Modeling subrational agents, such as humans or economic households, is inherently challenging due to the difficulty in calibrating reinforcement learning models or collecting data that involves human subjects. Existing work highlights the ability of Large Language Models (LLMs) to address complex reasoning tasks and mimic human communication, while simulation using LLMs as agents shows emergent social behaviors, potentially improving our comprehension of human conduct. In this paper, we propose to investigate the use of LLMs to generate synthetic human demonstrations, which are then used to learn subrational agent policies though Imitation Learning. We make an assumption that LLMs can be used as implicit computational models of humans, and propose a framework to use synthetic demonstrations derived from LLMs to model subrational behaviors that are characteristic of humans (e.g., myopic behavior or preference for risk aversion). We experimentally evaluate the ability of our framework to model sub-rationality through four simple scenarios, including the well-researched ultimatum game and marshmallow experiment. To gain confidence in our framework, we are able to replicate well-established findings from prior human studies associated with the above scenarios. We conclude by discussing the potential benefits, challenges and limitations of our framework.

Identifying how much a model ${\widehat{p}}_{\theta}(Y|X)$ knows about the stochastic real-world process $p(Y|X)$ it was trained on is important to ensure it avoids producing incorrect or "hallucinated" answers or taking unsafe actions. But this is difficult for generative models because probabilistic predictions do not distinguish between per-response noise (aleatoric uncertainty) and lack of knowledge about the process (epistemic uncertainty), and existing epistemic uncertainty quantification techniques tend to be overconfident when the model underfits. We propose a general strategy for teaching a model to both approximate $p(Y|X)$ and also estimate the remaining gaps between ${\widehat{p}}_{\theta}(Y|X)$ and $p(Y|X)$: train it to predict pairs of independent responses drawn from the true conditional distribution, allow it to "cheat" by observing one response while predicting the other, then measure how much it cheats. Remarkably, we prove that being good at cheating (i.e. cheating whenever it improves your prediction) is equivalent to being second-order calibrated, a principled extension of ordinary calibration that allows us to construct provably-correct frequentist confidence intervals for $p(Y|X)$ and detect incorrect responses with high probability. We demonstrate empirically that our approach accurately estimates how much models don't know across ambiguous image classification, (synthetic) language modeling, and partially-observable navigation tasks, outperforming existing techniques.

Gaussian, an approach known as denoising score matching [10-12]. In the context of molecule generation, the score is The combinatorial scaling of the available chemical closely related to atomic forces. Consider training data space with molecule size is one of the main challenges that comprise configurations sampled using molecular in the design of new molecules and materials. Generative dynamics or other methods from an underlying Boltzmann modelling aims to solve this by directly proposing distribution, x exp ( βU(x)) /Z. Here, x = structures with desirable properties, without exhaustively {r, z} is a set that represents a molecule, with r the enumerating and screening candidates. Recently, atomic positions and z the chemical elements, U(x) the diffusion-based models have achieved impressive results potential energy, β the inverse temperature, and Z the in molecular docking [1] and generation of linkers [2], partition function. In this case, when the elements z drug-like molecules [3, 4] and crystal structures [5, 6]. are fixed, the score of the data distribution s(x, 0) corresponds Diffusion models are trained to reverse a stochastic to the atomic force (defined as the negative gradient noising process, which gradually corrupts samples of of the potential energy) up to a multiplicative constant: training data until they are indistinguishable from samples drawn from an uninformative prior distribution, such as a standard Gaussian [7-9].

Prompt optimization aims to find the best prompt to a large language model (LLM) for a given task. LLMs have been successfully used to help find and improve prompt candidates for single-step tasks. However, realistic tasks for agents are multi-step and introduce new challenges: (1) Prompt content is likely to be more extensive and complex, making it more difficult for LLMs to analyze errors, (2) the impact of an individual step is difficult to evaluate, and (3) different people may have varied preferences about task execution. While humans struggle to optimize prompts, they are good at providing feedback about LLM outputs; we therefore introduce a new LLM-driven discrete prompt optimization framework that incorporates human-designed feedback rules about potential errors to automatically offer direct suggestions for improvement. Our framework is stylized as a genetic algorithm in which an LLM generates new candidate prompts from a parent prompt and its associated feedback; we use a learned heuristic function that predicts prompt performance to efficiently sample from these candidates. This approach significantly outperforms both human-engineered prompts and several other prompt optimization methods across eight representative multi-step tasks (an average 27.7% and 28.2% improvement to current best methods on GPT-3.5 and GPT-4, respectively). We further show that the score function for tasks can be modified to better align with individual preferences. We believe our work can serve as a benchmark for automatic prompt optimization for LLM-driven multi-step tasks. Datasets and Codes are available at https://github.com/yongchao98/PROMST. Project Page is available at https://yongchao98.github.io/MIT-REALM-PROMST.

To evaluate code large language models (LLMs), research has relied on a few small manually curated benchmarks, such as HumanEval and MBPP, which represent a narrow part of the real-world software domains. In this work, we introduce round-trip correctness (RTC) as an alternative evaluation method. RTC allows Code LLM evaluation on a broader spectrum of real-world software domains without the need for costly human curation. RTC rests on the idea that we can ask a model to make a prediction (e.g., describe some code using natural language), feed that prediction back (e.g., synthesize code from the predicted description), and check if this round-trip leads to code that is semantically equivalent to the original input. We show how to employ RTC to evaluate code synthesis and editing. We find that RTC strongly correlates with model performance on existing narrow-domain code synthesis benchmarks while allowing us to expand to a much broader set of domains and tasks which was not previously possible without costly human annotations.

