When LLMs are given contradictory passages, the generated response may appear decisive while masking how internal signals behave under conflicting evidence. Does it favor one passage, blend both, or ignore both? We study model behavior under inter-context contradiction by introducing a white-box tracing framework that extracts internal model signals based on hidden-state shifts, counterfactual attribution via passage removal, and token-level uncertainty during generation under two single-passage prompts and one dual-passage prompt with conflicting evidence. We evaluate four open-weight LLMs from the LLaMA (8B and 70B), OLMo, and Qwen on WikiContradict, a human-annotated QA benchmark with contradictory passages. Each dual-passage answer is labeled as committing to passage A or passage B, blending both, or fabricating an answer unsupported by either passage. Comparing these labels with the extracted signals shows that larger models reduce fabrication, but hidden-state and probability-level signals can disagree with the final answer. Moreover, fabricated answers often remain high-confidence, showing that confidence alone is unreliable for identifying ungrounded behavior under contradictory evidence. We further validate on ConflictBank, a larger misinformation-conflict QA dataset, showing consistent disagreement patterns across models. Finally, we test whether pre-generation white-box signals are informative beyond interpretation by training classifiers to predict output behavior. Classifiers using white-box signals outperform probability baselines, showing these signals contain diagnostic information about model behavior under contradictory evidence beyond log probabilities.

Recent controllable text generation (CTG) for sentiment control has largely focused on decoder-based large language models, making causal attention the dominant paradigm. While effective for fluent generation, these models still struggle to satisfy complex constraints and follow fine-grained sentiment signals specified by users. Existing sentiment-aware CTG methods typically simplify the problem by treating sentiment either as a coarse categorical label (e.g., positive or negative) or as a single fine-grained control signal applied to an entire document. Consequently, more challenging settings such as sentence-level sentiment control within long-form text remain underexplored. To address these limitations, we introduce SenseShift, an encoder-based framework for fine-grained sentence-level CTG. Unlike standard decoder architectures, SenseShift leverages bidirectional attention, quantized sentiment signals, and iterative mask infilling to generate local sentences conditioned on target sentiment intensity. Empirical evaluations on story and review generation demonstrate that SenseShift achieves stronger sentiment controllability while maintaining text quality and robustness to out-of-domain generation compared to larger decoder-based baselines.

This tutorial bridges psychology and NLP to clarify cognitive effects and biases in large language models, compares methodological assumptions across disciplines, and discusses practical evaluation challenges and open research directions.

Transformer-based Job Ranking Systems (JRSs) are vulnerable to societal biases inherited in unbalanced datasets. These biases often manifest as unjust job rankings, particularly disadvantaging candidates of different genders. Most bias mitigation techniques leverage candidates’ gender and align gender distributions within the embeddings of JRSs to mitigate bias. While such methods effectively align distributional properties and make JRSs agnostic to gender, they frequently fall short in addressing empirical fairness metrics, such as the performance gap across genders. In this study, we shift our attention from candidate gender to mitigate bias based on gendered language in job advertisements. We propose a novel neutrality score based on automatically discovered biased words in job ads and use it to re-rank the model’s decisions. We evaluate our method by comparing it with different bias mitigation strategies and empirically demonstrate that our proposed method not only improves fairness but can also enhance the model’s performance.

Multi-task learning (MTL) has shown considerable practical benefits, particularly when using language models (LMs). While this is commonly achieved by learning tasks under a joint optimization procedure, some methods, such as AdapterFusion, divide the problem into two stages: (i) task learning, where knowledge specific to a task is encapsulated within sets of parameters (e.g., adapters), and (ii) transfer, where this already learned knowledge is leveraged for a target task. This separation of concerns provides numerous benefits (e.g., promoting reusability). However, current two stage MTL introduces a substantial number of additional parameters. We address this issue by leveraging the usefulness of linearly scaling the output representations of source adapters for transfer learning. We introduce ScaLearn, a simple and highly parameter-efficient two-stage MTL method that capitalizes on the knowledge of the source tasks by learning a minimal set of scaling parameters that enable effective transfer to a target task. Our experiments on three benchmarks (GLUE, SuperGLUE, and HumSet) and two encoder LMs show that ScaLearn consistently outperforms strong baselines with a small number of transfer parameters (~0.35% of those of AdapterFusion). Remarkably, we observe that ScaLearn maintains its strong abilities even when further reducing parameters, achieving competitive results with only 8 transfer parameters per target task. Our proposed approach thus demonstrates the power of simple scaling as a promise for more efficient task transfer. Our code is available at https://github.com/CPJKU/ScaLearn.

Language models frequently inherit societal biases from their training data. Numerous techniques have been proposed to mitigate these biases during both the pre-training and fine-tuning stages. However, fine-tuning a pre-trained debiased language model on a downstream task can reintroduce biases into the model. Additionally, existing debiasing methods for downstream tasks either (i) require labels of protected attributes (e.g., age, race, or political views) that are often not available or (ii) rely on indicators of bias, which restricts their applicability to gender debiasing since they rely on gender-specific words. To address this, we introduce a novel debiasing regularization technique based on the class-wise variance of embeddings. Crucially, our method does not require attribute labels and targets any attribute, thus addressing the shortcomings of existing debiasing methods. Our experiments on encoder language models and three datasets demonstrate that our method outperforms existing strong debiasing baselines that rely on target attribute labels while maintaining performance on the target task.

Bias mitigation of Language Models has been the topic of many studies with a recent focus on learning separate modules like adapters for on-demand debiasing. Besides optimizing for a modularized debiased model, it is often critical in practice to control the degree of bias reduction at inference time, e.g., in order to tune for a desired performance-fairness trade-off in search results or to control the strength of debiasing in classification tasks. In this paper, we introduce Controllable Gate Adapter (ConGater), a novel modular gating mechanism with adjustable sensitivity parameters, %In addition to better perseverance of task performance and enhanced information removal, which allows for a gradual transition from the biased state of the model to the fully debiased version at inference time. We demonstrate ConGater performance by (1) conducting adversarial debiasing experiments with three different models on three classification tasks with four protected attributes, and (2) reducing the bias of search results through fairness list-wise regularization to enable adjusting a trade-off between performance and fairness metrics. Our experiments on the classification tasks show that compared to baselines of the same caliber, ConGater can maintain higher task performance while containing less information regarding the attributes. Our results on the retrieval task show that the fully debiased ConGater can achieve the same fairness performance while maintaining more than twice as high task performance than recent strong baselines. Overall, besides strong performance ConGater enables the continuous transitioning between biased and debiased states of models, enhancing personalization of use and interpretability through controllability.

Societal biases are reflected in large pre-trained language models and their fine-tuned versions on downstream tasks. Common in-processing bias mitigation approaches, such as adversarial training and mutual information removal, introduce additional optimization criteria, and update the model to reach a new debiased state. However, in practice, end-users and practitioners might prefer to switch back to the original model, or apply debiasing only on a specific subset of protected attributes. To enable this, we propose a novel modular bias mitigation approach, consisting of stand-alone highly sparse debiasing subnetworks, where each debiasing module can be integrated into the core model on-demand at inference time. Our approach draws from the concept of diff pruning, and proposes a novel training regime adaptable to various representation disentanglement optimizations. We conduct experiments on three classification tasks with gender, race, and age as protected attributes. The results show that our modular approach, while maintaining task performance, improves (or at least remains on-par with) the effectiveness of bias mitigation in comparison with baseline finetuning. Particularly on a two-attribute dataset, our approach with separately learned debiasing subnetworks shows effective utilization of either or both the subnetworks for selective bias mitigation.