Comprehensive evaluations of language models (LM) during both development and deployment phases are necessary because these models possess numerous capabilities (e.g., mathematical reasoning, legal support, or medical diagnostic) as well as safety risks (e.g., racial bias, toxicity, or misinformation). The average score across a wide range of benchmarks provides a signal that helps guide the use of these LMs in practice. Currently, holistic evaluations are costly due to the large volume of benchmark questions, making frequent evaluations impractical. A popular attempt to lower the cost is to compute the average score on a subset of the benchmark. This approach, unfortunately, often renders an unreliable measure of LM performance because the average score is often confounded with the difficulty of the questions in the benchmark subset. Item response theory (IRT) was designed to address this challenge, providing a reliable measurement by careful controlling for question difficulty. Unfortunately, question difficulty is expensive to estimate. Facing this challenge, we train a model that predicts question difficulty from its content, enabling a reliable measurement at a fraction of the cost. In addition, we leverage this difficulty predictor to further improve the evaluation efficiency through training a question generator given a difficulty level. This question generator is essential in adaptive testing, where, instead of using a random subset of the benchmark questions, informative questions are adaptively chosen based on the current estimation of LLM performance. Experiments on 22 common natural language benchmarks and 172 LMs show that this approach is more reliable and efficient compared to current common practice.

The rapid advancement of AI systems has raised widespread concerns about potential harms of frontier AI systems and the need for responsible evaluation and oversight. In this position paper, we argue that frontier AI companies should report both pre- and post-mitigation safety evaluations to enable informed policy decisions. Evaluating models at both stages provides policymakers with essential evidence to regulate deployment, access, and safety standards. We show that relying on either in isolation can create a misleading picture of model safety. Our analysis of AI safety disclosures from leading frontier labs identifies three critical gaps: (1) companies rarely evaluate both pre- and post-mitigation versions, (2) evaluation methods lack standardization, and (3) reported results are often too vague to inform policy. To address these issues, we recommend mandatory disclosure of pre- and post-mitigation capabilities to approved government bodies, standardized evaluation methods, and minimum transparency requirements for public safety reporting. These ensure that policymakers and regulators can craft targeted safety measures, assess deployment risks, and scrutinize companies' safety claims effectively.

We present a submission to the SemEval 2025 shared task on unlearning sensitive content from LLMs. Our approach employs negative preference optimization using low-rank adaptation. We show that we can utilize this combination to cheaply compute additional regularization terms, which help with unlearning stabilization. The results of our approach significantly exceed the shared task baselines.

In qualitative research, data transcription is often labor-intensive and time-consuming. To expedite this process, a workflow utilizing artificial intelligence (AI) was developed. This workflow not only enhances transcription speed but also addresses the issue of AI-generated transcripts often lacking compatibility with standard content analysis software. Within this workflow, automatic speech recognition is employed to create initial transcripts from audio recordings, which are then formatted to be compatible with content analysis software such as ATLAS.ti or MAXQDA. Empirical data from a study of 12 interviews suggests that this workflow can reduce transcription time by up to 46.2%. Furthermore, by using widely used standard software, this process is suitable for both students and researchers while also being adaptable to a variety of learning, teaching, and research environments. It is also particularly beneficial for non-native speakers. In addition, the workflow is GDPR-compliant and facilitates local, offline transcript generation, which is crucial when dealing with sensitive data.

Large language models (LLMs) offer significant potential for coding, yet fine-tuning (FT) with curated data is essential for niche languages like Verilog. Using proprietary intellectual property (IP) for FT presents a serious risk, as FT data can be leaked through LLM inference. This leads to a critical dilemma for design houses: seeking to build externally accessible LLMs offering competitive Verilog coding, how can they leverage in-house IP to enhance FT utility while ensuring IP protection? For the first time in the literature, we study this dilemma. Using LLaMA 3.1-8B, we conduct in-house FT on a baseline Verilog dataset (RTLCoder) supplemented with our own in-house IP, which is validated through multiple tape-outs. To rigorously assess IP leakage, we quantify structural similarity (AST/Dolos) and functional equivalence (Synopsys Formality) between generated codes and our in-house IP. We show that our IP can indeed be leaked, confirming the threat. As defense, we evaluate logic locking of Verilog codes (ASSURE). This offers some level of protection, yet reduces the IP's utility for FT and degrades the LLM's performance. Our study shows the need for novel strategies that are both effective and minimally disruptive to FT, an essential effort for enabling design houses to fully utilize their proprietary IP toward LLM-driven Verilog coding.

However, deliberative forums such as citizens' assemblies have shown promise in bypassing party polarization and fostering productive discussions on contentious political issues [3]. Unfortunately, most deliberations do not take place in carefully structured settings with nationally representative participants. Instead, they often occur within homogeneous groups [17]. When this happens, deliberation can lead to group polarization, where individuals become more extreme in their initial positions rather than engaging with opposing viewpoints [22]. This can be problematic if the goal of deliberation is to build common ground and consensus within a pluralistic electorate. Given that large language models (LLMs) have demonstrated some fidelity in accurately responding to opinion surveys [1, 20] and adopting different personas [12], we explore whether an LLM-powered tool can help introduce missing perspectives in group deliberation.

