Current studies have exposed the risk of Large Language Models (LLMs) generating harmful content by jailbreak attacks. However, they overlook that the direct generation of harmful content from scratch is more difficult than inducing LLM to calibrate benign content into harmful forms. In our study, we introduce a novel attack framework that exploits AdVersArial meTAphoR (AVATAR) to induce the LLM to calibrate malicious metaphors for jailbreaking. Specifically, to answer harmful queries, AVATAR adaptively identifies a set of benign but logically related metaphors as the initial seed. Then, driven by these metaphors, the target LLM is induced to reason and calibrate about the metaphorical content, thus jailbroken by either directly outputting harmful responses or calibrating residuals between metaphorical and professional harmful content. Experimental results demonstrate that AVATAR can effectively and transferable jailbreak LLMs and achieve a state-of-the-art attack success rate across multiple advanced LLMs.
Large language models are increasingly used in social sciences, but their training data can introduce lookahead bias and training leakage. A good chronologically consistent language model requires efficient use of training data to maintain accuracy despite time-restricted data. Here, we overcome this challenge by training a suite of chronologically consistent large language models, ChronoBERT and ChronoGPT, which incorporate only the text data that would have been available at each point in time. Despite this strict temporal constraint, our models achieve strong performance on natural language processing benchmarks, outperforming or matching widely used models (e.g., BERT), and remain competitive with larger open-weight models. Lookahead bias is model and application-specific because even if a chronologically consistent language model has poorer language comprehension, a regression or prediction model applied on top of the language model can compensate. In an asset pricing application predicting next-day stock returns from financial news, we find that ChronoBERT's real-time outputs achieve a Sharpe ratio comparable to state-of-the-art models, indicating that lookahead bias is modest. Our results demonstrate a scalable, practical framework to mitigate training leakage, ensuring more credible backtests and predictions across finance and other social science domains.
The adaptivity and maneuvering capabilities of Autonomous Underwater Vehicles (AUVs) have drawn significant attention in oceanic research, due to the unpredictable disturbances and strong coupling among the AUV's degrees of freedom. In this paper, we developed large language model (LLM)-enhanced reinforcement learning (RL)-based adaptive S-surface controller for AUVs. Specifically, LLMs are introduced for the joint optimization of controller parameters and reward functions in RL training. Using multi-modal and structured explicit task feedback, LLMs enable joint adjustments, balance multiple objectives, and enhance task-oriented performance and adaptability. In the proposed controller, the RL policy focuses on upper-level tasks, outputting task-oriented high-level commands that the S-surface controller then converts into control signals, ensuring cancellation of nonlinear effects and unpredictable external disturbances in extreme sea conditions. Under extreme sea conditions involving complex terrain, waves, and currents, the proposed controller demonstrates superior performance and adaptability in high-level tasks such as underwater target tracking and data collection, outperforming traditional PID and SMC controllers.
Recent advancements in language models (LMs) have demonstrated strong capabilities in semantic understanding and contextual modeling, which have flourished in generative speech enhancement (SE). However, many LM-based SE approaches primarily focus on semantic information, often neglecting the critical role of acoustic information, which leads to acoustic inconsistency after enhancement and limited generalization across diverse SE tasks. In this paper, we introduce LLaSE-G1, a LLaMA-based language model that incentivizes generalization capabilities for speech enhancement. LLaSE-G1 offers the following key contributions: First, to mitigate acoustic inconsistency, LLaSE-G1 employs continuous representations from WavLM as input and predicts speech tokens from X-Codec2, maximizing acoustic preservation. Second, to promote generalization capability, LLaSE-G1 introduces dual-channel inputs and outputs, unifying multiple SE tasks without requiring task-specific IDs. Third, LLaSE-G1 outperforms prior task-specific discriminative and generative SE models, demonstrating scaling effects at test time and emerging capabilities for unseen SE tasks. Additionally, we release our code and models to support further research in this area.
With the rapid advent of generative models, efficiently deploying these models on specialized hardware has become critical. Tensor Processing Units (TPUs) are designed to accelerate AI workloads, but their high power consumption necessitates innovations for improving efficiency. Compute-in-memory (CIM) has emerged as a promising paradigm with superior area and energy efficiency. In this work, we present a TPU architecture that integrates digital CIM to replace conventional digital systolic arrays in matrix multiply units (MXUs). We first establish a CIM-based TPU architecture model and simulator to evaluate the benefits of CIM for diverse generative model inference. Building upon the observed design insights, we further explore various CIM-based TPU architectural design choices. Up to 44.2% and 33.8% performance improvement for large language model and diffusion transformer inference, and 27.3x reduction in MXU energy consumption can be achieved with different design choices, compared to the baseline TPUv4i architecture.
