks. We introduce FailureScope, a behavioral-diagnosis method that clusters evaluation probes by their cross-model pass/fail patterns (leave-one-model-out, LOMO), and show it yields stable, interpretable failure taxonomies across three regimes usually studied separately: single-turn benchmarks, multi-turn dialogue, and adversarial agent attacks. On 2,664 single-turn tasks across 18 models, taxonomy-conditioned sampling reaches Kendall's tau = 0.81 at 50 tasks (versus 0.34 for random selection), and cross-model failure prediction reaches AUC 0.88. The same primitive recovers interpretable clusters on a 363-task multi-turn corpus and on 630 adversarial agent traces, where it exposes a meta-failure mode: a 73-100 percentage-point gap between LLM-judge ASR and real execution. Cluster cohesion remains strong across all three regimes, which we take as evidence that behavioral clustering is a portable diagnosis primitive that generalizes beyond any single benchmark. We release the pipeline, three annotated corpora, and the cross-regime taxonomies.

LLM wiki systems compile knowledge into pre-filled KV caches for efficient inference, but assume a static corpus -- an assumption that fails whenever the underlying information landscape evolves. We formalize Streaming Knowledge Compilation: given a document stream, a fixed token budget, and future queries unknown at ingestion time, maintain a compiled wiki that minimizes cumulative regret against an offline oracle with perfect foresight. The enabling insight is a materiality signal $ϕ_t(k,n)\in[0,1]$ that scores document importance for entity $k$ at time $t$, acting as a query-relevance surrogate for proactive pinning before queries arrive; we prove an $O(\sqrt{T\log K})$ regret bound where $\varepsilon=\mathbb{E}[|ϕ_t-\hatϕ_t|]$ is the only domain-specific quantity. We instantiate in two domains: finance, where $ϕ_t$ is abnormal stock volatility predicted by frozen Llama 3.1 8B classification head (AUROC = 0.728 on 76K articles, strict temporal split; $1.49\times$ higher realized forward volatility for predicted-material articles); and Wikipedia, where $ϕ_t$ is the Abnormal Edit Ratio (AER), a cross-sectionally normalized edit velocity -- showing the same algorithm generalizes beyond the finance domain. End-to-end QA evaluation on 173 matched pairs (finance) and 119 (Wikipedia) reveals a pervasive LLM-as-judge confound on post-training knowledge, establishing that regret analysis -- not absolute QA scores -- is the reliable evaluation metric for compiled knowledge systems. Finance cumulative regret converges to -20.0 (-0.12/step); Wikipedia to +16.0 (+0.13/step), with the positive sign confirming that Wikipedia edit content is genuinely post-training -- richer context consistently improves scores (No Wiki 3.80 vs. Oracle 4.74) -- and eliminates this confound. The $O(\sqrt{T\log K})$ guarantee applies to any domain where knowledge gaps can be predicted from streaming signals.

Large language models often express high confidence in answers that are wrong. Standard calibration remedies typically act globally or at the score level, reducing unwarranted confidence but also risking erosion of warranted confidence on correct answers. We introduce Probe-Conditioned Head Intervention (PCHI), an inference-time method that uses a frozen probe to detect likely wrong-but-confident responses and conditionally rescales downstream attention-head outputs during confidence generation. On Qwen3-4B-Instruct solving OpenMathInstruct problems with a structured binary confidence field, readout-token PCHI converts 82.2% of originally wrong-yes confidence readouts to $\texttt{no}$, while a joint intervention across upstream confidence-template tokens reduces ECE from 21.9% to 9.2% and damages only 5.1% of originally correct-yes readouts. The readout-token effect also appears on Gemma3-4B, though upstream interventions are weaker and more mask-dependent. These results show that verbalized overconfidence can be selectively reduced through conditionally applied internal intervention, partially decoupling the suppression of unwarranted confidence from the loss of warranted confidence.

