Black box large language models (LLMs) make information extraction (IE) easy to configure, but hard to trust. Unlike traditional information extraction pipelines, the information "extracted" is not guaranteed to be grounded in the document. To prevent this, this paper introduces the notion of a "safe passage": context generated by the LLM that is both grounded in the document and consistent with the extracted information. This is operationalized via a three-step pipeline, SafePassage, which consists of: (1) an LLM extractor that generates structured entities and their contexts from a document, (2) a string-based global aligner, and (3) a scoring model. Results show that using these three parts in conjunction reduces hallucinations by up to 85% on information extraction tasks with minimal risk of flagging non-hallucinations. High agreement between the SafePassage pipeline and human judgments of extraction quality mean that the pipeline can be dually used to evaluate LLMs. Surprisingly, results also show that using a transformer encoder fine-tuned on a small number of task-specific examples can outperform an LLM scoring model at flagging unsafe passages. These annotations can be collected in as little as 1-2 hours.

Large Language Models (LLMs) have shown extraordinary success across various text generation tasks; however, their potential for simple yet essential text classification remains underexplored, as LLM pre-training tends to emphasize generation over classification. While LLMs with instruction tuning can transform classification into a generation task, they often struggle to categorize nuanced texts. One such example is text revision, which involves nuanced edits between pairs of texts. Although simply fine-tuning LLMs for revision classification seems plausible, it requires a large amount of revision annotations, which are exceptionally expensive and scarce in the community. To address this issue, we introduce a plug-and-play layer-wise parameter-efficient fine-tuning (PEFT) framework, i.e., IR-Tuning, which fine-tunes a subset of important LLM layers that are dynamically selected based on their gradient norm distribution, while freezing those of redundant layers. Extensive experiments suggest that IR-Tuning surpasses several layer-wise PEFT baselines over diverse text revisions, while achieving fast convergence, low GPU memory consumption, and effectiveness on small revision corpora.

The alignment of large language models (LLMs) with human values increasingly relies on using other LLMs as automated judges, or ``autoraters''. However, their reliability is limited by a foundational issue: they are trained on discrete preference labels, forcing a single ground truth onto tasks that are often subjective, ambiguous, or nuanced. We argue that a reliable autorater must learn to model the full distribution of preferences defined by a target population. In this paper, we propose a general framework for calibrating probabilistic autoraters to any given preference distribution. We formalize the problem and present two learning methods tailored to different data conditions: 1) a direct supervised fine-tuning for dense, probabilistic labels, and 2) a reinforcement learning approach for sparse, binary labels. Our empirical results show that finetuning autoraters with a distribution-matching objective leads to verbalized probability predictions that are better aligned with the target preference distribution, with improved calibration and significantly lower positional bias, all while preserving performance on objective tasks.

Interest in generative Electrocardiogram-Language Models (ELMs) is growing, as they can produce textual responses conditioned on ECG signals and textual queries. Unlike traditional classifiers that output label probabilities, ELMs are more versatile, supporting domain-specific tasks (e.g., waveform analysis, diagnosis, prognosis) as well as general tasks (e.g., open-ended questions, dialogue). Retrieval-Augmented Generation (RAG), widely used in Large Language Models (LLMs) to ground LLM outputs in retrieved knowledge, helps reduce hallucinations and improve natural language generation (NLG). However, despite its promise, no open-source implementation or systematic study of RAG pipeline design for ELMs currently exists. To address this gap, we present the first open-source RAG pipeline for ELMs, along with baselines and ablation studies for NLG. Experiments on three public datasets show that ELMs with RAG consistently improves performance over non-RAG baselines and highlights key ELM design considerations. Our code is available at: https://github.com/willxxy/ECG-Bench.

Variational Auto-Encoders (VAEs) are known to generate blurry and inconsistent samples. One reason for this is the "prior hole" problem. A prior hole refers to regions that have high probability under the VAE's prior but low probability under the VAE's posterior. This means that during data generation, high probability samples from the prior could have low probability under the posterior, resulting in poor quality data. Ideally, a prior needs to be flexible enough to match the posterior while retaining the ability to generate samples fast. Generative models continue to address this tradeoff. This paper proposes to model the prior as an energy-based model (EBM). While EBMs are known to offer the flexibility to match posteriors (and also improving the ELBO), they are traditionally slow in sample generation due to their dependency on MCMC methods. Our key idea is to bring a variational approach to tackle the normalization constant in EBMs, thus bypassing the expensive MCMC approaches. The variational form can be approximated with a sampler network, and we show that such an approach to training priors can be formulated as an alternating optimization problem. Moreover, the same sampler reduces to an implicit variational prior during generation, providing efficient and fast sampling. We compare our Energy-based Variational Latent Prior (EVaLP) method to multiple SOTA baselines and show improvements in image generation quality, reduced prior holes, and better sampling efficiency.

