Large Language Models (LLMs) are known to be vulnerable to backdoor attacks, where triggers embedded in poisoned samples can maliciously alter LLMs' behaviors. In this paper, we move beyond attacking LLMs and instead examine backdoor attacks through the novel lens of natural language explanations. Specifically, we leverage LLMs' generative capabilities to produce human-readable explanations for their decisions, enabling direct comparisons between explanations for clean and poisoned samples. Our results show that backdoored models produce coherent explanations for clean inputs but diverse and logically flawed explanations for poisoned data, a pattern consistent across classification and generation tasks for different backdoor attacks. Further analysis reveals key insights into the explanation generation process. At the token level, explanation tokens associated with poisoned samples only appear in the final few transformer layers. At the sentence level, attention dynamics indicate that poisoned inputs shift attention away from the original input context during explanation generation. These findings enhance our understanding of backdoor mechanisms in LLMs and present a promising framework for detecting vulnerabilities through explainability.

Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language processing tasks, but their vulnerability to jailbreak attacks poses significant security risks. This survey paper presents a comprehensive analysis of recent advancements in attack strategies and defense mechanisms within the field of Large Language Model (LLM) red-teaming. We analyze various attack methods, including gradient-based optimization, reinforcement learning, and prompt engineering approaches. We discuss the implications of these attacks on LLM safety and the need for improved defense mechanisms. This work aims to provide a thorough understanding of the current landscape of red-teaming attacks and defenses on LLMs, enabling the development of more secure and reliable language models.

Image editing has advanced significantly with the development of diffusion models using both inversion-based and instruction-based methods. However, current inversion-based approaches struggle with big modifications (e.g., adding or removing objects) due to the structured nature of inversion noise, which hinders substantial changes. Meanwhile, instruction-based methods often constrain users to black-box operations, limiting direct interaction for specifying editing regions and intensity. To address these limitations, we propose BrushEdit, a novel inpainting-based instruction-guided image editing paradigm, which leverages multimodal large language models (MLLMs) and image inpainting models to enable autonomous, user-friendly, and interactive free-form instruction editing. Specifically, we devise a system enabling free-form instruction editing by integrating MLLMs and a dual-branch image inpainting model in an agent-cooperative framework to perform editing category classification, main object identification, mask acquisition, and editing area inpainting. Extensive experiments show that our framework effectively combines MLLMs and inpainting models, achieving superior performance across seven metrics including mask region preservation and editing effect coherence.

As Large Language Model (LLM)-based agents increasingly undertake real-world tasks and engage with human society, how well do we understand their behaviors? We (1) investigate how LLM agents' prosocial behaviors -- a fundamental social norm -- can be induced by different personas and benchmarked against human behaviors; and (2) introduce a behavioral and social science approach to evaluate LLM agents' decision-making. We explored how different personas and experimental framings affect these AI agents' altruistic behavior in dictator games and compared their behaviors within the same LLM family, across various families, and with human behaviors. The findings reveal substantial variations and inconsistencies among LLMs and notable differences compared to human behaviors. Merely assigning a human-like identity to LLMs does not produce human-like behaviors. Despite being trained on extensive human-generated data, these AI agents are unable to capture the internal processes of human decision-making. Their alignment with human is highly variable and dependent on specific model architectures and prompt formulations; even worse, such dependence does not follow a clear pattern. LLMs can be useful task-specific tools but are not yet intelligent human-like agents.

We create two German-only decoder models, LL\"aMmlein 120M and 1B, transparently from scratch and publish them, along with the training data, for the German NLP research community to use. The model training involved several key steps, including extensive data preprocessing, the creation of a custom German tokenizer, the training itself, as well as the evaluation of the final models on various benchmarks. Throughout the training process, multiple checkpoints were saved and analyzed using the SuperGLEBer benchmark to monitor the models' learning dynamics. Compared to state-of-the-art models on the SuperGLEBer benchmark, both LL\"aMmlein models performed competitively, consistently matching or surpassing models with similar parameter sizes. The results show that the models' quality scales with size as expected, but performance improvements on some tasks plateaued early, offering valuable insights into resource allocation for future model development.

Large language models (LLMs) have a transformative impact on a variety of scientific tasks across disciplines including biology, chemistry, medicine, and physics. However, ensuring the safety alignment of these models in scientific research remains an underexplored area, with existing benchmarks primarily focusing on textual content and overlooking key scientific representations such as molecular, protein, and genomic languages. Moreover, the safety mechanisms of LLMs in scientific tasks are insufficiently studied. To address these limitations, we introduce SciSafeEval, a comprehensive benchmark designed to evaluate the safety alignment of LLMs across a range of scientific tasks. SciSafeEval spans multiple scientific languages-including textual, molecular, protein, and genomic-and covers a wide range of scientific domains. We evaluate LLMs in zero-shot, few-shot and chain-of-thought settings, and introduce a "jailbreak" enhancement feature that challenges LLMs equipped with safety guardrails, rigorously testing their defenses against malicious intention. Our benchmark surpasses existing safety datasets in both scale and scope, providing a robust platform for assessing the safety and performance of LLMs in scientific contexts. This work aims to facilitate the responsible development and deployment of LLMs, promoting alignment with safety and ethical standards in scientific research.