Reconstruction-based methods are widely used for time series anomaly detection, where models are trained to reconstruct subsequences, and anomalies are identified through reconstruction errors. However, reported results are often hard to compare due to heterogeneous evaluation practices and underspecified inference procedures. In this paper, we revisit reconstruction-based anomaly detection in the univariate offline setting and study the role of the inference stride, which controls whether subsequences are processed as disjoint windows or with overlap. We propose a unified training, tuning, and multi-seed evaluation protocol on the curated TSB-AD benchmark, and study how overlapping inference affects anomaly detection performance for a range of reconstruction models, including PCA-based baselines, DLinear, an AutoEncoder, TimesNet, and Transformer variants. The results show that across all models, overlapping windows yield consistent improvements, with average relative gain up to +28%, and can alter method rankings. We further analyze variability across datasets, random seeds, and hyperparameter configurations. Finally, we complement the benchmark study with an evaluation on the full UCR archive using localization criteria aligned with sliding-window reconstruction. Overall, our results highlight that reconstruction-based anomaly detection performance depends not only on model architecture and training, but also on inference choices, motivating a clear and reproducible protocol. Our results show that reconstructionbased baselines achieve strong performance on both TSB-AD and UCR benchmarks, supporting them as competitive and practical approaches for univariate time series anomaly detection.

Traffic forecasting is a fundamental component of intelligent transportation systems, yet remains challenging in real-world settings due to irregular sensor distributions and the high computational cost of modeling large-scale spatiotemporal dependencies. In practical traffic networks, sensors are unevenly distributed across regions, leading to non-uniform spatial structures that limit the effectiveness and scalability of existing graph-based and attention-based models. To address these challenges, we propose PatchSTG, a patch-based spatiotemporal graph Transformer designed for efficient forecasting on irregular sensor networks. The key idea is to introduce a hierarchical spatial representation that partitions sensors into balanced, locality-preserving patches based on geographic information. On top of this structure, a dual attention encoder alternates between intra-patch attention for capturing local interactions and inter-patch attention for modeling global dependencies, reducing computational complexity from quadratic to near-linear scaling. We evaluate PatchSTG on real-world traffic data from Rhode Island and additional large-scale datasets. Experimental results demonstrate that the proposed model achieves stable and competitive forecasting performance across multiple horizons, while significantly improving computational efficiency. Ablation studies further validate the effectiveness of spatial partitioning and dual attention in capturing both local and long-range traffic dynamics. These results suggest that patch-based spatiotemporal modeling provides a scalable and effective framework for traffic forecasting under irregular spatial settings.

As Multimodal Large Language Models (MLLMs) face growing privacy risks and regulatory constraints, machine unlearning (MU) has emerged as a crucial solution for removing sensitive data while preserving model performance. However, existing MU methods typically rely on visual data of the target concepts, which is often unavailable due to strict data retention policies, thus creating a demand for source-free unlearning approaches that operate without access to the target data. In this work, we propose Source-free Proxy Anchor Concept Erasure (SPACE), the first source-free unlearning framework specialized for MLLMs. SPACE consists of two stages: (1) Text-Guided Proxy Anchor Selection (TPAS), which retrieves semantically aligned proxy anchors from the shared feature space. (2) Dual-Constraint Semantic Isolation (DCSI), which optimizes these anchors to indirectly erase target concepts. DCSI confines updates to the null space of retained knowledge, ensuring structural integrity. We theoretically prove that SPACE strictly bounds the perturbation on retained knowledge and maximizes feature spectral entropy, thereby maintaining the model's performance. Furthermore, extensive experiments across six datasets show that SPACE achieves performance comparable to that of state-of-the-art data-dependent methods, validating its effectiveness in source-free MU scenarios. The source code will be released.

Key-value (KV) cache quantization is widely used to reduce Large Language Model (LLM) inference memory, yet existing evaluations solely focus on measuring perplexity and accuracy without assessing the safety impact. In this study, we explore alignment preservation under KV cache quantization. Across eleven instruction-tuned models (3.8B-72B) and five benchmarks (1,894 prompts), we find that low-bit quantization can silently destroy safety alignment: Mistral-7B loses 15.2% of its refusals at only 1.03x perplexity, and no universal safe bit-width exists, with sharp model-specific phase transitions invisible to standard metrics. We identify that the root cause is geometric: safety features occupy a low-dimensional activation subspace 10^2-10^3x more vulnerable to quantization noise than the full representation space perplexity averages over. Inspired by this observation, we propose Per-Channel Reduction (PCR), a diagnostic that classifies each model into one of three mechanistic failure modes: outlier-crushes-safety, where safety lives in non-outlier channels collaterally damaged by outlier-driven scale factors; outlier-as-safety, where safety overlaps outlier channels and finer granularity cannot rescue it; and multi-layer dilution, where safety is distributed across many layers and per-layer fixes fail. PCR predicts the correct mitigation direction on all nine primary models and one held-out model from an independent family using 20 calibration prompts. PCR generalizes across unseen prompts, models, and production quantizers, including KIVI with up to 97.2% recovery, succeeding where attention-based allocation methods fail. The resulting training-free protocol, requiring approximately 35 GPU-minutes, recovers up to 97% of lost alignment at minimal memory overhead, addressing vulnerabilities confirmed in production vLLM serving with FP8 KV cache on NVIDIA GPUs.

