语言大模型 / LLM

The ``Pre-train, then fine-tune'' paradigm has revolutionized Natural Language Processing (NLP). In this context, transferable backdoors pose a severe threat to the Pre-trained Language Models (PLMs) supply chain, yet defensive research remains nascent, primarily relying on detecting anomalies in the output feature space. We identify a critical flaw that fine-tuning on downstream tasks inevitably modifies model parameters, shifting the output distribution and rendering pre-computed defense ineffective. To address this, we propose Patronus, a novel defense framework that shifts the defensive focus from output features to input-side invariance, exploiting the fact that adversarial triggers remain constant even as model weights change. To overcome the convergence challenges of discrete text optimization, Patronus introduces a multi-trigger contrastive search algorithm that effectively bridges gradient-based optimization with contrastive learning objectives. Furthermore, we employ a dual-stage mitigation strategy combining real-time input monitoring with model purification via adversarial training. Extensive experiments across 15 PLMs and nine tasks demonstrate that Patronus achieves $\geq98.3\%$ backdoor detection recall and reduces attack success rates to clean settings, significantly outperforming all state-of-the-art baselines in all settings. Code is available at https://github.com/zth855/Patronus.

Block-diffusion language models offer a promising path toward faster-than-autoregressive generation by combining block-wise autoregressive decoding with within-block parallel denoising. However, in the few-step regime needed for practical acceleration, standard confidence-thresholded decoding is often brittle: aggressive thresholds hurt quality, while conservative thresholds require unnecessary denoising steps. Existing approaches that address this issue either require additional training or incur extra test-time compute. We present S2D2, a training-free self-speculative decoding framework for block-diffusion language models. Our key observation is that a block-diffusion model becomes autoregressive when the block size is reduced to one, allowing the same pretrained model to act as both drafter and verifier. S2D2 inserts a speculative verification step into standard block-diffusion decoding and uses lightweight routing policies to decide when verification is worth its cost. This yields a hybrid decoding trajectory in which diffusion proposes tokens in parallel, while the autoregressive mode acts as a local sequence-level critic. Across three mainstream block-diffusion families, S2D2 consistently improves the accuracy-speed tradeoff over strong confidence-thresholding baselines. On SDAR, we observe up to $4.7\times$ speedup over autoregressive decoding, and up to $1.57\times$ over a tuned dynamic decoding baseline while improving accuracy by up to $4.5$ points. On LLaDA2.1-Mini, S2D2 remains complementary to built-in self-correction, including a conservative setting where it is $4.4\times$ faster than the static baseline with slightly higher accuracy.

Cross-lingual sentence encoders typically cover only a few hundred languages and often trade downstream quality for stronger alignment, limiting their adoption. We introduce OmniSONAR, a new family of omnilingual, cross-lingual and cross-modal sentence embedding models that natively embed text, speech, code, and mathematical expressions in a single semantic space, while delivering state-of-the-art downstream performance at the scale of thousands of languages, from high-resource to extremely low-resource varieties. To reach this scale without representation collapse, we use progressive training. We first learn a strong foundational space for 200 languages with an LLM-initialized encoder-decoder, combining token-level decoding with a novel split-softmax contrastive loss and synthetic hard negatives. Building on this foundation, we expand to several thousands language varieties via a two-stage teacher-student encoder distillation framework. Finally, we demonstrate the cross-modal extensibility of this space by seamlessly mapping 177 spoken languages into it. OmniSONAR halves cross-lingual similarity search error on the 200-language FLORES dataset and reduces error by a factor of 15 on the 1,560-language BIBLE benchmark. It also enables strong translation, outperforming NLLB-3B on multilingual benchmarks and exceeding prior models (including much larger LLMs) by 15 chrF++ points on 1,560 languages into English BIBLE translation. OmniSONAR also performs strongly on MTEB and XLCoST. For speech, OmniSONAR achieves a 43% lower similarity-search error and reaches 97% of SeamlessM4T speech-to-text quality, despite being zero-shot for translation (trained only on ASR data). Finally, by training an encoder-decoder LM, Spectrum, exclusively on English text processing OmniSONAR embedding sequences, we unlock high-performance transfer to thousands of languages and speech for complex downstream tasks.

