Research has shown the effectiveness of reasoning (e.g., Chain-of-Thought), planning (e.g., SelfAsk), and retrieval augmented generation strategies to improve the performance of Large Language Models (LLMs) on various tasks, such as question answering. However, using a single fixed strategy to answer different kinds of questions is suboptimal in performance and inefficient in terms of generated output tokens and performed retrievals. In our work, we propose a novel technique DyPlan, to induce a dynamic strategy selection process in LLMs, to improve performance and reduce costs in question-answering. DyPlan incorporates an initial decision step to select the most suitable strategy conditioned on the input question and guides the LLM's response generation accordingly. We extend DyPlan to DyPlan-verify, adding an internal verification and correction process to further enrich the generated answer. Experiments on three prominent multi-hop question answering (MHQA) datasets reveal how DyPlan can improve model performance by 7-13% while reducing the cost by 11-32% relative to the best baseline model.

It is staggering that words of the English language, which are on average represented by 5--6 bytes of ASCII, require as much as 24 kilobytes when served to large language models. We show that there is room for more information in every token embedding. We demonstrate that 1--3-layer transformers are capable of encoding and subsequently decoding standard English sentences into as little as a single 3-kilobyte token. Our work implies that even small networks can learn to construct valid English sentences and suggests the possibility of optimising large language models by moving from sub-word token embeddings towards larger fragments of text.

We introduce the "Belief State Transformer", a next-token predictor that takes both a prefix and suffix as inputs, with a novel objective of predicting both the next token for the prefix and the previous token for the suffix. The Belief State Transformer effectively learns to solve challenging problems that conventional forward-only transformers struggle with, in a domain-independent fashion. Key to this success is learning a compact belief state that captures all relevant information necessary for accurate predictions. Empirical ablations show that each component of the model is essential in difficult scenarios where standard Transformers fall short. For the task of story writing with known prefixes and suffixes, our approach outperforms the Fill-in-the-Middle method for reaching known goals and demonstrates improved performance even when the goals are unknown. Altogether, the Belief State Transformer enables more efficient goal-conditioned decoding, better test-time inference, and high-quality text representations on small scale problems.

Self-correction is one of the most amazing emerging capabilities of Large Language Models (LLMs), enabling LLMs to self-modify an inappropriate output given a natural language feedback which describes the problems of that output. Moral self-correction is a post-hoc approach correcting unethical generations without requiring a gradient update, making it both computationally lightweight and capable of preserving the language modeling ability. Previous works have shown that LLMs can self-debias, and it has been reported that small models, i.e., those with less than 22B parameters, are not capable of moral self-correction. However, there is no direct proof as to why such smaller models fall short of moral self-correction, though previous research hypothesizes that larger models are skilled in following instructions and understanding abstract social norms. In this paper, we empirically validate this hypothesis in the context of social stereotyping, through meticulous prompting. Our experimental results indicate that (i) surprisingly, 3.8B LLMs with proper safety alignment fine-tuning can achieve very good moral self-correction performance, highlighting the significant effects of safety alignment; and (ii) small LLMs are indeed weaker than larger-scale models in terms of comprehending social norms and self-explanation through CoT, but all scales of LLMs show bad self-correction performance given unethical instructions.

Recent approaches in machine learning often solve a task using a composition of multiple models or agentic architectures. When targeting a composed system with adversarial attacks, it might not be computationally or informationally feasible to train an end-to-end proxy model or a proxy model for every component of the system. We introduce a method to craft an adversarial attack against the overall multi-model system when we only have a proxy model for the final black-box model, and when the transformation applied by the initial models can make the adversarial perturbations ineffective. Current methods handle this by applying many copies of the first model/transformation to an input and then re-use a standard adversarial attack by averaging gradients, or learning a proxy model for both stages. To our knowledge, this is the first attack specifically designed for this threat model and our method has a substantially higher attack success rate (80% vs 25%) and contains 9.4% smaller perturbations (MSE) compared to prior state-of-the-art methods. Our experiments focus on a supervised image pipeline, but we are confident the attack will generalize to other multi-model settings [e.g. a mix of open/closed source foundation models], or agentic systems

Large Language Models (LLMs) are increasingly employed for simulations, enabling applications in role-playing agents and Computational Social Science (CSS). However, the reliability of these simulations is under-explored, which raises concerns about the trustworthiness of LLMs in these applications. In this paper, we aim to answer ``How reliable is LLM-based simulation?'' To address this, we introduce TrustSim, an evaluation dataset covering 10 CSS-related topics, to systematically investigate the reliability of the LLM simulation. We conducted experiments on 14 LLMs and found that inconsistencies persist in the LLM-based simulated roles. In addition, the consistency level of LLMs does not strongly correlate with their general performance. To enhance the reliability of LLMs in simulation, we proposed Adaptive Learning Rate Based ORPO (AdaORPO), a reinforcement learning-based algorithm to improve the reliability in simulation across 7 LLMs. Our research provides a foundation for future studies to explore more robust and trustworthy LLM-based simulations.