Large language models (LLMs) hold substantial promise for clinical decision support. However, their widespread adoption in medicine, particularly in healthcare, is hindered by their propensity to generate false or misleading outputs, known as hallucinations. In high-stakes domains such as women's health (obstetrics & gynaecology), where errors in clinical reasoning can have profound consequences for maternal and neonatal outcomes, ensuring the reliability of AI-generated responses is critical. Traditional methods for quantifying uncertainty, such as perplexity, fail to capture meaning-level inconsistencies that lead to misinformation. Here, we evaluate semantic entropy (SE), a novel uncertainty metric that assesses meaning-level variation, to detect hallucinations in AI-generated medical content. Using a clinically validated dataset derived from UK RCOG MRCOG examinations, we compared SE with perplexity in identifying uncertain responses. SE demonstrated superior performance, achieving an AUROC of 0.76 (95% CI: 0.75-0.78), compared to 0.62 (0.60-0.65) for perplexity. Clinical expert validation further confirmed its effectiveness, with SE achieving near-perfect uncertainty discrimination (AUROC: 0.97). While semantic clustering was successful in only 30% of cases, SE remains a valuable tool for improving AI safety in women's health. These findings suggest that SE could enable more reliable AI integration into clinical practice, particularly in resource-limited settings where LLMs could augment care. This study highlights the potential of SE as a key safeguard in the responsible deployment of AI-driven tools in women's health, leading to safer and more effective digital health interventions.
Tuberculosis (TB) is the leading cause of death from an infectious disease globally, with the highest burden in low- and middle-income countries. In these regions, limited healthcare access and high patient-to-provider ratios impede effective patient support, communication, and treatment completion. To bridge this gap, we propose integrating a specialized Large Language Model into an efficacious digital adherence technology to augment interactive communication with treatment supporters. This AI-powered approach, operating within a human-in-the-loop framework, aims to enhance patient engagement and improve TB treatment outcomes.
Mathematical formulas are a fundamental and widely used component in various scientific fields, serving as a universal language for expressing complex concepts and relationships. While state-of-the-art transformer models excel in processing and understanding natural language, they encounter challenges with mathematical notation, which involves a complex structure and diverse representations. This study focuses on the development of specialized training datasets to enhance the encoding of mathematical content. We introduce Math Mutator (MAMUT), a framework capable of generating equivalent and falsified versions of a given mathematical formula in LaTeX notation, effectively capturing the mathematical variety in notation of the same concept. Based on MAMUT, we have generated four large mathematical datasets containing diverse notation, which can be used to train language models with enhanced mathematical embeddings.
EXplainable machine learning (XML) has recently emerged to address the mystery mechanisms of machine learning (ML) systems by interpreting their 'black box' results. Despite the development of various explanation methods, determining the most suitable XML method for specific ML contexts remains unclear, highlighting the need for effective evaluation of explanations. The evaluating capabilities of the Transformer-based large language model (LLM) present an opportunity to adopt LLM-as-a-Judge for assessing explanations. In this paper, we propose a workflow that integrates both LLM-based and human judges for evaluating explanations. We examine how LLM-based judges evaluate the quality of various explanation methods and compare their evaluation capabilities to those of human judges within an iris classification scenario, employing both subjective and objective metrics. We conclude that while LLM-based judges effectively assess the quality of explanations using subjective metrics, they are not yet sufficiently developed to replace human judges in this role.
Strategic decision-making involves interactive reasoning where agents adapt their choices in response to others, yet existing evaluations of large language models (LLMs) often emphasize Nash Equilibrium (NE) approximation, overlooking the mechanisms driving their strategic choices. To bridge this gap, we introduce an evaluation framework grounded in behavioral game theory, disentangling reasoning capability from contextual effects. Testing 22 state-of-the-art LLMs, we find that GPT-o3-mini, GPT-o1, and DeepSeek-R1 dominate most games yet also demonstrate that the model scale alone does not determine performance. In terms of prompting enhancement, Chain-of-Thought (CoT) prompting is not universally effective, as it increases strategic reasoning only for models at certain levels while providing limited gains elsewhere. Additionally, we investigate the impact of encoded demographic features on the models, observing that certain assignments impact the decision-making pattern. For instance, GPT-4o shows stronger strategic reasoning with female traits than males, while Gemma assigns higher reasoning levels to heterosexual identities compared to other sexual orientations, indicating inherent biases. These findings underscore the need for ethical standards and contextual alignment to balance improved reasoning with fairness.