LLM agents can fail silently by asserting task completion when the environment state shows otherwise. We study this failure mode, false success, across two agent benchmarks: 9,876 tau2-bench trajectories from 8 model families and 1,879 AppWorld trajectories from 4 model families with text-independent ground truth. False success is common but varies by setting: 45--48% of failures in single-control tau2-bench domains, 3% in dual-control telecom, and 75.8% among AppWorld self-assessing coding-agent trajectories with explicit status claims. LLM judges fail reliably: no configuration across 5 judges, 5 prompt strategies, and full task specifications exceeds AUROC 0.65 on tau2-bench, and the same judges reach only 0.54 AUROC on AppWorld API-call traces. Judges rely on surface completion proxies -- confident closing language in tau2-bench and coarse action-sequence volume in AppWorld -- rather than verified state changes. Lightweight TF-IDF detectors achieve task-disjoint AUROC 0.83 on tau2-bench and 0.95 on AppWorld, recovering 4--8x more false successes than the best judge at the same flag rate with 3,300x lower latency. These results suggest that production monitoring should use lightweight, domain-calibrated detectors as triage signals rather than relying on LLM judges as the primary monitor for false success.

MLLMs frequently hallucinate objects inconsistent with visual inputs. This issue is typically attributed to the over-reliance on language priors, which can override the visual context. Recent training-free decoding strategies address this by penalizing language priors. However, these methods overlook the dual nature of language priors, where they can be both helpful and harmful depending on the alignment with visual evidence. In particular, blindly suppressing language priors often disrupts the model's semantic manifold, leading to performance degradation, a phenomenon we term Manifold Departure. To address this, we propose Manifold-Guided Adaptive Projection (MGAP), a geometry-aware, training-free decoding method that mitigates hallucinations while preserving representation structure. MGAP first constructs a language-prior subspace from blind hidden states via SVD. During decoding, MGAP projects each multimodal hidden state onto this subspace and applies a consistency-aware gate to adaptively attenuate only the projected prior component, yielding a subspace-selective update that largely preserves the orthogonal semantic components. Extensive experiments on POPE and CHAIR show that MGAP outperforms prior decoding baselines, achieving stronger hallucination suppression without sacrificing coherence.

A central objective in multimodal learning is to capture synergy: task-relevant information that arises only from the joint use of multiple modalities, and is not available from any single modality alone. While most approaches operate at the architectural level through larger or more complex fusion models, we propose a complementary axis: shaping the training objective itself. Standard training often emphasizes unimodal or redundant information, falling short on examples that require cross-modal reasoning. We formalize multimodal synergy through information theory and introduce the Synergistic Information Bottleneck (SynIB), a scalable objective that targets synergy directly. To prioritize learning synergy, SynIB motivates the model to predict accurately from all modalities while penalizing confidence when information from any modality is withheld. Alongside the standard task loss, the model runs forward passes with one modality masked at a time and is penalized for remaining confident, which would indicate reliance on unimodal cues rather than cross-modal interactions. We validate SynIB in two regimes. On synthetic XOR tasks where the ground-truth synergy is known by construction, standard training fails to recover it while SynIB does. On five real-world benchmarks, including three MultiBench affective tasks, Hateful Memes with CLIP-ViT and DeBERTa backbones, and a controllable irony extension of CREMA-D we introduce, SynIB improves accuracy on synergy-dependent examples by up to 7.8% and overall accuracy by up to 3.8%.

