Recent works in Natural Language Inference (NLI) and related tasks, such as automated fact-checking, employ atomic fact decomposition to enhance interpretability and robustness. For this, existing methods rely on resource-intensive generative large language models (LLMs) to perform decomposition. We propose JEDI, an encoder-only architecture that jointly performs extractive atomic fact decomposition and interpretable inference without requiring generative models during inference. To facilitate training, we produce a large corpus of synthetic rationales covering multiple NLI benchmarks. Experimental results demonstrate that JEDI achieves competitive accuracy in distribution and significantly improves robustness out of distribution and in adversarial settings over models based solely on extractive rationale supervision. Our findings show that interpretability and robust generalization in NLI can be realized using encoder-only architectures and synthetic rationales. Code and data available at https://jedi.nicpopovic.com

Large Language Models (LLMs) are increasingly deployed on converged Cloud and High-Performance Computing (HPC) infrastructure. However, as LLMs handle confidential inputs and are fine-tuned on costly, proprietary datasets, their heightened security requirements slow adoption in privacy-sensitive sectors such as healthcare and finance. We investigate methods to address this gap and propose Trusted Execution Environments (TEEs) as a solution for securing end-to-end LLM inference. We validate their practicality by evaluating these compute-intensive workloads entirely within CPU and GPU TEEs. On the CPU side, we conduct an in-depth study running full Llama2 inference pipelines (7B, 13B, 70B) inside Intel's TDX and SGX, accelerated by Advanced Matrix Extensions (AMX). We derive 12 insights, including that across various data types, batch sizes, and input lengths, CPU TEEs impose under 10% throughput and 20% latency overheads, further reduced by AMX. We run LLM inference on NVIDIA H100 Confidential Compute GPUs, contextualizing our CPU findings and observing throughput penalties of 4-8% that diminish as batch and input sizes grow. By comparing performance, cost, and security trade-offs, we show how CPU TEEs can be more cost-effective or secure than their GPU counterparts. To our knowledge, our work is the first to comprehensively demonstrate the performance and practicality of modern TEEs across both CPUs and GPUs for enabling confidential LLMs (cLLMs).

Under a unified operational definition, we define LLM memory as a persistent state written during pretraining, finetuning, or inference that can later be addressed and that stably influences outputs. We propose a four-part taxonomy (parametric, contextual, external, procedural/episodic) and a memory quadruple (location, persistence, write/access path, controllability). We link mechanism, evaluation, and governance via the chain write -> read -> inhibit/update. To avoid distorted comparisons across heterogeneous setups, we adopt a three-setting protocol (parametric only, offline retrieval, online retrieval) that decouples capability from information availability on the same data and timeline. On this basis we build a layered evaluation: parametric (closed-book recall, edit differential, memorization/privacy), contextual (position curves and the mid-sequence drop), external (answer correctness vs snippet attribution/faithfulness), and procedural/episodic (cross-session consistency and timeline replay, E MARS+). The framework integrates temporal governance and leakage auditing (freshness hits, outdated answers, refusal slices) and uncertainty reporting via inter-rater agreement plus paired tests with multiple-comparison correction. For updating and forgetting, we present DMM Gov: coordinating DAPT/TAPT, PEFT, model editing (ROME, MEND, MEMIT, SERAC), and RAG to form an auditable loop covering admission thresholds, rollout, monitoring, rollback, and change audits, with specs for timeliness, conflict handling, and long-horizon consistency. Finally, we give four testable propositions: minimum identifiability; a minimal evaluation card; causally constrained editing with verifiable forgetting; and when retrieval with small-window replay outperforms ultra-long-context reading. This yields a reproducible, comparable, and governable coordinate system for research and deployment.

