Machine unlearning (MU) for large language models (LLMs), commonly referred to as LLM unlearning, seeks to remove specific undesirable data or knowledge from a trained model, while maintaining its performance on standard tasks. While unlearning plays a vital role in protecting data privacy, enforcing copyright, and mitigating sociotechnical harms in LLMs, we identify a new vulnerability post-unlearning: unlearning trace detection. We discover that unlearning leaves behind persistent ''fingerprints'' in LLMs, detectable traces in both model behavior and internal representations. These traces can be identified from output responses, even when prompted with forget-irrelevant inputs. Specifically, even a simple supervised classifier can determine whether a model has undergone unlearning, using only its prediction logits or even its textual outputs. Further analysis shows that these traces are embedded in intermediate activations and propagate nonlinearly to the final layer, forming low-dimensional, learnable manifolds in activation space. Through extensive experiments, we demonstrate that unlearning traces can be detected with over 90% accuracy even under forget-irrelevant inputs, and that larger LLMs exhibit stronger detectability. These findings reveal that unlearning leaves measurable signatures, introducing a new risk of reverse-engineering forgotten information when a model is identified as unlearned, given an input query.

Adversarial threats against LLMs are escalating faster than current defenses can adapt. We expose a critical geometric blind spot in alignment: adversarial prompts exploit latent camouflage, embedding perilously close to the safe representation manifold while encoding unsafe intent thereby evading surface level defenses like Direct Preference Optimization (DPO), which remain blind to the latent geometry. We introduce ALKALI, the first rigorously curated adversarial benchmark and the most comprehensive to date spanning 9,000 prompts across three macro categories, six subtypes, and fifteen attack families. Evaluation of 21 leading LLMs reveals alarmingly high Attack Success Rates (ASRs) across both open and closed source models, exposing an underlying vulnerability we term latent camouflage, a structural blind spot where adversarial completions mimic the latent geometry of safe ones. To mitigate this vulnerability, we introduce GRACE - Geometric Representation Aware Contrastive Enhancement, an alignment framework coupling preference learning with latent space regularization. GRACE enforces two constraints: latent separation between safe and adversarial completions, and adversarial cohesion among unsafe and jailbreak behaviors. These operate over layerwise pooled embeddings guided by a learned attention profile, reshaping internal geometry without modifying the base model, and achieve up to 39% ASR reduction. Moreover, we introduce AVQI, a geometry aware metric that quantifies latent alignment failure via cluster separation and compactness. AVQI reveals when unsafe completions mimic the geometry of safe ones, offering a principled lens into how models internally encode safety. We make the code publicly available at https://anonymous.4open.science/r/alkali-B416/README.md.

There is an emerging effort to develop NLP for Indonesias 700+ local languages, but progress remains costly due to the need for direct engagement with native speakers. However, it is unclear what these language communities truly need from language technology. To address this, we conduct a nationwide survey to assess the actual needs of native speakers in Indonesia. Our findings indicate that addressing language barriers, particularly through machine translation and information retrieval, is the most critical priority. Although there is strong enthusiasm for advancements in language technology, concerns around privacy, bias, and the use of public data for AI training highlight the need for greater transparency and clear communication to support broader AI adoption.

Achieving political consensus is crucial yet challenging for the effective functioning of social governance. However, although frontier AI systems represented by large language models (LLMs) have developed rapidly in recent years, their capabilities in this scope are still understudied. In this paper, we introduce PoliCon, a novel benchmark constructed from 2,225 high-quality deliberation records of the European Parliament over 13 years, ranging from 2009 to 2022, to evaluate the ability of LLMs to draft consensus resolutions based on divergent party positions under varying collective decision-making contexts and political requirements. Specifically, PoliCon incorporates four factors to build each task environment for finding different political consensus: specific political issues, political goals, participating parties, and power structures based on seat distribution. We also developed an evaluation framework based on social choice theory for PoliCon, which simulates the real voting outcomes of different political parties to assess whether LLM-generated resolutions meet the requirements of the predetermined political consensus. Our experimental results demonstrate that even state-of-the-art models remain undersatisfied with complex tasks like passing resolutions by a two-thirds majority and addressing security issues, while uncovering their inherent partisan biases and revealing some behaviors LLMs show to achieve the consensus, such as prioritizing the stance of the dominant party instead of uniting smaller parties, which highlights PoliCon's promise as an effective platform for studying LLMs' ability to promote political consensus.

We introduce a novel dataset designed to benchmark the physical and spatial reasoning capabilities of Large Language Models (LLM) based on topology optimization, a method for computing optimal material distributions within a design space under prescribed loads and supports. In this dataset, LLMs are provided with conditions such as 2D boundary, applied forces and supports, and must reason about the resulting optimal material distribution. The dataset includes a variety of tasks, ranging from filling in masked regions within partial structures to predicting complete material distributions. Solving these tasks requires understanding the flow of forces and the required material distribution under given constraints, without access to simulation tools or explicit physical models, challenging models to reason about structural stability and spatial organization. Our dataset targets the evaluation of spatial and physical reasoning abilities in 2D settings, offering a complementary perspective to traditional language and logic benchmarks.