Accelerated acquisition of fMRI enables enhanced detection of neurovascular (BOLD) activity in the brain, but image reconstruction becomes challenging with high k-space undersampling: Task-evoked BOLD signals are small in magnitude, which traditional anatomical MRI reconstruction methods fail to recover, as they favor spatial accuracy over temporal fidelity. We present DD-INR, a Dynamics-Driven Implicit Neural Representation framework tailored for accelerated fMRI that benefits from incoherent time-varying sampling and a tailored spatiotemporal prior, outperforming traditional methods, demonstrated in simulation and in-vivo acquisition, both in terms of image quality and retrieval of activation patterns. DD-INR achieves this by splitting the fMRI data into a static background and a temporally varying dynamic component, representing only the dynamics with a dedicated INR, thereby focusing the model's capacity on activation-relevant changes while remaining compact. In general, DD-INR provides a promising framework for accelerated fMRI reconstruction, with the potential to improve the sensitivity and robustness of fMRI studies within practical scan time limits. The source code is available at https://github.com/JoosenLi/DD-INR.

Bone marrow smear review remains important for acute myeloid leukemia (AML) assessment, but manual single-cell interpretation is labor-intensive and patient-level diagnosis requires aggregation of many cellular observations. We present a cell-to-patient deep learning pipeline for AML-assisted diagnosis from bone marrow smear images. The study included 258 patients from six anonymized centers, including a main cohort of 169 patients from Centers 1-3 and an external validation cohort of 89 patients from Centers 4-6. A 16-category cell annotation vocabulary was used to describe the global cellular composition, including granulocytic, monocytic, erythroid, lymphoid, eosinophilic, and other cells. Rather than identifying strict AML blasts or leukemic blasts, the model targets an expert-defined composite category termed Composite Blast-like Cells (CBLC), comprising N, N1, M, M1, R, R1, J, and J1 according to the project-wide morphological standard. A fixed YOLO-based segmentation module detected cells, predicted contours were matched to expert polygon annotations by contour IoU, and standardized single-cell crops were generated. An EfficientNet-B0 classifier was trained through a two-stage GT-to-YOLO and YOLO-to-YOLO strategy with class-imbalance correction, center-border regularization, and morphology-assisted supervision. Cell-level predictions were aggregated into patient-level CBLC ratios for AML-oriented diagnostic support. The pipeline achieved stable internal validation and maintained external generalization, with ensemble weighted F1-scores of 0.9076, 0.8696, and 0.9124 on Centers 4, 5, and 6, respectively.

Deploying neural networks on unconventional hardware demands architectures that co-optimize task accuracy and platform-specific constraints such as energy cost, physical non-idealities, and numerical precision. Existing neural architecture search (NAS) methods are typically tailored to a single hardware family, limiting cross-platform comparison and generalization. We introduce Unconventional Hardware Neural Architecture Search (UH-NAS), a hardware-agnostic, LLM-guided NAS framework that integrates language models as evolutionary operators to co-optimize accuracy and inference energy. By exposing hardware as a swappable backend with per-platform energy models, physical constraints, and non-ideality simulators, UH-NAS enables fair system-level comparisons across various backends without modifying the search algorithm. Tested on optical MZI hardware, UH-NAS discovers more diverse, robust architectures than conventional baselines while outperforming existing LLM-to-NAS approaches. Additional ablations on architecture robustness under non-idealities and the role of system prompts highlight the importance of architecture-hardware co-design for emerging computing platforms.

Weed pressure in forage corn production causes yield losses of up to 31.5%, yet site-specific weed management (SSWM) systems built on UAV imagery and deep learning remain constrained by the scarcity of field-representative training datasets. We present USU-Corn-WeedDB, a publicly available UAV RGB image dataset collected from a commercial forage corn field in Cache Valley, Utah, designed to support multi-class weed detection under both supervised and semi-supervised learning frameworks. RGB imagery was acquired on 27 June 2025 using an Autel EVO II Dual 640T V2 drone at ~10m above ground level, yielding a ground sampling distance of approximately 0.48 cm/pixel. A total of 366 full-resolution images were tiled into 8,800 patches at 640 x 640-pixel resolution. Of these, 800 images were manually annotated for three weed species; common lambsquarters (Chenopodium album), redroot pigweed (Amaranthus retroflexus), and green foxtail (Setaria viridis) comprising 10,539 bounding-box instances, with the remaining 8,000 tiles retained as an unlabeled pool for semi-supervised experiments. This dataset reflects a natural class imbalance where redroot pigweed constitutes 53.86% of annotated instances, which was preserved intentionally to mirror real field conditions. To validate dataset utility, we trained 28 object detection models spanning five architecture families including YOLOv8, YOLOv9, YOLOv10, YOLO11, YOLO26, and RT-DETR under identical conditions without hyperparameter tuning. Test set mAP@0.5 ranged from 0.773 to 0.840, with lightweight models achieving competitive performance relevant to edge-deployed UAV systems. USU-Corn-WeedDB is publicly available at https://doi.org/10.5281/zenodo.20044178.