User profiling, as a core technique for user understanding, aims to infer structural attributes from user information. Large Language Models (LLMs) provide a promising avenue for user profiling, yet the progress is hindered by the lack of comprehensive benchmarks. To bridge this gap, we propose ProfileBench, an industrial benchmark derived from a real-world video platform, encompassing heterogeneous user data and a well-structured profiling taxonomy. However, the profiling task remains challenging due to the difficulty of collecting large-scale ground-truth labels, and the heterogeneous and noisy user information can compromise the reliability of LLMs. To approach label-free and reliable user profiling, we propose a Confidence-driven Profile reasoning framework Conf-Profile, featuring a two-stage paradigm. We first synthesize high-quality labels by leveraging advanced LLMs with confidence hints, followed by confidence-weighted voting for accuracy improvement and confidence calibration for a balanced distribution. The multiple profile results, rationales, and confidence scores are aggregated and distilled into a lightweight LLM. We further enhance the reasoning ability via confidence-guided unsupervised reinforcement learning, which exploits confidence for difficulty filtering, quasi-ground truth voting, and reward weighting. Experimental results demonstrate that Conf-Profile delivers substantial performance through the two-stage training, improving F1 by 13.97 on Qwen3-8B.

With the rapid advancement of large language model technology, there is growing interest in whether multi-feature approaches can significantly improve AI text detection beyond what single neural models achieve. While intuition suggests that combining semantic, syntactic, and statistical features should provide complementary signals, this assumption has not been rigorously tested with modern LLM-generated text. This paper provides a systematic empirical investigation of multi-hierarchical feature integration for AI text detection, specifically testing whether the computational overhead of combining multiple feature types is justified by performance gains. We implement MHFD (Multi-Hierarchical Feature Detection), integrating DeBERTa-based semantic analysis, syntactic parsing, and statistical probability features through adaptive fusion. Our investigation reveals important negative results: despite theoretical expectations, multi-feature integration provides minimal benefits (0.4-0.5% improvement) while incurring substantial computational costs (4.2x overhead), suggesting that modern neural language models may already capture most relevant detection signals efficiently. Experimental results on multiple benchmark datasets demonstrate that the MHFD method achieves 89.7% accuracy in in-domain detection and maintains 84.2% stable performance in cross-domain detection, showing modest improvements of 0.4-2.6% over existing methods.

RLVR has enhanced the reasoning capabilities of Large Language Models (LLMs) across various tasks. However, GRPO, a representative RLVR algorithm, suffers from a critical limitation: when all responses within a group are either entirely correct or entirely incorrect, the model fails to learn from these homogeneous responses. This is particularly problematic for homogeneously incorrect groups, where GRPO's advantage function yields a value of zero, leading to null gradients and the loss of valuable learning signals. To overcome this issue, we propose NGRPO (Negative-enhanced Group Relative Policy Optimization), an algorithm designed to convert homogeneous errors into robust learning signals. First, NGRPO introduces Advantage Calibration. This mechanism hypothesizes the existence of a virtual maximum-reward sample during advantage calculation, thereby altering the mean and variance of rewards within a group and ensuring that the advantages for homogeneously incorrect samples are no longer zero. Second, NGRPO employs Asymmetric Clipping, which relaxes the update magnitude for positive samples while imposing stricter constraints on that of negative samples. This serves to stabilize the exploration pressure introduced by the advantage calibration. Our experiments on Qwen2.5-Math-7B demonstrate that NGRPO significantly outperforms baselines such as PPO, GRPO, DAPO, and PSR-NSR on mathematical benchmarks including MATH500, AMC23, and AIME2025. These results validate NGRPO's ability to learn from homogeneous errors, leading to stable and substantial improvements in mathematical reasoning. Our code is available at https://github.com/nangongrui-ngr/NGRPO.