The advancement of Large Vision-Language Models (LVLMs) has increasingly highlighted the critical issue of their tendency to hallucinate non-existing objects in the images. To address this issue, previous works focused on using specially curated datasets or powerful LLMs (e.g., GPT-3.5) to rectify the outputs of LVLMs. However, these approaches require either expensive training/fine-tuning or API access to advanced LLMs to correct the model's output post-generation. In this paper, we tackle this challenge by introducing a framework called Mitigating hallucinAtion via classifieR-Free guIdaNcE (MARINE), which is both training-free and API-free, and can effectively and efficiently reduce object hallucinations during the generation process. Specifically, MARINE enriches the visual context of LVLMs by integrating existing open-source vision models, and employs classifier-free guidance to incorporate the additional object grounding features to improve the precision of LVLMs' generations. Through comprehensive evaluations across $6$ popular LVLMs with diverse evaluation metrics, we demonstrate the effectiveness of MARINE, which even outperforms existing fine-tuning-based methods. Remarkably, it not only reduces hallucinations but also improves the detailedness of LVLMs' generations, as assessed by GPT-4V.

Jailbreaks on Large language models (LLMs) have recently received increasing attention. For a comprehensive assessment of LLM safety, it is essential to consider jailbreaks with diverse attributes, such as contextual coherence and sentiment/stylistic variations, and hence it is beneficial to study controllable jailbreaking, i.e. how to enforce control on LLM attacks. In this paper, we formally formulate the controllable attack generation problem, and build a novel connection between this problem and controllable text generation, a well-explored topic of natural language processing. Based on this connection, we adapt the Energy-based Constrained Decoding with Langevin Dynamics (COLD), a state-of-the-art, highly efficient algorithm in controllable text generation, and introduce the COLD-Attack framework which unifies and automates the search of adversarial LLM attacks under a variety of control requirements such as fluency, stealthiness, sentiment, and left-right-coherence. The controllability enabled by COLD-Attack leads to diverse new jailbreak scenarios which not only cover the standard setting of generating fluent suffix attacks, but also allow us to address new controllable attack settings such as revising a user query adversarially with minimal paraphrasing, and inserting stealthy attacks in context with left-right-coherence. Our extensive experiments on various LLMs (Llama-2, Mistral, Vicuna, Guanaco, GPT-3.5) show COLD-Attack's broad applicability, strong controllability, high success rate, and attack transferability. Our code is available at https://github.com/Yu-Fangxu/COLD-Attack.

Human learning is sensitive to rule-like structure and the curriculum of examples used for training. In tasks governed by succinct rules, learning is more robust when related examples are blocked across trials, but in the absence of such rules, interleaving is more effective. To date, no neural model has simultaneously captured these seemingly contradictory effects. Here we show that this same tradeoff spontaneously emerges with "in-context learning" (ICL) both in neural networks trained with metalearning and in large language models (LLMs). ICL is the ability to learn new tasks "in context" - without weight changes - via an inner-loop algorithm implemented in activation dynamics. Experiments with pretrained LLMs and metalearning transformers show that ICL exhibits the blocking advantage demonstrated in humans on a task involving rule-like structure, and conversely, that concurrent in-weight learning reproduces the interleaving advantage observed in humans on tasks lacking such structure.

Text-to-SQL semantic parsing has made significant progress in recent years, with various models demonstrating impressive performance on the challenging Spider benchmark. However, it has also been shown that these models often struggle to generalize even when faced with small perturbations of previously (accurately) parsed expressions. This is mainly due to the linguistic form of questions in Spider which are overly specific, unnatural, and display limited variation. In this work, we use data augmentation to enhance the robustness of text-to-SQL parsers against natural language variations. Existing approaches generate question reformulations either via models trained on Spider or only introduce local changes. In contrast, we leverage the capabilities of large language models to generate more realistic and diverse questions. Using only a few prompts, we achieve a two-fold increase in the number of questions in Spider. Training on this augmented dataset yields substantial improvements on a range of evaluation sets, including robustness benchmarks and out-of-domain data.

We explored the viability of Large Language Models (LLMs) for triggering and personalizing content for Just-in-Time Adaptive Interventions (JITAIs) in digital health. JITAIs are being explored as a key mechanism for sustainable behavior change, adapting interventions to an individual's current context and needs. However, traditional rule-based and machine learning models for JITAI implementation face scalability and reliability limitations, such as lack of personalization, difficulty in managing multi-parametric systems, and issues with data sparsity. To investigate JITAI implementation via LLMs, we tested the contemporary overall performance-leading model 'GPT-4' with examples grounded in the use case of fostering heart-healthy physical activity in outpatient cardiac rehabilitation. Three personas and five sets of context information per persona were used as a basis of triggering and personalizing JITAIs. Subsequently, we generated a total of 450 proposed JITAI decisions and message content, divided equally into JITAIs generated by 10 iterations with GPT-4, a baseline provided by 10 laypersons (LayPs), and a gold standard set by 10 healthcare professionals (HCPs). Ratings from 27 LayPs indicated that JITAIs generated by GPT-4 were superior to those by HCPs and LayPs over all assessed scales: i.e., appropriateness, engagement, effectiveness, and professionality. This study indicates that LLMs have significant potential for implementing JITAIs as a building block of personalized or "precision" health, offering scalability, effective personalization based on opportunistically sampled information, and good acceptability.