Ensuring the reliability of Large Language Models (LLMs) under distribution drift requires inference-time adaptation. While inference-time alignment methods such as Best-of-$N$ and rejection sampling are widely used, they frame the task as a sampling-intensive, reward-guided search, leading to two key limitations: their performance is bounded by the base model's generation quality, and their reliance on imperfect reward models makes them vulnerable to reward hacking. To address these challenges, we introduce Gradient-Guided Reward Optimization (GGRO), a lightweight inference-time method that performs targeted, minimal intervention during decoding via gradient guidance. Specifically, GGRO monitors token-level entropy to identify high-uncertainty regions indicative of drift or misalignment. Upon detection, it responds by injecting nudging tokens, generated using gradient signals from an off-the-shelf reward model, to steer the generation trajectory rather than merely re-ranking samples. Experiments show that GGRO consistently improves inference-time alignment across safety, helpfulness, and reasoning benchmarks. It also increases coverage of high-quality responses and robustness to reward hacking, with minimal computational overhead. Code is available at https://github.com/lhk2004/GGRO.

Recent advances in neural text-to-speech (TTS) and multilingual speech generation have substantially improved synthetic speech quality, yet these gains remain unevenly distributed across the world's languages. Existing models are still dominated by a small set of high-resource languages, while many studies of low-resource TTS are simulated on artificially downsampled high-resource corpora that do not reflect the orthographic variation and limited phonetic coverage encountered in genuinely underrepresented settings. As such, we introduce OpenBibleTTS, which is a large-scale benchmark for low-resource speech synthesis spanning 37 underrepresented languages. Moreover, a systematic comparison of various TTS architectures and large-scale speech generation models is conducted across in-domain Biblical text and out-of-domain material. Results show that no single system dominates across languages and metrics: Gemini-TTS achieves the highest listener ratings on most evaluated languages, but monolingual EveryVoice models trained on OpenBibleTTS remain strongest for intelligibility and are preferred in several African languages, while open from-scratch systems degrade sharply on out-of-domain text, revealing a persistent gap between broad multilingual coverage and reliable synthesis quality in underserved linguistic communities. We complement automatic evaluation with subjective human judgments, and open-source all processed datasets, alignments, and trained models to support future low-resource TTS research.

Large audio language models (LALMs) are increasingly deployed in real-world applications, yet their safety alignment is still primarily evaluated on monolingual, text-based harmful prompts. This leaves their generalizability under multilingual and spoken settings, particularly code-switched speech, largely underexplored. To address this gap, we introduce SpeechJBB, an audio jailbreak dataset for benchmarking across multiple state-of-the-art LALMs. The extent of safety weaknesses is further probed by introducing an augmented setting where phonologically plausible pseudo-words are inserted around safety-critical terms to simulate localized obfuscation. Across models, code-switched harmful audio yields substantially high jailbreak success rates (JSR), with non-English monolingual and non-English code-switched pairs exhibiting the highest attack success. Pseudo-word insertion further reduces refusal rates, which demonstrates that natural-sounding obfuscation can effectively bypass safety policies.

Real-time inference of vision-language-action (VLA) models is essential for robotic control. While visual token pruning has shown strong potential for accelerating inference, most existing methods mainly base pruning decisions on shallow-layer cues and risk discarding visual information required by deep layers. To address this issue, we propose SAFE-Pruner, a plug-and-play pruning framework that incorporates attention cues of future layers into pruning decisions. Specifically, we identify semantic attention consistency, the tendency that VLA models concentrate their attention probability mass on the same semantic entity across execution steps. Based on this observation, we design a forward-looking strategy to forecast the token saliency in deep layers, which prevents the premature removal of critical tokens and leads to more stable acceleration. We further introduce an adaptive subtask division strategy to detect abrupt attention shifts, thereby improving forecasting accuracy and pruning reliability. Extensive experiments in simulation and real-world settings demonstrate that our method achieves up to 1.89x speedup with a minimal degradation in success rate of less than 1.7%, while outperforming state-of-the-art methods by up to 1.9%.

