This paper addresses the challenge of leveraging imperfect language models to guide human decision-making in the context of a grounded navigation task. We show that an imperfect instruction generation model can be complemented with an effective communication mechanism to become more successful at guiding humans. The communication mechanism we build comprises models that can detect potential hallucinations in instructions and suggest practical alternatives, and an intuitive interface to present that information to users. We show that this approach reduces the human navigation error by up to 29% with no additional cognitive burden. This result underscores the potential of integrating diverse communication channels into AI systems to compensate for their imperfections and enhance their utility for humans.
Although large language models (LLMs) perform impressively on many tasks, overconfidence remains a problem. We hypothesized that on multiple-choice Q&A tasks, wrong answers would be associated with smaller maximum softmax probabilities (MSPs) compared to correct answers. We comprehensively evaluate this hypothesis on ten open-source LLMs and five datasets, and find strong evidence for our hypothesis among models which perform well on the original Q&A task. For the six LLMs with the best Q&A performance, the AUROC derived from the MSP was better than random chance with p < 10^{-4} in 59/60 instances. Among those six LLMs, the average AUROC ranged from 60% to 69%. Leveraging these findings, we propose a multiple-choice Q&A task with an option to abstain and show that performance can be improved by selectively abstaining based on the MSP of the initial model response. We also run the same experiments with pre-softmax logits instead of softmax probabilities and find similar (but not identical) results.
Installing probabilistic world models into artificial agents opens an efficient channel for humans to communicate with and control these agents. In addition to updating agent policies, humans can modify their internal world models in order to influence their decisions. The challenge, however, is that currently existing world models are difficult for humans to adapt because they lack a natural communication interface. Aimed at addressing this shortcoming, we develop Language-Guided World Models (LWMs), which can capture environment dynamics by reading language descriptions. These models enhance agent communication efficiency, allowing humans to simultaneously alter their behavior on multiple tasks with concise language feedback. They also enable agents to self-learn from texts originally written to instruct humans. To facilitate the development of LWMs, we design a challenging benchmark based on the game of MESSENGER (Hanjie et al., 2021), requiring compositional generalization to new language descriptions and environment dynamics. Our experiments reveal that the current state-of-the-art Transformer architecture performs poorly on this benchmark, motivating us to design a more robust architecture. To showcase the practicality of our proposed LWMs, we simulate a scenario where these models augment the interpretability and safety of an agent by enabling it to generate and discuss plans with a human before execution. By effectively incorporating language feedback on the plan, the models boost the agent performance in the real environment by up to three times without collecting any interactive experiences in this environment.
Document Visual Question Answering (DocVQA) is a fast growing branch of document understanding. Despite the fact that documents contain sensitive or copyrighted information, none of the current DocVQA methods offers strong privacy guarantees. In this work, we explore privacy in the domain of DocVQA for the first time. We highlight privacy issues in state of the art multi-modal LLM models used for DocVQA, and explore possible solutions. Specifically, we focus on the invoice processing use case as a realistic, widely used scenario for document understanding, and propose a large scale DocVQA dataset comprising invoice documents and associated questions and answers. We employ a federated learning scheme, that reflects the real-life distribution of documents in different businesses, and we explore the use case where the ID of the invoice issuer is the sensitive information to be protected. We demonstrate that non-private models tend to memorise, behaviour that can lead to exposing private information. We then evaluate baseline training schemes employing federated learning and differential privacy in this multi-modal scenario, where the sensitive information might be exposed through any of the two input modalities: vision (document image) or language (OCR tokens). Finally, we design an attack exploiting the memorisation effect of the model, and demonstrate its effectiveness in probing different DocVQA models.
We investigate the problem of generating instructions to guide humans to navigate in simulated residential environments. A major issue with current models is hallucination: they generate references to actions or objects that are inconsistent with what a human follower would perform or encounter along the described path. We develop a model that detects these hallucinated references by adopting a model pre-trained on a large corpus of image-text pairs, and fine-tuning it with a contrastive loss that separates correct instructions from instructions containing synthesized hallucinations. Our final model outperforms several baselines, including using word probability estimated by the instruction-generation model, and supervised models based on LSTM and Transformer.
