This paper proposes predicting task difficulty for LLM agents without running expensive rollouts, studying the problem across 17 agentic benchmarks and showing that token-level entropy is a useful predictive signal.
The paper introduces TSR, a framework that decomposes social dialogue into strategic planning and linguistic execution, and LHRL-VGR, a reinforcement learning algorithm with variance-gated rewards. Fine-tuning a Qwen2.5-7B agent with this approach surpasses the GPT-4o baseline by 7.32% in goal completion on the SOTOPIA benchmark.
Introduces GROM, a gradient-free one-shot machine unlearning method that computes a closed-form additive weight update via ridge-regularized least squares, achieving state-of-the-art forgetting-utility trade-offs on benchmarks like TOFU and WMDP, and resisting quantization-based recovery attacks.
Introduces MoCA (Implicit Social Context Analysis), a new task and benchmark for modeling implicit social scenarios across affection, intent, and stance, along with a Conflict-Driven Abductive Reasoning (CoDAR) framework. Experiments show state-of-the-art multimodal LLMs struggle on this task, while CoDAR improves performance but still lags behind human reasoning.
This paper proposes a neuro-symbolic closed-loop architecture for laser powder bed fusion, where an in-loop ontology couples symbolic reasoning with statistical learning to control melt pool depth and eliminate overhang dross. Feasibility is demonstrated via a surrogate calibrated to the NIST AM-Bench benchmark.
This paper presents HallDetect, a lightweight and reference-free framework for hallucination detection that decomposes generated content into atomic claims and verifies them with a compact entailment model. It outperforms comparably resourced baselines on multiple benchmarks and provides a claim-to-span audit trail.
This arXiv paper presents a shape-constrained predictive modeling approach to estimate nanoparticle size and dispersity in continuous flow nanodrug production, reducing the need for extensive experimental screening.
This paper introduces M3R-Bench, a unified evidence-grounded benchmark for multimodal metaphor understanding with 1,000 image-text instances, and proposes M3R-Reasoner, an 8B-parameter model combining curriculum-based reasoning supervision and reinforcement learning that outperforms larger proprietary models.
This paper introduces HiLP, a hierarchical representation training method that adds multi-scale self-predictive learning to transformer pretraining, aiming to reduce compounding error and improve long-horizon reasoning and speculative decoding efficiency.
This paper presents an equipment-centric framework that uses deep learning-based vision and event-driven finite state machines to localize workpieces in hot forging factories, achieving high detection accuracy and low latency in operational settings.
This paper introduces Task-Conditional Flow Matching (TCFM), a framework for adapting multilingual text embedding models that selectively uses flow matching for translation tasks and other objectives for retrieval/classification. It achieves state-of-the-art results on the Indic Massive Text Embedding Benchmark.
This paper identifies an objective gap in Direct Forecasting for multivariate time series and proposes CvLoss, a plug-in structural regularizer that constrains forecast residuals on a cross-variable graph to improve consistency across synchronous and asynchronous interactions. Experiments show consistent improvements over competitive forecasting models.
CircuitSteer is a novel framework that uses sparse autoencoders to identify and manipulate multi-layer semantic circuits in LLMs, enabling more robust and fluency-preserving behavioral steering compared to single-layer methods like CAA.
This paper compares context biasing methods and speech LLMs for recognizing rare and new words in automatic speech recognition, reporting trade-offs in word error rate across read and non-read speech.
This paper introduces Spectral Aliasing Pretext (SAP), a self-supervised learning method for fault diagnosis in rotating machinery. By deliberately undersampling vibration signals and training a Transformer to reconstruct the original spectrum, SAP learns discriminative frequency-domain representations that achieve strong classification performance with limited labeled data.
This paper proposes EchoPrompt, a training-free detector for LLM-generated text that restores a latent prompt dependency by prepending a generic prefix and measuring likelihood gain differences between instruction-tuned and base models, achieving state-of-the-art zero-shot detection performance.
SEAM is a generator-agnostic framework that audits global consistency of explanations in scientific machine learning, detecting incompatible local explanations even when predictions are locally accurate and attributing failures to specific channels and overlaps. The paper presents theory and experiments across PDE systems, neural operators, and four open datasets.
This paper proposes a novel method to mitigate scoring bias in LLM-as-a-Judge by having LLMs randomly generate numbers to measure their latent numerical bias, then rectifying token generation probabilities accordingly. Experiments across four tasks show the method outperforms baselines and reveals that scoring bias varies across models, tasks, and score ranges.
This paper proposes a commutation theory for label-free reliability in vision-language figure reading, showing that consistency-based methods have a computable blind spot and introducing an Equivariance-Consistency Score enhanced by cyclic relabeling.
This paper studies using sparse PPMI graph averaging to refine Random Indexing embeddings, showing it improves accuracy on a fairytales analogy benchmark but trails neural baselines on text8 and SimLex-999.