AI evaluation results are produced at scale but reported inconsistently across leaderboards, model cards, benchmark papers, and company blogs. The cost is interpretive: readers cannot reliably compare results across sources, identify what a report omits, or trace an aggregate claim to its underlying evidence. Recent efforts address isolated components but leave three gaps: they cover only narrow slices of the evaluation lifecycle and do not compose into a single interpretable record; they specify static representations that do not differentiate the questions different stakeholders bring to the same evidence; and they remain proposals on paper, lacking the extraction infrastructure required for adoption at scale. We present \EvalCards{}, an operational reporting layer that composes benchmark metadata, evaluation run data, and model metadata into a unified record. We (1) derive a reporting schema from a structured review of 52 papers and 10 stakeholder interviews, (2) implement four interpretive signals (reproducibility, documentation completeness, provenance and risk, and score comparability), rendered through reader modes calibrated to research and non-research audiences, and (3) deploy a monitoring tool that applies \EvalCards{} across 5,816 models, 635 benchmarks, and 101,843 results, surfacing systematic gaps in current reporting practice.

Eye movements, including saccades, are widely regarded as highly sensitive and objective biomarkers of neurophysiologic states. Detecting saccadic signatures in neurologic diseases offers a rapid, portable alternative to brain imaging, avoiding access and cost barriers. Currently, there are no robust AI-enabled video-oculographic solutions (e.g., digital biomarkers) for screening, triaging, or localizing brain abnormalities due to privacy issues and scarce datasets. In this work, we propose the first fully synthetic, patient-free, multimodal eye movement generation pipeline for generalizable saccade analysis. Using this synthetic dataset, we trained a deep learning classifier to distinguish between normal and abnormal (hypometria and hypermetria) saccadic accuracies and evaluated its performance on real-world clinical data. The model achieved an AUROC of 0.76 and a sensitivity of 0.71, showing that the synthetic data has strong potential to generalize for clinical applications, including as a screening tool in at-home and emergency room settings or a tool for precise neuroanatomic localization.

Conditional generators provide a natural tool for controllable generation, including settings where the desired condition is a new composition of observed attributes or experimental factors. In many applications, especially in scientific domains, such models are attractive to explore conditions for which real samples are rare, expensive, or not yet observed. However, this creates a circularity for evaluation: standard conditional quality metrics require a reference target distribution, but in the extrapolative regime that distribution is unavailable by definition. We address this problem with a post-hoc, per-sample trust score for assessing conditional samples using only the training distribution. The score combines two estimable quantities: global realism, measuring compatibility with the real data manifold, and attribute-wise faithfulness, measuring whether a sample is closer to the requested attributes than to plausible alternatives. We show that the score can recover meaningful comparisons across extrapolated generations, under a mild coverage condition on the observed attributes. These comparisons enable effective filtering, ranking, and abstention of generations and can be used directly on off-the-shelf pretrained models. In biological imaging, selected samples preserve real morphological structure better and improve downstream predictive performance, while similar gains are observed on controlled vision benchmarks. Finally, we show how the score can be applied during generation, enabling abstention before full decoding. Code is available at https://github.com/berkerdemirel/faithful-cond-gen.

As AI assistants serve millions of users daily, evaluating user experience (UX) beyond general model capability has become increasingly important. We present UXBench, the first user-centric benchmark grounded in real user feedback signals for evaluating preference alignment and dialogue generation. The benchmark consists of three interconnected tasks, UX Judge, UX Eval, and UX Recovery, with 7,400 test instances extracted from over 70K interaction logs of a mainstream Chinese AI assistant. The dataset closely reflects real user distributions, covering 8 scenarios, 83 domains, and diverse failure patterns that pose severe challenges. Extensive experiments on 26 frontier language models provide novel insights into how well models perceive user experience and how improvements in model capability contribute to better dialogue engagement. Through comprehensive analysis of model behavior and performance gaps, we show that user feedback prediction is a learnable capability, where a reward model trained from in-the-wild feedback signals can achieve well-calibrated accuracy. We further document the systematic biases of LLM-as-a-judge evaluation protocols and compare typical response strategies that directly affect user experience. UXBench establishes a new evaluation landscape and calls for greater attention to tailored UX optimization, contributing to a user-centric scaling law that shapes the success of AI assistants.