Assistant AI agents should be capable of rapidly acquiring novel skills and adapting to new user preferences. Traditional frameworks like imitation learning and reinforcement learning do not facilitate this capability because they support only low-level, inefficient forms of communication. In contrast, humans communicate with progressive efficiency by defining and sharing abstract intentions. Reproducing similar capability in AI agents, we develop a novel learning framework named Communication-Efficient Interactive Learning (CEIL). By equipping a learning agent with an abstract, dynamic language and an intrinsic motivation to learn with minimal communication effort, CEIL leads to emergence of a human-like pattern where the learner and the teacher communicate progressively efficiently by exchanging increasingly more abstract intentions. CEIL demonstrates impressive performance and communication efficiency on a 2D MineCraft domain featuring long-horizon decision-making tasks. Agents trained with CEIL quickly master new tasks, outperforming non-hierarchical and hierarchical imitation learning by up to 50% and 20% in absolute success rate, respectively, given the same number of interactions with the teacher. Especially, the framework performs robustly with teachers modeled after human pragmatic communication behavior.
How do language models "think"? This paper formulates a probabilistic cognitive model called bounded pragmatic speaker, which can characterize the operation of different variants of language models. In particular, we show that large language models fine-tuned with reinforcement learning from human feedback (Ouyang et al., 2022) implements a model of thought that conceptually resembles a fast-and-slow model (Kahneman, 2011). We discuss the limitations of reinforcement learning from human feedback as a fast-and-slow model of thought and propose directions for extending this framework. Overall, our work demonstrates that viewing language models through the lens of cognitive probabilistic modeling can offer valuable insights for understanding, evaluating, and developing them.
Humans exploit prior knowledge to describe images, and are able to adapt their explanation to specific contextual information, even to the extent of inventing plausible explanations when contextual information and images do not match. In this work, we propose the novel task of captioning Wikipedia images by integrating contextual knowledge. Specifically, we produce models that jointly reason over Wikipedia articles, Wikimedia images and their associated descriptions to produce contextualized captions. Particularly, a similar Wikimedia image can be used to illustrate different articles, and the produced caption needs to be adapted to a specific context, therefore allowing us to explore the limits of a model to adjust captions to different contextual information. A particular challenging task in this domain is dealing with out-of-dictionary words and Named Entities. To address this, we propose a pre-training objective, Masked Named Entity Modeling (MNEM), and show that this pretext task yields an improvement compared to baseline models. Furthermore, we verify that a model pre-trained with the MNEM objective in Wikipedia generalizes well to a News Captioning dataset. Additionally, we define two different test splits according to the difficulty of the captioning task. We offer insights on the role and the importance of each modality and highlight the limitations of our model. The code, models and data splits are publicly available at Upon acceptance.
Despite significant progress over the past few years, ambiguity is still a key challenge in Facial Expression Recognition (FER). It can lead to noisy and inconsistent annotation, which hinders the performance of deep learning models in real-world scenarios. In this paper, we propose a new uncertainty-aware label distribution learning method to improve the robustness of deep models against uncertainty and ambiguity. We leverage neighborhood information in the valence-arousal space to adaptively construct emotion distributions for training samples. We also consider the uncertainty of provided labels when incorporating them into the label distributions. Our method can be easily integrated into a deep network to obtain more training supervision and improve recognition accuracy. Intensive experiments on several datasets under various noisy and ambiguous settings show that our method achieves competitive results and outperforms recent state-of-the-art approaches. Our code and models are available at https://github.com/minhnhatvt/label-distribution-learning-fer-tf.
This paper tackles the challenge of forensic medical image matching (FMIM) using deep neural networks (DNNs). FMIM is a particular case of content-based image retrieval (CBIR). The main challenge in FMIM compared to the general case of CBIR, is that the subject to whom a query image belongs may be affected by aging and progressive degenerative disorders, making it difficult to match data on a subject level. CBIR with DNNs is generally solved by minimizing a ranking loss, such as Triplet loss (TL), computed on image representations extracted by a DNN from the original data. TL, in particular, operates on triplets: anchor, positive (similar to anchor) and negative (dissimilar to anchor). Although TL has been shown to perform well in many CBIR tasks, it still has limitations, which we identify and analyze in this work. In this paper, we introduce (i) the AdaTriplet loss -- an extension of TL whose gradients adapt to different difficulty levels of negative samples, and (ii) the AutoMargin method -- a technique to adjust hyperparameters of margin-based losses such as TL and our proposed loss dynamically. Our results are evaluated on two large-scale benchmarks for FMIM based on the Osteoarthritis Initiative and Chest X-ray-14 datasets. The codes allowing replication of this study have been made publicly available at \url{https://github.com/Oulu-IMEDS/AdaTriplet}.