We introduce SEKI, a novel large language model (LLM)-based neural architecture search (NAS) method. Inspired by the chain-of-thought (CoT) paradigm in modern LLMs, SEKI operates in two key stages: self-evolution and knowledge distillation. In the self-evolution stage, LLMs initially lack sufficient reference examples, so we implement an iterative refinement mechanism that enhances architectures based on performance feedback. Over time, this process accumulates a repository of high-performance architectures. In the knowledge distillation stage, LLMs analyze common patterns among these architectures to generate new, optimized designs. Combining these two stages, SEKI greatly leverages the capacity of LLMs on NAS and without requiring any domain-specific data. Experimental results show that SEKI achieves state-of-the-art (SOTA) performance across various datasets and search spaces while requiring only 0.05 GPU-days, outperforming existing methods in both efficiency and accuracy. Furthermore, SEKI demonstrates strong generalization capabilities, achieving SOTA-competitive results across multiple tasks.
Chinese spelling correction (CSC) is a crucial task that aims to correct character errors in Chinese text. While conventional CSC focuses on character substitution errors caused by mistyping, two other common types of character errors, missing and redundant characters, have received less attention. These errors are often excluded from CSC datasets during the annotation process or ignored during evaluation, even when they have been annotated. This issue limits the practicality of the CSC task. To address this issue, we introduce the task of General Chinese Character Error Correction (C2EC), which focuses on all three types of character errors. We construct a high-quality C2EC benchmark by combining and manually verifying data from CCTC and Lemon datasets. We extend the training-free prompt-free CSC method to C2EC by using Levenshtein distance for handling length changes and leveraging an additional prompt-based large language model (LLM) to improve performance. Experiments show that our method enables a 14B-parameter LLM to be on par with models nearly 50 times larger on both conventional CSC and C2EC tasks, without any fine-tuning.
Large language models are increasingly used to process documents and facilitate question-answering on them. In our paper, we extract mechanistic circuits for this real-world language modeling task: context-augmented language modeling for extractive question-answering (QA) tasks and understand the potential benefits of circuits towards downstream applications such as data attribution to context information. We extract circuits as a function of internal model components (e.g., attention heads, MLPs) using causal mediation analysis techniques. Leveraging the extracted circuits, we first understand the interplay between the model's usage of parametric memory and retrieved context towards a better mechanistic understanding of context-augmented language models. We then identify a small set of attention heads in our circuit which performs reliable data attribution by default, thereby obtaining attribution for free in just the model's forward pass. Using this insight, we then introduce ATTNATTRIB, a fast data attribution algorithm which obtains state-of-the-art attribution results across various extractive QA benchmarks. Finally, we show the possibility to steer the language model towards answering from the context, instead of the parametric memory by using the attribution from ATTNATTRIB as an additional signal during the forward pass. Beyond mechanistic understanding, our paper provides tangible applications of circuits in the form of reliable data attribution and model steering.
Robust alignment guardrails for large language models are becoming increasingly important with their widespread application. In contrast to previous studies, we demonstrate that inference-time activation interventions can bypass safety alignments and effectively steer model generations towards harmful AI coordination for Llama 2. Our method applies fine-grained interventions at specific model subcomponents, particularly attention heads, using a simple binary choice probing strategy. These interventions then generalise to the open-ended generation setting effectively circumventing safety guardrails. We show that probing single attention heads is more effective than intervening on full layers and intervening on only four attention heads is comparable to supervised fine-tuning. We further show that only a few example completions are needed to compute effective steering directions, which is an advantage over classical fine-tuning. Our findings highlight the shortcomings of current alignment techniques. In addition, our results suggest that, at the attention head level, activations encode fine-grained linearly separable behaviors. Practically, the approach offers a straightforward methodology to steer large language model behaviour, which could be extended to diverse domains beyond safety requiring fine-grained control over the model output. The code and datasets for this study can be found on https://github.com/PaulDrm/targeted_intervention.
Mathematical reasoning is an increasingly important indicator of large language model (LLM) capabilities, yet we lack understanding of how LLMs process even simple mathematical tasks. To address this, we reverse engineer how three mid-sized LLMs compute addition. We first discover that numbers are represented in these LLMs as a generalized helix, which is strongly causally implicated for the tasks of addition and subtraction, and is also causally relevant for integer division, multiplication, and modular arithmetic. We then propose that LLMs compute addition by manipulating this generalized helix using the "Clock" algorithm: to solve $a+b$, the helices for $a$ and $b$ are manipulated to produce the $a+b$ answer helix which is then read out to model logits. We model influential MLP outputs, attention head outputs, and even individual neuron preactivations with these helices and verify our understanding with causal interventions. By demonstrating that LLMs represent numbers on a helix and manipulate this helix to perform addition, we present the first representation-level explanation of an LLM's mathematical capability.