This short paper introduces a stylometric interpretation method inspired by genome-wide association studies (GWAS). Each "gene" token's association with "phenotype" authorship is tested using logistic regression with multiple-comparison correction. Applied to English, German, and Russian corpora, the method detects statistically significant lexical markers distinctive of individual authors.

Work on `emergent misalignment' shows that finetuning LLMs on narrow tasks can induce broadly misaligned behavior. This supports the `persona selection' (PSM) hypothesis: during pre-training, LLMs learn to simulate different characters and perspectives, which can be elicited and refined during post-training. This paper investigates the converse phenomenon, `emergent alignment', and uses it to support and refine the PSM and motivate a novel desideratum for alignment. We finetune a helpful-only model on broad and narrow safety tasks. To create SFT samples, we follow the `Constitutional AI' (CAI) approach and use four constitutions which encode reasonable alignment strategies: deontology, consequentialism, virtue ethics, and aligning AIs as subordinate to human authority. For each of those models, we show that finetuning on two narrow safety sub-categories reliably induces emergent alignment over a representative set of general safety categories, and on safety subcategories that we directly filtered-out of the data sets used for narrow alignment. To test the `PSM' using a more fine-grained evaluation, we used a multidimensional `ethical persona' diagnostic. For each constitutionally finetuned (broad/narrow) model, we evaluate how well their behavior matches their expected signature profile. Our results show that our CAI models acquire their expected ``ethical persona'' -- e.g., the model narrowly fine-tuned on SFT samples created using the consequentialist constitution agrees significantly more with utilitarian than deontological beliefs. Yet our coarse and fine-grained evaluations show that there are significant differences across our (broad/narrow) finetuned CAI models in how well they project. We conclude that alignment strategies should be evaluated, not just on their (in-distribution) general safety performance, but also specifically on their degree of projectability.

Classification tasks require annotated data, which can often be expensive, time-consuming, or even unfeasible to collect. This is the case of the medical domain, where large datasets often have few annotated examples. To address this, we propose DecSelfMask (Decoder Self-learning by Masking), an approach to enhance decoder-only performance on classification tasks. We build on common self-learning approaches by leveraging a model to create training examples from unlabeled data to propose a novel relevance-guided masking strategy. We use relevance attribution methods to determine what portions of unannotated texts are relevant for a task. We then create self-supervised training examples by masking out those portions, training the model to reconstruct them via next-token-prediction. We hypothesize that those examples convey knowledge about the structure and semantics of unannotated data that can be useful for downstream performance. We test our approach on 136 tasks from a collection of 1.9M clinical notes from an Italian hospital. We quantify DecSelfMask's impact on downstream tasks on 5 models of different scales and families, including a probing analysis. Experiments show consistent gains, outperforming standard supervised fine-tuning approaches (+19.9 points in Macro F1), synthetic label generation (+12.5), and continual pretraining (+6.3), as well as common baselines.

Large language model agents increasingly rely on skills: reusable procedural documents encoding workflows, tool use, implementation patterns, validation checks, and domain rules. Skill rewriting is often treated as prompt compression, but shorter skills can make agents more expensive by removing sparse operational anchors that prevent exploration, debugging, and recovery. We study skill rewriting through this economic lens. Our controlled framework profiles skill structure, rewrites skills using information-preservation strategies, and evaluates the rewrites under fixed task instructions, environments, and verifiers. Experiments on SkillsBench reveal distinct quality--cost trade-offs across strategies: API/code anchoring, workflow guarding, and rule/formula anchoring benefit different task families, with no universally dominant template. In the main held-out evaluation, the learned policy reduces total cost by 7.0% and downstream agent-token cost by 6.0%; in frozen cross-model transfer, the corresponding reductions average 14.7% and 13.7%, while verifier quality is preserved. These results position skill design as cost-aware operational knowledge engineering rather than prompt compression. Resources: https://github.com/1Reminding/Skill_EE.