We study length generalization in sequence models on a composite problem involving both state tracking and associative recall. Prior work finds that recurrent networks handle state tracking well but struggle with recall, whereas Transformers excel at recall yet fail to extend state-tracking capabilities to longer sequences. Motivated by the complementary strengths of these architectures, we construct hybrid models integrating recurrent and attention-based components, and train them on the combined task to evaluate whether both capabilities can be preserved. Our results reveal that, in such hybrids, the Transformer component tends to exploit shortcut solutions, leading to poor length generalization. We identify this shortcut reliance as a key obstacle and propose a simple yet effective training strategy -- delaying the training of the attention layers -- that mitigates this effect and significantly improves length generalization performance. Our experiments show that this approach enables hybrid models to achieve near-perfect accuracy ($>90\%$) on hybrid sequences three times longer than those used during training.

We introduce TASER (Translation Assessment via Systematic Evaluation and Reasoning), a metric that uses Large Reasoning Models (LRMs) for automated translation quality assessment. TASER harnesses the explicit reasoning capabilities of LRMs to conduct systematic, step-by-step evaluation of translation quality. We evaluate TASER on the WMT24 Metrics Shared Task across both reference-based and reference-free scenarios, demonstrating state-of-the-art performance. In system-level evaluation, TASER achieves the highest soft pairwise accuracy in both reference-based and reference-free settings, outperforming all existing metrics. At the segment level, TASER maintains competitive performance with our reference-free variant ranking as the top-performing metric among all reference-free approaches. Our experiments reveal that structured prompting templates yield superior results with LRMs compared to the open-ended approaches that proved optimal for traditional LLMs. We evaluate o3, a large reasoning model from OpenAI, with varying reasoning efforts, providing insights into the relationship between reasoning depth and evaluation quality. The explicit reasoning process in LRMs offers interpretability and visibility, addressing a key limitation of existing automated metrics. Our results demonstrate that Large Reasoning Models show a measurable advancement in translation quality assessment, combining improved accuracy with transparent evaluation across diverse language pairs.

We introduce the coded-smoothing module, which can be seamlessly integrated into standard training pipelines, both supervised and unsupervised, to regularize learning and improve generalization with minimal computational overhead. In addition, it can be incorporated into the inference pipeline to randomize the model and enhance robustness against adversarial perturbations. The design of coded-smoothing is inspired by general coded computing, a paradigm originally developed to mitigate straggler and adversarial failures in distributed computing by processing linear combinations of the data rather than the raw inputs. Building on this principle, we adapt coded computing to machine learning by designing an efficient and effective regularization mechanism that encourages smoother representations and more generalizable solutions. Extensive experiments on both supervised and unsupervised tasks demonstrate that coded-smoothing consistently improves generalization and achieves state-of-the-art robustness against gradient-based adversarial attacks.

Training algorithms in deep learning usually treat a mini-batch of samples as a single object; they average gradients over the mini-batch, and then process the average in various ways. Computing other statistics beyond the average may have been seen as prohibitively resource intensive in automatic differentiation (AD) frameworks. We show that this is not the case. Generally, gradient statistics can be implemented through a surgery of the AD graph, which, in some cases, incur almost no computational and memory overheads compared to the mini-batch gradient computation. Additionally, we show that in certain classes of models, including transformers, JAX's vectorization transformation offers a viable implementation for prototyping and experimentation. We then revise our understanding of two nonlinear operations in optimization through the lens of per-example gradient transformations. We first study signSGD and show that the optimal placement of the sign operation in the gradient processing chain is crucial to success and can be predicted with a simple signal-to-noise ratio argument. Next we study per-example variations of the Adam preconditioner, and show that optimization is best served when the preconditioner is dominated by the mean rather than the variance of the gradient distribution - in contrast to conventional wisdom. Overall we demonstrate that per-example gradient information enables new analyses and possibilities for algorithm design.

Existing studies on bias mitigation methods for large language models (LLMs) use diverse baselines and metrics to evaluate debiasing performance, leading to inconsistent comparisons among them. Moreover, their evaluations are mostly based on the comparison between LLMs' probabilities of biased and unbiased contexts, which ignores the gap between such evaluations and real-world use cases where users interact with LLMs by reading model responses and expect fair and safe outputs rather than LLMs' probabilities. To enable consistent evaluation across debiasing methods and bridge this gap, we introduce BiasFreeBench, an empirical benchmark that comprehensively compares eight mainstream bias mitigation techniques (covering four prompting-based and four training-based methods) on two test scenarios (multi-choice QA and open-ended multi-turn QA) by reorganizing existing datasets into a unified query-response setting. We further introduce a response-level metric, Bias-Free Score, to measure the extent to which LLM responses are fair, safe, and anti-stereotypical. Debiasing performances are systematically compared and analyzed across key dimensions: the prompting vs. training paradigm, model size, and generalization of different training strategies to unseen bias types. We will publicly release our benchmark, aiming to establish a unified testbed for bias mitigation research.