We introduce Dynamic Information Sub-Selection (DISS), a novel framework of AI assistance designed to enhance the performance of black-box decision-makers by tailoring their information processing on a per-instance basis. Blackbox decision-makers (e.g., humans or real-time systems) often face challenges in processing all possible information at hand (e.g., due to cognitive biases or resource constraints), which can degrade decision efficacy. DISS addresses these challenges through policies that dynamically select the most effective features and options to forward to the black-box decision-maker for prediction. We develop a scalable frequentist data acquisition strategy and a decision-maker mimicking technique for enhanced budget efficiency. We explore several impactful applications of DISS, including biased decision-maker support, expert assignment optimization, large language model decision support, and interpretability. Empirical validation of our proposed DISS methodology shows superior performance to state-of-the-art methods across various applications.

Behavior change interventions are important to coordinate societal action across a wide array of important applications, including the adoption of electrified vehicles to reduce emissions. Prior work has demonstrated that interventions for behavior must be personalized, and that the intervention that is most effective on average across a large group can result in a backlash effect that strengthens opposition among some subgroups. Thus, it is important to target interventions to different audiences, and to present them in a natural, conversational style. In this context, an important emerging application domain for large language models (LLMs) is conversational interventions for behavior change. In this work, we leverage prior work on understanding values motivating the adoption of battery electric vehicles. We leverage new advances in LLMs, combined with a contextual bandit, to develop conversational interventions that are personalized to the values of each study participant. We use a contextual bandit algorithm to learn to target values based on the demographics of each participant. To train our bandit algorithm in an offline manner, we leverage LLMs to play the role of study participants. We benchmark the persuasive effectiveness of our bandit-enhanced LLM against an unaided LLM generating conversational interventions without demographic-targeted values.

We present a benchmark for large language models designed to tackle one of the most knowledge-intensive tasks in data science: writing feature engineering code, which requires domain knowledge in addition to a deep understanding of the underlying problem and data structure. The model is provided with a dataset description in a prompt and asked to generate code transforming it. The evaluation score is derived from the improvement achieved by an XGBoost model fit on the modified dataset compared to the original data. By an extensive evaluation of state-of-the-art models and comparison to well-established benchmarks, we demonstrate that the FeatEng of our proposal can cheaply and efficiently assess the broad capabilities of LLMs, in contrast to the existing methods. The reference implementation is available at https://github.com/FeatEng/FeatEng. The rapid evolution of LLMs has significantly expanded their capabilities in processing and generating human-like text. As these models become increasingly sophisticated, defining what constitutes a meaningful benchmark is becoming harder and harder, as it is much easier to distinguish between bad and good models than between good and better. Today, the limitations of LLMs are predominantly assessed using benchmarks focused on language understanding, world knowledge, code generation, or mathematical reasoning in separation. This setup, however, overlooks some critical capabilities that can be measured in scenarios requiring inregration of skills and verification of their instrumental value in complex, real-world problems. We argue that well-designed LLM benchmarks should embody the following qualities, each reflecting a fundamental aspect of problem-solving ability: 1. Practical Usability. We demand that tasks are grounded in real-world problems where solutions have high functional value. This ensures that improvements in the observed performance translates into tangible benefits, aligning with the pragmatist view on the instrumental value of knowledge and truth, meaning that the validity of an idea depends on its practical utility in achieving desired outcomes (James, 1907). We would value LLM's knowledge for its role in enabling reasoning, decision-making, and problem-solving. The benchmark should be designed to evaluate not only the breadth of a model's knowledge base but also, more importantly, its capacity to dynamically and effectively apply this knowledge within various functional contexts, similarly to how functionalism frames it (Block, 1980). We opt for assessing models concerning their ability to seamlessly combine various competencies, in contrast to measuring them in separation.

Machine unlearning (MU) has gained significant attention as a means to remove specific data from trained models without requiring a full retraining process. While progress has been made in unimodal domains like text and image classification, unlearning in multimodal models remains relatively underexplored. In this work, we address the unique challenges of unlearning in CLIP, a prominent multimodal model that aligns visual and textual representations. We introduce CLIPErase, a novel approach that disentangles and selectively forgets both visual and textual associations, ensuring that unlearning does not compromise model performance. CLIPErase consists of three key modules: a Forgetting Module that disrupts the associations in the forget set, a Retention Module that preserves performance on the retain set, and a Consistency Module that maintains consistency with the original model. Extensive experiments on the CIFAR-100 and Flickr30K datasets across four CLIP downstream tasks demonstrate that CLIPErase effectively forgets designated associations in zero-shot tasks for multimodal samples, while preserving the model's performance on the retain set after unlearning.