Due to their architecture and vast pre-training data, large language models (LLMs) demonstrate strong text classification performance. However, LLM output - here, the category assigned to a text - depends heavily on the wording of the prompt. While literature on prompt engineering is expanding, few studies focus on classification tasks, and even fewer address domains like psychology, where constructs have precise, theory-driven definitions that may not be well represented in pre-training data. We present an empirical framework for optimizing LLM performance for identifying constructs in texts via prompt engineering. We experimentally evaluate five prompting strategies -- codebook-guided empirical prompt selection, automatic prompt engineering, persona prompting, chain-of-thought reasoning, and explanatory prompting - with zero-shot and few-shot classification. We find that persona, chain-of-thought, and explanations do not fully address performance loss accompanying a badly worded prompt. Instead, the most influential features of a prompt are the construct definition, task framing, and, to a lesser extent, the examples provided. Across three constructs and two models, the classifications most aligned with expert judgments resulted from a few-shot prompt combining codebook-guided empirical prompt selection with automatic prompt engineering. Based on our findings, we recommend that researchers generate and evaluate as many prompt variants as feasible, whether human-crafted, automatically generated, or ideally both, and select prompts and examples based on empirical performance in a training dataset, validating the final approach in a holdout set. This procedure offers a practical, systematic, and theory-driven method for optimizing LLM prompts in settings where alignment with expert judgment is critical.

Understanding the dynamic evolution of opinions, such as U.S. public attitudes toward China, is essential for assessing geopolitical risks. However, existing LLM-based multiagent simulators predominantly rely on static rules and fixed datasets, limiting their ability to capture the dynamic, event-driven nature of macro-level opinion shifts in real-world settings. To address this limitation, we propose an Event-Steered Multi-Agent Simulator (ES-MAS), in which significant events and daily news continuously drive opinion evolution through dynamic interactions among agents. We first construct the China-U.S. Relation Evolution (CURE) dataset, covering 20 quarters from 2021 to 2025, including 258 major events and over 14,000 daily news articles, and providing a comprehensive temporal foundation for modeling opinion dynamics. Building upon the CURE dataset, we propose a Dual-Stream Data Integration Engine (DSDIE) that aligns simulations with historical timelines via macro-level events while enabling personalized information exposure based on individual agent profiles and contextual signals. Furthermore, we design a News-Driven Dynamic Interaction (NDDI) module, which adaptively groups agents with shared news interests into localized interaction contexts, facilitating bottom-up consensus formation while mitigating the risk of isolated information cocoons. Experimental results on the CURE dataset demonstrate that ES-MAS substantially outperforms existing simulators in reproducing real-world historical trends, offering a scalable and effective framework for modeling dynamic opinion evolution.

Mathematical benchmarks consisting of a range of mathematics problems are widely used to evaluate the reasoning abilities of large language models, yet little is known about how their structural properties influence model behaviour. In this work, we investigate two structural length variables, prompt length and solution length, and analyse how they relate to model performance on a newly constructed adversarial dataset of expert-authored mathematics problems. We find that both prompt and solution lengths correlate positively with increased model failure across models. We also include a secondary, exploratory analysis of cross-model disagreement. Under a difficulty-adjusted normalised analysis, both variables retain weak negative associations with realised model separation, slightly stronger for prompt length. Overall, our main robust finding is that structural length is linked to empirical difficulty in this dataset.

The rapid adoption of generative artificial intelligence (GenAI) in schools raises concerns about students' uncritical reliance on its outputs. Effective use of large language models (LLMs) requires not only technical knowledge but also the ability to monitor, evaluate, and regulate one's interaction with the system, processes closely tied to metacognitive regulation. These skills are still developing in middle school, making students particularly vulnerable to over-trust and premature acceptance of AI outputs. Because classroom time and teacher training resources are constrained, there is a pressing need to develop and evaluate AI literacy interventions that can be implemented under realistic school conditions. We report a controlled classroom study examining whether a two-hour AI literacy workshop improves students' interaction strategies and quality of final answers in LLM-supported science problem solving. A total of 116 students (grades 8-9; ages 13-15) completed six science investigation tasks using a generative AI system. Two days prior, the intervention group attended the workshop, which combined information about how LLMs work and fail with practical guidance on prompting and response evaluation; the control group received no training. Trained students showed less uncritical reliance on the system: they more often reformulated queries, asked follow-up questions, and more accurately judged response correctness, leading to better performance. In contrast, GenAI and metacognitive self-report scores did not predict performance, suggesting that effective use of generative AI depends less on self-reported measures and more on explicit training in interaction regulation. Overall, the results show that brief, scalable AI literacy instruction can meaningfully improve how middle-school students use generative AI in school-like learning activities.