The AgentSociety Challenge is the first competition in the Web Conference that aims to explore the potential of Large Language Model (LLM) agents in modeling user behavior and enhancing recommender systems on web platforms. The Challenge consists of two tracks: the User Modeling Track and the Recommendation Track. Participants are tasked to utilize a combined dataset from Yelp, Amazon, and Goodreads, along with an interactive environment simulator, to develop innovative LLM agents. The Challenge has attracted 295 teams across the globe and received over 1,400 submissions in total over the course of 37 official competition days. The participants have achieved 21.9% and 20.3% performance improvement for Track 1 and Track 2 in the Development Phase, and 9.1% and 15.9% in the Final Phase, representing a significant accomplishment. This paper discusses the detailed designs of the Challenge, analyzes the outcomes, and highlights the most successful LLM agent designs. To support further research and development, we have open-sourced the benchmark environment at https://tsinghua-fib-lab.github.io/AgentSocietyChallenge.
We propose a novel family of language models, Latent-Thought Language Models (LTMs), which incorporate explicit latent thought vectors that follow an explicit prior model in latent space. These latent thought vectors guide the autoregressive generation of ground tokens through a Transformer decoder. Training employs a dual-rate optimization process within the classical variational Bayes framework: fast learning of local variational parameters for the posterior distribution of latent vectors, and slow learning of global decoder parameters. Empirical studies reveal that LTMs possess additional scaling dimensions beyond traditional LLMs, yielding a structured design space. Higher sample efficiency can be achieved by increasing training compute per token, with further gains possible by trading model size for more inference steps. Designed based on these scaling properties, LTMs demonstrate superior sample and parameter efficiency compared to conventional autoregressive models and discrete diffusion models. They significantly outperform these counterparts in validation perplexity and zero-shot language modeling. Additionally, LTMs exhibit emergent few-shot in-context reasoning capabilities that scale with model and latent size, and achieve competitive performance in conditional and unconditional text generation.
This paper focuses on the piloting of CyberScholar, a Generative AI (GenAI) assistant tool that aims to provide feedback on writing K-12 contexts. The aim was to use GenAI to provide formative and summative feedback on students' texts in English Language Arts (ELA), Social Studies, and Modern World History. The trials discussed in this paper involved Grades 7, 8, 10, and 11 and were conducted in three schools in the Midwest and one in the Northwest of the United States. The tool used two main mechanisms: "prompt engineering" based on participant teachers' assessment rubric and "fine-tuning" a Large Language Model (LLM) from a customized corpus of teaching materials using Retrieval Augmented Generation. This paper focuses on CyberScholar's potential to enhance students' writing abilities and support teachers in diverse subject areas requiring written assignments.
Millions of melanocytic skin lesions are examined by pathologists each year, the majority of which concern common nevi (i.e., ordinary moles). While most of these lesions can be diagnosed in seconds, writing the corresponding pathology report is much more time-consuming. Automating part of the report writing could, therefore, alleviate the increasing workload of pathologists. In this work, we develop a vision-language model specifically for the pathology domain of cutaneous melanocytic lesions. The model follows the Contrastive Captioner framework and was trained and evaluated using a melanocytic lesion dataset of 42,512 H&E-stained whole slide images and 19,645 corresponding pathology reports. Our results show that the quality scores of model-generated reports were on par with pathologist-written reports for common nevi, assessed by an expert pathologist in a reader study. While report generation revealed to be more difficult for rare melanocytic lesion subtypes, the cross-modal retrieval performance for these cases was considerably better.
Vision-language models in pathology enable multimodal case retrieval and automated report generation. Many of the models developed so far, however, have been trained on pathology reports that include information which cannot be inferred from paired whole slide images (e.g., patient history), potentially leading to hallucinated sentences in generated reports. To this end, we investigate how the selection of information from pathology reports for vision-language modeling affects the quality of the multimodal representations and generated reports. More concretely, we compare a model trained on full reports against a model trained on preprocessed reports that only include sentences describing the cell and tissue appearances based on the H&E-stained slides. For the experiments, we built upon the BLIP-2 framework and used a cutaneous melanocytic lesion dataset of 42,433 H&E-stained whole slide images and 19,636 corresponding pathology reports. Model performance was assessed using image-to-text and text-to-image retrieval, as well as qualitative evaluation of the generated reports by an expert pathologist. Our results demonstrate that text preprocessing prevents hallucination in report generation. Despite the improvement in the quality of the generated reports, training the vision-language model on full reports showed better cross-modal retrieval performance.