Formal neural network verification -- proving that a network satisfies safety properties for *all* inputs in a specified domain -- is bounded in practice by GPU memory: standard implementations of bound-propagation algorithms (IBP, CROWN, $α$-CROWN) require weight and relaxation-coefficient matrices to reside entirely on one accelerator. We adapt two parallelism techniques originally developed for large-scale model training to the auto_LiRPA / $α,β$-CROWN verification framework. Tensor Parallelism (TP) shards both weight and $A$-matrices across GPUs, achieving ${\approx}2\times$ peak-memory reduction at $P{=}2$; soundness is confirmed on VNN-COMP 2022 MNIST-FC benchmarks, though bound tightness degrades with the number of sharded zones due to forced IBP substitution for intermediate bounds inside sharded zones. Fully Sharded Data Parallelism (FSDP) shards only weight matrices with a per-layer AllGather, producing bounds that are bitwise identical to the single-GPU baseline: baseline memory drops by 80--90%, peak memory by 34--39% on wide MLPs. FSDP integrates cleanly with complete verification ($β$-CROWN + Branch-and-Bound) and with convolutional layers (BoundConv); a complete unsat result is obtained for CIFAR-100 ResNet-large (VNN-COMP 2024) under FSDP. Across all experiments the memory bottleneck in $α$-CROWN+BaB mode proves to be per-neuron alpha tensors, not weight matrices, pointing to the key direction for future work.

Retrieval-augmented generation (RAG) enables agents to access external knowledge at inference time, but it primarily retrieves fragmented declarative evidence, leaving agents to repeatedly infer task procedures from passages, manuals, examples, logs, or trajectories. This raises a fundamental question: can skills extracted from external knowledge bases be installed into an agent, enabling it to rapidly approximate domain expertise? In this paper, we propose Anything2Skill, a taxonomy-guided framework that compiles heterogeneous external knowledge into reusable, retrievable, and executable skills for agents. Given a corpus of knowledge records, \textsc{Anything2Skill} first decomposes each record into evidence windows and performs plan-and-expand skill extraction under a skill-tree prior. The extracted candidates are then converted into structured skill contracts that specify invocation conditions, contraindications, action moves, workflow steps, constraints, output specifications, supporting evidence, and confidence scores. To construct a deployable procedural memory, Anything2Skill manages the extracted skills in a persistent SkillBank through taxonomy-aware compilation, registry-level reconciliation, lifecycle tracking, versioned updates, and visible skill-tree projection. At inference time, agents retrieve both task-specific passages from the original knowledge base and relevant procedural skills from the SkillBank, allowing RAG to provide declarative evidence while compiled skills provide reusable procedural guidance. Experiments on qsv and GitHub-CLI show that Anything2Skill combined with RAG achieves 98.85\% and 94.10\% success rates, respectively, substantially outperforming RAG-only agents. These results suggest that compiling latent procedural knowledge into explicit skills is an effective way to extend retrieval-augmented agents from knowledge access toward capability reuse.

The Robot Operating System 2 (ROS 2) is a widely used middleware for robotic systems, characterized by a publish-subscribe (pub-sub) communication mechanism in which computation is structured as callbacks dispatched by ROS 2 executors. Despite its popularity, the pub-sub pattern in ROS 2 is inherently nondeterministic: the order in which these callbacks run is nondeterministic even within a single executor, and distributed deployments add further nondeterminism from the interleaving of messages across nodes and from network latency. Such nondeterminism often leads to concurrency issues and makes it virtually impossible to analyze for safeness and provide guarantees. We present a framework that is able to convert an unmodified ROS 2 application and run it under Lingua Franca (LF), a coordination language for deterministic execution using logical time, so that the same input always produces the same deterministic execution order. We first describe which ROS 2 features can be executed deterministically under logical time. Such features enable the possibility to establish an automatic conversion framework to extract information from a ROS 2 application and directly convert it into an LF program. The rich features of LF, such as logical-time delays, federated execution across processes, and fault handling, can then be applied to make the ROS 2 application be executed in a deterministic and timing-predictable manner without changing the ROS 2 code. We evaluate the framework on a synthetic example and on the Autoware reference system. We show that the order in which callbacks are executed differs in default ROS 2, while also having end-to-end latencies that vary across executions. In contrast, our LF-controlled ROS 2 system produces a deterministic execution order and consistent end-to-end latencies.