Speech Large Language Models (SpeechLLMs) process spoken input directly, retaining cues such as accent and perceived gender that were previously removed in cascaded pipelines. This introduces speaker identity dependent variation in responses. We present a large-scale intersectional evaluation of accent and gender bias in three SpeechLLMs using 2,880 controlled interactions across six English accents and two gender presentations, keeping linguistic content constant through voice cloning. Using pointwise LLM-judge ratings, pairwise comparisons, and Best-Worst Scaling with human validation, we detect recurring directional disparities. Eastern European-accented speech receives lower helpfulness scores, particularly for female-presenting voices. Responses remain polite but differ in helpfulness. While LLM judges capture the directional trend of these biases, human evaluators exhibit significantly higher sensitivity, showing stronger accent-level contrasts.

In this paper, we investigate the potential of open-source Large Language Models (LLMs) for grading Unified Modeling Language (UML) class diagrams. In contrast to existing work, which primarily evaluates proprietary LLMs, we focus on non-proprietary models, making our approach suitable for universities where transparency and cost are critical. Additionally, existing studies assess performance over complete diagrams rather than individual criteria, offering limited insight into how automated grading aligns with human evaluation. To address these gaps, we propose a grading pipeline in which student-generated UML class diagrams are independently evaluated by both teaching assistants (TAs) and LLMs. Grades are then compared at the level of individual criteria. We evaluate this pipeline through a quantitative study of 92 UML class diagrams from a software design course, comparing TA grades against assessments produced by six open-source LLMs. Performance is measured across individual criteria, highlighting areas where LLMs diverge from human graders. Our results show per-criterion accuracy of up to 88.56\% and a Pearson correlation coefficient of up to 0.78, representing a substantial improvement over previous work while using only open-source models. The models achieve performance close to that of a TA, suggesting a possible path toward a mixed-initiative grading system, where TAs are aided in their grading. Our findings demonstrate that open-source LLMs can effectively support UML class diagram grading by explicitly identifying alignment with grading criteria. The proposed pipeline provides a practical approach to managing increasing workloads with growing student counts.

Social media platforms frequently impose restrictive policies to moderate user content, prompting the emergence of creative evasion language strategies. This paper presents a multi-agent framework based on Large Language Models (LLMs) to simulate the iterative evolution of language strategies under regulatory constraints. In this framework, participant agents, as social media users, continuously evolve their language expression, while supervisory agents emulate platform-level regulation by assessing policy violations. To achieve a more faithful simulation, we employ a dual design of language strategies (constraint and expression) to differentiate conflicting goals and utilize an LLM-driven GA (Genetic Algorithm) for the selection, mutation, and crossover of language strategies. The framework is evaluated using two distinct scenarios: an abstract password game and a realistic simulated illegal pet trade scenario. Experimental results demonstrate that as the number of dialogue rounds increases, both the number of uninterrupted dialogue turns and the accuracy of information transmission improve significantly. Furthermore, a user study with 40 participants validates the real-world relevance of the generated dialogues and strategies. Moreover, ablation studies validate the importance of the GA, emphasizing its contribution to long-term adaptability and improved overall results.

We revisit a classic "chicken-and-egg" problem in reinforcement learning: to safely improve a policy, the value function must be accurate on the state-visitation distribution of the updated policy. That distribution over states is unknown and cannot be sampled for the purposes of training the value function. Conservative updates solve this problem, but at the cost of shrinking the policy update. This paper explores an alternative solution, Approximate Next Policy Sampling (ANPS), which addresses the problem by modifying the training distribution rather than constraining the policy update. ANPS is satisfied if the distribution of the training data approximates that of the next policy. To demonstrate the feasibility and efficacy of ANPS, we introduce Stable Value Approximate Policy Iteration (SV-API). SV-API modifies the standard approximate policy iteration loop to hold the target policy fixed while an iteratively updated behavioral policy gathers relevant experience. It only commits to a new policy once a convergence criterion has been met. If certain stability criteria are met, the update is guaranteed to be safe; otherwise, it remains no less safe than standard approximate policy iteration. Applying SV-API to PPO yields Stable Value PPO (SV-PPO), which matches or improves performance on high-dimensional discrete (Atari) and continuous control benchmarks while executing substantially larger target policy updates. These results demonstrate the viability of ANPS as a new solution to this classic challenge in RL.

