As large language models (LLMs) see greater use in academic and commercial settings, there is increasing interest in methods that allow language models to generate texts aligned with human preferences. In this paper, we present an initial exploration of language model optimization for human preferences from direct outcome datasets, where each sample consists of a text and an associated numerical outcome measuring the reader's response. We first propose that language model optimization should be viewed as a causal problem to ensure that the model correctly learns the relationship between the text and the outcome. We formalize this causal language optimization problem, and we develop a method--causal preference optimization (CPO)--that solves an unbiased surrogate objective for the problem. We further extend CPO with doubly robust CPO (DR-CPO), which reduces the variance of the surrogate objective while retaining provably strong guarantees on bias. Finally, we empirically demonstrate the effectiveness of (DR-)CPO in optimizing state-of-the-art LLMs for human preferences on direct outcome data, and we validate the robustness of DR-CPO under difficult confounding conditions.
In this paper, we present an exploration of LLMs' abilities to problem solve with physical reasoning in situated environments. We construct a simple simulated environment and demonstrate examples of where, in a zero-shot setting, both text and multimodal LLMs display atomic world knowledge about various objects but fail to compose this knowledge in correct solutions for an object manipulation and placement task. We also use BLIP, a vision-language model trained with more sophisticated cross-modal attention, to identify cases relevant to object physical properties that that model fails to ground. Finally, we present a procedure for discovering the relevant properties of objects in the environment and propose a method to distill this knowledge back into the LLM.
In daily life, we encounter a variety of sounds, both desirable and undesirable, with limited control over their presence and volume. Our work introduces "Listen, Chat, and Edit" (LCE), a novel multimodal sound mixture editor that modifies each sound source in a mixture based on user-provided text instructions. LCE distinguishes itself with a user-friendly chat interface and its unique ability to edit multiple sound sources simultaneously within a mixture, without needing to separate them. Users input open-vocabulary text prompts, which are interpreted by a large language model to create a semantic filter for editing the sound mixture. The system then decomposes the mixture into its components, applies the semantic filter, and reassembles it into the desired output. We developed a 160-hour dataset with over 100k mixtures, including speech and various audio sources, along with text prompts for diverse editing tasks like extraction, removal, and volume control. Our experiments demonstrate significant improvements in signal quality across all editing tasks and robust performance in zero-shot scenarios with varying numbers and types of sound sources.
Large Language Models (LLMs) are widely used to evaluate natural language generation tasks as automated metrics. However, the likelihood, a measure of LLM's plausibility for a sentence, can vary due to superficial differences in sentences, such as word order and sentence structure. It is therefore possible that there might be a likelihood bias if LLMs are used for evaluation: they might overrate sentences with higher likelihoods while underrating those with lower likelihoods. In this paper, we investigate the presence and impact of likelihood bias in LLM-based evaluators. We also propose a method to mitigate the likelihood bias. Our method utilizes highly biased instances as few-shot examples for in-context learning. Our experiments in evaluating the data-to-text and grammatical error correction tasks reveal that several LLMs we test display a likelihood bias. Furthermore, our proposed method successfully mitigates this bias, also improving evaluation performance (in terms of correlation of models with human scores) significantly.
Large language models (LLMs) have demonstrated impressive abilities in generating unstructured natural language according to instructions. However, their performance can be inconsistent when tasked with producing text that adheres to specific structured formats, which is crucial in applications like named entity recognition (NER) or relation extraction (RE). To address this issue, this paper introduces an efficient method, G&O, to enhance their structured text generation capabilities. It breaks the generation into a two-step pipeline: initially, LLMs generate answers in natural language as intermediate responses. Subsequently, LLMs are asked to organize the output into the desired structure, using the intermediate responses as context. G&O effectively separates the generation of content from the structuring process, reducing the pressure of completing two orthogonal tasks simultaneously. Tested on zero-shot NER and RE, the results indicate a significant improvement in LLM performance with minimal additional efforts. This straightforward and adaptable prompting technique can also be combined with other strategies, like self-consistency, to further elevate LLM capabilities in various structured text generation tasks.