Large language models (LLMs) excel at generating long-form responses, but evaluating their factuality remains challenging due to complex inter-sentence dependencies within the generated facts. Prior solutions predominantly follow a decompose-decontextualize-verify pipeline but often fail to capture essential context and miss key relational facts. In this paper, we introduce VeriFact, a factuality evaluation framework designed to enhance fact extraction by identifying and resolving incomplete and missing facts to support more accurate verification results. Moreover, we introduce FactRBench , a benchmark that evaluates both precision and recall in long-form model responses, whereas prior work primarily focuses on precision. FactRBench provides reference fact sets from advanced LLMs and human-written answers, enabling recall assessment. Empirical evaluations show that VeriFact significantly enhances fact completeness and preserves complex facts with critical relational information, resulting in more accurate factuality evaluation. Benchmarking various open- and close-weight LLMs on FactRBench indicate that larger models within same model family improve precision and recall, but high precision does not always correlate with high recall, underscoring the importance of comprehensive factuality assessment.

Adverse weather conditions challenge safe transportation, necessitating robust real-time weather detection from traffic camera imagery. We propose a novel framework combining CycleGAN-based domain adaptation with efficient contrastive learning to enhance weather classification, particularly in low-light nighttime conditions. Our approach leverages the lightweight SigLIP-2 model, which employs pairwise sigmoid loss to reduce computational demands, integrated with CycleGAN to transform nighttime images into day-like representations while preserving weather cues. Evaluated on an Iowa Department of Transportation dataset, the baseline EVA-02 model with CLIP achieves a per-class overall accuracy of 96.55\% across three weather conditions (No Precipitation, Rain, Snow) and a day/night overall accuracy of 96.55\%, but shows a significant day-night gap (97.21\% day vs.\ 63.40\% night). With CycleGAN, EVA-02 improves to 97.01\% per-class accuracy and 96.85\% day/night accuracy, boosting nighttime performance to 82.45\%. Our Vision-SigLIP-2 + Text-SigLIP-2 + CycleGAN + Contrastive configuration excels in nighttime scenarios, achieving the highest nighttime accuracy of 85.90\%, with 94.00\% per-class accuracy and 93.35\% day/night accuracy. This model reduces training time by 89\% (from 6 hours to 40 minutes) and inference time by 80\% (from 15 seconds to 3 seconds) compared to EVA-02. By narrowing the day-night performance gap from 33.81 to 8.90 percentage points, our framework provides a scalable, efficient solution for all-weather classification using existing camera infrastructure.

This study investigates the relationship between deep learning (DL) model accuracy and expert agreement in classifying crash narratives. We evaluate five DL models -- including BERT variants, USE, and a zero-shot classifier -- against expert labels and narratives, and extend the analysis to four large language models (LLMs): GPT-4, LLaMA 3, Qwen, and Claude. Our findings reveal an inverse relationship: models with higher technical accuracy often show lower agreement with human experts, while LLMs demonstrate stronger expert alignment despite lower accuracy. We use Cohen's Kappa and Principal Component Analysis (PCA) to quantify and visualize model-expert agreement, and employ SHAP analysis to explain misclassifications. Results show that expert-aligned models rely more on contextual and temporal cues than location-specific keywords. These findings suggest that accuracy alone is insufficient for safety-critical NLP tasks. We argue for incorporating expert agreement into model evaluation frameworks and highlight the potential of LLMs as interpretable tools in crash analysis pipelines.

Compositionality, or correctly recognizing scenes as compositions of atomic visual concepts, remains difficult for multimodal large language models (MLLMs). Even state of the art MLLMs such as GPT-4o can make mistakes in distinguishing compositions like "dog chasing cat" vs "cat chasing dog". While on Winoground, a benchmark for measuring such reasoning, MLLMs have made significant progress, they are still far from a human's performance. W e show that compositional reasoning in these models can be improved by elucidating such concepts via data, where a model is trained to prefer the correct caption for an image over a close but incorrect one. W e introduce SCRAMBLe: Synthetic Compositional Reasoning Augmentation of MLLMs with Binary preference Learning, an approach for preference tuning open-weight MLLMs on synthetic preference data generated in a fully automated manner from existing image-caption data. SCRAMBLe holistically improves these MLLMs' compositional reasoning capabilities which we can see through significant improvements across multiple vision language composition-ality benchmarks, as well as smaller but significant improvements on general question answering tasks. As a sneak peek, SCRAMBLe tuned Molmo-7B model improves on Winoground from 49.5% to 54.8% (best reported to date), while improving by 1% on more general visual question answering tasks.

Survey researchers face two key challenges: the rising costs of probability samples and missing data (e.g., non-response or attrition), which can undermine inference and increase the use of convenience samples. Recent work explores using large language models (LLMs) to simulate respondents via persona-based prompts, often without labeled data. We study a more practical setting where partial survey responses exist: we fine-tune LLMs on available data to impute self-reported vote choice under both random and systematic nonresponse, using the German Longitudinal Election Study. We compare zero-shot prompting and supervised fine-tuning against tabular classifiers (e.g., CatBoost) and test how different convenience samples (e.g., students) used for fine-tuning affect generalization. Our results show that when data are missing completely at random, fine-tuned LLMs match tabular classifiers but outperform zero-shot approaches. When only biased convenience samples are available, fine-tuning small (3B to 8B) open-source LLMs can recover both individual-level predictions and population-level distributions more accurately than zero-shot and often better than tabular methods. This suggests fine-tuned LLMs offer a promising strategy for researchers working with non-probability samples or systematic missing-ness, and may enable new survey designs requiring only easily accessible subpopulations.