How does the choice of training data influence an AI model? This broad question is of central importance to interpretability, privacy, and basic science. At its technical core is the data deletion problem: after a reasonable amount of precomputation, quickly predict how the model would behave in a given situation if a given subset of training data had been excluded from the learning algorithm. We present a data deletion scheme capable of predicting model outputs with vanishing error $\varepsilon$ and failure probability $δ$ in the deep learning setting. Our precomputation and prediction algorithms are only $\tilde{O}(\log(1/δ)/\varepsilon^2)$ factors slower than regular training and inference, respectively. The storage requirements are those of $\tilde{O}(\log(1/δ)/\varepsilon^2)$ models. Our proof is based on an assumption that we call stability. In contrast to the assumptions made by prior work, stability appears to be fully compatible with learning powerful AI models. In support of this, we show that stability is satisfied in a minimal set of experiments with microgpt. Our code is available at https://github.com/SamSpo1/microgpt-sketch. At a technical level, our work is based on a new method for locally sketching an arithmetic circuit by computing higher-order derivatives in random complex directions. Forward-mode automatic differentiation allows cheap computation of these derivatives.

Conformal prediction provides distribution-free prediction intervals with finite-sample coverage guarantees, and recent work by Snell \& Griffiths reframes it as Bayesian Quadrature (BQ-CP), yielding powerful data-conditional guarantees via Dirichlet posteriors over thresholds. However, BQ-CP fundamentally requires the i.i.d. assumption. Meanwhile, weighted conformal prediction handles distribution shift via importance weights but remains frequentist, producing only point-estimate thresholds. We propose \textbf{Weighted Bayesian Conformal Prediction (WBCP)}, which generalizes BQ-CP to arbitrary importance-weighted settings by replacing the uniform Dirichlet $\Dir(1,\ldots,1)$ with a weighted Dirichlet $\Dir(\neff \cdot \tilde{w}_1, \ldots, \neff \cdot \tilde{w}_n)$, where $\neff$ is Kish's effective sample size. We prove four theoretical results: (1)~$\neff$ is the unique concentration parameter matching frequentist and Bayesian variances; (2)~posterior standard deviation decays as $O(1/\sqrt{\neff})$; (3)~BQ-CP's stochastic dominance guarantee extends to per-weight-profile data-conditional guarantees; (4)~the HPD threshold provides $O(1/\sqrt{\neff})$ improvement in conditional coverage. We instantiate WBCP for spatial prediction as \emph{Geographical BQ-CP}, where kernel-based spatial weights yield per-location posteriors with interpretable diagnostics. Experiments on synthetic and real-world spatial datasets demonstrate that WBCP maintains coverage guarantees while providing substantially richer uncertainty information.

Statistical inference with bandit data presents fundamental challenges owing to adaptive sampling, which violates the independence assumptions underlying classical asymptotic theory. Recent work has identified stability~\citep{laiwei82} as a sufficient condition for valid inference under adaptivity. This paper first provides a refined stability condition, stated in terms of the iterates of an online algorithm, and shows that a large class of regularized stochastic-mirror-descent-style algorithms satisfy it. This refined condition allows us to strengthen the asymptotic results of~\citet{laiwei82} in several ways. First, we derive a non-asymptotic Berry--Esseen bound for the empirical reward estimates under adaptive sampling. Second, we derive matching non-asymptotic upper and lower bounds on the regret of the proposed algorithm, yielding a precise characterization of its regret. Third, we show that these regularized algorithms preserve asymptotic normality and valid inference under a prescribed level of adversarial corruption. Finally, we show that regularization is necessary rather than incidental: Lai--Wei stability is incompatible with the optimal $O(\sqrt{T})$ regret rate -- the rate attained by unregularized algorithms such as EXP3 -- so that a controlled, polylogarithmic inflation in regret is the price of valid inference.