Recent large-scale Text-To-Image (T2I) models such as DALLE-3 demonstrate great potential in new applications, but also face unprecedented fairness challenges. Prior studies revealed gender biases in single-person image generation, but T2I model applications might require portraying two or more people simultaneously. Potential biases in this setting remain unexplored, leading to fairness-related risks in usage. To study these underlying facets of gender biases in T2I models, we propose a novel Paired Stereotype Test (PST) bias evaluation framework. PST prompts the model to generate two individuals in the same image. They are described with two social identities that are stereotypically associated with the opposite gender. Biases can then be measured by the level of conformation to gender stereotypes in generated images. Using PST, we evaluate DALLE-3 from 2 perspectives: biases in gendered occupation and biases in organizational power. Despite seemingly fair or even anti-stereotype single-person generations, PST still unveils gendered occupational and power associations. Moreover, compared to single-person settings, DALLE-3 generates noticeably more masculine figures under PST for individuals with male-stereotypical identities. PST is therefore effective in revealing underlying gender biases in DALLE-3 that single-person settings cannot capture. Our findings reveal the complicated patterns of gender biases in modern T2I models, further highlighting the critical fairness challenges in multimodal generative systems.
When communicating routes in natural language, the concept of {\em acquired spatial knowledge} is crucial for geographic information retrieval (GIR) and in spatial cognitive research. However, NLP navigation studies often overlook the impact of such acquired knowledge on textual descriptions. Current navigation studies concentrate on egocentric local descriptions (e.g., `it will be on your right') that require reasoning over the agent's local perception. These instructions are typically given as a sequence of steps, with each action-step explicitly mentioning and being followed by a landmark that the agent can use to verify they are on the right path (e.g., `turn right and then you will see...'). In contrast, descriptions based on knowledge acquired through a map provide a complete view of the environment and capture its overall structure. These instructions (e.g., `it is south of Central Park and a block north of a police station') are typically non-sequential, contain allocentric relations, with multiple spatial relations and implicit actions, without any explicit verification. This paper introduces the Rendezvous (RVS) task and dataset, which includes 10,404 examples of English geospatial instructions for reaching a target location using map-knowledge. Our analysis reveals that RVS exhibits a richer use of spatial allocentric relations, and requires resolving more spatial relations simultaneously compared to previous text-based navigation benchmarks.
In this paper, we investigate the effectiveness of various LLMs in interpreting tabular data through different prompting strategies and data formats. Our analysis extends across six benchmarks for table-related tasks such as question-answering and fact-checking. We introduce for the first time the assessment of LLMs' performance on image-based table representations. Specifically, we compare five text-based and three image-based table representations, demonstrating the influence of representation and prompting on LLM performance. Our study provides insights into the effective use of LLMs on table-related tasks.
Multi-modal learning has emerged as an increasingly promising avenue in vision recognition, driving innovations across diverse domains ranging from media and education to healthcare and transportation. Despite its success, the robustness of multi-modal learning for visual recognition is often challenged by the unavailability of a subset of modalities, especially the visual modality. Conventional approaches to mitigate missing modalities in multi-modal learning rely heavily on algorithms and modality fusion schemes. In contrast, this paper explores the use of text-to-image models to assist multi-modal learning. Specifically, we propose a simple but effective multi-modal learning framework GTI-MM to enhance the data efficiency and model robustness against missing visual modality by imputing the missing data with generative transformers. Using multiple multi-modal datasets with visual recognition tasks, we present a comprehensive analysis of diverse conditions involving missing visual modality in data, including model training. Our findings reveal that synthetic images benefit training data efficiency with visual data missing in training and improve model robustness with visual data missing involving training and testing. Moreover, we demonstrate GTI-MM is effective with lower generation quantity and simple prompt techniques.
On-chip photonic processors for neural networks have potential benefits in both speed and energy efficiency but have not yet reached the scale at which they can outperform electronic processors. The dominant paradigm for designing on-chip photonics is to make networks of relatively bulky discrete components connected by one-dimensional waveguides. A far more compact alternative is to avoid explicitly defining any components and instead sculpt the continuous substrate of the photonic processor to directly perform the computation using waves freely propagating in two dimensions. We propose and demonstrate a device whose refractive index as a function of space, $n(x,z)$, can be rapidly reprogrammed, allowing arbitrary control over the wave propagation in the device. Our device, a 2D-programmable waveguide, combines photoconductive gain with the electro-optic effect to achieve massively parallel modulation of the refractive index of a slab waveguide, with an index modulation depth of $10^{-3}$ and approximately $10^4$ programmable degrees of freedom. We used a prototype device with a functional area of $12\,\text{mm}^2$ to perform neural-network inference with up to 49-dimensional input vectors in a single pass, achieving 96% accuracy on vowel classification and 86% accuracy on $7 \times 7$-pixel MNIST handwritten-digit classification. This is a scale beyond that of previous photonic chips relying on discrete components, illustrating the benefit of the continuous-waves paradigm. In principle, with large enough chip area, the reprogrammability of the device's refractive index distribution enables the reconfigurable realization of any passive, linear photonic circuit or device. This promises the development of more compact and versatile photonic systems for a wide range of applications, including optical processing, smart sensing, spectroscopy, and optical communications.