Methods designed to remove such knowledge must do so from all prompt directions, in a multi-lingual capacity and without degrading general model performance. To this end, we introduce the targeted angular reversal (TARS) method of knowledge removal from LLMs. The TARS method firstly leverages the LLM in combination with a detailed prompt to aggregate information about a selected concept in the internal representation space of the LLM. It then refines this approximate concept vector to trigger the concept token with high probability, by perturbing the approximate concept vector with noise and transforming it into token scores with the language model head. The feed-forward weight vectors in the LLM which operate directly on the internal representation space, and have the highest cosine similarity with this refined targeting vector, are then replaced by a reversed targeting vector, thus limiting the ability of the concept to propagate through the model. The modularity of the TARS method allows for a sequential removal of concepts from Llama 3.1 8B, such as the famous literary detective Sherlock Holmes, and the planet Saturn. It is demonstrated that the probability of triggering target concepts can be reduced to 0.00 with as few as 1 TARS edit, whilst simultaneously removing the knowledge bi-directionally. Moreover, knowledge is shown to be removed across all languages despite only being targeted in English. Importantly, TARS has minimal impact on the general model capabilities, as after removing 5 diverse concepts in a modular fashion, there is minimal KL divergence in the next token probabilities of the LLM on large corpora of Wikipedia text (median of 0.002).

Adsorption energy is a key reactivity descriptor in catalysis, enabling efficient screening for optimal catalysts. However, determining adsorption energy typically requires evaluating numerous adsorbate-catalyst configurations. Current algorithmic approaches rely on exhaustive enumeration of adsorption sites and configurations, which makes the process computationally intensive and does not inherently guarantee the identification of the global minimum energy. In this work, we introduce Adsorb-Agent, a Large Language Model (LLM) agent designed to efficiently identify system-specific stable adsorption configurations corresponding to the global minimum adsorption energy. Adsorb-Agent leverages its built-in knowledge and emergent reasoning capabilities to strategically explore adsorption configurations likely to hold adsorption energy. By reducing the reliance on exhaustive sampling, it significantly decreases the number of initial configurations required while improving the accuracy of adsorption energy predictions. We evaluate Adsorb-Agent's performance across twenty representative systems encompassing a range of complexities. The Adsorb-Agent successfully identifies comparable adsorption energies for 83.7% of the systems and achieves lower energies, closer to the actual global minimum, for 35% of the systems, while requiring significantly fewer initial configurations than conventional methods. Its capability is particularly evident in complex systems, where it identifies lower adsorption energies for 46.7% of systems involving intermetallic surfaces and 66.7% of systems with large adsorbate molecules. These results demonstrate the potential of Adsorb-Agent to accelerate catalyst discovery by reducing computational costs and improving the reliability of adsorption energy predictions.

Recent advances in Multi-modal Large Language Models (MLLMs), such as LLaVA-series models, are driven by massive machine-generated instruction-following data tuning. Such automatic instruction collection pipelines, however, inadvertently introduce significant variability in data quality. This paper introduces a novel instruction curation algorithm, derived from two unique perspectives, human and LLM preference alignment, to compress this vast corpus of machine-generated multimodal instructions to a compact and high-quality form: (i) For human preference alignment, we have collected a machine-generated multimodal instruction dataset and established a comprehensive set of both subjective and objective criteria to guide the data quality assessment critically from human experts. By doing so, a reward model was trained on the annotated dataset to internalize the nuanced human understanding of instruction alignment. (ii) For LLM preference alignment, given the instruction selected by the reward model, we propose leveraging the inner LLM used in MLLM to align the writing style of visual instructions with that of the inner LLM itself, resulting in LLM-aligned instruction improvement. Extensive experiments demonstrate that we can maintain or even improve model performance by compressing synthetic multimodal instructions by up to 90%. Impressively, by aggressively reducing the training instructions from 158k to 14k (9$\times$ smaller), our model consistently outperforms its full-size dataset counterpart across various MLLM benchmarks. Our project is available at https://github.com/DCDmllm/Align2LLaVA.

Readability-controlled text simplification (RCTS) rewrites texts to lower readability levels while preserving their meaning. RCTS models often depend on parallel corpora with readability annotations on both source and target sides. Such datasets are scarce and difficult to curate, especially at the sentence level. To reduce reliance on parallel data, we explore using instruction-tuned large language models for zero-shot RCTS. Through automatic and manual evaluations, we examine: (1) how different types of contextual information affect a model's ability to generate sentences with the desired readability, and (2) the trade-off between achieving target readability and preserving meaning. Results show that all tested models struggle to simplify sentences (especially to the lowest levels) due to models' limitations and characteristics of the source sentences that impede adequate rewriting. Our experiments also highlight the need for better automatic evaluation metrics tailored to RCTS, as standard ones often misinterpret common simplification operations, and inaccurately assess readability and meaning preservation.