Accurate International Classification of Diseases (ICD) coding is critical for clinical documentation, billing, and healthcare analytics, yet it remains a labour-intensive and error-prone task. Although large language models (LLMs) show promise in automating ICD coding, their challenges in base model selection, input contextualization, and training data redundancy limit their effectiveness. We propose a modular framework for ICD-10 Clinical Modification (ICD-10-CM) code prediction that addresses these challenges through principled model selection, redundancy-aware data sampling, and structured input design. The framework integrates an LLM-as-judge evaluation protocol with Plackett-Luce aggregation to assess and rank open-source LLMs based on their intrinsic comprehension of ICD-10-CM code definitions. We introduced embedding-based similarity measures, a redundancy-aware sampling strategy to remove semantically duplicated discharge summaries. We leverage structured discharge summaries from Taiwanese hospitals to evaluate contextual effects and examine section-wise content inclusion under universal and section-specific modelling paradigms. Experiments across two institutional datasets demonstrate that the selected base model after fine-tuning consistently outperforms baseline LLMs in internal and external evaluations. Incorporating more clinical sections consistently improves prediction performance. This study uses open-source LLMs to establish a practical and principled approach to ICD-10-CM code prediction. The proposed framework provides a scalable, institution-ready solution for real-world deployment of automated medical coding systems by combining informed model selection, efficient data refinement, and context-aware prompting.

Open-weight versions of large language models (LLMs) are rapidly advancing, with state-of-the-art models like DeepSeek-V3 now performing comparably to proprietary LLMs. This progression raises the question of whether small open-weight LLMs are capable of effectively replacing larger closed-source models. We are particularly interested in the context of biomedical question-answering, a domain we explored by participating in Task 13B Phase B of the BioASQ challenge. In this work, we compare several open-weight models against top-performing systems such as GPT-4o, GPT-4.1, Claude 3.5 Sonnet, and Claude 3.7 Sonnet. To enhance question answering capabilities, we use various techniques including retrieving the most relevant snippets based on embedding distance, in-context learning, and structured outputs. For certain submissions, we utilize ensemble approaches to leverage the diverse outputs generated by different models for exact-answer questions. Our results demonstrate that open-weight LLMs are comparable to proprietary ones. In some instances, open-weight LLMs even surpassed their closed counterparts, particularly when ensembling strategies were applied. All code is publicly available at https://github.com/evidenceprime/BioASQ-13b.

In this paper, we introduce LLMCHECKER, a model-checking-based verification method to verify the probabilistic computation tree logic (PCTL) properties of an LLM text generation process. We empirically show that only a limited number of tokens are typically chosen during text generation, which are not always the same. This insight drives the creation of $α$-$k$-bounded text generation, narrowing the focus to the $α$ maximal cumulative probability on the top-$k$ tokens at every step of the text generation process. Our verification method considers an initial string and the subsequent top-$k$ tokens while accommodating diverse text quantification methods, such as evaluating text quality and biases. The threshold $α$ further reduces the selected tokens, only choosing those that exceed or meet it in cumulative probability. LLMCHECKER then allows us to formally verify the PCTL properties of $α$-$k$-bounded LLMs. We demonstrate the applicability of our method in several LLMs, including Llama, Gemma, Mistral, Genstruct, and BERT. To our knowledge, this is the first time PCTL-based model checking has been used to check the consistency of the LLM text generation process.

Recent advances in diffusion models have significantly improved image and video synthesis. In addition, several concept control methods have been proposed to enable fine-grained, continuous, and flexible control over free-form text prompts. However, these methods not only require intensive training time and GPU memory usage to learn the sliders or embeddings but also need to be retrained for different diffusion backbones, limiting their scalability and adaptability. To address these limitations, we introduce Text Slider, a lightweight, efficient and plug-and-play framework that identifies low-rank directions within a pre-trained text encoder, enabling continuous control of visual concepts while significantly reducing training time, GPU memory consumption, and the number of trainable parameters. Furthermore, Text Slider supports multi-concept composition and continuous control, enabling fine-grained and flexible manipulation in both image and video synthesis. We show that Text Slider enables smooth and continuous modulation of specific attributes while preserving the original spatial layout and structure of the input. Text Slider achieves significantly better efficiency: 5$\times$ faster training than Concept Slider and 47$\times$ faster than Attribute Control, while reducing GPU memory usage by nearly 2$\times$ and 4$\times$, respectively.