Many real-world eligibility problems, ranging from medical diagnosis to tax planning, can be mapped to decision problems expressed in natural language, wherein a model must make a binary choice based on user features. Large-scale domains such as legal codes or frequently updated funding opportunities render human annotation (e.g., web forms or decision trees) impractical, highlighting the need for agents that can automatically assist in decision-making. Since relevant information is often only known to the user, it is crucial that these agents ask the right questions. As agents determine when to terminate a conversation, they face a trade-off between accuracy and the number of questions asked, a key metric for both user experience and cost. To evaluate this task, we propose BeNYfits, a new benchmark for determining user eligibility for multiple overlapping social benefits opportunities through interactive decision-making. Our experiments show that current language models struggle with frequent hallucinations, with GPT-4o scoring only 35.7 F1 using a ReAct-style chain-of-thought. To address this, we introduce ProADA, a novel approach that leverages program synthesis to assist in decision-making by mapping dialog planning to a code generation problem and using gaps in structured data to determine the best next action. Our agent, ProADA, improves the F1 score to 55.6 while maintaining nearly the same number of dialog turns.

With the increasing prevalence of Web-based platforms handling vast amounts of user data, machine unlearning has emerged as a crucial mechanism to uphold users' right to be forgotten, enabling individuals to request the removal of their specified data from trained models. However, the auditing of machine unlearning processes remains significantly underexplored. Although some existing methods offer unlearning auditing by leveraging backdoors, these backdoor-based approaches are inefficient and impractical, as they necessitate involvement in the initial model training process to embed the backdoors. In this paper, we propose a TAilored Posterior diffErence (TAPE) method to provide unlearning auditing independently of original model training. We observe that the process of machine unlearning inherently introduces changes in the model, which contains information related to the erased data. TAPE leverages unlearning model differences to assess how much information has been removed through the unlearning operation. Firstly, TAPE mimics the unlearned posterior differences by quickly building unlearned shadow models based on first-order influence estimation. Secondly, we train a Reconstructor model to extract and evaluate the private information of the unlearned posterior differences to audit unlearning. Existing privacy reconstructing methods based on posterior differences are only feasible for model updates of a single sample. To enable the reconstruction effective for multi-sample unlearning requests, we propose two strategies, unlearned data perturbation and unlearned influence-based division, to augment the posterior difference. Extensive experimental results indicate the significant superiority of TAPE over the state-of-the-art unlearning verification methods, at least 4.5$\times$ efficiency speedup and supporting the auditing for broader unlearning scenarios.

Training large models is both resource-intensive and time-consuming, making it crucial to understand the quantitative relationship between model performance and hyperparameters. In this paper, we present an empirical law that describes how the pretraining loss of large language models evolves under different learning rate schedules, such as constant, cosine, and step decay schedules. Our proposed law takes a multi-power form, combining a power law based on the sum of learning rates and additional power laws to account for a loss reduction effect induced by learning rate decay. We extensively validate this law on various model sizes and architectures, and demonstrate that after fitting on a few learning rate schedules, the law accurately predicts the loss curves for unseen schedules of different shapes and horizons. Moreover, by minimizing the predicted final pretraining loss across learning rate schedules, we are able to find a schedule that outperforms the widely used cosine learning rate schedule. Interestingly, this automatically discovered schedule bears some resemblance to the recently proposed Warmup-Stable-Decay (WSD) schedule (Hu et al, 2024) but achieves a slightly lower final loss. We believe these results could offer valuable insights for understanding the dynamics of pretraining and designing learning rate schedules to improve efficiency.

In real-world text classification tasks, negative texts often contain a minimal proportion of negative content, which is especially problematic in areas like text quality control, legal risk screening, and sensitive information interception. This challenge manifests at two levels: at the macro level, distinguishing negative texts is difficult due to the high similarity between coarse-grained positive and negative samples; at the micro level, the issue stems from extreme class imbalance and a lack of fine-grained labels. To address these challenges, we propose transforming the coarse-grained positive-negative (PN) classification task into an imbalanced fine-grained positive-unlabeled (PU) classification problem, supported by theoretical analysis. We introduce a novel framework, Balanced Fine-Grained Positive-Unlabeled (BFGPU) learning, which features a unique PU learning loss function that optimizes macro-level performance amidst severe imbalance at the micro level. The framework's performance is further boosted by rebalanced pseudo-labeling and threshold adjustment. Extensive experiments on both public and real-world datasets demonstrate the effectiveness of BFGPU, which outperforms other methods, even in extreme scenarios where both macro and micro levels are highly imbalanced.