Online mirror descent (OMD) is a fundamental algorithmic paradigm that underlies many algorithms in optimization, machine learning and sequential decision-making. The OMD iterates are defined as solutions to optimization subproblems which, oftentimes, can be solved only approximately, leading to an inexact version of the algorithm. Nonetheless, existing OMD analyses typically assume an idealized error free setting, thereby limiting our understanding of performance guarantees that should be expected in practice. In this work we initiate a systematic study into inexact OMD, and uncover an intricate relation between regularizer smoothness and robustness to approximation errors. When the regularizer is uniformly smooth, we establish a tight bound on the excess regret due to errors. Then, for barrier regularizers over the simplex and its subsets, we identify a sharp separation: negative entropy requires exponentially small errors to avoid linear regret, whereas log-barrier and Tsallis regularizers remain robust even when the errors are only polynomial. Finally, we show that when the losses are stochastic and the domain is the simplex, negative entropy regains robustness-but this property does not extend to all subsets, where exponentially small errors are again necessary to avoid suboptimal regret.

Concolic testing for neural networks alternates concrete execution with constraint solving to search for inputs that flip model decisions. We present a concolic tester for Transformer classifiers that uses SHAP estimates to rank pending path predicates by their impact on the current prediction. To support self-attention with multiple heads in execution backed by SMT solving, we implement attention semantics in pure Python that are compatible with the solver and make the softmax boundary explicit by concretizing exponentiation arguments. We evaluate our method on CIFAR-10 across three compact Transformer classifiers, ResNet18, and VGG16 under a one-pixel budget and a 900s horizon. Across the 500 model--input pairs in this matched comparison, our method achieves 60% success, compared with 15% for a differential evolution baseline that treats the model as a black box. In the primary two-layer Transformer branch-ordering study, SHAP-based predicate prioritization raises success from 56% to 60% and reduces median attack time by 51%. These results show that influence-guided path exploration can make concolic testing a practical way to find adversarial examples in Transformer models.

World Model, the algorithmic simulator of the real-world environment which biological agents experience and act upon, has been an emerging topic in recent years due to the rising need to develop virtual agents with artificial (general) intelligence. There has been much discussion on what a world model really is, how to build it, how to use it, and how to evaluate it. In this essay, starting from the imagination in the famed Sci-Fi classic Dune, and drawing inspiration from the concept of ``hypothetical thinking'' in psychology literature, we argue the primary goal of a world model to be {\it simulating all actionable possibilities of the real world for purposeful reasoning and acting}. We examine the key design dimensions of world modeling: data, representation, architecture, learning objective, and usage, surveying existing approaches and analyzing their tradeoffs. Building on this examination, we propose a new Generative Latent Prediction (GLP) architecture for a general-purpose world model, based on stateful, hierarchical, multi-level, and mixed continuous/discrete representations, and a generative and self-supervised learning framework, with an outlook of a Physical, Agentic, and Nested (PAN) AGI system enabled by such a model.

Minimally redundant representations are typically learned by minimizing feature covariance. However, covariance-based methods fail to eliminate all dependencies/redundancies, as linearly uncorrelated variables can still exhibit nonlinear relationships. To address this, we introduce ADM, a differentiable algorithm that minimizes statistical dependence between feature dimensions through an adversarial game: auxiliary networks identify dependencies, while the encoder removes them. We prove that mutual independence is achieved at the global optimum, empirically verify convergence, and study three potential applications: extending PCA to nonlinear decorrelation, improving generalization in image classification, and preventing dimensional collapse in self-supervised learning. By promoting statistically independent representations, ADM paves the way for learning more robust, compressed, and generalizable representations across diverse applications.

We study conditional generation in diffusion models under hard constraints, where generated samples must satisfy prescribed events with probability one. Such constraints arise naturally in safety-critical applications and in rare-event simulation, where soft or reward-based guidance methods offer no guarantee of constraint satisfaction. Building on a probabilistic interpretation of diffusion models, we develop a principled conditional diffusion guidance framework based on Doob's h-transform, martingale representation and quadratic variation process. Specifically, the resulting guided dynamics augment a pretrained diffusion with an explicit drift correction involving the logarithmic gradient of a conditioning function, without modifying the pretrained score network. Leveraging martingale and quadratic-variation identities, we propose two novel off-policy learning algorithms based on a martingale loss and a martingale-covariation loss to estimate h and its gradient using only trajectories from the pretrained model. We provide non-asymptotic guarantees for the resulting conditional sampler in both total variation and Wasserstein distances, explicitly characterizing the impact of score approximation and guidance estimation errors. Numerical experiments demonstrate the effectiveness of the proposed methods in enforcing hard constraints and generating rare-event samples. The code of the numerical experiments can be found at https://github.com/ZhengyiGuo2002/CDG_Finance.